This disclosure relates in general to the field of visual computing, and more particularly, though not exclusively, to a ubiquitous visual computing witness.
The process of gathering facts, evidence, and/or information often involves numerous sources, such as eye witnesses, physical environments, electronic environments (e.g., surveillance systems), and so forth. Existing approaches for gathering facts and evidence, however, suffer from various shortcomings. In many cases, for example, it can be challenging or impossible to even identify the respective sources of relevant data. Further, because existing approaches are not automated, they typically require information to be manually collected from each individual source. Moreover, the respective sources are often under the ownership and/or control of multiple different entities, which further complicates the information gathering process. Finally, while advancements in technology have led to widespread penetration of numerous types of electronic devices that could potentially be leveraged for information gathering purposes (e.g., cameras, sensors, mobile devices), there is no efficient mechanism for identifying electronic devices with relevant information and collecting information from those devices.
The present disclosure is best understood from the following detailed description when read with the accompanying figures. It is emphasized that, in accordance with the standard practice in the industry, various features are not necessarily drawn to scale, and are used for illustration purposes only. Where a scale is shown, explicitly or implicitly, it provides only one illustrative example. In other embodiments, the dimensions of the various features may be arbitrarily increased or reduced for clarity of discussion.
This patent application claims the benefit of the filing date of U.S. Provisional Patent Application Ser. No. 62/611,536, filed on Dec. 28, 2017, and entitled “VISUAL FOG,” the content of which is hereby expressly incorporated by reference.
The following disclosure provides many different embodiments, or examples, for implementing different features of the present disclosure. Specific examples of components and arrangements are described below to simplify the present disclosure. These are, of course, merely examples and are not intended to be limiting. Further, the present disclosure may repeat reference numerals and/or letters in the various examples. This repetition is for the purpose of simplicity and clarity and does not in itself dictate a relationship between the various embodiments and/or configurations discussed. Different embodiments may have different advantages, and no particular advantage is necessarily required of any embodiment.
Example embodiments that may be used to implement the features and functionality of this disclosure will now be described with more particular reference to the attached FIGURES.
Visual Fog Introduction
Existing approaches to large-scale visual computing, however, suffer from numerous limitations. In particular, existing visual computing approaches are implemented using rigid designs that utilize resources inefficiently (e.g., processing, bandwidth, and storage resources) and provide limited functionality. For example, using existing approaches, visual data is typically captured by devices at the edge of a network and simply funneled to the cloud for processing and storage, thus relying heavily on the cloud infrastructure. Due to the large size of visual data, however, this approach typically consumes significant network bandwidth and requires substantial processing and storage resources in the cloud. As a result, existing approaches often suffer from high latency and inefficient resource utilization, and may also be inaccurate, unreliable, inflexible, and incapable of scaling efficiently.
Accordingly, this disclosure describes various embodiments of a visual fog computing system 100 for performing large-scale visual computing in an efficient and reliable manner. For example, rather than relying exclusively or primarily on cloud resources 130 for visual computing tasks, visual fog system 100 leverages both cloud 130 and edge 110 resources, which may be collectively referred to as the “fog.” In this manner, visual fog system 100 can leverage all available “fog” resources to perform visual computing tasks more efficiently, thus improving resource utilization, latency, accuracy, precision, and reliability. Moreover, as described further throughout this disclosure, visual fog system 100 can be implemented using a flexible design that supports ad-hoc queries and is highly scalable, thus rendering it suitable for many visual computing applications and use cases.
In the illustrated embodiment of
Edge resources 110 may include any equipment, devices, and/or components deployed or connected near the “edge” of a communication network. In the illustrated embodiment, for example, edge resources 110 include end-user devices 112a,b (e.g., desktops, laptops, mobile devices), Internet-of-Things (IoT) devices 114, and gateways or routers 116, as described further below. Edge resources 110 may communicate with each other and/or with other remote networks and resources (e.g., cloud resources 130) through one or more communication networks 150, such as local area network 150a and/or wide area network 150b. Moreover, in the illustrated embodiment, edge resources 110 collectively include a plurality of visual sensors 120 (e.g., cameras) for capturing visual representations and data associated with their surroundings. In some embodiments, for example, certain end-user devices 112 and/or IoT devices 114 may include one or more cameras and/or other types of visual sensors 120. Visual sensors 120 may include any type of visual or optical sensors, such as cameras, ultraviolet (UV) sensors, laser rangefinders (e.g., light detection and ranging (LIDAR)), infrared (IR) sensors, electro-optical/infrared (EO/IR) sensors, and so forth.
End-user devices 112 may include any device that enables or facilitates interaction with a user in visual fog system 100, including, for example, desktop computers, laptops, tablets, mobile phones and other mobile devices, and wearable devices (e.g., smart watches, smart glasses, headsets), among other examples.
IoT devices 114 may include any device capable of communicating and/or participating in an Internet-of-Things (IoT) system or network. IoT systems may refer to new or improved ad-hoc systems and networks composed of a variety of different devices (e.g., IoT devices 114) interoperating and synergizing for a particular application or use case. Such ad-hoc systems are emerging as more and more products and equipment evolve to become “smart,” meaning they are controlled or monitored by computer processors and are capable of communicating with other devices. For example, an IoT device 114 may include a computer processor and/or communication interface to allow interoperation with other components of visual fog system 100, such as with cloud resources 130 and/or other edge resources 110. IoT devices 114 may be “greenfield” devices that are developed with IoT capabilities from the ground-up, or “brownfield” devices that are created by integrating IoT capabilities into existing legacy devices that were initially developed without IoT capabilities. For example, in some cases, IoT devices 114 may be built from sensors and communication modules integrated in or attached to “things,” such as equipment, toys, tools, vehicles, living things (e.g., plants, animals, humans), and so forth. Alternatively, or additionally, certain IoT devices 114 may rely on intermediary components, such as edge gateways or routers 116, to communicate with the various components of system 100.
IoT devices 114 may include various types of sensors for monitoring, detecting, measuring, and generating sensor data and signals associated with characteristics of their environment. In some embodiments, for example, certain IoT devices 114 may include visual sensors 120 (e.g., cameras) for capturing visual representations and data associated with their surroundings. IoT devices 114 may also include other types of sensors configured to detect characteristics such as movement, weight, physical contact, temperature, wind, noise, light, position, humidity, radiation, liquid, specific chemical compounds, battery life, wireless signals, computer communications, and bandwidth, among other examples. Sensors can include physical sensors (e.g., physical monitoring components) and virtual sensors (e.g., software-based monitoring components). IoT devices 114 may also include actuators to perform various actions in their respective environments. For example, an actuator may be used to selectively activate certain functionality, such as toggling the power or operation of a security system (e.g., alarm, camera, locks) or household appliance (e.g., audio system, lighting, HVAC appliances, garage doors), among other examples.
Indeed, this disclosure contemplates use of a potentially limitless universe of IoT devices 114 and associated sensors/actuators. IoT devices 114 may include, for example, any type of equipment and/or devices associated with any type of system 100 and/or industry, including transportation (e.g., automobile, airlines), industrial manufacturing, energy (e.g., power plants), telecommunications (e.g., Internet, cellular, and television service providers), retail, medical (e.g., healthcare, pharmaceutical), and/or food and beverage, among others. In the transportation industry, for example, IoT devices 114 may include equipment and devices associated with aircrafts, automobiles, or vessels, such as navigation systems, autonomous flight or driving systems, traffic monitoring and/or planning systems, parking systems, and/or any internal mechanical or electrical components that are monitored by sensors (e.g., engines). IoT devices 114 may also include equipment, devices, and/or infrastructure associated with industrial manufacturing and production, shipping (e.g., cargo tracking), communications networks (e.g., gateways, routers, servers, cellular towers), server farms, electrical power plants, wind farms, oil and gas pipelines, water treatment and distribution, wastewater collection and treatment, and weather monitoring (e.g., temperature, wind, and humidity sensors), among other examples. IoT devices 114 may also include, for example, any type of “smart” device or system, such as smart entertainment systems (e.g., televisions, audio systems, videogame systems), smart household or office appliances (e.g., heat-ventilation-air-conditioning (HVAC) appliances, refrigerators, washers and dryers, coffee brewers), power control systems (e.g., automatic electricity, light, and HVAC controls), security systems (e.g., alarms, locks, cameras, motion detectors, fingerprint scanners, facial recognition systems), and other home automation systems, among other examples. IoT devices 114 can be statically located, such as mounted on a building, wall, floor, ground, lamppost, sign, water tower, or any other fixed or static structure. IoT devices 114 can also be mobile, such as devices in vehicles or aircrafts, drones, packages (e.g., for tracking cargo), mobile devices, and wearable devices, among other examples. Moreover, any type of edge resource 110 may also be considered as an IoT device 114, including end-user devices 112 and edge gateways 116, among other examples.
Edge gateways and/or routers 116 may be used to facilitate communication to and from edge resources 110. For example, gateways 116 may provide communication capabilities to existing legacy devices that were initially developed without any such capabilities (e.g., “brownfield” IoT devices 114). Gateways 116 can also be utilized to extend the geographical reach of edge resources 110 with short-range, proprietary, or otherwise limited communication capabilities, such as IoT devices 114 with Bluetooth or ZigBee communication capabilities. For example, gateways 116 can serve as intermediaries between IoT devices 114 and remote networks or services, by providing a front-haul to the IoT devices 114 using their native communication capabilities (e.g., Bluetooth, ZigBee), and providing a back-haul to other networks 150 and/or cloud resources 130 using another wired or wireless communication medium (e.g., Ethernet, Wi-Fi, cellular). In some embodiments, a gateway 116 may be implemented by a dedicated gateway device, or by a general-purpose device, such as another IoT device 114, end-user device 112, or other type of edge resource 110. In some instances, gateways 116 may also implement certain network management and/or application functionality (e.g., visual computing functionality, IoT application and management functionality), either separately or in conjunction with other components, such as cloud resources 130 and/or other edge resources 110.
Cloud resources 130 may include any resources or services that are hosted remotely over a network, which may otherwise be referred to as in the “cloud.” In some embodiments, for example, cloud resources 130 may be remotely hosted on servers in a datacenter (e.g., application servers, database servers). Cloud resources 130 may include any resources, services, and/or functionality that can be utilized by or for edge resources 110, including but not limited to, visual computing applications and services, IoT application and management services, data storage, computational services (e.g., data analytics, searching, diagnostics and fault management), security services (e.g., surveillance, alarms, user authentication), mapping and navigation, geolocation services, network or infrastructure management, payment processing, audio and video streaming, messaging, social networking, news, and weather, among other examples.
Communication networks 150a,b may be used to facilitate communication between components of system 100. In the illustrated embodiment, for example, edge resources 110 are connected to local area network (LAN) 150a in order to facilitate communication with each other and/or other remote networks or resources, such as wide area network (WAN) 150b and/or cloud resources 130. In various embodiments, visual fog system 100 may be implemented using any number or type of communication network(s) 150, including local area networks, wide area networks, public networks, the Internet, cellular networks, Wi-Fi networks, short-range networks (e.g., Bluetooth or ZigBee), and/or any other wired or wireless communication networks or mediums.
In general, edge resources 110 (and in particular IoT devices 114) may generate an extremely large volume and variety of data. As one example, edge resources 110 with visual sensors 120 may generate large volumes of visual data, such as video and/or images. Edge resources 110 typically offload this data to the cloud 130 for processing and/or storage. Cloud resources 130, however, may not necessarily be suited to handle the rapidly growing volume, variety, and velocity of data generated by IoT devices 114 and other edge resources 110. For example, cloud-based processing may not be ideal in certain circumstances, such as processing time-sensitive or highly confidential data, or when faced with network bandwidth constraints, among other examples. Accordingly, in some embodiments, visual fog system 100 may leverage “edge” processing to augment the performance and capabilities of the cloud 130 using edge resources 110. Edge processing is an approach that involves processing certain data at the network edge (e.g., using edge resources 110), near where the data is generated, rather than simply funneling large volumes of data to the cloud for processing and storage. Certain data may still be sent to the cloud, as appropriate, such as for deeper analysis and/or long-term storage. Edge processing may be used to complement the shortcomings of cloud-based processing (e.g., when cloud-based processing is inefficient, ineffective, and/or unsecure), and thus improve the handling of the growing volume, variety, and velocity of data generated by IoT devices 114 and/or other edge resources 110. For example, in some cases, processing data near its source (e.g., in the network edge) rather than in the cloud may improve performance and/or avoid system failures or disasters. Edge processing may also conserve network bandwidth, which may be particularly beneficial when facing bandwidth constraints and/or limited network connectivity.
In some cases, the collective use of both edge 110 and cloud 130 resources may be referred to as “fog” computing, as functionality of the “cloud” 130 is effectively extended by the edge resources 110, thus forming a “fog” over the network edge. Moreover, in some embodiments, devices 110 in the “fog” may connect and/or communicate with each other using an interconnection standard or protocol, such as the open interconnect consortium (OIC) standard specification 1.0, released by the Open Connectivity Foundation™ (OCF) on Dec. 23, 2015, which enables devices to discover and connect with each other; Thread, a networking protocol for Internet-of-Things (IoT) devices used in “smart” home automation and similar deployments, developed by an alliance of organizations named the “Thread Group”; the optimized link state routing (OLSR) protocol; and/or the better approach to mobile ad-hoc networking (B.A.T.M.A.N.), among other examples.
Moreover, in some embodiments, fog computing may be leveraged by visual fog system 100 for large-scale visual computing applications. For example, in some embodiments, the components of visual fog system 100 (e.g., edge resources 110, cloud resources 130) may be implemented with some or all aspects of the visual computing functionality described throughout this disclosure in connection with the remaining FIGURES.
Any, all, or some of the computing devices of system 100 may be adapted to execute any operating system, including Linux or other UNIX-based operating systems, Microsoft Windows, Windows Server, MacOS, Apple iOS, Google Android, or any customized and/or proprietary operating system, along with virtual machines adapted to virtualize execution of a particular operating system.
While
Additional embodiments associated with the implementation of a visual fog computing system 100 are described further in connection with the remaining FIGURES. Accordingly, it should be appreciated that visual fog system 100 of
Example Internet-of-Things (IoT) Implementations
Often, IoT devices are limited in memory, size, or functionality, allowing larger numbers to be deployed for a similar cost to smaller numbers of larger devices. However, an IoT device may be a smart phone, laptop, tablet, or PC, or other larger device. Further, an IoT device may be a virtual device, such as an application on a smart phone or other computing device. IoT devices may include IoT gateways, used to couple IoT devices to other IoT devices and to cloud applications, for data storage, process control, and the like.
Networks of IoT devices may include commercial and home automation devices, such as water distribution systems, electric power distribution systems, pipeline control systems, plant control systems, light switches, thermostats, locks, cameras, alarms, motion sensors, and the like. The IoT devices may be accessible through remote computers, servers, and other systems, for example, to control systems or access data.
The future growth of the Internet and like networks may involve very large numbers of IoT devices. Accordingly, in the context of the techniques discussed herein, a number of innovations for such future networking will address the need for all these layers to grow unhindered, to discover and make accessible connected resources, and to support the ability to hide and compartmentalize connected resources. Any number of network protocols and communications standards may be used, wherein each protocol and standard is designed to address specific objectives. Further, the protocols are part of the fabric supporting human accessible services that operate regardless of location, time or space. The innovations include service delivery and associated infrastructure, such as hardware and software; security enhancements; and the provision of services based on Quality of Service (QoS) terms specified in service level and service delivery agreements. As will be understood, the use of IoT devices and networks, such as those introduced in
The network topology may include any number of types of IoT networks, such as a mesh network provided with the network 256 using Bluetooth low energy (BLE) links 222. Other types of IoT networks that may be present include a wireless local area network (WLAN) network 258 used to communicate with IoT devices 204 through IEEE 802.11 (Wi-Fi®) links 228, a cellular network 260 used to communicate with IoT devices 204 through an LTE/LTE-A (4G) or 5G cellular network, and a low-power wide area (LPWA) network 262, for example, a LPWA network compatible with the LoRaWan specification promulgated by the LoRa alliance, or a IPv6 over Low Power Wide-Area Networks (LPWAN) network compatible with a specification promulgated by the Internet Engineering Task Force (IETF). Further, the respective IoT networks may communicate with an outside network provider (e.g., a tier 2 or tier 3 provider) using any number of communications links, such as an LTE cellular link, an LPWA link, or a link based on the IEEE 802.15.4 standard, such as Zigbee®. The respective IoT networks may also operate with use of a variety of network and internet application protocols such as Constrained Application Protocol (CoAP). The respective IoT networks may also be integrated with coordinator devices that provide a chain of links that forms cluster tree of linked devices and networks.
Each of these IoT networks may provide opportunities for new technical features, such as those as described herein. The improved technologies and networks may enable the exponential growth of devices and networks, including the use of IoT networks into as fog devices or systems. As the use of such improved technologies grows, the IoT networks may be developed for self-management, functional evolution, and collaboration, without needing direct human intervention. The improved technologies may even enable IoT networks to function without centralized controlled systems. Accordingly, the improved technologies described herein may be used to automate and enhance network management and operation functions far beyond current implementations.
In an example, communications between IoT devices 204, such as over the backbone links 202, may be protected by a decentralized system for authentication, authorization, and accounting (AAA). In a decentralized AAA system, distributed payment, credit, audit, authorization, and authentication systems may be implemented across interconnected heterogeneous network infrastructure. This allows systems and networks to move towards autonomous operations. In these types of autonomous operations, machines may even contract for human resources and negotiate partnerships with other machine networks. This may allow the achievement of mutual objectives and balanced service delivery against outlined, planned service level agreements as well as achieve solutions that provide metering, measurements, traceability and trackability. The creation of new supply chain structures and methods may enable a multitude of services to be created, mined for value, and collapsed without any human involvement.
Such IoT networks may be further enhanced by the integration of sensing technologies, such as sound, light, electronic traffic, facial and pattern recognition, smell, vibration, into the autonomous organizations among the IoT devices. The integration of sensory systems may allow systematic and autonomous communication and coordination of service delivery against contractual service objectives, orchestration and quality of service (QoS) based swarming and fusion of resources. Some of the individual examples of network-based resource processing include the following.
The mesh network 256, for instance, may be enhanced by systems that perform inline data-to-information transforms. For example, self-forming chains of processing resources comprising a multi-link network may distribute the transformation of raw data to information in an efficient manner, and the ability to differentiate between assets and resources and the associated management of each. Furthermore, the proper components of infrastructure and resource based trust and service indices may be inserted to improve the data integrity, quality, assurance and deliver a metric of data confidence.
The WLAN network 258, for instance, may use systems that perform standards conversion to provide multi-standard connectivity, enabling IoT devices 204 using different protocols to communicate. Further systems may provide seamless interconnectivity across a multi-standard infrastructure comprising visible Internet resources and hidden Internet resources.
Communications in the cellular network 260, for instance, may be enhanced by systems that offload data, extend communications to more remote devices, or both. The LPWA network 262 may include systems that perform non-Internet protocol (IP) to IP interconnections, addressing, and routing. Further, each of the IoT devices 204 may include the appropriate transceiver for wide area communications with that device. Further, each IoT device 204 may include other transceivers for communications using additional protocols and frequencies.
Finally, clusters of IoT devices may be equipped to communicate with other IoT devices as well as with a cloud network. This may allow the IoT devices to form an ad-hoc network between the devices, allowing them to function as a single device, which may be termed a fog device. This configuration is discussed further with respect to
The fog 320 may be considered to be a massively interconnected network wherein a number of IoT devices 302 are in communications with each other, for example, by radio links 322. As an example, this interconnected network may be facilitated using an interconnect specification released by the Open Connectivity Foundation™ (OCF). This standard allows devices to discover each other and establish communications for interconnects. Other interconnection protocols may also be used, including, for example, the optimized link state routing (OLSR) Protocol, the better approach to mobile ad-hoc networking (B.A.T.M.A.N.) routing protocol, or the OMA Lightweight M2M (LWM2M) protocol, among others.
Three types of IoT devices 302 are shown in this example, gateways 304, data aggregators 326, and sensors 328, although any combinations of IoT devices 302 and functionality may be used. The gateways 304 may be edge devices that provide communications between the cloud 300 and the fog 320, and may also provide the backend process function for data obtained from sensors 328, such as motion data, flow data, temperature data, and the like. The data aggregators 326 may collect data from any number of the sensors 328, and perform the back-end processing function for the analysis. The results, raw data, or both may be passed along to the cloud 300 through the gateways 304. The sensors 328 may be full IoT devices 302, for example, capable of both collecting data and processing the data. In some cases, the sensors 328 may be more limited in functionality, for example, collecting the data and allowing the data aggregators 326 or gateways 304 to process the data.
Communications from any IoT device 302 may be passed along a convenient path (e.g., a most convenient path) between any of the IoT devices 302 to reach the gateways 304. In these networks, the number of interconnections provide substantial redundancy, allowing communications to be maintained, even with the loss of a number of IoT devices 302. Further, the use of a mesh network may allow IoT devices 302 that are very low power or located at a distance from infrastructure to be used, as the range to connect to another IoT device 302 may be much less than the range to connect to the gateways 304.
The fog 320 provided from these IoT devices 302 may be presented to devices in the cloud 300, such as a server 306, as a single device located at the edge of the cloud 300, e.g., a fog device. In this example, the alerts coming from the fog device may be sent without being identified as coming from a specific IoT device 302 within the fog 320. In this fashion, the fog 320 may be considered a distributed platform that provides computing and storage resources to perform processing or data-intensive tasks such as data analytics, data aggregation, and machine-learning, among others.
