The technology disclosed generally relates to detecting location and positioning of a mobile device, and more particularly relates to application of visual processing and inertial sensor data to positioning and guidance technologies.
The subject matter discussed in this section should not be assumed to be prior art merely as a result of its mention in this section. Similarly, a problem mentioned in this section or associated with the subject matter provided as background should not be assumed to have been previously recognized in the prior art. The subject matter in this section merely represents different approaches, which in and of themselves can also correspond to implementations of the claimed technology.
Autonomous robots have long been the stuff of science fiction fantasy. One technical challenge in realizing the truly autonomous robot is the need for the robot to be able to identify where they are, where they have been and plan where they are going. Traditional SLAM techniques have improved greatly in recent years, however, there remains considerable technical challenge to providing fast accurate and reliable positional awareness to robots and self-guiding mobile platforms.
With the recent proliferation of virtual reality headsets such as the Oculus Rift™, PlayStation™ VR, Samsung Gear™ VR, the HTC Vive™ and others, a new class of devices—one that is not autonomous but rather worn by a human user—that would benefit from fast, accurate and reliable positional information has arisen. Many technical challenges remain however in the field of enabling machines and devices to identify where they are, where they have been and plan where they are going. On especially challenging area involves recognizing a location and obstructions accurately and quickly. A variety of different approaches have been tried. For example RFID/WiFi approaches have proven to be expensive and of limited accuracy. Depth sensor based approaches have been found to be high cost and suffer from power drain and interference issues. Marker based approaches require markers placed within the work area—limiting the useful area in which the device can operate. Visual approaches currently are slow leading to failure when used in fast motion applications. Such approaches can also suffer scale ambiguity. Yet these implementations failed to live up to the standards required for widespread adoption.
The challenge of providing fast reliable affordable positional awareness to devices heretofore remained largely unsolved.
In the drawings, like reference characters generally refer to like parts throughout the different views. Also, the drawings are not necessarily to scale, with an emphasis instead generally being placed upon illustrating the principles of the technology disclosed. In the following description, various implementations of the technology disclosed are described with reference to the following drawings, in which:
The following detailed description is made with reference to the figures. Sample implementations are described to illustrate the technology disclosed, not to limit its scope, which is defined by the claims. Those of ordinary skill in the art will recognize a variety of equivalent variations on the description that follows.
This document describes positional awareness techniques employing visual-inertial sensory data gathering and analysis hardware with reference to specific example implementations. The discussion is organized as follows. First, an introduction describing some of the problems addressed by various implementations will be presented. Then, a high-level description of one implementation will be discussed at an architectural level. Next, the processes used by some implementations to efficiently process image and inertial data are discussed. Lastly, the technology disclosed will be illustrated with reference to particular applications of (i) Robots and self-guided autonomous platforms, (ii) virtual reality headsets and wearable devices, and (iii) augmented reality headsets and wearable devices. The references to specific examples are intended to be illustrative of the approaches disclosed herein rather than limiting.
Improvements in the use of sensors, techniques and hardware design can enable specific implementations to provide improved speed and accuracy, however, such improvements come with an increased number of parameters and significant memory and computational requirements. Conventional approaches to automatic guidance have largely focused on single sensor input. Camera based approaches have been relatively accurate, but suffer speed limitations (most hardware provide 30 fps, 60 fps at most), and are computationally expensive since these approaches process every pixel. Inertial guidance based approaches suffer from drift of the zero or origin point. Further, these approaches require expensive hardware in order to achieve useful results. WIFI and RFID approaches based on older technology exist, however, these have shown themselves to be limited in capability. Depth sensor based approaches are expensive. Further, these approaches require active sensing, so the computational cost is relatively high. Finally, the device's active sensing can pose interference issues.
To overcome the computational burden of processing large amounts of image data all the time, inertial data can be used to estimate changes in the environment due to changes in pose of the machine under guidance. To overcome the drift problems associated with inertial sensors, images can be captured and processed to correct and update pose estimates made based upon inertial data. Further, stereo imaging sensors comprised of RGB and grayscale camera combinations can provide stereo imaging capabilities, at lower cost points than stereo RGB systems. Yet further, using low-end sensors to construct a sensor, e.g., cameras having resolution of 640×480, obviates the cost of high-end image sensors. Still further, use of a low-power Control Unit to perform certain sensor based processing, instead of a powerful processor of a host or the machine under guidance, enables use of the system at reduced cost relative to conventional approaches. Implementations can be deployed in a variety of usage scenarios, including robot or other mobile platform guidance, Virtual Reality/Augmented Reality (VR/AR) headsets, goggles or other wearable devices, and others.