In some examples, the IoT devices 302 may be configured using an imperative programming style, e.g., with each IoT device 302 having a specific function and communication partners. However, the IoT devices 302 forming the fog device may be configured in a declarative programming style, allowing the IoT devices 302 to reconfigure their operations and communications, such as to determine needed resources in response to conditions, queries, and device failures. As an example, a query from a user located at a server 306 about the operations of a subset of equipment monitored by the IoT devices 302 may result in the fog 320 device selecting the IoT devices 302, such as particular sensors 328, needed to answer the query. The data from these sensors 328 may then be aggregated and analyzed by any combination of the sensors 328, data aggregators 326, or gateways 304, before being sent on by the fog 320 device to the server 306 to answer the query. In this example, IoT devices 302 in the fog 320 may select the sensors 328 used based on the query, such as adding data from flow sensors or temperature sensors. Further, if some of the IoT devices 302 are not operational, other IoT devices 302 in the fog 320 device may provide analogous data, if available.
Other example groups of IoT devices may include remote weather stations 414, local information terminals 416, alarm systems 418, automated teller machines 420, alarm panels 422, or moving vehicles, such as emergency vehicles 424 or other vehicles 426, among many others. Each of these IoT devices may be in communication with other IoT devices, with servers 404, with another IoT fog device or system (not shown, but depicted in
As can be seen from
Clusters of IoT devices, such as the remote weather stations 414 or the traffic control group 406, may be equipped to communicate with other IoT devices as well as with the cloud 400. This may allow the IoT devices to form an ad-hoc network between the devices, allowing them to function as a single device, which may be termed a fog device or system (e.g., as described above with reference to
The IoT device 550 may include a processor 552, which may be a microprocessor, a multi-core processor, a multithreaded processor, an ultra-low voltage processor, an embedded processor, or other known processing element. The processor 552 may be a part of a system on a chip (SoC) in which the processor 552 and other components are formed into a single integrated circuit, or a single package, such as the Edison™ or Galileo™ SoC boards from Intel. As an example, the processor 552 may include an Intel® Architecture Core™ based processor, such as a Quark™, an Atom™, an i3, an i5, an i7, or an MCU-class processor, or another such processor available from Intel® Corporation, Santa Clara, Calif. However, any number other processors may be used, such as available from Advanced Micro Devices, Inc. (AMD) of Sunnyvale, Calif., a MIPS-based design from MIPS Technologies, Inc. of Sunnyvale, Calif., an ARM-based design licensed from ARM Holdings, Ltd. or customer thereof, or their licensees or adopters. The processors may include units such as an A5-A10 processor from Apple® Inc., a Snapdragon™ processor from Qualcomm® Technologies, Inc., or an OMAP™ processor from Texas Instruments, Inc.
The processor 552 may communicate with a system memory 554 over an interconnect 556 (e.g., a bus). Any number of memory devices may be used to provide for a given amount of system memory. As examples, the memory may be random access memory (RAM) in accordance with a Joint Electron Devices Engineering Council (JEDEC) design such as the DDR or mobile DDR standards (e.g., LPDDR, LPDDR2, LPDDR3, or LPDDR4). In various implementations, the individual memory devices may be of any number of different package types such as single die package (SDP), dual die package (DDP) or quad die package (Q17P). These devices, in some examples, may be directly soldered onto a motherboard to provide a lower profile solution, while in other examples the devices are configured as one or more memory modules that in turn couple to the motherboard by a given connector. Any number of other memory implementations may be used, such as other types of memory modules, e.g., dual inline memory modules (DIMMs) of different varieties including but not limited to microDlMMs or MiniDIMMs.
To provide for persistent storage of information such as data, applications, operating systems and so forth, a storage 558 may also couple to the processor 552 via the interconnect 556. In an example, the storage 558 may be implemented via a solid state disk drive (SSDD). Other devices that may be used for the storage 558 include flash memory cards, such as SD cards, microSD cards, xD picture cards, and the like, and USB flash drives. In low power implementations, the storage 558 may be on-die memory or registers associated with the processor 552. However, in some examples, the storage 558 may be implemented using a micro hard disk drive (HDD). Further, any number of new technologies may be used for the storage 558 in addition to, or instead of, the technologies described, such resistance change memories, phase change memories, holographic memories, or chemical memories, among others.
The components may communicate over the interconnect 556. The interconnect 556 may include any number of technologies, including industry standard architecture (ISA), extended ISA (EISA), peripheral component interconnect (PCI), peripheral component interconnect extended (PCIx), PCI express (PCIe), or any number of other technologies. The interconnect 556 may be a proprietary bus, for example, used in a SoC based system. Other bus systems may be included, such as an I2C interface, an SPI interface, point to point interfaces, and a power bus, among others.
The interconnect 556 may couple the processor 552 to a mesh transceiver 562, for communications with other mesh devices 564. The mesh transceiver 562 may use any number of frequencies and protocols, such as 2.4 Gigahertz (GHz) transmissions under the IEEE 802.15.4 standard, using the Bluetooth® low energy (BLE) standard, as defined by the Bluetooth® Special Interest Group, or the ZigBee® standard, among others. Any number of radios, configured for a particular wireless communication protocol, may be used for the connections to the mesh devices 564. For example, a WLAN unit may be used to implement Wi-Fi™ communications in accordance with the Institute of Electrical and Electronics Engineers (IEEE) 802.11 standard. In addition, wireless wide area communications, e.g., according to a cellular or other wireless wide area protocol, may occur via a WWAN unit.
The mesh transceiver 562 may communicate using multiple standards or radios for communications at different range. For example, the IoT device 550 may communicate with close devices, e.g., within about 10 meters, using a local transceiver based on BLE, or another low power radio, to save power. More distant mesh devices 564, e.g., within about 50 meters, may be reached over ZigBee or other intermediate power radios. Both communications techniques may take place over a single radio at different power levels, or may take place over separate transceivers, for example, a local transceiver using BLE and a separate mesh transceiver using ZigBee.
A wireless network transceiver 566 may be included to communicate with devices or services in the cloud 500 via local or wide area network protocols. The wireless network transceiver 566 may be a LPWA transceiver that follows the IEEE 802.15.4, or IEEE 802.15.4g standards, among others. The IoT device 550 may communicate over a wide area using LoRaWAN™ (Long Range Wide Area Network) developed by Semtech and the LoRa Alliance. The techniques described herein are not limited to these technologies, but may be used with any number of other cloud transceivers that implement long range, low bandwidth communications, such as Sigfox, and other technologies. Further, other communications techniques, such as time-slotted channel hopping, described in the IEEE 802.15.4e specification may be used.
Any number of other radio communications and protocols may be used in addition to the systems mentioned for the mesh transceiver 562 and wireless network transceiver 566, as described herein. For example, the radio transceivers 562 and 566 may include an LTE or other cellular transceiver that uses spread spectrum (SPA/SAS) communications for implementing high speed communications. Further, any number of other protocols may be used, such as Wi-Fi® networks for medium speed communications and provision of network communications.
The radio transceivers 562 and 566 may include radios that are compatible with any number of 3GPP (Third Generation Partnership Project) specifications, notably Long Term Evolution (LTE), Long Term Evolution-Advanced (LTE-A), and Long Term Evolution-Advanced Pro (LTE-A Pro). It can be noted that radios compatible with any number of other fixed, mobile, or satellite communication technologies and standards may be selected. These may include, for example, any Cellular Wide Area radio communication technology, which may include e.g. a 5th Generation (5G) communication systems, a Global System for Mobile Communications (GSM) radio communication technology, a General Packet Radio Service (GPRS) radio communication technology, or an Enhanced Data Rates for GSM Evolution (EDGE) radio communication technology, a UMTS (Universal Mobile Telecommunications System) communication technology, In addition to the standards listed above, any number of satellite uplink technologies may be used for the wireless network transceiver 566, including, for example, radios compliant with standards issued by the ITU (International Telecommunication Union), or the ETSI (European Telecommunications Standards Institute), among others. The examples provided herein are thus understood as being applicable to various other communication technologies, both existing and not yet formulated.
A network interface controller (NIC) 568 may be included to provide a wired communication to the cloud 500 or to other devices, such as the mesh devices 564. The wired communication may provide an Ethernet connection, or may be based on other types of networks, such as Controller Area Network (CAN), Local Interconnect Network (LIN), DeviceNet, ControlNet, Data Highway+, PROFIBUS, or PROFINET, among many others. An additional NIC 568 may be included to allow connect to a second network, for example, a NIC 568 providing communications to the cloud over Ethernet, and a second NIC 568 providing communications to other devices over another type of network.
The interconnect 556 may couple the processor 552 to an external interface 570 that is used to connect external devices or subsystems. The external devices may include sensors 572, such as accelerometers, level sensors, flow sensors, optical light sensors, camera sensors, temperature sensors, a global positioning system (GPS) sensors, pressure sensors, barometric pressure sensors, and the like. The external interface 570 further may be used to connect the IoT device 550 to actuators 574, such as power switches, valve actuators, an audible sound generator, a visual warning device, and the like.
In some optional examples, various input/output (I/O) devices may be present within, or connected to, the IoT device 550. For example, a display or other output device 584 may be included to show information, such as sensor readings or actuator position. An input device 586, such as a touch screen or keypad may be included to accept input. An output device 584 may include any number of forms of audio or visual display, including simple visual outputs such as binary status indicators (e.g., LEDs) and multi-character visual outputs, or more complex outputs such as display screens (e.g., LCD screens), with the output of characters, graphics, multimedia objects, and the like being generated or produced from the operation of the IoT device 550.
A battery 576 may power the IoT device 550, although in examples in which the IoT device 550 is mounted in a fixed location, it may have a power supply coupled to an electrical grid. The battery 576 may be a lithium ion battery, or a metal-air battery, such as a zinc-air battery, an aluminum-air battery, a lithium-air battery, and the like.
A battery monitor/charger 578 may be included in the IoT device 550 to track the state of charge (SoCh) of the battery 576. The battery monitor/charger 578 may be used to monitor other parameters of the battery 576 to provide failure predictions, such as the state of health (SoH) and the state of function (SoF) of the battery 576. The battery monitor/charger 578 may include a battery monitoring integrated circuit, such as an LTC4020 or an LTC2990 from Linear Technologies, an ADT7488A from ON Semiconductor of Phoenix Ariz., or an IC from the UCD90xxx family from Texas Instruments of Dallas, Tex. The battery monitor/charger 578 may communicate the information on the battery 576 to the processor 552 over the interconnect 556. The battery monitor/charger 578 may also include an analog-to-digital (ADC) convertor that allows the processor 552 to directly monitor the voltage of the battery 576 or the current flow from the battery 576. The battery parameters may be used to determine actions that the IoT device 550 may perform, such as transmission frequency, mesh network operation, sensing frequency, and the like.
A power block 580, or other power supply coupled to a grid, may be coupled with the battery monitor/charger 578 to charge the battery 576. In some examples, the power block 580 may be replaced with a wireless power receiver to obtain the power wirelessly, for example, through a loop antenna in the IoT device 550. A wireless battery charging circuit, such as an LTC4020 chip from Linear Technologies of Milpitas, Calif., among others, may be included in the battery monitor/charger 578. The specific charging circuits chosen depend on the size of the battery 576, and thus, the current required. The charging may be performed using the Airfuel standard promulgated by the Airfuel Alliance, the Qi wireless charging standard promulgated by the Wireless Power Consortium, or the Rezence charging standard, promulgated by the Alliance for Wireless Power, among others.
The storage 558 may include instructions 582 in the form of software, firmware, or hardware commands to implement the techniques described herein. Although such instructions 582 are shown as code blocks included in the memory 554 and the storage 558, it may be understood that any of the code blocks may be replaced with hardwired circuits, for example, built into an application specific integrated circuit (ASIC).
In an example, the instructions 582 provided via the memory 554, the storage 558, or the processor 552 may be embodied as a non-transitory, machine readable medium 560 including code to direct the processor 552 to perform electronic operations in the IoT device 550. The processor 552 may access the non-transitory, machine readable medium 560 over the interconnect 556. For instance, the non-transitory, machine readable medium 560 may include storage units such as optical disks, flash drives, or any number of other hardware devices. The non-transitory, machine readable medium 560 may include instructions to direct the processor 552 to perform a specific sequence or flow of actions, for example, as described with respect to the flowchart(s) and diagram(s) of operations and functionality described throughout this disclosure.
Example Computing Architectures
Processor 600 can execute any type of instructions associated with algorithms, processes, or operations detailed herein. Generally, processor 600 can transform an element or an article (e.g., data) from one state or thing to another state or thing.
Code 604, which may be one or more instructions to be executed by processor 600, may be stored in memory 602, or may be stored in software, hardware, firmware, or any suitable combination thereof, or in any other internal or external component, device, element, or object where appropriate and based on particular needs. In one example, processor 600 can follow a program sequence of instructions indicated by code 604. Each instruction enters a front-end logic 606 and is processed by one or more decoders 608. The decoder may generate, as its output, a micro operation such as a fixed width micro operation in a predefined format, or may generate other instructions, microinstructions, or control signals that reflect the original code instruction. Front-end logic 606 may also include register renaming logic and scheduling logic, which generally allocate resources and queue the operation corresponding to the instruction for execution.
Processor 600 can also include execution logic 614 having a set of execution units 616a, 616b, 616n, etc. Some embodiments may include a number of execution units dedicated to specific functions or sets of functions. Other embodiments may include only one execution unit or one execution unit that can perform a particular function. Execution logic 614 performs the operations specified by code instructions.
After completion of execution of the operations specified by the code instructions, back-end logic 618 can retire the instructions of code 604. In one embodiment, processor 600 allows out of order execution but requires in order retirement of instructions. Retirement logic 620 may take a variety of known forms (e.g., re-order buffers or the like). In this manner, processor 600 is transformed during execution of code 604, at least in terms of the output generated by the decoder, hardware registers and tables utilized by register renaming logic 610, and any registers (not shown) modified by execution logic 614.
Although not shown in
Processors 770 and 780 are shown including integrated memory controller (IMC) units 772 and 782, respectively. Processor 770 also includes as part of its bus controller units point-to-point (P-P) interfaces 776 and 778; similarly, second processor 780 includes P-P interfaces 786 and 788. Processors 770, 780 may exchange information via a point-to-point (P-P) interface 750 using P-P interface circuits 778, 788. As shown in
Processors 770, 780 may each exchange information with a chipset 790 via individual P-P interfaces 752, 754 using point to point interface circuits 776, 794, 786, 798. Chipset 790 may optionally exchange information with the coprocessor 738 via a high-performance interface 739. In one embodiment, the coprocessor 738 is a special-purpose processor, such as, for example, a high-throughput MIC processor, a network or communication processor, compression engine, graphics processor, GPGPU, embedded processor, matrix processor, or the like.
A shared cache (not shown) may be included in either processor or outside of both processors, yet connected with the processors via P-P interconnect, such that either or both processors' local cache information may be stored in the shared cache if a processor is placed into a low power mode.
Chipset 790 may be coupled to a first bus 716 via an interface 796. In one embodiment, first bus 716 may be a Peripheral Component Interconnect (PCI) bus, or a bus such as a PCI Express bus or another third generation I/O interconnect bus, although the scope of this disclosure is not so limited.
As shown in
All or part of any component of
Embodiments of the mechanisms disclosed herein may be implemented in hardware, software, firmware, or a combination of such implementation approaches. Certain embodiments may be implemented as computer programs or program code executing on programmable systems comprising at least one processor, a storage system (including volatile and non-volatile memory and/or storage elements), at least one input device, and at least one output device.
Program code, such as code 730 illustrated in
The program code may be implemented in a high level procedural or object oriented programming language to communicate with a processing system. The program code may also be implemented in assembly or machine language, if desired. In fact, the mechanisms described herein are not limited in scope to any particular programming language. In any case, the language may be a compiled or interpreted language.
One or more aspects of at least one embodiment may be implemented by representative instructions stored on a machine-readable medium which represents various logic within the processor, which when read by a machine causes the machine to fabricate logic to perform the techniques described herein. Such representations, known as “IP cores” may be stored on a tangible, machine readable medium and supplied to various customers or manufacturing facilities to load into the fabrication machines that actually make the logic or processor.
Such machine-readable storage media may include, without limitation, non-transitory, tangible arrangements of articles manufactured or formed by a machine or device, including storage media such as hard disks, any other type of disk including floppy disks, optical disks, compact disk read-only memories (CD-ROMs), compact disk rewritable's (CD-RWs), and magneto-optical disks, semiconductor devices such as read-only memories (ROMs), random access memories (RAMS) such as dynamic random access memories (DRAMs), static random access memories (SRAMs), erasable programmable read-only memories (EPROMs), flash memories, electrically erasable programmable read-only memories (EEPROMs), phase change memory (PCM), magnetic or optical cards, or any other type of media suitable for storing electronic instructions.
Accordingly, embodiments of this disclosure also include non-transitory, tangible machine-readable media containing instructions or containing design data, such as Hardware Description Language (HDL), which defines structures, circuits, apparatuses, processors and/or system features described herein. Such embodiments may also be referred to as program products.
Visual Fog Architecture
In the illustrated embodiment, fog node 810 includes various application programming interfaces (APIs) that provide fundamental capabilities for fog node 810, such as auxiliary API 820, primitive vision API 830, and storage API 840. In some embodiments, for example, these APIs may be used or implemented by lower-level algorithm developers.
Auxiliary API 820 provides various fundamental functionality for fog node 810, such as security 822a, communication 822b, compression 822c (e.g., codecs), and so forth.
Primitive vision API 830 provides fundamental vision processing capabilities for fog node 810. For example, primitive vision API 830 provides access to a plurality of vision kernels 832 that can be used to perform primitive vision operations (e.g., person or object detection, facial recognition). Primitive vision API 830 may also provide access to various machine learning and/or neural network frameworks (e.g., Caffe, OpenCV, TensorFlow).
Storage API 840 provides storage capabilities for fog node 810. In some embodiments, for example, storage API 840 may include a variety of databases 842 for storing different types of visual data, such as graph databases, relational databases, array-based databases (e.g., TileDB), and so forth. In some embodiments, for example, the particular database used to store certain visual data may depend on the type of data, such as raw visual data or pixels, compressed visual data, visual metadata, and so forth.
Moreover, fog node 810 further includes a vision application API 850 that provides higher-level vision functionality, which may be used or implemented by developers of vision applications. For example, vision application API 850 may include a privacy policy 852 that defines the requisite privacy treatment for all data and devices associated with a visual fog network. Vision application API 850 may also include a vision kernel management service 854 that provides access to a variety of primitive vision operations or vision kernels. In some embodiments, for example, vision kernel management service 854 may retrieve vision kernels from a vision kernel repository. For example, if a particular vision application employs person detection functionality, vision kernel management service 854 may retrieve the appropriate vision kernel for performing person detection using the available hardware of the particular fog node 810.
Fog node 810 further includes a vision analytics API 860 and query API 870, which may be used by end-users or operators to perform visual analytics and visual queries. For example, vision analytics API 860 may perform inline (e.g. real-time) and/or offline processing of visual data, application launching, scheduling, resource monitoring, and so forth. Vision analytics API 860 may also include a vision application management service 862 that provides access to a variety of vision applications (e.g., people searching/tracking, object detection/tracking, and so forth). In some embodiments, for example, vision application management service 862 may retrieve vision applications from a vision application repository. In this manner, if an end-user wants to perform a people search, vision application management service 862 may retrieve an appropriate vision application for people searching. In some embodiments, for example, a people search vision application may use vision kernels that perform person detection followed by facial recognition. The end-user, however, can utilize the people search vision application without any knowledge of the underlying vision kernels or vision operations used to implement the application.
Moreover, query API 870 provides an interface that enables end-users to submit visual search requests or queries. In some embodiments, for example, query API 870 may support flexible visual queries in a variety of syntaxes, such as natural language, functional syntax (e.g., using logical operators), relational syntax, and so forth. In some embodiments, query API 870 may further include a query primitive repository 874 that contains the primitive operations that are supported for visual queries. Moreover, query API 870 may include a query compiler 872 for compiling the visual queries into visual processing dataflows that can be executed by visual fog nodes.
For example,
Separately, visual search queries containing user-defined vision functions (UVFs) 1104a-c are received from end-users 1102 of visual fog 1100. A UVF 1104 received from an end-user 1102 is first processed by a compiler 1110 in order to generate a vision dataflow graph for executing the UVF. Accordingly, the vision dataflow graph is then executed by the distributed runtime environment 1120 in order to generate a result for the UVF 1104. In some embodiments, for example, the distributed runtime environment 1120 may determine the result using existing visual metadata that has already been generated (e.g., from the initial or real-time processing of the original visual data), and/or by performing further analysis on the visual data (e.g., by executing a particular vision application 1150). The result obtained from execution of the UVF 1104 may then be provided back to the requesting end-user 1102.
Further, in various embodiments, the distributed runtime environment 1120 may perform the described visual data processing (e.g., initial pre-processing and/or UVF processing) by scheduling or distributing vision workloads across the available fog devices or resources 1140 (e.g., cloud servers 1140a, cameras 1140b, mobile devices, IoT devices, gateways, and/or other fog/edge devices).
In the illustrated embodiment, for example, the vision program 1203 leverages a distributed runtime environment 1214 to process visual data 1217 captured in visual fog 1200. The distributed runtime environment 1214, for example, can perform visual data processing using the collection of available fog devices 1216 in visual fog 1200.