Examples of robot applications that benefit from employing positional awareness techniques such as described herein include:
In each of the scenarios listed above, the robot utilizes the techniques described herein in order to track its own location and to recognize the objects that it encounters. Also, since the robot performs many complex tasks, each with real-time constraints, it is beneficial that the sensing be done rapidly to accelerate the perception pipeline. To overcome the computational burden imposed by this processing, implementations offload some computation from the main processor to the visual-inertial sensor module. In addition, since it is a mobile robot, which carries limited battery, energy consumption is a major challenge. Accordingly, some implementations offload some computational tasks from the main processor to a low-power sensor module, thereby enabling implementations to achieve overall energy efficiency. Since cost is an issue in mobile robots, because lowering the cost of the robot makes the robot affordable to more customers, cost reduction is another factor for sensor design. Accordingly, some implementations employ one low-cost grayscale sensor that is used for localization tasks, and one colored sensor for recognition tasks. This design point enables these implementations to significantly reduce the cost over a stereo colored sensor designs without sacrificing performance.
Virtual Reality (VR) and Augmented Reality (AR) scenarios require a wearable headset to track its own location, and maybe to recognize the objects that it encounters. In order to track its location, the wearable headset is equipped with a positional self-aware device that senses its own movement through a stereo inertial hardware sensor. Accordingly, the sensor generates reliable inertial data so that the tracking and mapping pipeline that follows can accurately infer the device's—and hence the headset's—location.
In implementations in which the device is embedded within another device, e.g., robot, mobile platform, wearable computer, AR/VR headset, goggles, wrist or other watches, etc., limited computational resources are available, while the workload of robot guidance, or AR/VR processing demands real-time performance, sensing is done rapidly to accelerate the perception processing pipeline. Accordingly, some implementations achieve these goals by offloading some computation from the main processor to the sensor module.
In addition, in AR/VR applications the mobile embedded device carries limited battery power, making energy consumption a challenge. Accordingly, some implementations offload some computation from the main processor to the low-power sensor module, in order to achieve overall energy efficiency.
Yet further, cost is an issue in many AR/VR applications because as the cost of the device is lowered, the potential to reach more customers is expanded. Hence cost is another factor for the sensor module design. Accordingly, some implementations use one low-cost grayscale sensor for localization tasks, and one colored sensor for recognition tasks. This design can provide significantly reduced cost over a stereo colored sensor design without sacrificing performance.
Examples of systems, apparatus, and methods according to the disclosed implementations are described in a robot guidance, VR and AR wearable device contexts with image and inertial data. In other instances, the technology disclosed can be applied to autonomous vehicle guidance technology, navigation, telecommunications systems, financial systems, security trading, banking, business intelligence, marketing, mining, energy, etc. and using sonar, audio, and LIDAR data. Other services are possible, such that the following examples should not be taken as definitive or limiting either in scope, context, or setting.
The technology disclosed relates to improving utilization of computing resources such as computational power and memory use during processing of image and inertial data inside a single input-multiple data (SIMD) architecture. The technology disclosed can be implemented in the context of any computer-implemented system including an reduced instruction set (RISC) system, emulated hardware environment, or the like. Moreover, this technology can be implemented using two or more separate and distinct computer-implemented systems that cooperate and communicate with one another. This technology can be implemented in numerous ways, including as a process, a method, an apparatus, a system, a device, a computer readable medium such as a computer readable storage medium that stores computer readable instructions or computer program code, or as a computer program product comprising a computer usable medium having a computer readable program code embodied therein.
The technology disclosed can be implemented in the context of any computer-implemented system like a NEON ARM VFP9-S processor, an ARM core processor, or a compatible processor implementation.
In addition, the technology disclosed can be implemented using a variety of different imaging sensors and technologies, including RGB, grayscale, binary (e.g., digital image subjected to threshold intensity level), IR, sonar, LIDAR or combinations thereof.
System Overview
Stereo Visual-Inertial Sensor
The Inertial component 120 includes an Inertial Measurement engine 105 that implements a time stamping processor 107 that time stamps sets of inertial data from an inertial sensor (not shown in
Visual-Inertial Sensor
An IMU 202 provides raw sensor data for agent localization pipeline, which consumes IMU data at a high frequency (>200 Hz) to generate agent positional information in real-time. In an implementation, the localization pipeline combines information from IMU 202 which runs at relatively high frequency to provide frequent updates of less accurate information, and cameras 208, 210, which run at relatively lower frequency, 30 Hz, to provide more accurate information with less frequency.
The Control Unit 100 performs control of the sensors, IMU 202 and Cameras 208, 210, time stamping sensor data from the sensors, performs pre-computation in order to accelerate the localization pipeline, and packages raw data for sending over USB 103 to a host.
The USB interface 103 enables the visual-inertial sensor 200 to interact with a host. The host (not shown in
Note that one implementation averages the aligned images. In other implementations, other techniques are used. Also note that in another implementation an image quality measurement sub-step is included. So if the output image is too dark or still not sharp or clear enough, the image will be rejected and not passed to the rest of the pipeline.
Sensor Data Generation and Processing
In an embodiment, IMU raw data is corrected on the CU 100, thereby enabling implementations that do not require extra processing from the host processor, therefore accelerating the sensor pre-processing pipeline.