In some embodiments, for example, the distributed runtime environment 1214 may be used to perform initial pre-processing on captured visual data 1217 in real-time (e.g., when the visual data is first captured rather than in response to a query). The resulting visual data or metadata 1217 generated by the distributed runtime environment 1214 may then be stored in a database or data storage 1218.
Moreover, a layperson or end-user 1204 may subsequently submit a declarative query 1205 associated with visual data 1217 captured by visual fog 1200. The declarative query 1205 is processed by a visual question answering (VQA) system 1206, which uses a compiler or interpreter to generate a dataflow 1209 for responding to the query. In some cases, for example, it may be possible to respond to query 1205 using existing visual metadata 1217 that has already been generated (e.g., during the initial or real-time processing of the original visual data 1217 and/or during the processing associated with prior queries 1205). In other cases, however, further processing may need to be performed on the visual data 1217 in order to respond to the query 1205. In either case, an appropriate dataflow 1209 for responding to the query 1205 may be generated, and the resulting dataflow 1209 may be further partitioned into one or more underlying vision processing workloads 1211. Moreover, based on the resource availability 1215 of fog devices 1216 in the distributed runtime environment 1214, a schedule 1213 for distributing the workloads 1211 across the available fog devices 1216 may be generated. Accordingly, the respective workloads 1211 may then be distributed across the fog devices 1216 based on the generated schedule 1213, and each fog device 1216 may execute its respective workload(s) 1211. In this manner, the dataflow 1209 for responding to the query 1205 is executed by the various fog devices 1216 using a distributed approach. A response to the query 1205 may then be provided to the end-user 1204, and the resulting visual metadata 1217 may be stored in database 1218 for responding to subsequent queries.
Visual Question Answering (VQA)
In some embodiments, for example, a VQA framework may support flexible or ad hoc visual search queries using a variety of syntaxes, such as natural language, functional syntax (e.g., using logical operators), relational syntax, and so forth. Accordingly, when a visual search query is received from a user, the query may be compiled into a visual processing dataflow that can be distributed across and executed by the various fog nodes in a visual fog architecture. In this manner, end-users can perform complex searches on large sets of visual data without any knowledge of the underlying architecture or processing required to execute the searches.
Moreover, in some embodiments, users or developers may be capable of defining custom vision functions that can be used in visual search queries, referred to as user-defined vision functions (UVFs). As an example, a UVF could be defined for visually equivalency, or performing “equal” operations on visual data. Many ad hoc visual queries, for example, require information related to the same object or person to be identified or grouped together. Identifying the same object or person across different images or video streams, however, can be challenging. In some embodiments, for example, this task may require feature extraction to be performed across multiple cameras. The respective features extracted from each camera often differ, however, and not all cameras have the same field of view, and thus certain features may be successfully extracted from some cameras but not others. Accordingly, in some embodiments, a user may implement a UVF to define how visual equivalency or “equal” operations are to be performed on visual data. In some embodiments, for example, a UVF for visual equivalency may define objects as “equal” if their feature vectors are “close enough” to each other, meaning the feature vectors must be sufficiently similar but do not have to be an exact match. Further, if feature vectors from different cameras are missing certain features, only the partial features will be compared and the “close enough” definition will be scaled accordingly.
In some cases, for example, the visual processing dataflow 1308 may leverage existing visual metadata that has already been generated and stored on data storage 1314. For example, an inline analytics framework 1310 may be used to perform initial visual data processing in real-time (e.g., when visual data is first captured rather than in response to a query), and an offline analytics framework 1312 may be used to perform further visual data processing required for responding to search queries. Moreover, both the inline and offline analytics frameworks 1310, 1312 may store their resulting visual metadata on data storage 1314 for use in responding to subsequent visual search queries. Accordingly, in some cases, the visual processing dataflow 1308 for a particular query 1302 may leverage existing visual metadata that has already been generated and stored on data storage 1314. In other cases, however, further processing may be required to respond to the query 1302, and thus the visual processing dataflow 1308 may leverage the offline analytics framework 1312 to perform additional processing. In either case, the visual processing pipeline or dataflow 1308 generated by compiler 1304 is executed by the runtime environment in order to generate a response to the visual query 1302.
In the illustrated embodiment, for example, UVFs 1402 are provided to the compiler 1400 via a declarative API 1412. The compiler 1400 may then generate a graph of high-level vision operations 1415 that are required to execute the UVFs 1402, which may in turn be used to generate a vision dataflow graph 1417. In some embodiments, for example, the vision dataflow graph 1417 may be a directed acyclic graph (DAG) that represents the visual processing pipeline required to execute the particular UVFs 1402. Moreover, the compiler 1400 may use dataflow de-duplication to optimize the vision dataflow graph 1417, for example, by merging redundant portions of the dataflows of multiple UVFs 1402 to eliminate the redundancies.
In some embodiments, for example, compiler 1400 may generate the vision dataflow graph 1417 using information from the underlying vision modules 1418 (e.g., hardware-specific information required for scheduling workloads on heterogeneous hardware). The compiler 1400 may also generate a number of database API calls to obtain visual data and/or metadata required to execute the UVFs 1402. In various embodiments, these database API calls may either be part of, or separate from, the vision dataflow graph 1417. Moreover, in some embodiments, the compiler 1400 may generate different results depending on the available visual metadata.
In this manner, the resulting vision dataflow graph 1417 generated by compiler 1400 can subsequently be executed by the runtime environment in order to generate the results for responding to UVFs 1402.
Runtime
The visual fog paradigm envisions tens of thousands (or more) heterogeneous, camera-enabled edge devices distributed across the Internet and/or other large-scale networks, providing live sensing for a myriad of different visual processing applications, given task parallelism and data parallelism. The scale, computational demands, and bandwidth needed for visual computing pipelines necessitates intelligent offloading to distributed computing infrastructure, including the cloud, Internet gateway devices, and the edge devices themselves.
In some embodiments, for example, visual processing may be scheduled or distributed across available fog devices based on various criteria, including device connectivity, device resource capabilities, device resource availability, workload type, privacy constraints, and so forth. Further, machine learning can be leveraged to optimize scheduling decisions.
Workload deployment and/or migration can be implemented using a hot-pluggable runtime environment with universal plugin APIs. For example, conventional workload deployment/migration can be expensive, as it may require sending the runtime environment and toolchains to the assigned nodes. With hot-pluggable runtimes, however, workloads are hot-swappable (e.g., stop runtime, replace plugin, start runtime).
Moreover, a plugin or vision kernel repository can be used to facilitate workload deployment. For example, a cloud-based or distributed repository may be used to manage a collection of device and implementation abstractions for each supported vision capability. In this manner, the repository can distribute the appropriate plugins or vision kernels to fog nodes based on their respective workload assignments.
Incremental processing may be leveraged by a visual fog runtime to maintain the state of any prior processing that has already been performed on visual data, enabling the results of the prior processing to be leveraged for subsequent visual processing and queries. For example, the results of any processing performed on visual data may be represented as visual metadata, which may be stored for later use to avoid performing duplicative processing for subsequent visual queries. In this manner, when a visual query or UVF is received, the dataflow generated by a compiler may vary depending on the available metadata that has already been generated and can be reused.
Metadata pre-provisioning can be used to reduce vision query latency by pre-processing visual data to complete common or frequent types of processing in advance. In some embodiments, for example, a machine learning model may be used to optimize the types of pre-processing that is performed. For example, based on patterns of queries of the same type or that involve similar types of processing, machine learning may be used to model the relationships of diverse queries, while also taking other modalities into account (e.g., weather, traffic). For example, metadata can be pre-provisioned by pre-scheduling certain types of processing in advance based on the recent history of vision queries and UVFs. In this manner, patterns of common or similar vision workloads can trigger pre-processing on newly captured visual data for those types of workloads to reduce query latency.
Similarly, stream prioritization or prefetching can be used to perform low-latency visual data loading or fetching based on historical trends and/or workflows. For example, the vision processing history can be used to prioritize certain data streams and/or pre-fetch data from memory for a particular application to improve query latency. Compared to metadata pre-provisioning, which involves expedited processing that is performed in advance, stream prioritization involves obtaining or moving visual data to a location where it will likely be needed (e.g., from a camera to certain processing nodes).
Cached visual analytics can be used to optimize visual processing using cached workflows, similar to incremental processing. For example, based on cached information regarding particular visual streams that have already been obtained and processed, along with the type of processing or workloads performed on those streams, subsequent vision processing dataflows may omit certain processing steps that have previously been performed and whose results have been cached. For example, a visual analytics application involves a number of primitive vision operations. The volume of computation can be reduced, however, by caching visual analytics results and reusing them for subsequent operations when possible. For example, when executing a visual analytics application, cached visual metadata resulting from prior processing can be searched to avoid duplicative computation. In some embodiments, for example, cached visual analytics may be implemented as follows:
1. Each primitive vision operation is tagged or labeled using a cache tag;
2. For each instance or stream of visual data (e.g., each stored video), any corresponding visual metadata that has already been generated is stored in a metadata database or cache;
3. If there is a cache tag hit for a particular primitive vision operation with respect to a particular instance or stream of visual data, then the particular primitive vision operation can be omitted and instead the existing visual metadata can be used; and
4. If there is a cache tag miss, however, the particular primitive vision operation is executed and the resulting metadata is cached in the metadata database for subsequent use.
Tensor factorization can also be used for distributed neural network inferencing in order to address the overfitting problem. For example, representative weights of consecutive neural network layers can utilize tensor factorization to “smooth out” the model.
Accordingly, in order to address the heterogeneity of devices with different types of hardware and/or software, the illustrated embodiments implement device-centric scheduling using a vision capabilities repository. In some embodiments, for example, the vision capabilities repository may include multiple implementations of a particular vision capability that are optimized for different hardware and/or software environments. In this manner, vision workloads can be scheduled or distributed across fog devices based on their respective types of resources and capabilities, along with per-resource telemetry information that identifies resource availability.
The basic principle is to abstract capabilities (e.g., face detection, gesture recognition) from their underlying kernels/implementations (e.g., SIFT-based implementations, deep neural network implementations). This type of abstraction provides the flexibility to deploy an arbitrary vision capability on a per-device basis. For example, using resource-based scheduling, heterogeneous resource types of different fog devices can be considered as a whole in order to determine the optimal task-to-device mapping across the various fog devices, and also identify the corresponding vision capability implementations that each device should use for its assigned tasks. Moreover, resource telemetry can be used to monitor resource availability of fog devices on a per-resource basis (e.g., CPU, GPU, FPGA, ASIC, and so forth) to further facilitate intelligent scheduling decisions. Further, the vision capability repository hosts collections of implementations of different vision capabilities, and may also provide a request-response service that allows a device to request an available implementation of a particular vision capability.
In this manner, device-centric scheduling can be used to improve end-to-end (E2E) performance (e.g., latency and bandwidth efficiency) and scalability for visual fog computing.
In some embodiments, users 1502 may submit search queries for visual data captured by cameras 1510a. Moreover, in order to respond to those queries efficiently, scheduling server 1504 may schedule or distribute vision processing workloads across the various fog devices 1510. In some embodiments, for example, scheduling server 1504 may perform intelligent scheduling decisions based on various criteria, such as the types of resources in the fog (e.g., the heterogeneous types of resources of the various fog devices 1510), resource telemetry information (e.g., the availability of fog resources on a per-resource-type basis), and the implementations of vision capabilities that are available in the vision capability repository 1506.
An example embodiment of the scheduling process, for example, is described below in connection with
The flowchart may begin at block 1602 by collecting the available vision capability implementations. In some embodiments, for example, the scheduling server continuously synchronizes the collection of available implementations of vision capabilities from the vision capability repository.
The flowchart may then proceed to block 1604 to collect the resource telemetry of fog devices. In some embodiments, for example, the scheduling server may collect the resource availability of all fog devices on a per-resource-type basis. For example, the scheduling server may collect information regarding the resource availability of CPUs, GPUs, FPGAs, ASICs, and/or any other resource type across all fog devices.
In this manner, based on the available vision capability implementations collected at block 1602, and the resource telemetry information collected at block 1604, the scheduling server can subsequently schedule vision workloads based on the optimal task-to-device mapping in the visual fog paradigm.
For example, the flowchart may then proceed to block 1606 to determine whether a new vision workload has been received from a user. In some embodiments, for example, a user may submit a new visual query, which may require a new vision workload to be scheduled or distributed across the fog devices.
If it is determined at block 1606 that a new vision workload has NOT been received, the flowchart may then proceed back to block 1602 to continue synchronizing the available vision capability implementations and collecting resource telemetry information until a new vision workload is received.
If it is determined at block 1606 that a new vision workload has been received, the flowchart may then proceed to block 1608 to re-schedule all pending workloads. In some embodiments, for example, receiving a new vision workload for a user may trigger the scheduling server to re-schedule all pending workloads to ensure the collective workloads are distributed across the fog devices in the most efficient manner possible (e.g., based on the optimal task-to-device mapping).
In some embodiments, for example, scheduling may be performed based on various criteria, such as the types of fog resources that are available, telemetry information for those resources, and the vision capability implementations that are available for those fog resources.
In some embodiments, for example, a schedule that adheres to the constraints of multiple resource types can be determined using integer linear programming (ILP). Integer linear programming (ILP) is a mathematical optimization or feasibility technique for solving or optimizing a mathematical model represented by linear relationships. In particular, ILP can be used to optimize a linear objective function, subject to additional linear equality and linear inequality constraints. As an example, an ILP problem can be expressed as follows:
Moreover, this ILP model can be used to determine an optimal schedule f that satisfies a specified objective (e.g., total network utilization), while also adhering to other additional constraints (e.g., device resource constraints). In the above ILP model, for example, x presents the collection of possible schedules f, K is the length of x, the objective term presents a scheduling objective to be minimized (e.g., total network utilization), and the inequality/equality constraints present any additional constraints (e.g., device, resource, network, mapping, and/or privacy constraints). A device resource constraint, for example, can be presented as an inequality constraint of the ILP model. For example, in order to take into account constraints of multiple resource types, they can be expended into multiple inequalities in the form of Ax≤b in the ILP model above.
Accordingly, based on the scheduling decisions, the scheduling server assigns each fog device zero or more tasks. In some embodiments, for example, a task may be specified in a tuple of the form t=(p, r), where p denotes the vision capability and r denotes resource type (e.g., p=face detection, r=Movidius processor).
The flowchart may then proceed to block 1610 to determine if an updated workload schedule is available. For example, after a new vision workload is received and the pending workloads are re-scheduled, the scheduling server may have an updated or improved workload schedule that needs to be distributed to the fog devices. In some embodiments, however, the scheduling server may only update the workload schedule if the newly generated schedule is better or more efficient than the current workload schedule.
If it is determined at block 1610 that the workload schedule has NOT been updated, the flowchart may then proceed back to block 1602 to continue synchronizing the available vision capability implementations and collecting resource telemetry until the current workload schedule is eventually updated.
However, if it is determined at block 1610 that an updated workload schedule is available, the flowchart may then proceed to block 1612 to push the updated schedule to all fog devices.
The flowchart may then proceed to block 1614 to receive requests from fog devices for vision capability implementations. For example, each fog device may query the vision capability repository to request implementations of vision capabilities for the tasks assigned to the particular fog device. In some embodiments, for example, the request from a particular fog device may identify each of its assigned tasks t.
The flowchart may then proceed to block 1616 to identify the appropriate vision capability implementations for each fog device. In some embodiments, for example, the vision capability repository may be a dictionary of key-value pairs in the form of (task t, implementation i), where an implementation i can be distributed in various forms (e.g., a dynamic linking library in C/C++). Accordingly, based on the task(s) t specified in the request from a particular fog device, the vision capability repository identifies the corresponding implementation(s) i for that fog device. In some embodiments, for example, the vision capability repository identifies the optimal implementation of each vision capability requested by a fog device based on the available resources of that fog device.
The flowchart may then proceed to block 1618 to distribute the identified vision capability implementations to each fog device. In this manner, each fog device can then perform its assigned tasks using the appropriate vision capability implementations.
At this point, the flowchart may be complete. In some embodiments, however, the flowchart may restart and/or certain blocks may be repeated. For example, in some embodiments, the flowchart may restart at block 1602 to continue scheduling vision workloads.
For example, end-users may subsequently submit visual search queries and/or user-defined vision functions (UVFs) 1709 associated with the visual data captured by the visual fog system. Accordingly, the UVFs 1709 are provided to a UVF compiler 1710, which compiles the UVFs 1709 into a vision dataflow graph 1711 that can be used to execute the UVFs. For example, the vision dataflow graph 1711 is provided to a distributed UVF execution framework 1712, which distributes or schedules workloads associated with the vision dataflow graph 1711 across the available fog nodes in the visual fog architecture.
After the workloads finish executing, the distributed UVF execution framework 1712 generates an output 1713 resulting from execution of the UVFs 1709. For example, the output 1713 may include, or may be derived from, a filtered stream of visual data and/or metadata 1707 generated by execution of the UVFs 1709. Moreover, in some embodiments, the resulting stream of visual data and/or metadata 1707 may then be stored in data storage 1706 for responding to subsequent visual search queries or UVFs.
Storage
As the volume of visual data generated in the real-world continues to grow, it is becoming increasingly common for visual data to be processed automatically by computers rather than manually reviewed by humans. Due to the increasing volume of visual data, however, data access has become a bottleneck in visual data processing, as existing visual data storage approaches suffer from various deficiencies.
To illustrate, image classification is a common visual data operation that uses a neural network to identify the contents of an image. For example, in machine learning, a convolutional neural network (CNN) is a type of feed-forward artificial neural network where the input is generally assumed to be an image. CNNs are commonly used for image classification, where the goal is to determine the contents of an image with some level of confidence. For example, a CNN is first trained for a specific classification task using a set of images whose object classes or features have been labeled, and the CNN can then be used to determine the probability of whether other images contain the respective object classes.
Visual data (e.g., images, video) must first be loaded from a storage system before it can be processed by a CNN. In the past, the data access latency has typically been less than the CNN vision processing latency, allowing the data access to be performed during the CNN processing. However, as hardware and software optimizations continue to improve the performance of CNN vision processing algorithms, the data access latency of existing solutions has become the bottleneck. Moreover, existing solutions typically store visual data in its original format rather than a format designed to aid with visual data processing, which further hinders performance.
Existing solutions are also unable to efficiently search visual data. For example, given a large data set (e.g., millions of images), the ability to efficiently identify a subset of relevant images using a query is important. The output of a CNN used for image classification typically includes a vector of values corresponding to the probability of various objects existing in an image. However, existing solutions typically use this information for the task at hand and then discard it, requiring the processing to be repeated for subsequent use. For example, a CNN used to process an image with a dog and a cat may provide a probability for both, but if the goal was to find images with dogs, the information about cats is typically lost or discarded, thus preventing future use. In this manner, a subsequent search for images that contain cats would typically require the CNN to be run again on each image.
Accordingly,
In the illustrated embodiment, for example, storage architecture 1800 includes a request server 1802 for receiving visual search queries from a client API 1801, a metadata database 1804, a visual compute library 1806, and a persistent data storage 1810, as explained further below.
In some embodiments, for example, storage architecture 1800 may provide a unified API 1801 for visual data access (e.g., for both visual data and metadata). For example, visual data is commonly stored directly as files or in various types of databases (e.g., key-value, relational, and/or graph databases). Visual metadata is typically stored in databases, for example, while images and videos are typically stored as files. Moreover, different types of file systems and databases provide API functions in various programming and/or query languages in order to enable users to access and store data. Accordingly, in some embodiments, visual storage architecture 1800 may be implemented with a unified API (e.g., JSON-based) that supports multi-modal queries for retrieving any type of visual data from any storage source. In some embodiments, for example, the unified API could be used to retrieve and/or combine visual metadata and the original visual data from different storage locations. The unified API may also allow certain types of processing to be performed on visual data before it is returned to the requesting user. Further, the unified API may allow users to explicitly recognize visual entities such as images, feature vectors, and videos, and may simplify access to those visual entities based on their relationship with each other and with other entities associated with a particular vision application.
Moreover, in some embodiments, a multi-tier lazy data storage approach may be used to store visual data more efficiently (e.g., using long- or short-term storage in different portions of the distributed edge-to-cloud network). For example, multiple storage tiers may be used to store visual data in different locations and for varying amounts of time based on the type or importance of the visual data. In some embodiments, for example, video cameras may store all video captured within the past day, gateways may store video with motion activities within the past week, and the cloud may store video associated with certain significant events within the past year.
Similarly, intelligent placement and aging of visual data across the storage tiers may further improve the data storage efficiency (e.g., determining where to store the visual data within the distributed edge-to-cloud system, when the data should be moved from hot to warm to cold storage, and so forth). For example, visual data and metadata can be distinguished and segregated based on data access patterns. Moreover, analysis friendly storage formats can be used to enable data to be read faster when needed for vision processing. These various data formats may be used to form the hot, warm, and cold tiers of data that can be mapped to various heterogeneous memory and storage technologies, based on the intended use and lifetime of the data. For example, storage tiers can be used to represent hot, cold, and optionally warm data. Hot data is accessed frequently; warm data is accessed occasionally; and cold data is accessed rarely (if ever). Accordingly, cold data may be stored on slower hardware since low access latency for retrieval of the data is less important. In this manner, intelligent decisions can be used to determine when and which portions of visual data should remain in the hot tiers and when it should be migrated to colder tiers, and which storage format should be used. For example, regions of interest may remain in hot storage in the analysis friendly format much longer than the entire image/video.