Time Stamping:
The time stamping processor 107 time stamps each set of inertial measurement data that the control unit 100 receives from the IMU sensor 202 data, in order to assure that the visual-inertial sensor 200 maintains a temporally accurate stream of sensor data. Such rigorous attention to maintaining the integrity of the sensor data stream enables implementations to provide agent localization that works reliably. Time-stamping raw data by the visual-inertial sensor obviates the need for complex synchronization tasks.
Bias Correction:
The bias correction processor 109 corrects IMU data readout from the timestamped inertial data. Due to manufacturing imperfections, IMU sensors usually have bias problems such that its measurements contain errors. A bias error, if not removed from the measurement, is integrated twice as part of the mechanization process. In this case, a constant bias (error) in acceleration becomes a linear error in velocity and a quadratic error in position. A constant bias in attitude rate (gyro) becomes a quadratic error in velocity and a cubic error in position. The bias can be derived from the offline factory sensor calibration stage. This calibration information in CU 100 to perform bias correction task on CU 100.
Scale Correction:
The scale correction processor 111 applies stored scale factor information to the corrected inertial data. Scale factor error is the relation between input and output. If the input is 100%, the expected output is 100%. The actual output is the result of a linear effect, where the output is proportional to the input but scaled. For example, if the input is 10 m/s2, but there is a 2% scale factor error, the output measurement is 10.2 m/s2. The scale factor can be derived from the offline factory sensor calibration stage. This calibration information in CU 100 to perform scale correction task on CU 100.
Misalignment Correction:
The mis-alignment correction processor 113 corrects misalignments of sensory elements of the inertial measurement sensor. There are three gyroscopes and three accelerometers are mounted orthogonal to each other. The mountings, however, have errors and so are not perfectly 90 degrees. This leads to a correlation between sensors. For example, assume one axis is pointed perfectly up and the IMU is level. The accelerometer on this axis is measuring gravity. If the other two axes were perfectly orthogonal, they do not measure any of the effect of gravity. If there is a non-orthogonality, the other axes also measure gravity, leading to a correlation in the measurements. The effect of non-orthogonality occurs within sensor sets (between accelerometers or gyroscopes), between sensor sets or between the sensor sets and the enclosure (package misalignment). Careful manufacturing, as well as factory calibration, can help minimize this error source. Continuous estimation and correction during system operation is also an approach used to minimize this effect. Package misalignment (between the IMU 202 and the enclosure) can be removed by performing a bore-sighting estimation to determine the offset between the IMU 202 measurement frame and the sensor (objective) frame. The misalignment numbers can be derived from the offline factory sensor calibration stage. This calibration information in CU 100 to perform misalignment correction task on CU 100.
Sensor Visual Data Generation and Processing
Image Undistortion:
The image undistortion processor 114 corrects distortion in the image data in the captured frames. The image distortion is generally referred to an optical aberration that deforms and bends physically straight lines and makes them appear curvy in images. Optical distortion occurs as a result of optical design. In order to achieve reliable computer vision results, image undistortion processor 114 can un-distort the image before further processing is performed. This can be achieved by using a lookup table of the size of the input image, and performing a remapping operation to undistort the whole image.
Feature Undistortion:
In cases when the remaining portions of the processing pipeline do not require the whole image, but only the feature points within the image, the feature undistortion processor 116 perform a feature undistortion operation on the CU. In detail, this operation runs after the feature extraction stage, and undistorts each feature point.
Feature Detection:
The Shi-Tomasi processor 115 performs feature detection upon image frames. Features are “interesting” parts of an image. The Shi-Tomasi feature detection includes methods that aim at computing abstractions of image information and making local decisions at every image point whether there is an image feature of a given type at that point or not. The resulting features will be subsets of the image domain, often in the form of isolated points. Some implementations perform the feature detection on the CU 100 to relieve the host from performing such tasks, and to accelerate the feature detection process. Accordingly, in an implementation, processing includes:
Feature Description:
The feature description engine 117 performs feature description on detected features. The feature description includes methods to uniquely identify each detected points in an image. Feature description can be used to compare and match feature points between different images. Some implementations perform the feature description on the CU 100 to relieve the host from performing such tasks, and to accelerate the feature description process.
One implementation of feature description engine 117 uses a SIMD-accelerated ORB descriptor to describe features. The description of a feature can be used for matching purposes and describing a feature's uniqueness. The ORB descriptor approach was selected for its relative rotational invariance and immunity to Gaussian image noise. One example of an ORB feature detector and binary descriptor can be found at “ORB feature detector and binary descriptor”, which is incorporated herein by reference for all purposes. For further information on ORB Descriptor, reference may be had to Ethan Rublee, et al., “ORB: an efficient alternative to SIFT or SURF”, which is incorporated herein by reference for all purposes.
2D Feature Correspondence Generation:
The optical flow feature correspondence processor 118 performs 2D feature correspondence generation for the features. The feature correspondence computation is used to identify the feature points that appear in both the left and the right cameras. Once feature correspondence is identified for any two feature points, triangulation can be applied to the feature points to derive the depth of the point in space. This depth information is employed by processes later in the localization pipeline. Some implementations perform the feature correspondence generation on the CU 100 to relieve the host from performing such tasks, and to accelerate the feature correspondence generation.