Metadata database 1804 is used to store metadata in a manner that facilitates efficient searches of visual data. For example, when performing image classification using a CNN, the resulting image-object relationships or probabilities can be stored as metadata, and the metadata can be used for subsequent searches of the images, thus eliminating the need to repeatedly process the images for each search. For example,
In some embodiments, storage architecture 1800 may store visual metadata as a property graph to identify relationships between visual data, such as images that contain the same object or person, images taken in the same location, and so forth. For example,
Moreover, in some embodiments, metadata database 1804 of storage architecture 1800 may be implemented as a persistent memory graph database (PMGD) to enable visual metadata to be searched more efficiently. For example, using persistent memory (PM) technology, a graph database containing the visual metadata can be stored both in-memory and persistently. In this manner, a persistent memory graph database (PMGD) can be designed to leverage a memory hierarchy with data structures and transactional semantics that work with the PM caching architecture, reduce write requests (addressing PM's lower write bandwidth compared to DRAM), and reduce the number of flushes and memory commits. This approach enables a graph database of visual metadata to be searched efficiently to identify relevant visual data.
Further, feature vector storage optimizations may be used to achieve fast searching of visual metadata. For example, feature vectors can be generated by various vision algorithms to identify regions or features of interest in visual data (e.g., faces, people, objects), and they are typically represented as vectors of n-dimensional floating-point values. Finding the nearest neighbor for a given feature vector is a common operation that is computationally expensive, especially at the cloud scale due to billions of potential feature vectors (e.g., a feature vector for each interesting region of each image or video frame). Accordingly, in some embodiments, feature vectors may be represented and stored as visual metadata using an efficient format. For example, visual metadata may be stored using an analysis-friendly array format that indicates where the feature vectors reside, and an index may be built on interesting dimensions within the metadata storage to narrow the search space.
Storage architecture 1800 also includes a separate data storage 1810 for storing the visual data itself, such as images or videos. Segregating the metadata and visual data in this manner enables each type of data to be mapped to the most suitable hardware in a heterogeneous system, thus providing flexibility for the request server 1802 to identify the most efficient way to handle a visual data request.
Moreover, storage architecture 1800 is also capable of storing visual data on data storage 1810 using an analytic image format designed to aid in visual processing. In the illustrated embodiment, for example, visual compute library (VCL) 1806 of storage architecture 1800 is designed to handle processing on analytic image formats 1807 in addition to traditional formats 1808. For example, visual compute library 1806 can implement an analytic image format 1807 using an array-based data management system such as TileDB, as described further with respect to
Deep learning neural networks, such as CNNs, are frequently used for image processing, including object/edge detection, segmentation, and classification, among other examples. Images are typically read from disk during both training and inferencing, for example, using background threads to pre-fetch images from disk and overlap the disk fetch and decode times with the other compute threads. However, compute cycles may still be wasted reading the images from disk and decompressing/decoding the images to prepare them for processing, thus reducing the overall throughput (e.g., images/second) of an image processing system.
Moreover, traditional lossy image formats (e.g., JPEG) are designed to compress image data by discarding high-frequency information that is not perceptible by humans. While the discarded information may be meaningless to humans, however, it can improve the accuracy and performance of deep learning neural networks used for image processing.
For example, images can be compressed either in a lossless or lossy manner. Lossless image compression preserves all the information in the image, while lossy compression takes advantage of visual perception and statistical properties to achieve better compression rates, but results in some data being lost. The JPEG compression algorithm is a commonly used lossy algorithm that is often used for images on the web. The JPEG algorithm is based on discrete cosine transforms (DCT), and discards high-frequency details that are not perceptible to the human eye, which results in much smaller image file sizes. However, in cases where exact image reproduction is required, or when the image will be edited multiple times, lossless compression is preferred. For example, PNG is an image file format that supports lossless compression using a bitmap image. With PNG, images are transformed using a filter type on a per-line basis, and then compressed using the DEFLATE algorithm. There are numerous other image formats with similar technologies behind them that are suitable for different applications and use cases. While a traditional lossless image format (e.g., PNG) could be used to retain all image data for image processing purposes, that comes at the cost of a lower compression rate.
Further, images stored using traditional formats (e.g., JPEG and PNG) must be converted into an internal array format before any processing can begin. For example, before any operations can be performed on images stored using traditional formats, the entire image file must be read from disk and decoded into an internal array format. In analytics, however, operations such as resizing and cropping are often performed before any sort of learning or understanding happens, thus rendering traditional image formats inefficient for image processing and analytics.
Accordingly, traditional image formats (e.g., JPEG and PNG) are designed for human consumption, and performing operations on them is often time-consuming and inefficient. Moreover, lossy image formats (e.g., JPEG) discard information that may be useful in machine learning, and thus are not well-suited for image processing. Moreover, while existing database management systems could be used to store images, they are not designed for image data and thus do not store image data efficiently.
The analytic image format 2200 of
In some embodiments, for example, analytic image format 2200 can be implemented by defining how the pixel data of an image is stored and accessed in an array-based format (e.g., using an array-based data storage manager such as TileDB). In this manner, image format 2200 enables efficiency in processing large images, which reduces the overall time for image analytics. As visual understanding algorithms get faster and the hardware to perform the algorithms gets better, the time to retrieve and process the images is becoming more and more significant. However, by using analytic image format 2200, storage and retrieval of images does not become a bottleneck in the visual processing pipeline.
For example, analytic image format 2200 allows an image to be stored as a lossless compressed array of pixel values. Accordingly, when image data is needed for processing, the image data does not need to be decoded before being processed, as required for traditional image formats. This improves the speed at which data is retrieved and made usable, yet still provides some level of compression. While this approach requires images to be written to the analytic image format 2200 prior to training or inference, the additional write overhead is minimal.
Moreover, because TileDB outperforms many array database managers for both sparse and dense data access, it is an ideal choice for implementing analytic image format 2200. In other embodiments, however, analytic image format 2200 can be implemented using any other type of array-based data manager or data format. The use of a fast, enhanced array storage system such as TileDB enables image format 2200 to eliminate slow reads of images from disk, and remove the in-loop conversion of traditional image formats to arrays.
Image format 2200 is also beneficial in applications where subarray accesses are common, such as accessing regions of interest in an image. For example, an array data manager such as TileDB can be used to improve the speed of common operations that are needed for image analytics, such as resize and crop, by enabling fast subarray accesses.
In some embodiments, the optimal tile size for analytic operations can be dynamically determined for each image. For example, in order to determine the optimal tile size for a particular image, a random portion of the image may be selected and then processed using different tile sizes and compression algorithms in order to determine the ideal tile size and compression for that image. Moreover, since image processing operations are often postponed until the data is actually needed, there is a period of time available to carry out the experimentation without impacting performance.
An image that does not fit perfectly into tiles of the selected tile size will have partially empty tiles that are padded with empty characters, as depicted in
For high-resolution images, image format 2200 improves the speed of common operations such as reading and writing, as well as the speed of operations used in image analytics, such as cropping and resizing. For example, storing images using image format 2200 improves read performance, as the images are compressed but not encoded, and thus do not need to be decoded when they are read from the file system. In addition, image format 2200 enables fast access to subarrays of image pixels, making cropping a simple matter of reading a particular subarray rather than reading the entire image and then cropping it to the appropriate size.
For example,
With respect to traditional image formats, for example, raw pixel data 5002 is encoded 5004, compressed 5006, and then stored 5010. With respect to analytic image formats, however, raw pixel data 5002 is compressed 5008 and then stored 5010, but the encoding step is omitted. While the resulting analytic image format may result in a larger file size on disk, the latency of data access operations (e.g., writes) and other image operations may be reduced.
Moreover, the read processing flow for traditional and analytic image formats may be implemented as the reverse of the write processing flow 5000. For example, with respect to traditional image formats, the encoded/compressed data is read from disk, decompressed, and then decoded into the original image. With respect to analytic image formats, the compressed data is read from disk and then decompressed into the original image, but the decoding step is omitted since the encoding step was omitted during the write processing flow 5000.
TABLE 1 illustrates an example analytic image format schema. In some embodiments, for example, the analytic image format schema of TABLE 1 could be implemented using an array-based database manager (e.g., TileDB) to store images as dense arrays.
The schema of TABLE 1 specifies parameters about the array that can be used to arrange the image data. Moreover, some parameters of the analytic image format are fixed, while others are determined on a per-image basis. For example, images have only two dimensions, a height and a width, thus fixing the number of dimensions as well as the names of the dimensions. The number of attributes is set to one, which means each cell holds the blue, green, and red (BGR) values for the corresponding pixel. All three values are generally read together, as a pixel is defined by all three values. In other embodiments, however, the color values may be stored separately. The intra-tile and array-level tile ordering is fixed to be row major. Row major order means that data is written and read from left to right in rows within a tile, and tiles are written and read in the same manner. This information allows the array database to efficiently perform subarray reads.
The dimensions and domain of the array depend on the resolution of the original image and therefore are calculated dynamically on a per-image basis. Since images often do not have an evenly divisible number of pixels in one or both dimensions, this occasionally results in the dimensions of an array not matching the original resolution of the image. This is reflected in TABLE 1, where the array height is one pixel larger than the image height. To make up the difference between an image dimension and an array domain, the image is padded with empty characters. An example of this can be seen in
Tile extents depend on the array dimensions and are calculated once the array dimensions are known. All tiles have the same height and width. The optimal number of tiles may vary based on image content and resolution, and thus in some embodiments, the optimal number of tiles may be determined on a per-image basis. For example, in order to determine the best tile size, a portion of the image may be randomly selected and tested using different tile sizes and compression algorithms to determine the best combination for that image. Since all operations are postponed until the data is actually needed, there is a period of time to carry out the experimentation that does not affect the performance. In other embodiments, however, a predefined minimum number of tiles per dimension (e.g., 4 tiles per dimension) may be used as a basis to determine tile height and width.
The compression algorithm used to compress the analytic image data has a fixed default (e.g., the LZ4 compression algorithm), but other compression algorithms can be set manually.
When a user creates an analytic image using VCL 5100, the analytic image schema is automatically set using the parameters described above in TABLE 1. VCL 5100 then creates a layer of abstraction with function calls of TileDB 5102 (e.g., the array-database manager used in the illustrated embodiment) combined with specialized transformation operations to provide an interface to the analytic image. VCL 5100 also extends the abstraction layer to OpenCV 5104, providing support for PNG and JPEG image formats. VCL 5100 uses OpenCV 5104 to perform both I/O and transformation operations on images that are stored in either PNG or JPEG format. For images stored in the analytic format, VCL 5100 handles the transformation operations and uses TileDB 5102 for I/O operations.
To initially store an image in the analytic format, the raw pixel data of an image is passed to VCL 5100 in some manner (e.g., as a path to a PNG or JPEG file stored on disk, an OpenCV matrix, a buffer of encoded pixel data, a buffer of raw pixel data, and so forth). This data is converted to a raw pixel buffer in order to write to the analytic format. Since the TileDB array schema for images has already been set at this point (e.g., using the parameters of TABLE 1), the TileDB functions can be used to write the data to disk.
Reading an image in the analytic format requires the metadata to be read first to determine the original image resolution. This ensures that only image data is read and that empty characters are ignored. The raw analytic-format or TileDB data is read into a buffer, keeping the data in the order in which it was written, which is referred to as “tile order” (e.g., as illustrated in
Crop, another frequently used operation in image processing, is used to retrieve a region of interest within an image for processing. Rather than reading the entire image and then selecting a sub-region (as is required for traditional image formats), the analytic or TileDB crop function uses the crop parameters to specify a subarray of the analytic image data. The subarray is then the only portion of the image that is read.
Resize, another frequently used operation in image processing, is used to resize the dimensions of an image (e.g., to either a smaller or larger size). The TileDB resize occurs after the image has been read, but while the data is still in tile order. VCL 5100 implements a version of resize for TileDB that uses a bilinear interpolation, following the OpenCV default. For example, in a linear interpolation, a new value is calculated based on two points; bilinear interpolation does this in two different directions and then takes a linear interpolation of the results. These points are identified by (row, column) in the original image. Given the data is in tile order, it is necessary to identify which tile each point is part of in order to locate the value of that point in the buffer. The resulting resized image buffer is in image order, although other approaches may be used to keep it in tile order.
Compression/Compressive Learning
The performance of large-scale visual processing systems can be improved using efficient compression algorithms and techniques for storing and processing visual data. The compression approaches of existing visual processing solutions, however, suffer from various deficiencies. For example, existing solutions require visual data to be fully decompressed before any processing can be performed (e.g., using deep learning neural networks). Moreover, existing solutions typically compress and store images individually, thus failing to leverage the potential compressive benefits of collections of similar or related images with redundant visual data.
Accordingly, this disclosure presents various embodiments for compressing and processing visual data more efficiently. In some embodiments, for example, neural networks can be designed to operate on compressed visual data directly, thus eliminating the need to decompress visual data before it can be processed. Moreover, context-aware compression techniques can be used to compress visual data and/or visual metadata more efficiently. For example, context-aware compression can be used to compress distinct instances of redundant visual data more efficiently, such as a group of images taken close in time, at the same location, and/or of the same object. Similarly, context-aware compression can be used to compress visual metadata more efficiently (e.g., using a context-aware lossless compression codec). In some embodiments, for example, visual metadata could be compressed by pre-training a convolutional neural network (CNN) to classify visual metadata, replacing long strings of visual metadata with shorter symbols (e.g., pre-defined human codes), performing multi-scale de-duplication on the visual metadata, and finally compressing the resulting visual metadata using a compression algorithm (e.g., the LZ77 lossless compression algorithm or another similar alternative).
In distributed visual analytics systems, image and video is often compressed before transmission (e.g., from the pixel domain to a compressed domain), and subsequently decompressed after transmission (e.g., back to the pixel domain) before any processing can be performed, such as deep learning using neural networks. As an example, image and video captured by edge devices may be compressed and transmitted to the cloud, and then decompressed by the cloud before any further processing begins.
This approach suffers from various disadvantages. First, extra computation is required to fully decompress the visual data before it can be processed, thus significantly increasing the total processing time (e.g., by up to 100% in some cases). For example, before processing can be performed, the visual data must be fully decompressed back to the pixel domain using hardware or software decoding. Accordingly, given that not all processors include built-in video decompression accelerators, decompression may incur an additional cost for video analytics.
Next, extra bandwidth is required to transmit the decompressed data between separate processing components (e.g., between a decompression engine and an analysis engine), thus significantly increasing bandwidth usage (e.g., by up to 20 times in some cases).
Moreover, the requirement to fully decompress visual data prior to processing precludes the ability to leverage a fully distributed neural network in the edge-to-cloud sense. For example, the use of distributed analytics to process visual data exclusively in the pixel domain requires the visual data to be analyzed at multiple scales.
Further, relying on the cloud to perform processing on visual data captured by edge devices often results in wasted transmission bandwidth, as many images or videos transmitted from the edge to the cloud may not contain any objects or features of interest. In many cases, for example, it could be possible to perform object detection and classification closer to the network edge (e.g., near the sensors that capture the visual data) using lower complexity analytics algorithms, potentially saving the transmission cost of insignificant or unimportant data.
Accordingly,
In some embodiments, for example, multi-domain cascade CNN 2400 may be a cascaded CNN that includes multiple decision stages. For example, in a first or early decision stage, a subset of the compressed domain visual data or features may be used (e.g., motion vectors) to attempt to generate an early decision. If the visual data cannot be detected or classified in the early stage, additional compressed domain data (e.g., motion prediction residuals) may be provided as input to a subsequent or late decision stage. Finally, for improved accuracy and/or in the event the late decision stage is unsuccessful, the visual data may be fully decompressed and a final decision stage may be performed using the decompressed visual data.
In the illustrated embodiment, for example, CNN 2400 includes an early decision stage (illustrated in
In some embodiments, for example, compressed visual data 2402 provided as input to CNN 2400 may first be partially decoded to separate and extract different syntax elements (e.g., motion vectors, macroblock (MB) coding modes, quantized prediction residuals), thus producing a subset of partial compression data 2404.
As shown in
If the early decision stage is unable to detect or classify the partial compression data 2404 using the first stage CNN 2405a, CNN 2400 may proceed to a late decision stage, as shown in
Finally, for improved accuracy and/or in the event the late decision stage is unsuccessful (e.g., the late decision stage is unable to detect or classify the additional compression data 2410 using the second stage CNN 2405b), CNN 2400 may proceed to a final decision stage, as shown in
Accordingly, the collective stages of multi-domain cascade CNN 2400 are depicted in
The described embodiments of multi-domain cascade CNN 2400 provide numerous advantages. First, visual data (e.g., images or video) does not need to be fully decompressed before its contents can be analyzed using deep learning neural networks, thus reducing memory usage and computation typically required for decoding or decompressing the visual data. Next, the cascading approach of CNN 2400 avoids the need to transmit certain compressed data to the cloud, such as when an early decision can be reached by an edge or fog node, thus improving bandwidth usage. Finally, a large portion of the overall analysis often occurs in the early decision stage, which typically involves a simplified CNN or machine learning model, thus reducing the overall computational complexity.
As discussed above, many visual analytics systems require visual data to be fully decompressed before any visual processing can be performed (e.g., using deep learning neural networks), which is an approach that suffers from various inefficiencies, including higher processing latency, additional transmission bandwidth, and so forth. Accordingly, this disclosure presents various embodiments of a deep learning neural network that is capable of analyzing compressed visual data directly. In particular, the described embodiments present a multi-domain CNN that uses butterfly operations to enable visual data processing in either the pixel domain or the compressed domain.
To illustrate, existing deep learning CNNs (e.g., inception or ResNet CNN models) typically repeat an inner module multiple times, and the inner module aggregates the results from multiple convolution layers and/or the original input at the end (analogous to a bottleneck). For example,
With respect to inner module 2800 of
Note that this approach, however, does not require multiple butterfly layers to be stacked within a single inner module, as the CNN does not have to perform a complete inverse DCT. For example, the goal of multiple convolution layers is to extract/transform the input data to a feature space where the fully connected layers can easily separate different clusters. Accordingly, the butterfly layers do not have to perform a complete inverse DCT, and instead, they can simply be designed to aid in extracting and transforming the input data into the feature space. In this manner, a complete or entire stack or organized butterfly layers does not need to be included in the CNN.
Moreover, the weights of each butterfly can be adjusted during the training phase, and thus the decision of whether to use the butterfly layers and/or how much to rely on them will be adjusted automatically.
In some embodiments, the butterfly operation may be performed in multiple stages. In each stage, for example, the butterfly operation may generate two outputs or channels using separate addition and subtraction operations (e.g., by computing the sum of two points over a large distance and the difference of two points over a large distance). For example, during a particular stage, the 1st and 4th points may be added together to compute their sum (1st point+4th point), and also subtracted to compute their difference (1st point-4th point). The points may then be shifted up cyclically and the process may be repeated for the next stage. For example, after each stage, the 4th point becomes the 3rd point, the 3rd point becomes the 2nd point, the 2nd point becomes the 1st point, and the 1st point becomes the 4th point. After the points are shifted, the next stage of the butterfly operation is performed by repeating the addition and subtraction on the 1st on 4th points (e.g., using the new ordering of points).
In
Accordingly, the butterfly operations can be incorporated into a CNN in this manner in order to enable processing of visual data in both the pixel domain and compressed domain (e.g., DCT domain), thus eliminating the requirement of fully decompressing visual data before analyzing its contents using a deep learning neural network. For example, rather than explicitly performing an inverse DCT transform to fully decompress visual data before processing it using a CNN, the CNN can instead be implemented using butterfly layers to inherently incorporate decompression functionality into the CNN, thus enabling the CNN to be provided with compressed data as input.
Many visual analytics systems require visual data to be decompressed before any processing can be performed, such as processing by a deep learning neural network. To illustrate,
In the illustrated embodiment of
In some embodiments, for example, 3D CNN 3200 may be designed to operate directly on compressed visual data (e.g., video frames) represented in the DCT domain using a 3D matrix. For example, in some embodiments, the DCT block indices may be represented by the x and y dimensions of the 3D matrix, while the DCT transform magnitude vectors may be organized along the z dimension. In this manner, the convolutional kernels in the first layer of the new CNN architecture can be implemented using 3D filters designed to better capture the spatial and frequency domain correlations and features of the compressed data, thus improving the performance of the CNN operation in the DCT domain.
The majority of common video and image encoding schemes use discrete cosine transforms (DCT) to convert spatial pixel intensities to frequency domain representations. The illustrated embodiment is based on the observation that once image data is split into 4×4 pixel blocks and passed through a transform such as DCT, the transformed data has different correlation properties than the original data. For example, with respect to a DCT transform, the DC coefficients of adjacent blocks are often strongly correlated, while the corresponding higher frequency AC coefficients of adjacent blocks may be similarly correlated.
Accordingly,
Next, as shown in
The illustrated embodiment provides numerous advantages, including the ability to directly process compressed visual data in an efficient manner, thus eliminating the need to decompress the data before analyzing its contents (e.g., using a deep learning neural network). In this manner, the overall computational complexity of visual analytics can be reduced. Moreover, because compressed or DCT domain data is quantized and thus represented using a more compact form than the original visual data (e.g., video frame), the overall CNN complexity may be further reduced compared to a conventional pixel-domain CNN. For example, with respect to visual data (e.g., images or video) compressed in certain compression formats such as JPEG or M-JPEG, the DCT coefficients are quantized, and typically the highest frequency components may be zeroed out by the quantization. Thus, the total volume of non-zero data processed by the CNN is reduced compared to the original image data. Accordingly, based on the data volume reduction of the compressed data (e.g., due to DCT coefficient quantization), the CNN complexity may be further reduced, and the training speed of convergence may improve.
Today, many people rely on the cloud for storing or backing up their photos. Typically, photos are stored as individually compressed files or units. In the current computing era, however, that approach is often inefficient. For example, people increasingly use their mobile devices to take photos, and each new generation of mobile devices are updated with cameras that support more and more megapixels, which results in larger volumes of photos that require more storage space. Moreover, people often capture multiple photos of the same object or scene during a single occasion, which often results in a close temporal correlation among those photos, along with substantial redundancy. Accordingly, due to the redundancy across similar photos, individually compressing and storing each photo can be an inefficient approach. For example, traditionally, each photo is compressed and saved independently using a particular image compression format, such as JPEG. By compressing each photo individually, however, current approaches fail to leverage the inter-picture correlations between groups of similar photos, and thus more storage space is required to store the photos. For example, two photos that are nearly identical would still require double the storage of a single photo.
Accordingly, in the illustrated embodiment, groups of similar or related photos are compressed and stored more efficiently. For example, context information associated with photos is extracted and used to identify similar or related photos, and similar photos are then compressed jointly as a group. The contextual information, for example, could be used to identify a group of pictures from a single user that were taken very close in time and/or at the same location. As another example, the contextual information could be used to identify a group of pictures taken by different users but at the same location. Accordingly, the identified group of similar photos may be compressed using video coding in order to leverage the inter-photo correlations and ultimately compress the photos more efficiently. In this manner, compressing related or correlated images using video compression rather than standard image compression can significantly reduce the storage space required for the photos (e.g., 2-5 times less storage space in some cases). Accordingly, this approach can be used to save or reduce storage in the cloud.
The flowchart may begin at block 3902 by first obtaining a new photo. In some cases, for example, the new photo could be captured by the camera of a mobile device. In other cases, however, any type of device or camera may be used to capture the photo.
The flowchart may then proceed to block 3904 to collect context information associated with the new photo. For example, when a photo is newly captured (e.g., by a mobile device), corresponding context information associated with the photo is collected, such as a timestamp, GPS coordinates, device orientation and motion states, and so forth.
The flowchart may then proceed to block 3906 to determine if a matching master photo can be identified for the new photo. In some embodiments, for example, the context information of the new photo is compared to the context information of other previously captured master photos to determine whether the new photo is closely correlated to any of the existing master photos. For example, if the photo is taken in the same location, within a certain amount of time, and with little phone movement compared to a master photo, it is likely that the new photo is highly correlated with the master photo. Further, in some embodiments, image feature matching techniques can then be applied to confirm the photo correlation. In some embodiments, for example, a scale-invariant feature transform (SIFT) may be used to determine whether a pair of photos are sufficiently correlated or matching.
If a matching master photo is identified at block 3906, the flowchart may then proceed to block 3908 to encode the new photo with the matching master photo. In some embodiments, for example, a video codec (e.g., H.264) may be used to compress the new photo as an inter-frame associated with the master photo. For example, video codecs typically provide inter-frame encoding, which effectively utilizes the temporal correlation between similar images to improve the coding efficiency.
In some embodiments, a master photo may include any photo that is compressed without reference to other parent or related images, while a slave photo may include any photo that is compressed with reference to a master or parent image (e.g., using inter-frame mode of a video codec). Accordingly, a slave photo must efficiently record or correlate relevant information of its master photo, so that when the slave photo needs to be decoded for display of the entire image, the associated master photo can be quickly identified.
If a matching master photo is NOT identified at block 3906, the flowchart may then proceed to block 3910 to encode the new photo by itself. For example, when the new photo does not match any of the existing master photos, the new photo is encoded without referencing any other photos, and the flowchart may then proceed to block 3912 to designate the new photo as a master photo, allowing it to potentially be compressed with other subsequently captured photos.
At this point, the flowchart may be complete. In some embodiments, however, the flowchart may restart and/or certain blocks may be repeated. For example, in some embodiments, the flowchart may restart at block 3902 to continue obtaining and compressing newly captured photos.
Privacy/Security
In distributed visual processing systems, it is important to implement effective privacy and security policies to protect sensitive visual data of underlying users or subjects (e.g., images or video with people's faces). Accordingly, in some embodiments, the visual fog architecture described throughout this disclosure may be implemented using a variety of privacy and security safeguards.
In some embodiments, for example, privacy-preserving distributed visual processing may be used in order to schedule or distribute vision workloads across available fog nodes in an efficient manner, while also adhering to any applicable privacy and/or security constraints.
Similarly, a multi-tiered storage approach may be used to store visual data in different locations and/or for different durations of time, depending on the particular level of sensitivity of the data. For example, the cloud may be used for long term storage of less sensitive or high-level visual data or metadata, while edge devices (e.g., on premise gateways) may be used for storage of highly sensitive visual data.
Moreover, certain vision operations may be implemented using privacy-preserving approaches. For example, for some vision applications (e.g., automated demographics identification), feature extraction and recognition may be implemented using cameras and sensors that capture top-down views rather than intrusive frontal views.
As another example, gateway cloud authentication may be used to securely authenticate gateways and/or other fog devices to the cloud using JSON web tokens.
As another example, wallets or distributed keys, along with MESH or GOSSIP based communication protocol, can be used to provide improved and more secure key management solutions.
Stream multiplexing may be used in application layer routing for streaming media, for example, by multiplexing visual sensors over multiple channels and introducing entropy to make channel prediction more difficult. For example, additional security can be provided by introducing entropy and other noise (e.g., chaff signals) designed to complicate channel prediction, thus thwarting efforts of malicious actors to pick up on video feeds.
As another example, a self-sovereign blockchain can be used to provide multi-tenant device identification. For example, the blockchain can be used to handle the orchestration and acceptance of device identities across multiple visual fog networks (e.g., even for legacy systems), thus allowing devices to assert their identity without relying on third party or centralized services. A self-sovereign blockchain can similarly be used for other purposes, such as managing a collection of distributed computing algorithms.
As another example, blockchain lifecycle management (e.g., managing the instantiation and lifecycle of blockchains) can be used to provide an additional level of security on blockchains used in a visual fog architecture. For example, blockchain lifecycle management can be used to ensure that a particular blockchain is implemented correctly and behaves as expected.
As another example, stakeholder management can be used to provide a set of protocols and frameworks to allow self-interests to be asserted, while arbitrating against conflicts in an equitable way.
In some embodiments, for example, computer vision and/or facial recognition technology may be used to identify human demographics. For example, demographics could be identified based on frontal and/or side facial features extracted using computer vision facial recognition technology. The use of frontal facial recognition technology in public, however, may implicate potential privacy concerns. Moreover, demographic identification is crucial across different domains and should not be limited to only frontal-based sensors and recognition techniques, particularly in the Internet-of-Things (IoT) era, which is projected to have over 20 billion connected devices by year 2020. Further, when limited to frontal-based vision sensors, it may be challenging to develop a demographics identification system that overcomes the person occlusion problem, while also providing wide processing viewing angles.
Accordingly, in the illustrated embodiment of
The process of training the demographics classification model is illustrated by blocks 4021-4024. At block 4021, a training database of top-view human data must first be obtained or generated. In some embodiments, for example, the training database may include data captured by top-view sensors 4015, such as camera images, infrared images, point clouds, and so forth. At block 4022, features that are typically representative of human demographics are then selected/trained from the database using feature extraction methodologies, such as principal component analysis (PCA), discrete cosine transforms (DCT), machine learning (e.g., deep learning using a neural network), and so forth. At block 4023, the selected/trained features are then provided as input to a process used to train a demographics classification model. At block 4024, the trained demographics model is then saved in the fog 4020 for subsequent use during the demographics identification process, as described further below.
The process of identifying human demographics is illustrated by blocks 4025-4029. At block 4025, sensor data is captured by edge devices 4010 using one or more top-view sensor(s) 4015, such as a vision camera, infrared camera, LiDAR sensor, and so forth. The raw sensor data (e.g., RGB images, thermal images, point clouds) is then transmitted from the edge 4010 to the fog 4020 in order to perform data pre-processing in the fog 4020 (e.g., on-premises), such as data transformations, de-noising, and so forth. At block 4026, person detection is then performed on the pre-processed input stream. In some embodiments, for example, the pre-processed input stream is analyzed to determine if a person is captured in the underlying visual data. As an example, pre-processed image data from a top-view camera may be analyzed to determine if the image contains a person, and if so, the portion of the image that contains the person may be extracted. At block 4027, features that are typically representative of human demographics are then selected or extracted from the detected person using feature extraction/machine learning techniques. At block 4028, the extracted features from block 4027 and the pre-trained demographics model from block 4024 are then used by a demographics classifier to classify the demographic attributes of the detected person. At block 4029, demographic information associated with the detected person is then identified based on the output of the demographics classifier.
The described embodiments of top-view demographics identification provide numerous advantages. As an example, the described embodiments enable demographic information to be accurately identified based on top-down views of humans captured using a single- or multi-sensor approach. Compared to a frontal view approach, for example, a top-down or aerial perspective provides a wider angle of view for processing, reduces the problem of blocking or occlusion of people captured by the sensors, and preserves depth information associated with people and features captured and processed by the system. In addition, the described embodiments are less privacy-intrusive, as they only capture top views of people rather than other more intrusive views, such as frontal views. The described embodiments also identify demographic information based on permanent or lasting anthropometry features rather than features that may change or vary. Moreover, unlike motion-based detection approaches, the described embodiments are operable using only static views or images and do not require continuous image sequences or videos. Further, the described embodiments can be leveraged for a variety of use cases and applications, including retail, digital surveillance, smart buildings, and/or other any other applications involving human sensing, person identification, person re-identification (e.g., detecting/tracking/re-identifying people across multiple monitored areas), and so forth.
The flowchart may begin at block 5302 by obtaining sensor data from a top-view sensing device. A top-view sensing device, for example, may be used to capture sensor data associated with the environment below the top-view sensing device (e.g., from a top-down perspective). In some embodiments, the top-view sensing device may include a plurality of sensors, including a camera, infrared sensor, heat sensor, laser-based sensor (e.g., LiDAR), and so forth.
The flowchart may then proceed to block 5304 to perform preprocessing on the sensor data, such as data transformations, filtering, noise reduction, and so forth. In some embodiments, for example, the raw sensor data may be transmitted to and/or obtained by a processor that is used to perform the preprocessing. For example, the preprocessing may be performed by an edge processing device at or near the network edge (e.g., near the top-view sensing device), such as an on-premise edge gateway.
The flowchart may then proceed to block 5306 to generate a visual representation of the environment below the top-view sensing device. The visual representation, for example, may be generated using the sensor data captured by the top-view sensing device (e.g., camera images, infrared images, point clouds, and so forth). In some embodiments, for example, the visual representation may be a three-dimensional (3D) representation or mapping of the environment from a top-down perspective. Moreover, in some embodiments, the visual representation may be generated at or near the network edge (e.g., near the top-view sensing device). For example, in some embodiments, an edge processing device (e.g., an on-premise edge gateway) may be used to generate the visual representation.
The flowchart may then proceed to block 5308 to determine whether a person is detected in visual representation. For example, if a person was located under the top-view sensing device when the sensor data was captured, then the visual representation generated using the sensor data may include a representation of the person from a top-view perspective. Accordingly, the visual representation may be analyzed (e.g., using image processing techniques) to determine whether it contains a person. In some embodiments, for example, the person detection may be performed at or near the network edge (e.g., near the top-view sensing device) by an edge processing device (e.g., an on-premise edge gateway).
If it is determined at block 5308 that a person is NOT detected in the visual representation, the flowchart may proceed back to block 5302 to continue obtaining and processing sensor data until a person is detected.
If it is determined at block 5308 that a person is detected in the visual representation, however, the top-view representation of the person may be extracted from the visual representation, and the flowchart may then proceed to block 5310 to identify one or more features associated with the person. In some embodiments, for example, the top-view representation of the person may be analyzed to identify or extract anthropometric features associated with the person (e.g., features or measurements associated with the size and proportions of the person). For example, in some embodiments, the anthropometric features may be identified by performing feature extraction using an image processing technique, such as a discrete cosine transform (DCT), principal component analysis (PCA), machine learning technique, and so forth. Moreover, in some embodiments, the feature identification or extraction may be performed at or near the network edge (e.g., near the top-view sensing device) by an edge processing device (e.g., an on-premise edge gateway).
The flowchart may then proceed to block 5312 to identify demographic information associated with the person (e.g., age, gender, race) based on the identified features. In some embodiments, for example, a machine learning model may be trained to recognize demographic information based on human anthropometric features. In this manner, the machine learning model can be used to classify the identified features of the person to recognize the associated demographic information.
In some embodiments, the demographics identification may be performed at or near the network edge (e.g., near the top-view sensing device) by an edge processing device (e.g., an on-premise edge gateway). Moreover, in some embodiments, the edge processing device may transmit the demographics information (e.g., using a communication interface) to a cloud processing device to perform further analytics, such as generating a heat map or a people profile.
At this point, the flowchart may be complete. In some embodiments, however, the flowchart may restart and/or certain blocks may be repeated. For example, in some embodiments, the flowchart may restart at block 5302 to continue obtaining and processing sensor data from a top-view sensing device.
In visual computing, multi-target multi-camera tracking (MTMCT) and target re-identification (ReID) are some of the most common workloads across different use cases. MTMCT involves tracking multiple objects across multiple views or cameras, while ReID involves re-identifying an object (e.g., by extracting robust features) even after the object undergoes significant changes in appearance. For example, in retail, MTMCT is often used to track shoppers within a store, while ReID may be used to extract and summarize robust features of shoppers so they can later be re-identified (e.g., using MTMCT) in different circumstances, such as when a shopper has a significant change in appearance or visits a different store.
Currently, there are no coherent end-to-end (E2E) solutions for performing MTMCT and ReID that are scalable to large-scale visual computing systems (e.g., with tens of thousands of camera streams or more). In particular, bandwidth limitations render it challenging to deploy such a system in a conventional cloud computing paradigm where cameras send continuous video streams to the cloud for processing. For example, due to the large volume of video data generated by such systems, it is not feasible to funnel all of that data to the cloud for processing. On the other hand, it is unlikely that edge devices near the source of the video data are capable of processing a complete visual processing workload in real time.
Moreover, privacy is also a challenge in scaling out such a system, as sending visual data to the cloud for processing may implicate privacy concerns. For example, in order to preserve customer privacy, many retailers will not allow any video or images to be transmitted out of their stores.
Accordingly,
In some embodiments, for example, vision workloads may be scheduled and executed across visual fog nodes based on specified privacy constraints. As an example, privacy constraints for an MTMCT and/or ReID workload may require tasks that output pictures with faces to remain on-premises (e.g., neither the tasks nor their output are assigned or transmitted beyond the premise or to the cloud), be anonymized (e.g., face-blurred), and/or be deployed only on devices with enhanced link security.
In some embodiments, for example, rather than funneling every bit of visual data to the cloud for processing, intelligent decisions can be made regarding how visual data and workloads are processed and distributed across a visual computing system. Based on the privacy requirements of a particular visual application, for example, a privacy boundary can be defined within the end-to-end paradigm of a visual computing system in order to achieve performance efficiency while also preserving privacy.
In some embodiments, for example, job partitioning can be used to partition a visual analytics workload into a directed acrylic graph (DAG) with vertices that represent primitive visual operations and edges that represent their dependencies. In this manner, the graph can be used to represent the various tasks and associated dependencies for a particular workload. Moreover, a privacy policy can be defined separately for each dependency. Similarly, a device connectivity graph can be used to represent the various devices and their connectivity in the edge-to-cloud paradigm, and a privacy level agreement (PLA) can be established for each edge of connectivity in the graph. In this manner, the edge-to-cloud architecture can be implemented to include a coherent management interface that performs end-to-end workload distribution without compromising privacy. For example, using the job partitioning approach described above, workload distribution effectively becomes a mapping problem of assigning the tasks of a workload onto devices in the edge-to-cloud paradigm. In some embodiments, for example, a global scheduler can be used to determine an optimal mapping between tasks and devices in order to maximize performance while preserving privacy constraints.
In the illustrated example, privacy is treated as an explicit constraint when performing task-to-device mapping to deploy the workload. In some embodiments, for example, workloads can be represented in linear forms to enable the mapping problem to be solved efficiently using state of the art integer linear programming (ILP) solvers.
In some embodiments, for example, when scheduling a particular workload on an edge-to-cloud network, the workload and the edge-to-cloud network may each be represented using a graph, such as a directed acrylic graph (DAG). For example, the workload and its underlying tasks may be represented by a workload or task dependency graph GT=(VT, ET), where each vertex v∈VT represents a task, and each edge (u, v)∈ET represents a dependency between task u and task v. Similarly, the edge-to-cloud network may be represented by a network or device connectivity graph GD=(VD, ED), where each vertex v∈VD represents a device in the network, and each edge (u, v)∈ED represents the connectivity from device u to device v.
Moreover, the privacy policy (PP) for each task dependency in the workload graph may be defined using a PP function p: ET→, such that the smaller the number (), the more vulnerable the data transmission. Similarly, the privacy level agreement (PLA) for each connectivity link in the device connectivity graph may be defined using a PLA function s: ED→, such that the smaller the number (), the more secure the link.
In this manner, based on the privacy policy (PP) and privacy level agreement (PLA) functions, a privacy constraint (PC) can be defined as s(d)≤p(e), ∀e∈ET, d∈f(e), where f: ET→xi=0kED is the mapping function from a particular workload to the edge-to-cloud paradigm. Essentially, f maps an edge in a workload graph to a path in an edge-to-cloud connectivity graph. For example, in the context of visual fog computing, f is a scheduling function that determines the particular fog devices that the tasks of a workload should be assigned to, along with the particular network connectivity links between pairs of fog devices that should be used for the data transmissions. Accordingly, the above privacy constraint (PC) requires the privacy level agreement (PLA) of a particular connectivity link to be capable of accommodating the privacy policy (PP) of a particular data transmission sent over that connectivity link. For example, in some embodiments, a data transmission of PP level 1 (unrestricted access) can only map to a link of PLA level 1 (high security), while a data transmission of PP level 2 (privacy-preserving) can map to connectivity links of PLA level 1 (high security) and PLA level 2 (limited security).
Moreover, in some embodiments, a visual fog schedule that adheres to the above privacy constraint (PC) can be determined using integer linear programming (ILP). Integer linear programming (ILP) is a mathematical optimization or feasibility technique for solving or optimizing a mathematical model represented by linear relationships. In particular, ILP can be used to optimize a linear objective function, subject to additional linear equality and linear inequality constraints. In some cases, for example, an ILP problem can be expressed as follows:
Moreover, this ILP model can be used to determine an optimal schedule f that satisfies a specified objective (e.g., total network utilization), while also adhering to other additional constraints, such as a privacy constraint and any other device, network, or mapping constraints. For example, when using the example ILP model above to perform visual fog scheduling, x presents the collection of possible schedules f, K is the length of x, the objective term presents a scheduling objective to be minimized (e.g., total network utilization), and the inequality/equality constraints present any additional constraints, such as device, network, mapping, and/or privacy constraints. The above privacy constraint (PC), for example, can be presented as an inequality constraint of the ILP problem.
The flowchart may begin at block 5402 by identifying a new workload. In some embodiments, for example, the new workload may include a plurality of tasks associated with processing sensor data captured by one or more sensors. For example, in some embodiments, the sensor data may be visual data captured by one or more vision-based sensors (e.g., a camera, infrared sensor, and/or laser-based sensor).
The flowchart may then proceed to block 5404 to generate a workload graph based on the workload. In some embodiments, for example, the workload graph may include information associated with the underlying tasks of the workload, along with the task dependencies among those tasks.
The flowchart may then proceed to block 5406 to generate or identify a device connectivity graph. In some embodiments, for example, the device connectivity graph may include device connectivity information associated with a plurality of processing devices, such as edge, cloud, and/or intermediary network processing devices. The device connectivity information, for example, may include information associated with the device connectivity links among the plurality of processing devices.
The flowchart may then proceed to block 5408 to identify a privacy policy associated with the workload and/or its underlying tasks. In some embodiments, for example, the privacy policy may comprise privacy requirements associated with the task dependencies among the workload tasks.
The flowchart may then proceed to block 5410 to identify privacy level information associated with the plurality of processing devices. In some embodiments, for example, the privacy level information may include privacy levels provided by the device connectivity links among the plurality of processing devices. Moreover, in some embodiments, the privacy level information may be specified by a privacy level agreement.
The flowchart may then proceed to block 5412 to identify a privacy constraint for workload scheduling based on the privacy policy and the privacy level information. In some embodiments, for example, the privacy constraint may require the privacy level of a particular connectivity link to be capable of accommodating the privacy policy of any task dependency mapped to that connectivity link for data transmission.
The flowchart may then proceed to block 5414 to determine a workload schedule. The workload schedule, for example, may include a mapping of the workload onto the plurality of processing devices. Moreover, in some embodiments, the workload schedule may be determined based on the privacy constraint, the workload graph, and the device connectivity graph. For example, in some embodiments, the workload schedule may be determined by solving an integer linear programming model based on the privacy constraint, the workload graph, and the device connectivity graph (e.g., as described in connection with
In some embodiments, the resulting workload schedule may then be distributed to the plurality of processing devices (e.g., via a communication interface) in order to execute the workload.
At this point, the flowchart may be complete. In some embodiments, however, the flowchart may restart and/or certain blocks may be repeated. For example, in some embodiments, the flowchart may restart at block 5402 to continue scheduling new workloads.