One optical flow feature correspondence processor 118 implementation employs optical flow methods to calculate the motion between two image frames, taken at times t and t+Δt at each voxel position. One such method, called a differential method, is based on local Taylor series approximations of the image signal, using partial derivatives with respect to the spatial and temporal coordinates. Accordingly, in an implementation, processing includes:
IMU-Camera Coordinate Transformation:
In some implementations, the IMU 202 and the cameras 208, 210 do not reside at the same physical location, there is a distance between the IMU 202 and the cameras 208, 210. Accordingly, in order to enable later processes in the localization pipeline to treat the IMU 202 and the cameras 208, 210 as being co-located, on implementation determines a transformation matrix between the IMU 202 and the cameras 208, 210, which can be achieved from an offline production or post-production calibration stage. In CU 100, this transformation matrix is stored locally, and applied to the IMU data. This technique enables later processes to be able to treat the IMU 202 and the cameras 208, 210 to be co-located.
Referring now to Referring now to
The computing environment may also include other removable/non-removable, volatile/nonvolatile computer storage media. For example, a hard disk drive may read or write to non-removable, nonvolatile magnetic media. A magnetic disk drive may read from or write to a removable, nonvolatile magnetic disk, and an optical disk drive may read from or write to a removable, nonvolatile optical disk such as a CD-ROM or other optical media. Other removable/non-removable, volatile/nonvolatile computer storage media that can be used in the exemplary operating environment include, but are not limited to, magnetic tape cassettes, flash memory cards, digital versatile disks, digital video tape, solid state RAM, solid state ROM, and the like. The storage media are typically connected to the system bus through a removable or non-removable memory interface.
In an embodiment, the processor 332 is a NEON ARM processor implementing a single input-multiple data (SIMD) architecture as a reduced instruction set computer (RISC) architecture. Depending on implementation, however, processor 332 can alternatively be a realized using a specific purpose microcontroller, peripheral integrated circuit element, a CSIC (customer-specific integrated circuit), an ASIC (application-specific integrated circuit), a logic circuit, a digital signal processor, a programmable logic device such as an FPGA (field-programmable gate array), a PLD (programmable logic device), a PLA (programmable logic array), an RFID processor, smart chip, or any other device or arrangement of devices that are capable of implementing the actions of the processes of the technology disclosed.
Communications interface 342 can include hardware and/or software that enables communication between visual inertial positioning system 300 and other systems controlling or enabling customer hardware and applications (hereinafter, a “host system” or “host”) such as for example, a robot or other guided mobile platform, an autonomous vehicle, a virtual reality-augmented reality wearable device (VR/AR headset) or the like (not shown in
Instructions defining FASAT process 344 are stored in memory 334, and these instructions, when executed, perform analysis on image frames captured by the cameras 208, 210 and inertial data captured by the IMU 202 connected to visual inertial positioning system 300. In one implementation, FASAT process 344 includes various logical processes, such as a feature extractor 352 that receives a raw image and determines a salient points' representation of objects in the image thereby representing the geometry understanding of the objects from a machine's perspective view. In some implementations, feature extractor 352 analyzes images (e.g., image frames captured via cameras 208, 210) to detect edges of an object therein and/or other information about the object's location. A sensor fusion tracking process 354 uses feature extraction results and inertial data from IMU 202 to generate pose accurately and rapidly. A smart interaction map 356 enables using a known map of obstructions to localize the sensor 200. The map is built using mapping functionality of mapping process 392, which is described in further detail herein below. A Re-localizer process 358 recovers device positional awareness when the device has lost track of device position. A system diagnostic and response (SDAR) 366 manages of current localizing state of the device and provide response strategy.
A mapping process 392 generates a hybrid occupancy grid that maps the space and objects recognized by the feature extractor 352. The hybrid occupancy grid includes (i) a point cloud representation of points in space located in the image frames and (ii) one or more x-y plane occupancy grids arranged at heights to intersect points on the extracted features.
In some implementations, other processing 394 analyzes audio or ultrasonic signals (e.g., audio signals captured via sonar or audio sensors comprising non-tactile input 382) to localize objects and obstructions by, for example, time distance of arrival, multilateration or the like. (“multilateration is a navigation technique based on the measurement of the difference in distance to two or more stations at known locations that broadcast signals at known times. Audio signals place the object on a known surface, and the strength and variation of the signals can be used to detect object's presence. If both audio and image information is simultaneously available, both types of information can be analyzed and reconciled to produce a more detailed and/or accurate path analysis.