Accordingly, in the illustrated embodiments, a multi-tenant fog node can use a self-sovereign device identity in order to allow the node owner to retain an assertion of ownership even when the fog node participates in, or roams to, other fog networks. In some embodiments, for example, a self-sovereign identity blockchain may be used to register the identities of fog nodes or devices. A blockchain, for example, may be a dynamic list of records or blocks that are linked and/or secured using cryptographic approaches. In some embodiments, for example, each block in a blockchain may include a hash pointer linking to a previous block, a timestamp, transaction data, and so forth. Accordingly, in some embodiments, a blockchain can be used as a distributed ledger for recording transactions in an efficient, verifiable, and/or permanent manner. In visual computing, for example, before adding a device identifier for a new fog node, a blockchain may optionally be used to verify that the identifier has not been previously asserted by another node. Further, the public key used to verify the device identity of the fog node may also be contributed to the blockchain, allowing the device to later prove it is the rightful owner of its identity.
A new fog device 4430 that is seeking to be used in multiple fog networks 4410, but is not exclusive to any particular fog network, may not have sufficient resources or capabilities to create and maintain virtual sandbox environments for each of the fog networks. Moreover, each fog network 4410 may have a large set of its own local fog devices that are exclusive to that network and do not roam into other fog networks. Accordingly, reusing device identifiers may not pose a significant problem of duplicative identifiers until a new device 4430 with a conflicting identity roams into a particular fog network.
There is often a cost associated with changing the identity of a device, however, as credentials, access tokens, and application logic may be linked to the device identity. Moreover, the respective owners of devices with conflicting identifies have a self-interest in resolving the conflict (e.g., to avoid ownership disputes), but without bearing the cost. For example, the conflicting devices may respectively view each other as “foreign,” and thus each device may want the other “foreign” device to bear the cost of an identity change. Accordingly, to resolve the opposing self-interests of devices with conflicting identities, a blockchain 4420 may be used to provide a fair algorithm for giving preference to a device for its use of an identity. In some embodiments, for example, the device that first registered a particular identity with the blockchain 4420 is given preference in the event of a conflict.
Moreover, the blockchain may have a policy for preventing duplicative identity registrations, for example, by first checking for duplicates and only allowing registration if no duplicates exist. For example, duplicative identity detection may be performed by blockchain processing nodes as a requirement for vetting transaction blocks used for identity registration. In the illustrated call flow 4500, for example, each node performs the following steps:
(1) receive transaction request from new device: TXn+1={S1, “A71C3”}, where S1=Sign_Kalice(“A71C3”);
(2) compute hash H1=SHA256(“A71C3”);
(3) search hash tree of transaction attributes, where Bx-poss=Search(TxTree, H1);
(4) IF Bx-poss=“H1” THEN return ERROR_DUP_FOUND;
(5) ELSE IF Bx-poss=“ ” THEN add TXn+1 to the current block where CurrentBlock=[TXn+1, TXn, TXn−1, . . . , TXn−m];
(6) compute new current block hash BH=SHA256([TXn+1, TXn, TXn−1, . . . , TXn−m]);
(7) write BH to the blockchain at Bcurr-pos (current position); and
(8) insert the tuple (H1, BH, Bx-poss) into TxTree.
In some embodiments, however, a less restrictive policy may be used, such as a policy that does not check for duplicates during identity or name registration, and instead relies on dispute resolution to resolve duplicative identities. For example, at the time a device is on-boarded onto a new fog network, the blockchain can be consulted to determine if the identifier has previously been used, and if so, conflict resolution can be performed. The advantages of a less restrictive policy include improved performance and the ability to support mass registration workloads, among other examples.
In this manner, based on the illustrated embodiments of
Distributed computing interoperability depends on agreement among participating nodes regarding the particular algorithms used to process information at each node. In some cases, for example, algorithm agreement among nodes may depend on a central authority that manages a registry or database of algorithm identifiers. In this manner, distributed nodes must rely on the registry for selection of the appropriate algorithms, otherwise interoperability is not achieved.
This dependence on central authorities can lead to service disruptions, however, such as when a registry goes offline, a registry is slow to publish new algorithm identifiers (e.g., thus slowing the pace at which new algorithms can be deployed), a central authority becomes the target of politicizations (e.g., registration requests are held in ransom for processing fees, political favors, and/or other forms of manipulation that are not tied to the economics of the distributed computing application), and so forth. For example, these approaches are often highly centralized and may involve international or governmental institutions, which may be prone to politicizations and/or government regulation (e.g., net neutrality). Moreover, since agreement on which algorithms to use is fundamental to distributed computing, a centralized approach for managing algorithm identifiers can create an artificial bottleneck or choking point, and entities seeking to impose regulation or control can effectively leverage the centralized design to restrict or prevent interoperability among distributed computing nodes.
Accordingly, in the illustrated embodiments of
Accordingly, the described embodiments allow anyone that discovers a useful distributed computing algorithm to make that algorithm known and available to a large community. Blockchain networks, for example, are presumed to be large in number and open to large communities of users. In this manner, members of the community can build distributed computing systems without being hindered by bureaucratic roadblocks and oversight. As a result, the time between algorithm development and practical deployment can be minimized.
As an example, if a useful distributed computing algorithm 4430 is invented, discovered, and/or improved upon in a first fog network (e.g., fog network A 4810a), the first fog network may register the new algorithm in a self-sovereign blockchain 4420 used for algorithm identification. The blockchain processing nodes of the blockchain 4420 may then progressively vet the algorithm in order to provide progressively stronger assurances regarding its legitimacy (e.g., based on the computational properties and outcome of the algorithm). Moreover, a second fog network (e.g., fog network B 4810b) may subsequently be notified of the availability of the new algorithm, and may determine whether the new algorithm has been adequately vetted (e.g., by consulting the vetting status of the algorithm in the blockchain 4420). If the second fog network is satisfied with the vetting of the new algorithm, the second fog network may agree to use the algorithm. For example, in some embodiments, after the algorithm has been adequately vetted, the first fog network and second fog network may agree to begin using the new algorithm.
In some embodiments, the algorithm registration and vetting process may involve: (1) registration of a self-sovereign algorithm identifier (SSAI); (2) peer-review of a human-readable description of the algorithm; (3) machine analysis of a machine-readable representation of the algorithm (e.g., analysis by a logic processor to identify safe behavioral properties); and (4) execution of a machine-executable implementation of the algorithm (e.g., execution in a sandbox environment used to analyze expected behavior). Moreover, once a certain threshold (e.g., a majority) of blockchain processing nodes or evaluators achieve similar vetting results, the algorithm identity and its vetting criteria/results are recorded in a block of the blockchain 4420.
In the illustrated call flow 4900, for example, each blockchain processing node performs the following steps:
(1) receive transaction request from new device: TXn+1={S1, “91E21”}, where S1=Sign_Kalice(“91E21”, “Human-readable-description”, “Machine-readable-description”, “Machine-executable-implementation”);
(2) optional algorithm vetting (e.g., peer-review of a human-readable algorithm description, logical analysis of a machine-readable algorithm description/representation, sandbox execution of a machine-executable algorithm form);
(3) compute hash H1=SHA256(“91E21”);
(4) search hash tree of transaction attributes, where Bx-poss=Search(TxTree, H1);
(5) IF Bx-poss=“H1” THEN return ERROR_DUP_FOUND;
(6) ELSE IF Bx-poss=“ ” THEN add TXn+1 to the current block, where CurrentBlock=[TXn−1, TXn, TXn−1, . . . , TXn−m];
(7) compute new current block hash BH=sHA256([TXn+1, TXn, TXn−1, . . . , TXn−m]);
(8) write BH to the blockchain at Bcurr-pos (current position); and
(9) insert the tuple (H1, BH, Bx-poss) into TxTree.
Once the vetting process completes, the blockchain contains a vetted and registered instance of the algorithm and its associated identifier. In this manner, distributed computing nodes may then begin using the algorithm (e.g., based on the algorithm identifier and optionally its machine-readable and/or machine-executable forms).
Applications
The visual fog architecture and embodiments described throughout this disclosure can be used for a variety of large-scale visual computing applications and use cases, such as digital security and surveillance, business automation and analytics (e.g., retail and enterprise), transportation (e.g., traffic monitoring, navigation, parking, infrastructure planning, security or amber alerts), education, video broadcasting and playback, artificial intelligence, and so forth.
As an example, the described embodiments could be used to implement wearable cameras for first responders that are capable of automatically detecting events or emergency situations and performing certain responsive measures, such as notifying the appropriate personnel, triggering recording of the event by related or nearby cameras, and so forth.
As another example, the described embodiments could be used to implement a digital surveillance and security (DSS) system with people search or facial recognition capabilities across visual data streams from multiple different cameras, sensors, and/or locations.
As another example, the described embodiments could be used to implement a digital surveillance and security (DSS) system with license plate identification and fraud detection capabilities (e.g., identifying a car with a license plate that does not match the corresponding vehicle record, identifying multiple cars with same license plate, and so forth).
As another example, the described embodiments could be used to provide customer insights and analytics (e.g., for retail shoppers), such as an intra-store shopper trip summary (e.g., a list of products or departments interacted with by a shopper), an inter-store shopper trip summary (e.g., identifying repeat customers by differentiating between new and returning customers as they enter a store with a single or multiple locations), and so forth.
Similarly, the described embodiments could be used to provide visualization of customer or shopper insights and analytics (e.g., visualizing a graph representation of visual metadata for human consumption).
As another example, the described embodiments could be used to perform automated demographics identification in a privacy-preserving manner (e.g., using top-view cameras or sensors for demographic mapping of gender, age, race, and so forth).
As another example, the described embodiments could be used to perform heat mapping in retail stores or other brick-and-mortar environments to generate a representation of the crowd (e.g., using top-view sensors or cameras and/or multi-modal crowd emotion heat mapping). In some embodiments, for example, heat mapping could be leveraged for optimization of store layouts, among other examples.
As another example, the described embodiments could be used to implement multi-modal real-time customer reviews. For example, customer reviews and/or customer satisfaction information could be collected and analyzed in real-time using multi-sensory data, which can be translated into quantitative customer-to-customer reviews for any products or in-store activities of a particular store or brick-and-mortar environment.
Similarly, the described embodiments could be used to implement multi-modal retailer-shopper double review, which may focus on collection and analysis of both product reviews from customers and customer reviews from retailers.
As another example, the described embodiments could be used for automated customer satisfaction analysis. For example, visual data could be used to measure customer satisfaction at check-out based on non-verbal communication or body language. In this manner, customer satisfaction can be automatically inferred without requiring manual customer feedback (e.g., via a button or survey).
As another example, the described embodiments could be used to monitor the effectiveness of employee-customer interactions. For example, visual data could be used to measure and track the effectiveness of communication between customers and salespeople with respect to finding desired products or items. In some embodiments, for example, visual data could be used to track users within a store, identify customer-employee contact and interactions, and monitor the employee and/or customer responses.
As another example, the described embodiments could be used to provide dynamic ambience environments by identifying contextual information (e.g., relationships or actions) within a group of people. For example, visual data could be used to identify individuals and their associated contextual information to determine whether they are part of the same group (e.g., based on physical proximity and/or corresponding movement), and if so, to identify various parameters or characteristics of the group (e.g., a family shopping together in a store).
As another example, the described embodiments could be used to implement double auction real-time bidding (RTB). In some embodiments, for example, visual data could be used to implement multi-shopper, multi-bidder real-time bidding (RTB) for brick-and-mortar retailers.
As another example, the described embodiments could be used to monitor and detect changes to store layouts based on visual data and/or sensors.
As another example, the described embodiments could be used for robotic inventory tracking and logistics (e.g., using stationary and/or moving cameras to track inventory of retail stores, warehouses, offices, and so forth).
As another example, the described embodiments could be used for robotic equipment inspection (e.g., using computer vision technology to inspect the safety and/or health of equipment in a factory, plant, warehouse, store, office, and so forth).
As another example, the described embodiments could be used to provide automated tipping recommendations, for example, based on multi-sensory inputs and/or visual data reflective of factors that typically impact customer tipping behavior.
As another example, the described embodiments could be used for workplace automation, such as workplace quality control, employee monitoring, and so forth. In some embodiments, for example, visual data could be used to analyze employee emotions in order to improve productivity.
As another example, the described embodiments could be used for education and/or automated learning (e.g., using visual data to analyze student behavior in the classroom or at home in order to provide further assistance when appropriate).
As another example, the described embodiments could be used for video playback, such as user-centric video rendering, focused replays, and so forth. For example, user-centric video rendering could be used to perform focused rendering on 360-degree video by analyzing what the user is focusing on, and performing no or low-resolution processing on portions of the video that are outside the focus area of the user (e.g., for virtual-reality (VR) and/or augmented-reality (AR) applications). As another example, focused video replays could be used to automatically focus the rendering of a video replay on an area of interest, such as the portion of a sports replay where most players are located.
As another example, the described embodiments could be used to train artificial intelligence systems. In some embodiments, for example, visual data could be used to automatically generate ground truth information that can be used to train artificial intelligence or machine learning models, such as deep learning neural networks.
As another example, the described embodiments could be used to implement a ubiquitous visual computing witness, as described further below in connection with
These examples are merely illustrative of the limitless universe of visual applications and use cases that can be implemented using the visual fog architecture described throughout this disclosure.
Ubiquitous Visual Computing Witness
Existing approaches for evidence gathering suffer from various drawbacks. For example, there are no automated solutions for evidence gathering. Moreover, for electronic devices that continuously generate and/or stream data (e.g., sensors, cameras), the data retention policies can be rather short due to resource constraints. For example, data may only be retained for a couple of hours for vehicle dashboard/surveillance cameras due to limited storage capacity on an SD card. Further, manual approaches to evidence gathering often face various challenges. In particular, ownership of the data can present problems, especially in the absence of a shared authority that brokers access to the data.
Accordingly, this disclosure describes various embodiments of a ubiquitous visual computing witness. In particular, this disclosure proposes a novel new application of a ubiquitous witness implemented using an infrastructure of visual fog computing and visual fog networking for various example use cases. Further, the described ubiquitous witness addresses the challenges associated with data retention and data aggregation in connection with evidence gathering. In particular, the ubiquitous witness provides real-time data collection using visual fog computing. For example, when an interesting event (e.g., anomalous, unusual, rare) occurs, a snapshot of local data is locked (e.g., securely stored) by the subject device that detected the event, thus preventing the data from being overwritten. Further, the subject that detected the event notifies other relevant subjects (e.g., nearby subjects in many cases) in real time to lock their respective counterpart data snapshots. Further, the data retention problem can be addressed using quality of service (QoS) designations to prioritize which data to store versus expire (e.g., visual data versus other forms of data). Further, with visual fog computing, the data may be preprocessed to save bandwidth for the next step of data aggregation. In particular, the ubiquitous witness provides automated data aggregation using visual fog networking. For example, once data has been retained and optionally preprocessed by the various subject devices, the collection of data and metadata distributed across the respective subject devices is aggregated using visual fog networking and/or information-centric networking (ICN), thus allowing the respective data snapshots to be associated together and properly stored by the devices or nodes in the visual fog paradigm.
In this manner, the ubiquitous witness provides various advantages in connection with evidence gathering. For example, the central or key evidence associated with an incident is unimpeded by data retention policies, as the relevant subject devices are notified in real time to collect and lock their respective data snapshots. As another example, information-centric networking (ICN) and/or event-based data routing can be leveraged to provide a more efficient approach for collecting, aggregating, and/or routing data. In some embodiments, for example, relevant data can be routed by location, time, triggering event, source, destination (e.g., the responsible authorities), and/or any salient attributes describing key features of the event or incident.
The illustrated dataflow begins at block 5602 by determining whether an incident has been detected. In various embodiments, for example, an incident may be detected by processing sensor data using any suitable approach, such as identifying unexpected or outstanding changes in sensor data (e.g., data from internal sensors of vehicles and/or other types of equipment), performing event detection and recognition using computer-vision and/or machine learning techniques, and so forth.
Once an incident has been detected, that will trigger local data collection by the detecting device (block 5604) along with nearby data collection by any surrounding devices (block 5608), and the incident will also be given a name (block 5610).
For example, at block 5604, the detecting device performs local data collection, which may involve performing local analytics in order to acquire and/or generate data and/or metadata associated with the incident. The metadata, for example, may include or otherwise identify regions of interest (ROIs), geolocation names (e.g., addresses, cities, states, zip codes, landmarks, countries), tagging or labels on the scene of the incident derived using computer vision techniques (e.g., image classification or segmentation), weather, and/or any other contextual or circumstantial information associated with the incident, among other examples. Further, the relevant data may be designated with a high-priority QoS in order to ensure that the data is retained.
At block 5606, mechanisms are then employed to inform nearby devices of the incident, which enables those devices to lock down their data as well as perform any other actions that may be appropriate in the particular circumstances and/or use case. For example, in the context of a neighborhood watch, ZigBee and/or LTE may be used to notify nearby devices of an incident for data collection purposes, while LTE and/or dedicated short-range communications (DSRC) may be used in the context of automobile incidents. Regardless of the underlying wireless communication infrastructure, an alert message is routed from the device that detected the incident to its neighbors in order to trigger the process of locking down the relevant data of the nearby or neighboring devices (block 5608).
At block 5610, the incident is then given a “name” for identification, routing, and/or networking purposes. In some embodiments, for example, the incident name may be derived using an arbitrary combination of information associated with the incident, such as location, time, event, type of incident, priority/importance/fatalities, image/video captured of the event, and so forth. In this manner, the concatenation of this descriptive information can serve as the event name 5612 for purposes of requesting the underlying event data within an information-centric network (ICN), as described further below.
At block 5818, for example, an information-centric network (ICN) may be used to request, route, and aggregate the relevant data (e.g., local device data 5614 and/or nearby device data 5616) based on the incident name (e.g., the attributes associated with the data). In this manner, the network can then route or be queried for data associated with a particular incident of interest. For example, evidence of theft in a neighborhood watch may be requested and subsequently directed to a nearby police station 5822, while evidence of automobile speeding may be requested and subsequently directed to the department of motor vehicles (DMV) 5824 and/or an insurance company.
Moreover, the data associated with an incident may also be stored by certain network resources 5820 in order to satisfy subsequent queries relating to the incident. For example, at block 5826, users may query for evidence associated with an incident using the various attributes in the incident name. For example, a user might only know the approximate time and place of an incident for purposes of querying the network, and thus the network can disseminate the query to the relevant data stores, and those with relevant data can then reply.
In this manner, a ubiquitous witness can be implemented by leveraging visual fog computing to perform data processing and/or pre-processing at edge and/or fog nodes, followed by using visual fog networking and/or information-centric networking (ICN) to associate and/or aggregate the collection of evidence for an incident (e.g., data and metadata from various sources), timely route the evidence to the appropriate destination(s), and/or store the evidence in the appropriate data stores. Further, in some embodiments, a ubiquitous witness may be implemented within a network topology that leverages “converged node” routers, as described further below in connection with
Currently, however, the data generated by the various sensors in vehicles and roadside infrastructure is often isolated, as it is often stored in separate vendor- or application-specific clouds. There are various reasons for this, including but not limited to the following: (1) sensors within a single vehicle may be deployed by different manufacturers and their respective data may be routed to vendor-owned clouds; (2) obtaining data from nearby vehicles is often difficult, as the vehicles may be manufactured by different vendors that have different ways of capturing and storing information, including potentially different proprietary formats for sensory data; and (3) the entities involved may or may not be available or traceable in the first place.
Accordingly, with respect to existing solutions, sensor fusion exists strictly within each individual vehicle and is strictly siloed by each sensor vendor. Thus, sensor fusion can only be leveraged to aggregate image data collected from the field of view of each vehicle in isolation. Moreover, while existing video aggregation solutions can collect streams from multiple cameras (e.g., infrastructure-based solutions such as video surveillance or augmented reality), those solutions are ineffective when the sensors are owned by multiple different manufacturers, their “black box” data is generally proprietary, and there is very little cross-organization access or coordination. Accordingly, video aggregation and stitching using cameras in roadside infrastructure is typically limited to specific functions (e.g., traffic management services, awareness of road conditions, and/or surveillance). These respective approaches fall short in addressing anomaly detection in a globally distributed fashion, on the fly (e.g., in real time), and from multiple parties, particularly with respect to the fusion of data from both static and mobile entities.
Accordingly, this disclosure presents a solution for multi-dimensional anomaly event reconstruction for smart automated systems, such as autonomous vehicles and smart cities. In particular, the described embodiments enable multi-dimensional reconstruction of detected anomalies, where the multi-dimensional aspect refers to the ability to recreate an event from many different perspectives using multi-modal sensory information from multiple sources. These sources can include sensors, cameras, and/or other audio/visual elements, which may reside in multiple vehicles, roadside infrastructure, and/or the “black box” of information required by regulation for each vehicle. The objective is to solicit enough reconstruction data to recreate the event (which occurred at a particular point in time and space) in multiple dimensions (from different vantage points), while simultaneously scoping the solicitation within a bounded region of relevance.
Anomaly detection and event reconstruction is a multi-stage process that begins when a potential anomaly is detected. In some embodiments, for example, an anomaly may be detected by analyzing sensor data from one or more sources using any suitable processing techniques (e.g., computer vision, machine learning, artificial neural networks). The detection and identification of the anomaly (e.g., time, location and/or type of anomaly) may then serve as inputs to the event reconstruction process. In some embodiments, for example, the event reconstruction process may involve (1) digital witness identification, (2) data gathering, and (3) multi-dimensional event reconstruction, as described below.
Once the anomaly is detected, the first stage of the response is identifying the parties involved, including those involved directly in the anomaly and those that are digital witnesses to the anomaly, both of which may possess data needed to reconstruct the anomaly.