In some implementations, other processing 394 determines paths to track and predict device movements in space based upon the hybrid occupancy grid generated by mapping process 392. Some implementations 394 includes an augmented reality (AR)/virtual reality (VR) environment that provides integration of virtual objects reflecting real objects (e.g., virtual presence of friend 915 in
Presentation interface 378, audio presentation 379, non-tactile input 382, and communications interface 342 can be used to facilitate user interaction via device 200 with Visual inertial positioning system 300. These components can be of highly customized design, generally conventional design or combinations thereof as desired to provide any type of user interaction. In some implementations, results of analyzing captured images using inertial measuring unit 202 and cameras 208, 210 and FASAT program 352 can be interpreted as representing objects and obstacles in 3D space. For example, a robot equipped with visual-inertial sensor 200 can perform path planning and/or obstacle avoidance across a surface that has been analyzed using FASAT program 352, and the results of this analysis can be interpreted as an occupancy map by some other program executing on processor 332 (e.g., a motion planner, localization and tracking process, or other application). Thus, by way of illustration, a robot might use sweeping of cameras 208, 210 across a room in order to “map” a space currently imaged to a hybrid point grid that can be used by a host device such as a monitor, VR headset or the like via presentation interface 378, to provide visual input of the area that the robot is “seeing”. Smart interaction map 356 may use the representation of space built by mapping 392 to plan a path for a robot or mobile platform through the space, e.g., to improve localization and tracking of the robot or platform through the space.
It will be appreciated that Visual-inertial positioning system 300 is illustrative and that variations and modifications are possible. Visual inertial positioning systems can be implemented in a variety of form factors, including “cloud” computing systems of servers and networks, desktop systems, laptop systems, tablets, smart phones or personal digital assistants, and so on. A particular implementation may include other functionality not described herein for clarity sake. In some implementations, one or more cameras and two or more microphones may be built into the visual-inertial system 300 or may be supplied as separate components. Further, an image or audio analyzer can be implemented using only a subset of Visual inertial positioning system 300 components (e.g., as a processor executing program code, an ASIC, or a fixed-function digital signal processor, with suitable I/O interfaces to receive image data and output analysis results).
While Visual inertial positioning system 300 is described herein with reference to particular blocks, it is to be understood that the blocks are defined for convenience of description and are not intended to imply a particular physical arrangement of component parts. Further, the blocks need not correspond to physically distinct components. To the extent that physically distinct components are used, connections between components (e.g., for data communication) can be wired and/or wireless as desired. Thus, for example, execution of feature extractor 352 by processor 332 can cause processor 332 to operate inertial measurement unit 202 and cameras 208, 210 to capture images and/or audio signals of an object traveling across and in contact with a surface to detect its entrance by analyzing the image and/or audio data.
Tracking
Tracking refers to capabilities rendered by system hardware and functional processes that enable a controlled device (robot, mobile platform, or VR/AR headset, goggles, or other hardware) to continuously localize itself and have positional awareness e.g., the sensor system can determine where it is in the real world.
Architecture for Tracking System
Feature Extraction
A feature extractor (FE) 402 represents the geometric understanding of a space from a machine's perspective view. Feature extractor 402 receives raw image information from control unit 100 and provides a salient points' representation hybrid point cloud map to a sensor fusion tracker (SFT) 411. One implementation of a feature extractor 402 performs a variety of actions including image preprocessing, feature detection, and feature descriptor preparation.
Image processing tasks include performing Gaussian blur and gamma correction on raw image data.
Feature Detection: Optical Flow
Optical flow gives 2D-2D correspondence between previous image and a current image. Feature extractor 402 generates a hierarchy of levels; each level generation is aligned with optical flow needs, and therefore need only be computed once.
Processed images are stored in a multiple hierarchical patch and/or undistorted patch. In an example implementation, as levels in the hierarchy increase, the more blur exists in the image. At each level in the hierarchy, the size of the image is reduced to ¼ of the size of the image from which it is derived, e.g. the size of the image in the previous level. For example, if an image on the first (“zero level”) has a size 640×480, then the size of the corresponding image on the next higher level (“first level”) has a size 320×240, and an next corresponding image at the next higher level (“second level”) has a size 160×120. If the same feature point appears at more than one level, then that feature point is determined by the system to be a strong feature point. In the foregoing example, a scale factor of 2 is used, however, any scale factor can be selected, e.g., a scale factor of 1.6 provides the original 640×480 image with a next level image having a resolution of 400×300.
One feature extractor 402 implementation employs optical flow methods to calculate the motion between two image frames, taken at times t and t+Δt at each voxel position. One such method, called a differential method, is based on local Taylor series approximations of the image signal, using partial derivatives with respect to the spatial and temporal coordinates. Accordingly, in an implementation, processing includes:
Feature Detection: Shi-Tomasi
One implementation of feature extraction processor 402 uses NEON-accelerated Shi-Tomasi feature detection with spatial binning in order to obtain features from regions to cover as much as possible for the full image. The feature extractor 402 uses IMU and the pose information received from sensor fusion tracker 411 in order to dynamically decide the regions to track and the parameters to use. Features are “interesting” parts of an image. Accordingly, in an implementation, processing includes:
Feature Descriptor: ORB Descriptor
One implementation of feature extractor 402 uses a SIMD-accelerated ORB descriptor to describe features. The description of a feature can be used for matching purposes and describing a feature's uniqueness. The ORB descriptor approach was selected for its relative rotational invariance and immunity to Gaussian image noise. One example of an ORB feature detector and binary descriptor can be found at “ORB feature detector and binary descriptor”, which is incorporated herein by reference for all purposes. For further information on ORB Descriptor, reference may be had to Ethan Rublee, et al., “ORB: an efficient alternative to SIFT or SURF”, which is incorporated herein by reference for all purposes.
Bag-of-Words
One implementation of feature extractor 402 performs bag-of-words extraction, determining from one image a byte array, sparse histogram over vocabulary. Accordingly, in an implementation, Bag-of-words processing includes using a list of feature descriptors in one image to extract Bag-of-words feature for one image. For example, in an image having 400 features with 400 corresponding descriptors, and a bag-of-word feature with the size of 256 bits. The resulting Bag-of-words is a 256 bit array, and one image only needs/has one 256 bit array. Bag-of-words is useful in performing relocalization as described herein below in further detail. Bag-of-words can indicate among A, B, C images, an image D has a matching score individually D-A, D-B, D-C. Say for example that D-B has the highest score, then the camera taking image D is closest to the camera pose used when taking image B, not A or C.
One implementation of feature extractor 402 adaptively adjusts a threshold that is applied to a number of features needed in order for the system to keep track of a moving object. Such intelligent thresholds include a threshold that is adaptively adjusted based upon device movement, sensor readings, situational or environmental variables (e.g., low light, fog, bright light, and so forth) or combinations thereof. Once sufficient number of features (e.g., above the threshold) are identified, the system will continue to track an object. Otherwise, the system does not track the object. Feature extractor 402 will adaptively adjust the threshold based on device movement and other sensor levels.
SFT: Generating Pose Information from Feature Extraction Results and Inertial Data
Sensor fusion based tracker (SFT) 411 determines how to smartly use feature extraction results and inertial data from IMU 202 to generate pose accurately and rapidly. Sensor fusion based tracker 411 performs initialization for the system state, feature extractor initial state, interaction with map state, and other states. Sensor fusion based tracker 411 has logic to decide how to initialize the state including use how many first sample IMUs, how many images to skip, and any other initialization steps for specific hardware.
In an implementation, sensor fusion based tracker 411 performs IMU buffer management. Timestamps of IMU data frames are corrected and aligned. Additionally, IMU information that has not been processed yet is maintained in a buffer.
Propagation:
Implementations can employ extended Kalman filtering (EKF), shown in a general nonlinear model form by equations (1), to extrapolate an initial pose using inertial data from the multi-axis IMU, to generate a propagated pose:
{dot over (x)}=f(x)+w
z=h(x)+n (1)
where x denotes a state vector, f denotes a nonlinear function, h denotes observation nonlinear function, n denotes measurement noise, z denotes an observation vector, and w denotes a noise vector.
A linearized version of the discrete-time model of equation (1) is useful for covariance propagation and update, as well as gain computation is illustrated by equations (2):
{tilde over (X)}
l+1≃φl{tilde over (X)}l+wl (2)
{tilde over (r)}
l
≃H
l
{tilde over (X)}
l
+n
l
In an implementation, IMU data alone is used to perform propagation of the EKF. This enables certain implementations to provide fast results and at relatively low computation costs. An update process described below is employed to correct for “drifting” that occurs when only inertial information is used in the EKF to determine pose. Propagation can be performed in a number of ways: (i) the propagation happens right after an inertial measurement taken by the IMU arrives; or (ii) delayed propagation happens in batch right after an image taken by the visual sensors arrives but before the update procedure described below is performed. Of course, wheel odometry and/or user-defined control can also be used to perform propagation. In one implementation, a “plug-and-play” decision engine will decide if, when, and how to do multiple types of propagation.
Update:
In an implementation, an update process is performed using image information together with IMU buffer samples, which can correct the pose which may be drifted from propagation. In a single observation implementation, based on the current image observation, the information needed for propagation, e.g., features, poses, map points, etc. is prepared. Then 2D-3D correspondence information for the optical flow tracked 2D features is obtained by directly using 2D-2D correspondences from optical flow tracking results. For the remaining 3D map points, smartly search over 3D with adaptive parameters by: (i) if the remaining number of points is small, e.g., below a threshold, perform a linear search, otherwise, (ii) if the remaining number of points is fairly big, e.g., above a threshold, perform a log(n) search with the use of a kd-tree or octree. During search, use previous time period velocity/acceleration to predict a pose so as to narrow the search region. Next, perform update, and generate pose for the current frame using only “good outstanding” points. Alternatively, perform the update twice: first only with high weights “good outstanding” points, then with all the map points. As used herein, the term “good outstanding” points denotes high quality points, e.g., those points having smaller reprojection error and/or being tracked many times in a good state, which can be selected by application of some heuristics.
In a multiple observation implementation, based on multiple previous image observation, information from multiple poses is used in the extended Kalman filter described herein with reference to equations (1) and (2). Every feature's whole life is used in the extended Kalman filter.