The second stage involves data gathering from the relevant parties that have been identified. The data gathering stage can be implemented using any suitable approach and/or technology, including a cloud-based solution that uses IP-based routing, and/or an edge-based solution that leverages information-centric networking (ICN) or ICN-based routing, among other examples. Further, in some embodiments, a network topology implemented with “converged node” routers may be leveraged to facilitate the data gathering process, as described further below in connection with
The third stage then uses the gathered data to perform multi-dimensional event reconstruction. In some embodiments, for example, three-dimensional (3D) event reconstruction may be used to recreate the event from a variety of different types of sensory data (e.g., image, audio, video), using different media-appropriate processing algorithms to assimilate all the sensory information in a time-synchronized fashion.
Accordingly, this disclosure presents a novel solution for general anomaly detection and event reconstruction, which can be leveraged for a variety of use cases, including connected and autonomous vehicles that live within a smart transportation infrastructure. In this manner, the sensing capabilities of multiple vehicles and the surrounding infrastructure can be leveraged for anomaly detection and event reconstruction purposes.
Moreover, the illustrated example portrays a snapshot in time and space of an automotive anomaly that involves a collision between two vehicles. The collision may be detected and recorded as an anomalous incident (e.g., with details of time and location) by any of the vehicles involved in the collision (either directly involved or indirectly involved as witnesses) and/or the roadside infrastructure or RSUs. In various embodiments, for example, an anomaly may be detected by processing sensor data using any suitable approach, such as identifying unexpected or outstanding changes in sensor data (e.g., data from internal sensors of vehicles and/or other types of equipment), performing event detection and recognition using computer-vision and/or machine learning techniques, and so forth.
Once an anomaly has been detected, the exact sequence of events leading up to the collision and slightly beyond can then be reconstructed, as described further below. For example,
With respect to IP-based embodiments, for example, the event reconstruction process begins by identifying all entities that may have useful data for the reconstruction process. In many cases, for example, it can be assumed that potentially relevant entities are those that reside within a region of interest referred to as the “anomaly coverage area” 5730 during a particular time window in which the anomaly occurred. The process of identifying these entities or “witnesses” can be triggered in multiple ways, including by vehicle(s) involved in the incident and/or nearby roadside infrastructure.
In some embodiments, for example, the process of identifying relevant entities or “witnesses” may be triggered by a vehicle that detects an anomaly, such as a vehicle directly involved in an incident or a nearby vehicle that merely witnesses the incident. As noted above, for example, many vehicles come equipped with sensors, cameras, and some level of automation, and the scope and sophistication of that technology is continuously increasing due to the ongoing development of connected and autonomous vehicles. Accordingly, a vehicle may leverage those various technologies to detect an anomalous incident or event (e.g., based on a sudden shock, air bag deployment, shock of the impact with another vehicle, unusual audio or video, and so forth). On detection of the anomalous event, the vehicle system triggers an alert and generates an automatic push request to send all sensor logs (including camera video streams) within a time window that extends pre- and post-incident to the vehicle's cloud systems (e.g., clouds of the vehicle manufacturer, insurer, municipal authorities, and so forth). The cloud, upon receipt of data associated with the anomalous event, may then employ an analytics engine to perform a search of the visual/sensory data using timestamp information to find license plates and other identifying information of the surrounding vehicles and/or infrastructure. The license plate information may then be mapped to the corresponding vehicle owners using a vehicle database (e.g., a publicly available database maintained by the Department of Motor Vehicles (DMV) in the U.S. or the equivalent in other countries). The cloud engine then sends out a request for information to the identified vehicles and their owners to reconstruct the event using an application server. Alternatively, the vehicles involved in the incident can process their own logs within their own local analytics engine even before the upload, and then through the local analytics engine, identify other vehicles and witnesses that were within the anomaly coverage area 5730.
Alternatively, the process of identifying relevant entities or “witnesses” may be triggered by roadside infrastructure located within the anomaly coverage area 5730. For example, roadside infrastructure may include various types of roadside units (RSUs) with edge and fog computing capabilities (e.g., storage, processing, communication/routing, sensors/cameras), such as traffic lights, street lights, lampposts, road signs, and so forth. In this manner, roadside infrastructure within the anomaly coverage area 5730 may detect, witness, or otherwise be alerted to an anomalous incident, and thus may trigger an alert or response to the incident. For example, the infrastructure equipment may send sensory data (e.g., video streaming feeds) from a time window that extends pre- and post-incident to an appropriate authority and/or cloud destination (e.g., the roadside unit (RSU) infrastructure cloud) to provide proof of witness. In this manner, the event reconstruction process is triggered by the roadside infrastructure, as described further below in connection with
As an example, once the data solicitation requests have been sent out, the responsive data may subsequently be aggregated at the entity that initially triggered the process, such as the RSU cloud or its associated municipal authority, the vehicle manufacturer, an insurance provider, and so forth. Further, in some cases, the same anomaly may be reconstructed by multiple different participants, such as the municipal authorities that own the RSUs (e.g., to improve traffic safety measures), the vehicle manufacturers (e.g., to improve their connected vehicle technology), and the insurance providers (e.g., to understand societal accident patterns), largely using the same process described above. After the data is gathered, it must then be stitched together to re-create the event. In various embodiments, for example, event reconstruction may be performed by consolidating sensor data from multiple sources and perspectives into a multi-dimensional representation using any suitable processing techniques, such as sensor fusion, computer vision, artificial neural networks, machine learning, and so forth.
In some embodiments, the various clouds may provide application programming interfaces (APIs) that allow other clouds or entities to query for and receive data in a format that can be parsed and decoded by the recipients. Further, each individual database within each individual cloud may be responsible for determining the legitimacy of the entities that request or query for data, including whether they have permission to receive the responsive data. Similarly, the receiver of the data may be responsible for determining the authenticity of data received in response to its query. In various embodiments, these types of safeguards can be implemented using an offline process, or they can take place in real-time if the requisite systems and authorizations for data access have been properly deployed and configured.
First, an ICN network is capable of routing data by name (which is often a series of attributes related to the data, such as its purpose, owner, location, and/or other contextual information) rather than by IP address, which enables location-independent routing. This capability can be leveraged through the use of a “fuzzy request,” which requests anomaly-related data based on a name that includes the time and location of the anomaly, as well as a delta of uncertainty, among other potential characteristics or attributes. In this manner, the fuzzy request acts as a query to the network itself, but without having to know the identities or IP addresses of the entities that detected the anomaly. Further, the delta of uncertainty allows responses from any entities that were located within a scoped physical region of interest—and that collected qualifying data of interest—at a time when the anomaly occurred plus or minus a threshold margin of error before and after the anomalous event.
Second, an ICN network offers caching natively at the routing layer, meaning as data passes through a router it may optionally be cached for later usage (e.g., when the source of the data is unavailable due to mobility, interference, disconnectivity, an energy-saving sleep mode, and so forth). This has the side effect of making data available in multiple locations and potentially much closer to the requesters, which saves precious and often limited network resources. As a result, when a fuzzy request is solicited, any witnesses or nodes with data that was collected within the approximate time and location of the anomaly may respond to the fuzzy request, regardless of whether the data was created or merely cached at that node.
The entities that have the capability to capture, create, process and/or store sensory data and make it available in a form that can be shared at the network routing layer are known as the “producers” or “publishers” in an ICN network, while the entities that request information are known as “consumers” or “subscribers.” The physical network itself is formed through the interconnection of the underlying participants using certain connectivity solution(s) (e.g., LTE, DSRC, Wi-Fi, 5G), and thus the network continuously changes as its participants join and leave (e.g., as vehicles connect to and disconnect from base stations while moving from one cell to another).
This framework enables distributed storage of anomaly information across vehicles and roadside infrastructure in the anomaly coverage area, while also allowing “subscriber” nodes (e.g., the DMV, insurance company investigation systems, vehicle manufacturer post-accident services) to pull together all pieces of the stored view of the anomaly and reconstruct the scene. Given that ICN does not require the underlying participants to be specifically known or identifiable, but instead merely requires the name and/or attributes of relevant data to be known, the process of gathering anomaly data from numerous sources is much more straightforward. Scene reconstruction can then take place in a cascaded and cooperative manner, as described further below.
For example, upon detecting the anomaly, the RSU sends a “scoped interest” packet over the ICN network (block 5902) using multicast in order to allow the packet to be heard and/or received by the respective entities that are within the coverage range of the network. In typical ICN networks, for example, a subscriber sends an “interest packet” that contains the precise name of the desired content or data. In this case, however, because the subscriber does not know the exact name of the content, the scoped interest packet merely identifies certain attributes of the desired content. Thus, the structure of a scoped interest packet differs from traditional ICN interest packets, as a scoped interest packet is essentially a “fuzzy” request for data that satisfies certain criteria rather than a request for data with a precise name. In some embodiments, for example, a scoped interest packet may request data based on an approximate location in space and time, such as data captured within certain deltas relative to a reference location x,y,z (RefLoc) and a reference time t (RefTime).
The naming convention for such an anomaly re-construction process is designed and implemented a-priori to facilitate the search. For example, each participating publisher follows the same naming convention such that the object names of the sensory data being recorded are named consistently, such as using the following format:
The geographical location could be a standardized format that is similar to the naming mechanism for map tiles used to construct vehicular maps. The date and timestamp reflect the local date and time for the particular geographical location. The entity name could reflect a unique identity for a vehicle or the infrastructure equipment. The query contains the geographical location, date, and timestamp in the same format as the naming convention to facilitate an efficient search.
In some embodiments, a scoped interest packet may also identify and/or include a “function” for each node to apply to any data collected in response to the interest packet, such as a particular processing function for aggregating and/or compressing responsive data received at the intermediate hops before it reaches the original subscriber, thus improving network efficiency.
The scoped interest packet is received by multiple vehicles or nodes (blocks 5904 and 5906), and each node searches for responsive data and may also forward the packet along to other nodes (blocks 5908 and 5910).
The process flow of an individual node is further illustrated and described in connection with
However, unlike a conventional ICN where the transaction ends once a node provides matching content in response to an interest packet, here the node forwards the interest packet to other potentially relevant nodes (block 6014) and also stores the packet in its pending interest table to avoid processing duplicative interest packets for the same query or incident. In order to bound the region where the interest packet propagates, a “scoped” region of interest is defined in the interest packet, such that the interest packet continues to propagate within the scoped region until the requisite “proximity” is exceeded. In some cases, for example, the scoped region may be defined as a certain radius around the anomalous event, such as a radius that corresponds to the maximum range or visibility of cameras that could potentially capture the event. If a particular node does not find any matching content for a given “scoped interest” packet, that is an indication that the query has reached its geographical boundary, and thus the node does not forward the packet any further (block 6012).
As responsive data packets are received by the various nodes and end points in response to the forwarded scoped interest packet, each node may wait for a certain timer T to expire in order to gather all responses before sending its own response to the scoped interest packet (block 6016). In this manner, the ultimate response contains a comprehensive collection of data from many content providers rather than only a single content provider. This approach enables a many-to-one type of transaction at the network layer, where a single request associated with a scoped interest packet can be satisfied by multiple providers who in turn may forward the scoped interest packet along, although whether each provider also accumulates responses from multiple providers may be dependent on the type of query and depth of the network. For example, given that many responses may be received at the intermediate nodes, the nodes may choose to aggregate and/or filter the received responses before sending them along with their own response (block 6018).
The flowchart begins at block 6102 by accessing sensor data captured by one or more sensors of a first device. The first device, for example, may be a processing device with one or more processors, sensors, and/or communication interfaces. The sensors can include any type of device used for capturing data associated with a surrounding or operating environment, such as vision sensors for generating visual representations of a surrounding environment (e.g., cameras, depth sensors, ultraviolet (UV) sensors, laser rangefinders (e.g., light detection and ranging (LIDAR)), infrared (IR) sensors, electro-optical/infrared (EO/IR) sensors), internal sensors of vehicles and other types of equipment, and so forth. In some embodiments, for example, the first device may be associated with a vehicle and the sensors may include a combination of vision sensors and internal mechanical sensors. Alternatively, the first device may be associated with a surveillance system and the sensors may include one or more vision sensors.
The flowchart then proceeds to block 6104 to determine that an incident occurred within the vicinity of the first device. In some embodiments, for example, the incident may be detected based on an analysis of the sensor data captured by the sensors of the first device. For example, an incident may be detected by processing sensor data using any suitable approach, such as identifying unexpected or outstanding changes in sensor data (e.g., data from internal sensors of vehicles and/or other types of equipment), performing event detection and recognition using computer-vision and/or machine learning techniques, and so forth. In some embodiments, for example, an anomaly may be identified during analysis of the sensor data, and it may be determined that the anomaly is indicative of a particular incident or event. For example, during steady-state operation, sensor signals may be monitored and/or recorded to establish a baseline, and an anomaly may be identified when the sensor signals deviate from the baseline, which may be determined using statistical analysis, thresholds, rates, and so forth. Alternatively, or additionally, data from visual sensors may be analyzed using image and/or video processing techniques to identify an incident and/or event captured in the visual data (e.g., using feature recognition, image classification, artificial intelligence, machine learning, artificial neural networks, and so forth). Further, in some embodiments, detection of an incident may be triggered based on the cumulative detection of multiple anomalies, incidents, and/or events over a period of time.
Alternatively, the first device may be notified of the incident by another device. For example, in some embodiments, the first device may receive a notification associated with the incident over an information-centric network (ICN), where the notification comprises a request for data associated with the incident, and where the request for data comprises an indication of a plurality of attributes associated with the incident, which the first device can use to identify relevant data associated with the incident.
The flowchart then proceeds to block 6106 to identify and preserve data associated with the incident. In particular, the sensor data captured by the first device may be analyzed in order to identify relevant sensor data that is associated with the incident. In some cases, for example, sensor data captured by the first device around the approximate time and location of the incident may be deemed relevant to the incident. Accordingly, the relevant sensor data may be preserved, such as by designating it with a high priority and/or quality of service (QoS) designation, and storing it on a suitable memory or storage device. Alternatively, in some embodiments, a snapshot of all available sensor data may be preserved in response to an incident, with different designations for the most relevant data versus the remaining data. In this manner, the full collection of sensor data will be available for subsequent processing and/or analysis (e.g., offline), if needed.
The flowchart then proceeds to block 6108 to notify and request data from other devices within the vicinity. For example, the first device may broadcast/multicast a notification with a request for data associated with the incident to other devices within the vicinity of the first device. In various embodiments, for example, the notification may be transmitted over an information-centric network (ICN) and/or an IP-based network. Further, in some embodiments, the notification may identify various attributes associated with the incident, and/or a name or identifier associated with the incident, which the receiving devices can use to identify relevant data associated with the incident.
The flowchart then proceeds to block 6110 to determine whether to continue forwarding the request. For example, when the notification is received by other devices within the vicinity of the first device, each receiving device may determine whether it has any relevant data associated with the incident, as well as whether to continue forwarding the notification. In some embodiments, for example, each receiving device may search for relevant data (e.g., sensor data) based on the attributes of the incident that were provided in the notification, such as the time and/or location of the incident. Moreover, if relevant data is identified by a particular device, that device may forward the notification and request for data to other devices within its vicinity. However, if no relevant data is identified by that device, the device may decline to forward the notification.
The flowchart then proceeds to block 6112 to consolidate, aggregate, and/or compress the relevant data identified and/or received by the respective devices. For example, as each device identifies its own relevant data and also receives relevant data from other devices that it forwarded the notification to, each device may consolidate, aggregate, and/or compress its collection of data before providing data in response to the notification. In this manner, all relevant data from the respective devices is eventually collected and consolidated in response to the notification associated with the incident.
The flowchart then proceeds to block 6114 to reconstruct the incident based on the consolidated data. In various embodiments, for example, incident reconstruction may be performed by generating a multi-dimensional representation of the incident from the consolidated data using any suitable processing techniques, such as sensor fusion, computer vision, artificial neural networks, machine learning, and so forth.
The flowchart then proceeds to block 6116 to send the consolidated data and/or reconstructed incident to an appropriate entity, such as a police department, DMV, insurance company, vehicle manufacturer, or any other person or entity that needs the data.
At this point, the flowchart may be complete. In some embodiments, however, the flowchart may restart and/or certain blocks may be repeated. For example, in some embodiments, the flowchart may restart at block 6102 to continue detecting, reconstructing, and/or reporting incidents of interest.
Converged Node
In the Internet-of-Things (IoT), there are increasing numbers of sensors connecting to networks. Many of them, such as cameras, are generating high-volume time-sensitive data streams. Because these sensors are joining at the very edges of a network, their data typically flows in the direction from the edge towards the cloud, which is the reverse direction from how the network was originally provisioned. Moreover, the size of a single data stream can be significant from a bandwidth perspective, and thus often results in latency constraints for real-time applications. Furthermore, many of these flows are continuous in nature. Thus, if many high-volume streams simultaneously arrive at and converge upon the same router, which is often the case in typical hierarchical and/or wireless network topologies, it is possible that these data flows will not fit either in memory, local storage, and/or along the next hop in the network (which may lead to denial of service if the router is overloaded with traffic beyond its capacity). Even when the flows are not high-volume or continuous, if the number of flows is extremely high, the aggregated bandwidth requirements still might exceed the available resources (both on and off platform).
These scenarios present a problem of data implosion. Current congestion control techniques are inadequate, as they only partially address the problem when there are simply too many data streams competing for router buffer queue resources in the network. For example, a normal layer 3 (L3) router has several strategies to manage an oversized incoming data stream: (1) simply do not forward packets that do not fit (e.g., drop them); or (2) indirectly signal the source, as well as intermediate routers, of the mismatch in required bandwidth versus available bandwidth (e.g., explicit congestion notification (ECN) in IP headers). At the transport layer (e.g., TCP), retransmission timeouts are used to detect congestion to reduce the flow of packets until the congestion is eased. At the application layer, a sender of data has the opportunity to (1) ask or negotiate configuration details up front before the data exchange begins, or (2) infer from routing hints along the path how to fit the data over the route and/or transform individual data streams (e.g., by taking a data flow out of its original format and converting it into a format that would fit the most constrained part of the path through the system), which can be achieved by dynamically renegotiating with the application how to fit the data over the network. Congestion can also be addressed by adding more resources, such as more bandwidth using a bigger and more expensive backbone for the aggregated flows, but that is often a very expensive and impractical solution given the flexibility and potential scale of IoT.
The existing congestion control techniques at the application, transport, and routing layer are insufficient, as the IoT capacity problem exists due to the increase in flows versus existing capacity, rather than a temporary congestion problem that relates solely to peak or busy hour traffic. Further, adding network capacity without incurring significant costs is not possible with the existing solutions.
Although congestion control algorithms seek to minimize data over-run at traditional router nodes in the network, congestion ultimately leads to a growing queue of packets, which increases end-to-end delay. When a queue overflows, meaning there is no more room to store arriving data, a router begins to drop packets, which may have a big impact on flows that require the delivery of in-order packets or even a modicum of impact on the delivery of flows that require reliable (but possibly out-of-order) delivery of packets, which might need to wait a roundtrip for a retransmission of any packet losses. For example, delay and/or packet loss can be highly disruptive, not to mention unacceptable for a camera mounted on a connected car whose data is meant to be archived as the black box or ground truth for anomalous car or environmental incidents.
Accordingly, this disclosure presents various embodiments of a “converged node” that is designed to efficiently route/process/cache visual data streams flowing in an N-to-1 manner in edge and fog computing systems. The converged node enables the continued transmission of data even when there is a vast difference between the amount of incoming data and resources available for the outgoing data. This is possible due to the collective processing and caching of contextually-related streams, such as when multiple cameras capture images from different but overlapping perspectives.
The basic idea of the converged node is to extend the functionality of a traditional L3 router to comprehend and prevent the implosion of IoT data flows (e.g., the many-to-1 nature of data flowing upstream from a dense camera deployment in a smart environment). Because multiple incoming streams are often contextually related (e.g., in space, time, and possibly other ways), converged node IoT routers offer a new strategy: process, analyze, and transform the streams together. For example, a converged node effectively takes multiple incoming streams and transforms them into a single output stream by performing transformative compute functions on them, such as analytics, machine learning, artificial intelligence, and so forth, which serves as a new kind of “compression.” The new emergent stream is then stored for re-use by contributing and/or nearby stakeholders.
The converged node implicates several new requirements for L3 routers in order to service IoT deployments that have a need for supporting reverse data flows at scale (e.g., high-volume, continuous, low-latency, and/or high-density flows) while avoiding the data implosion problem. For example, with respect to a converged node, the network data flows converge on the node and may be combined or merged from multiple incoming streams into a single outgoing stream (while in-flight and en-route to somewhere else). The implications of this convergence of flows are that the routing, processing, and compute sub-systems can be co-designed in hardware to support low-latency handling of the data (e.g., via an internal shared data bus or shared memory). The extended capabilities of the L3 router, as implemented by the new converged node, are described further below.
The functionality of a converged node can be implemented in software, although additional efficiencies can be gained in a hardware solution (e.g., using accelerators and/or FPGAs). For example, a hardware version can employ zero-copy design, which means it would avoid copying/moving data to each sub-system for consideration, and instead moves the functionality/service/operation to the data itself, subscribing to the philosophy of “moving the compute to the data.”
The converged node provides a solution to network congestion that is scalable in a cost-efficient manner and provides greater performance than existing solutions. For example, converged node routers can provide significant cost and performance benefits for network operators and other entities that deploy servers for edge computing in real networks, particularly with respect to video and visual data flows. Additionally, this solution is highly applicable to contexts where the end-to-end reverse flows pass through multiple converged nodes, such as when data is organized by geographic region and thus passes up a hierarchical or multi-tiered topology (e.g., layered administrative domains or clouds), which is often the case for video surveillance, energy monitoring, and healthcare use cases, among other examples.