SIM: Smart map interaction and localize
Smart interaction with map (SIM) processor 414 enables smart interaction by a device being guided (i.e., receiving positional awareness information from visual-inertial sensor system 300) with a known map in order to localize itself. Accordingly, one implementation of a smart interaction map processor 414 processing includes:
SDAR: Manage Localization State and Response Strategy
System diagnostics and response (SDAR) processor 403 performs system level management of current localizing state of the visual-inertial sensor 200 and provides response strategy. In one implementation, the system diagnostics and response processor 403 monitors tracked map points and corresponding reprojection error, monitors the longevity of untracked map points, and suggests actions to mapping module, monitors device speed/acceleration/jitter/drift, monitors map “activity” (i.e., is it updated regularly?), performs new added map contribute enough to the overall mapping (i.e., so it's not doing useless hovering?), monitors keyrig topology (i.e., stretch enough or too much?), suggests when and where the device should add keyrig, and monitors feature extractor pipeline
Relocalize a Lost Device
Re-locator 422 recovers a device that has lost track of its positional from this state. Once re-locator 422 returns the lost device back on track, the device can resume normal interacting with smart interaction map processor 414, and/or can contribute to map building, using the techniques described herein below.
Relocalization
Geometric information obtained by 3D feature-based tracking is used to construct a hybrid grid that represents the features in a map useful in guiding a mobile device. The process of performing this task is described next with reference to examples depicted by
Mapping
In implementations, geometric information obtained by a 3D feature-based tracking process is represented using a descriptive point cloud representation. The descriptive point cloud is a set of sparse 3D points, where each point encodes a 3D geometric location, an uncertainty of the geometric location, and a set of 2D location plus appearance-based descriptors, each observed by a keyrig. A keyrig contains (i) a Timestamp where the images in the keyrig is captured; (ii) a pose (i.e. the transformation from the visual-inertial sensor's 200 coordinate to the map coordinate at the current time) and its uncertainty; and (iii) a sequence of IMU readings starting from a first temporal point before the timestamp to a second temporal point after the timestamp. Some implementations further include a flag to specify whether the visual-inertial sensor 200 is moving or static. This flag can be set from information obtained from the IMU readings.
Hybrid Representation
Implementations represent the spatial information perceived by a 3D feature-based mapping system and optionally an active ranging sensor (placed in parallel to the ground plane) using a hybrid point grid representation. One example implementation of a hybrid point grid representation combines a descriptive point cloud and a multi-layer 2D occupancy grid map.
The occupancy grid map can be a plane oriented substantially perpendicular to the direction of gravity. Each layer has a specific height. (The layer on the floor is typically 0 according to one convention.).
A 3 degrees of freedom transform that relates the descriptive point cloud coordinates to the gravity-based coordinates. For example, if the normal vector of the ground plane (i.e., gravity) is known, first rotate the point cloud so that the z-axis of the point cloud is parallel to the gravity, using a 3-by-3 rotation matrix R=R_pitch*R_roll. Then, (optionally) translate the point by [0, 0. T_z] so that the z=0 is on the ground plane. Note, a normal 3-by-3 rotation matrix has 3 degrees of freedom (R_yaw,* R_pitch*R_roll), however in this case are only 2 degrees. A normal 3-by-1 translation has 3 degrees of freedom ([t_x, t_y, t_z]), but here there is only 1 degree.
Creating a Point Cloud for Features Representation
Implementations incrementally create a descriptive point cloud from the results of 3D feature-based tracking by adding keyrigs according to a criterion, such as in a following example. Result of 3D feature based tracking is received in sets of image data that include feature points and pose information. The pose information includes a location of the mobile device and view of the camera that captured the image data. These sets are referred to collectively as keyrigs. Accordingly, an implementation builds a descriptive point cloud by selecting keyrigs using one or more criteria such as for example the following:
If the keyrig being considered for addition contains features not included in the current descriptive point cloud, triangulate the new features from images captured from the device at this timestamp to obtain the points in the device's coordinate frame. Add the new points to the map by transforming the points from the device's coordinate frame to the map's coordinate frame. Noteworthy is that some implementations include one or more of (i) triangulating new feature points across images from a current/same keyrig (e.g. between left and right cameras), (ii) triangulating new feature points across images from two different keyrigs, wherein the two different keyrigs are not necessarily in sequence (e.g. left camera from keyrig 1 to left camera from keyrig 10), and (iii) triangulating new feature points from images in keyrigs chosen based upon a criterion (e.g., smart selection of an “optimal” keyrig). The transform is determined using feature-based tracking or other methods.
Refining a Point Cloud
In one implementation, a descriptive point cloud from the 3D feature-based tracking is refined by obtaining an “optimal” (i) pose of each keyrig and (ii) 3D locations of all the points that minimizes weighted sum of re-projection errors of all the points projected to their observed keyrigs, and (iii) optionally, the accelerometer readings from keyrigs with static flag must be [0, 0, 1] to force the z-axis of the map coordinate to be closer to gravity direction, while fulfilling the following constraint (if applicable): the pose moves along a plane (if the device is mounted on a vehicle with planar motion).