The converged node also provides benefits relating to time-sensitive capabilities, such as the utility of high-resolution timestamps (e.g., both to meet time-constrained delivery and time-sensitive coordination across multiple streams), implementing time coordinated compute (TCC) capabilities, and so forth.
As noted above, a converged node can be implemented by extending the functionality of a traditional router, such as an L3 router. The process flow within this extended “converged node” router may be as follows: (1) comprehend and manage the serious data implosion problem for IoT edge and fog systems; (2) process multiple contextually-related streams simultaneously; (3) look beyond packet headers to analyze if and how a stream is interesting or notable; (4) look beyond packet headers to analyze if and how the streams are related and might be “compressed”, in the broader sense of the term; (5) identify a functional strategy to emerge with a single new stream going forward, which may employ different functions for different scenarios, necessitating the ability to dynamically request and fetch functions/services/methods; (6) cache the new stream for re-use by contextually-related stakeholders, particularly those that are mobile, wireless, low-power, and/or proximate, for greater resource efficiency; (7) route the new stream in the reverse direction, which is where it was headed; (8) provide the option to route the new stream in other directions, e.g., within a local edge cloud and/or across neighboring clouds; (9) co-design the routing, computing and caching sub-systems to minimize time delay on platform; (10) employ zero-copy HW design techniques to implement efficient hand-off of data between sub-systems; and (11) leverage time-synchronization capabilities in the platform to synchronize across multiple data streams, in preparation for the analysis/transformation phase of the N-to-1 process.
There are various ways in which streams might be considered to be contextually related to each other. For example, in some embodiments, a broad correlation between the name of the content and different streams may be used. In an ICN context, all the streams that respond to a specific ICN Interest packet asking for data with a particular content “name” are considered contextually-related. The system could re-use the entries in the Pending Interest Table (PIT) as the entries in a directory for “contextually-related” groups. In a non-ICN context, the system still can leverage the name of the content or name-like attributes to organize streams into contextually-related groups, such as the URL/URI of the content, or the transport packet header tuples (e.g., sender and receiver addresses and port numbers), or the streams originating from the same subnet address (e.g., same subnet address pre-fix). The use of name-based techniques is beneficial over using deep-packet inspection of the content itself, which is a highly resource-intensive process. Alternatively, in some embodiments, in the visual realm, separate from ICN, streams within a “region of interest” (e.g., proximity in space and time) may be organized into contextually-related groups of packets.
In some embodiments, for example, a converged node may include the following capabilities on top of traditional router functionality: (1) analyzing data in data flows beyond packet headers; (2) defining what it means to be contextually-related (e.g., within a geographic region of interest, owned by same vendor, running the same version of malware software, and so forth); (3) collectively processing N contextually-related incoming data flows; (4) recommending a function to perform across these flows to “compress” them before they are routed/cached; (5) transforming the N streams into a single, new, less-resource-consuming data flow; (6) potentially caching/storing the new stream locally, such as for request and retrieval by nearby proximate sensors that may have contributed to the new stream and/or by those who have an interest in the results; and (7) routing the new data stream to its next intended upstream destination, which may be the northern direction in which the data was originally flowing, but may also include a broader dissemination, such as in the East-West direction to peer clouds or in the southern direction to interested parties.
While these functions are generally described as being performed at a single converged node in the reverse data flow path (e.g., as shown in
The converged node differs from previous L3 router solutions in that it requires the router to look inside the data stream beyond the packet header but rather focus on the content “packets payload,” which is the more traditional focus of an L3 router. Additionally, the converged node is capable of performing its described functionality while the data streams are in-flight versus post-facto. The converged node router performs an N-to-1 transformation, which may represent a range of processing capabilities, including but not limited to compression, encryption, transcoding, labeling, aggregation/grouping some flows into larger flows based on contextual commonality, sub-sampling, combination (e.g., stitching), and analytics (e.g., which broadly refers to any type of analysis, whether it is statistical analysis, machine learning (ML), deep learning (DL) or some other form of artificial intelligence or machine learning).
With respect to the analytics functionality of the converged node, the process may take the data out of its original data format and convert it into another format. This might be the case with summarization of phenomena within the data stream, such as object recognition and object counting, or with the labeling of phenomena within the data stream, to create a new stream of metadata that tracks the appearance of interesting events in the visual field of view.
A key benefit of the proposed method is that it applies to any dense IoT deployment suffering from N-to-1 data implosion, whether or not the streams are video/visual in nature. Furthermore, this technique couples nicely with information-centric network architectures, although it is not dependent on them.
There are numerous novel concepts associated with the converged node, including: comprehending and managing the serious data implosion problem for IoT edge and fog systems; processing multiple contextually-related streams simultaneously; looking beyond packet headers to analyze if and how a stream is interesting or notable; looking beyond packet headers to analyze if and how the streams are related and might be “compressed” (in a broader sense of the term); identifying a functional strategy to emerge with a single new stream going forward, which may employ different functions for different scenarios, necessitating the ability to dynamically request and fetch functions/services/methods; caching the new stream for re-use by contextually-related stakeholders, particularly those that are mobile, wireless, low-power, and/or proximate, for greater resource efficiency; routing the new stream in the reverse direction, which is where it was headed; providing the option to route the new stream in other directions, such as within a local edge cloud and/or across neighboring clouds; co-designing the routing, computing, and caching sub-systems to minimize time delay on the platform; employing zero-copy hardware design techniques to implement efficient hand-off of data between sub-systems; and leveraging time-synchronization capabilities in the platform to synchronize across multiple data streams, in preparation for the analysis/transformation phase of the N-to-1 process.
The flowcharts and block diagrams in the FIGURES illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various aspects of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order or alternative orders, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.
The foregoing disclosure outlines features of several embodiments so that those skilled in the art may better understand various aspects of the present disclosure. Those skilled in the art should appreciate that they may readily use the present disclosure as a basis for designing or modifying other processes and structures for carrying out the same purposes and/or achieving the same advantages of the embodiments introduced herein. Those skilled in the art should also realize that such equivalent constructions do not depart from the spirit and scope of the present disclosure, and that they may make various changes, substitutions, and alterations herein without departing from the spirit and scope of the present disclosure.
All or part of any hardware element disclosed herein may readily be provided in a system-on-a-chip (SoC), including a central processing unit (CPU) package. An SoC represents an integrated circuit (IC) that integrates components of a computer or other electronic system into a single chip. The SoC may contain digital, analog, mixed-signal, and radio frequency functions, all of which may be provided on a single chip substrate. Other embodiments may include a multi-chip-module (MCM), with a plurality of chips located within a single electronic package and configured to interact closely with each other through the electronic package. In various other embodiments, the computing functionalities disclosed herein may be implemented in one or more silicon cores in Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs), and other semiconductor chips.
As used throughout this specification, the term “processor” or “microprocessor” should be understood to include not only a traditional microprocessor (such as Intel's° industry-leading x86 and x64 architectures), but also graphics processors, matrix processors, and any ASIC, FPGA, microcontroller, digital signal processor (DSP), programmable logic device, programmable logic array (PLA), microcode, instruction set, emulated or virtual machine processor, or any similar “Turing-complete” device, combination of devices, or logic elements (hardware or software) that permit the execution of instructions.
Note also that in certain embodiments, some of the components may be omitted or consolidated. In a general sense, the arrangements depicted in the figures should be understood as logical divisions, whereas a physical architecture may include various permutations, combinations, and/or hybrids of these elements. It is imperative to note that countless possible design configurations can be used to achieve the operational objectives outlined herein. Accordingly, the associated infrastructure has a myriad of substitute arrangements, design choices, device possibilities, hardware configurations, software implementations, and equipment options.
In a general sense, any suitably-configured processor can execute instructions associated with data or microcode to achieve the operations detailed herein. Any processor disclosed herein could transform an element or an article (for example, data) from one state or thing to another state or thing. In another example, some activities outlined herein may be implemented with fixed logic or programmable logic (for example, software and/or computer instructions executed by a processor) and the elements identified herein could be some type of a programmable processor, programmable digital logic (for example, a field programmable gate array (FPGA), an erasable programmable read only memory (EPROM), an electrically erasable programmable read only memory (EEPROM)), an ASIC that includes digital logic, software, code, electronic instructions, flash memory, optical disks, CD-ROMs, DVD ROMs, magnetic or optical cards, other types of machine-readable mediums suitable for storing electronic instructions, or any suitable combination thereof.
In operation, a storage may store information in any suitable type of tangible, non-transitory storage medium (for example, random access memory (RAM), read only memory (ROM), field programmable gate array (FPGA), erasable programmable read only memory (EPROM), electrically erasable programmable ROM (EEPROM), or microcode), software, hardware (for example, processor instructions or microcode), or in any other suitable component, device, element, or object where appropriate and based on particular needs. Furthermore, the information being tracked, sent, received, or stored in a processor could be provided in any database, register, table, cache, queue, control list, or storage structure, based on particular needs and implementations, all of which could be referenced in any suitable timeframe. Any of the memory or storage elements disclosed herein should be construed as being encompassed within the broad terms ‘memory’ and ‘storage,’ as appropriate. A non-transitory storage medium herein is expressly intended to include any non-transitory special-purpose or programmable hardware configured to provide the disclosed operations, or to cause a processor to perform the disclosed operations. A non-transitory storage medium also expressly includes a processor having stored thereon hardware-coded instructions, and optionally microcode instructions or sequences encoded in hardware, firmware, or software.
Computer program logic implementing all or part of the functionality described herein is embodied in various forms, including, but in no way limited to, hardware description language, a source code form, a computer executable form, machine instructions or microcode, programmable hardware, and various intermediate forms (for example, forms generated by an HDL processor, assembler, compiler, linker, or locator). In an example, source code includes a series of computer program instructions implemented in various programming languages, such as an object code, an assembly language, or a high-level language such as OpenCL, FORTRAN, C, C++, JAVA, or HTML for use with various operating systems or operating environments, or in hardware description languages such as Spice, Verilog, and VHDL. The source code may define and use various data structures and communication messages. The source code may be in a computer executable form (e.g., via an interpreter), or the source code may be converted (e.g., via a translator, assembler, or compiler) into a computer executable form, or converted to an intermediate form such as byte code. Where appropriate, any of the foregoing may be used to build or describe appropriate discrete or integrated circuits, whether sequential, combinatorial, state machines, or otherwise.
In one example, any number of electrical circuits of the FIGURES may be implemented on a board of an associated electronic device. The board can be a general circuit board that can hold various components of the internal electronic system of the electronic device and, further, provide connectors for other peripherals. More specifically, the board can provide the electrical connections by which the other components of the system can communicate electrically. Any suitable processor and memory can be suitably coupled to the board based on particular configuration needs, processing demands, and computing designs. Other components such as external storage, additional sensors, controllers for audio/video display, and peripheral devices may be attached to the board as plug-in cards, via cables, or integrated into the board itself. In another example, the electrical circuits of the FIGURES may be implemented as stand-alone modules (e.g., a device with associated components and circuitry configured to perform a specific application or function) or implemented as plug-in modules into application specific hardware of electronic devices.
Note that with the numerous examples provided herein, interaction may be described in terms of two, three, four, or more electrical components. However, this has been done for purposes of clarity and example only. It should be appreciated that the system can be consolidated or reconfigured in any suitable manner. Along similar design alternatives, any of the illustrated components, modules, and elements of the FIGURES may be combined in various possible configurations, all of which are within the broad scope of this specification. In certain cases, it may be easier to describe one or more of the functionalities of a given set of flows by only referencing a limited number of electrical elements. It should be appreciated that the electrical circuits of the FIGURES and its teachings are readily scalable and can accommodate a large number of components, as well as more complicated/sophisticated arrangements and configurations. Accordingly, the examples provided should not limit the scope or inhibit the broad teachings of the electrical circuits as potentially applied to a myriad of other architectures.
Numerous other changes, substitutions, variations, alterations, and modifications may be ascertained to one skilled in the art and it is intended that the present disclosure encompass all such changes, substitutions, variations, alterations, and modifications as falling within the scope of the appended claims.
Example Implementations
The following examples pertain to embodiments described throughout this disclosure.
One or more embodiments may include an apparatus, comprising: a memory to store sensor data captured by one or more sensors associated with a first device; and a processor comprising circuitry to: access the sensor data captured by the one or more sensors associated with the first device; determine that an incident occurred within a vicinity of the first device; identify a first collection of sensor data associated with the incident, wherein the first collection of sensor data is identified from the sensor data captured by the one or more sensors; preserve, on the memory, the first collection of sensor data associated with the incident; and notify one or more second devices of the incident, wherein the one or more second devices are located within the vicinity of the first device.
In one example embodiment of an apparatus, the circuitry to determine that the incident occurred within the vicinity of the first device is further to: analyze the sensor data captured by the one or more sensors; identify an anomaly associated with the sensor data; and detect the incident based on the anomaly associated with the sensor data.
In one example embodiment of an apparatus, the circuitry is further to: receive a notification associated with the incident over an information-centric network, wherein the notification comprises a request for data associated with the incident, and wherein the request for data comprises an indication of a plurality of attributes associated with the incident; and determine whether the sensor data captured by the one or more sensors corresponds to the plurality of attributes associated with the incident.
In one example embodiment of an apparatus, the plurality of attributes associated with the incident comprises a time and a location of the incident.
In one example embodiment of an apparatus, the circuitry is further to generate an identifier for the incident, wherein the identifier is generated based on one or more attributes associated with the incident.
In one example embodiment of an apparatus, the circuitry to notify the one or more second devices of the incident is further to: broadcast a notification associated with the incident over an information-centric network, wherein the notification comprises a request for data associated with the incident, and wherein the notification is received by the one or more second devices.
In one example embodiment of an apparatus, the circuitry is further to: receive one or more second collections of sensor data associated with the incident, wherein the one or more second collections of sensor data are received from the one or more second devices; and generate a consolidated collection of sensor data associated with the incident, wherein the consolidated collection of sensor data is generated from the first collection of sensor data and the one or more second collections of sensor data.
In one example embodiment of an apparatus, the circuitry is further to send the consolidated collection of sensor data to an entity associated with the incident.
In one example embodiment of an apparatus, the circuitry is further to generate a reconstructed representation of the incident based on the consolidated collection of sensor data.
In one example embodiment of an apparatus, the incident comprises an automobile accident, and wherein the first device is associated with: a vehicle involved in the automobile accident; or roadside infrastructure located near the automobile accident.
In one example embodiment of an apparatus, the first device is associated with a surveillance system, and wherein the one or more sensors comprise one or more cameras.
One or more embodiments may include a system, comprising: a plurality of sensors to capture sensor data associated with an operating environment of a first device; a communication interface to communicate with one or more second devices; and a processor to: determine that an incident occurred within a vicinity of the first device; identify a first collection of sensor data associated with the incident, wherein the first collection of sensor data is identified from the sensor data captured by the plurality of sensors; preserve the first collection of sensor data associated with the incident; and notify the one or more second devices of the incident, wherein the one or more second devices are located within the vicinity of the first device.
In one example embodiment of a system, the processor to determine that the incident occurred within the vicinity of the first device is further to detect the incident based on an analysis of the sensor data captured by the plurality of sensors.
In one example embodiment of a system, the processor is further to: receive a notification associated with the incident over an information-centric network, wherein the notification comprises a request for data associated with the incident, and wherein the request for data comprises an indication of a plurality of attributes associated with the incident; and determine whether the sensor data captured by the plurality of sensors corresponds to the plurality of attributes associated with the incident.
In one example embodiment of a system, the processor to notify the one or more second devices of the incident is further to: broadcast a notification associated with the incident over an information-centric network, wherein the notification comprises a request for data associated with the incident, and wherein the notification is received by the one or more second devices.
In one example embodiment of a system, the processor is further to: receive one or more second collections of sensor data associated with the incident, wherein the one or more second collections of sensor data are received from the one or more second devices; and generate a consolidated collection of sensor data associated with the incident, wherein the consolidated collection of sensor data is generated from the first collection of sensor data and the one or more second collections of sensor data.
In one example embodiment of a system, the processor is further to send the consolidated collection of sensor data to an entity associated with the incident.
In one example embodiment of a system, the processor is further to generate a reconstructed representation of the incident based on the consolidated collection of sensor data.
One or more embodiments may include at least one machine accessible storage medium having instructions stored thereon, wherein the instructions, when executed on a machine, cause the machine to: access sensor data captured by one or more sensors associated with a first device; detect an incident that occurs within a vicinity of the first device, wherein the incident is detected based on an analysis of the sensor data; identify a first collection of sensor data associated with the incident, wherein the first collection of sensor data is identified from the sensor data captured by the one or more sensors; preserve the first collection of sensor data associated with the incident; and notify one or more second devices of the incident, wherein the one or more second devices are located within the vicinity of the first device.
In one example embodiment of a storage medium, the instructions further cause the machine to: receive a notification associated with the incident over an information-centric network, wherein the notification comprises a request for data associated with the incident, and wherein the request for data comprises an indication of a plurality of attributes associated with the incident; and determine whether the sensor data captured by the one or more sensors corresponds to the plurality of attributes associated with the incident.
In one example embodiment of a storage medium, the instructions further cause the machine to: broadcast a notification associated with the incident over an information-centric network, wherein the notification comprises a request for data associated with the incident, and wherein the notification is received by the one or more second devices; receive one or more second collections of sensor data associated with the incident, wherein the one or more second collections of sensor data are received from the one or more second devices; generate a consolidated collection of sensor data associated with the incident, wherein the consolidated collection of sensor data is generated from the first collection of sensor data and the one or more second collections of sensor data; and send the consolidated collection of sensor data to an entity associated with the incident.
In one example embodiment of a storage medium, the instructions further cause the machine to generate a reconstructed representation of the incident based on the consolidated collection of sensor data.
One or more embodiments may include a method, comprising: accessing sensor data captured by one or more sensors associated with a first device; detecting an incident that occurs within a vicinity of the first device, wherein the incident is detected based on an analysis of the sensor data; identifying a first collection of sensor data associated with the incident, wherein the first collection of sensor data is identified from the sensor data captured by the one or more sensors; preserving the first collection of sensor data associated with the incident; and notifying one or more second devices of the incident, wherein the one or more second devices are located within the vicinity of the first device.
In one example embodiment of a method, the method further comprises: receiving a notification associated with the incident over an information-centric network, wherein the notification comprises a request for data associated with the incident, and wherein the request for data comprises an indication of a plurality of attributes associated with the incident; and determining whether the sensor data captured by the one or more sensors corresponds to the plurality of attributes associated with the incident.
In one example embodiment of a method, the method further comprises: broadcasting a notification associated with the incident over an information-centric network, wherein the notification comprises a request for data associated with the incident, and wherein the notification is received by the one or more second devices; receiving one or more second collections of sensor data associated with the incident, wherein the one or more second collections of sensor data are received from the one or more second devices; generating a consolidated collection of sensor data associated with the incident, wherein the consolidated collection of sensor data is generated from the first collection of sensor data and the one or more second collections of sensor data; and sending the consolidated collection of sensor data to an entity associated with the incident.
This patent application claims the benefit of the filing date of U.S. Provisional Patent Application Ser. No. 62/611,536, filed on Dec. 28, 2017, and entitled “VISUAL FOG,” the content of which is hereby expressly incorporated by reference.
Number | Name | Date | Kind |
---|---|---|---|
9432430 | Klenz | Aug 2016 | B1 |
10096038 | Ramirez | Oct 2018 | B2 |
10235880 | Seo | Mar 2019 | B2 |
20050162525 | Koshikawa | Jul 2005 | A1 |
20130006469 | Green | Jan 2013 | A1 |
20140300739 | Mimar | Oct 2014 | A1 |
20140368601 | deCharms | Dec 2014 | A1 |
20150381942 | Brewer | Dec 2015 | A1 |
20160042767 | Araya | Feb 2016 | A1 |
20160182707 | Gabel | Jun 2016 | A1 |
20160358129 | Walton | Dec 2016 | A1 |
20180184236 | Faraone | Jun 2018 | A1 |
20180330610 | Wu | Nov 2018 | A1 |
20190042867 | Chen | Feb 2019 | A1 |
20190042870 | Chen | Feb 2019 | A1 |
20190042900 | Smith | Feb 2019 | A1 |
20190043201 | Strong | Feb 2019 | A1 |
20190045207 | Chen | Feb 2019 | A1 |
20190124495 | Subramaniam | Apr 2019 | A1 |
Entry |
---|
Dabiri, Sina, et al., “Transport-domain application of widely used data sources in the smart transportation: A survey”, arXiv:1803.10902v2, accessed at: https://arxiv.org/vc/arxiv/papers/1803/1803.10902v2.pdf, version date: Jun. 12, 2018, 52 pages. |
Khalastchi, Eliahu, “Anomaly Detection in Unmanned Vehicles: Thesis”, The Maverick Group, Department of Computer Science, Bar Ilan University, Ramat Gan, Israel, 2010, 52 pages. |
Weber, Marc et al., “Embedded Hybrid Anomaly Detection for Automotive CAN Communication”, 9th European Congress on Embedded Real Time Software and Systems (ERTS 2018), Jan. 2018, Toulouse, France, accessed at: https://hal.archives-ouvertes.fr/hal-01716805, 11 pages. |
Number | Date | Country | |
---|---|---|---|
20190043351 A1 | Feb 2019 | US |
Number | Date | Country | |
---|---|---|---|
62611536 | Dec 2017 | US |