In one implementation, refining a descriptive point cloud from the 3D feature-based tracking is performed with limited computation resources by the following processing:
Creating a Hybrid Point Grid
In one implementation, an initial hybrid point grid representation is created from a descriptive point cloud using the following processing:
Refining a Hybrid Point Grid
In one implementation, the hybrid point grid representation is refined using information from active sensors using the following processing:
In one implementation, the refining hybrid point grid representation from an updated descriptive point cloud is performed using the following processing:
Estimating a Trajectory for a Guided Device
In one implementation, an offline trajectory of a device between two consecutive keyrigs (i.e. from keyrig k to keyrig k+1) in a descriptive point cloud is estimated using the following processing:
Estimating a Transform for a Descriptive Point Cloud and Gravity Based Frame
In one implementation, the 2 degrees of freedom coordinate transform between the descriptive point cloud to a coordinate where z-axis is along the gravity direction is estimated using any of the following methods:
Robot Implementation
In each of the scenarios listed above, the robot utilizes the technology disclosed herein in order to track its own location and to recognize the objects that it encounters. Also, since the robot performs many complex tasks, each with real-time constraints, it is beneficial that the sensing be done rapidly to accelerate the perception pipeline. In implementations, this can be achieved by offloading some computation from the main processor to the visual-inertial sensor 200. In addition, since it is a mobile robot, which carries limited storage capacity battery, energy consumption is a design point. In implementations, some computational tasks are off loaded from the main processor to the visual-inertial sensor 200 to reduce power consumption, thereby enabling implementations to achieve overall energy efficiency. Cost is an issue in mobile robots, since lowering the cost of the robot makes the robot affordable to more customers. Hence cost can be another factor for sensor and guidance system design. In implementations, one low-cost grayscale sensor is used for localization tasks, and one colored sensor for recognition tasks. This design point enables these implementations to significantly reduce the cost over a stereo colored sensor designs without sacrificing performance.
In
Obtain Real Time Image and Inertial Information
In order to track its location, the robot senses its own movement through visual-inertial sensor 200. The visual-inertial sensor 200 generates reliable inertial data so that the tracking and mapping pipeline that follows can accurately infer the robot's location.
Detect and Describe Features in Captured Images
Visual-inertial sensor 200 determines feature points 801, 811, 841, 851, 822, and so forth for the walls, corners and door 823 of room 800 from the information in the captured image frames. In some implementations, Shi-Tomasi feature detection is employed to determine the feature points 801, 811, 841, 851, 822 from the image frames. Features are assigned descriptors using ORB feature description. Optical flow techniques are used to determine 2D correspondences in the images, enabling matching together features in different images.
Map Feature Descriptors to a Hybrid Point Grid
The visual-inertial sensor 200 equipped robot 825 can build a descriptive point cloud 845 of the obstacles in room 800 enabling the robot 825 to circumnavigate obstacles and self-localize within room 800. Visual-inertial sensor 200 creates, updates, and refines descriptive point cloud 845 using feature descriptors determined for room features indicated by points 801, 811, 841, 851, 822 using the technology disclosed herein above under the Mapping sections. As depicted schematically in
Now with renewed reference to
The descriptive point cloud 845 and occupancy grid 855 comprise a hybrid point grid that enables the robot 825 to plan paths of travel through room 800, using the occupancy grid 855 and self-localize relative to features in the room 800 using the descriptive point cloud 845.
Using the Occupancy Grid and Path Planning
When the robot is activated in a previously mapped environment, the robot uses the technology described herein above in the Tracking sections to self-locate within the descriptive point cloud 845. The descriptive point cloud 845 and occupancy grid 855 comprise a hybrid point grid representation that is key to enabling robot action (i.e. moving on the floor) using passive sensors because the robot uses the occupancy grid 855 in order to plan a trajectory 856 from its current location to another location in the map using the technology described herein above in the Mapping sections. A person or entity can also command the robot to go to a specific point in the occupancy grid 855. While traveling, the robot uses the descriptive point cloud 845 to localize itself within the map as described herein above in the Tracking sections. The robot can update the map using the techniques described herein above in the Mapping sections. Further, some implementations equipped with active sensors (e.g. sonar, lidar) can update the map using information from these sensors as well.
VR/AR Implementations
In both AR and VR usage scenarios, the headset, goggles or other VR/AR presentation device employs the visual-inertial sensor 200 to track its own location, and in some applications recognizes the objects that it encounters. In order to track its location, the VR/AR presentation device needs to sense its own movement, which is accomplished using the visual-inertial sensor 200 technology described herein. Accordingly, the visual-inertial sensor 200 generates reliable inertial data so that the tracking and mapping pipeline that follows can accurately infer the VR/AR presentation device's location.
In
Headset 925 can be communicatively coupled to a source of content for presenting VR presentation 915 to the wearer. While depicted as a wireless connection to a hotspot coupled to a network gateway that further couples to a source of content for VR presentation 915, headset 925 can be coupled to presentation content using any of a wide variety of mechanisms, wired or wireless. Further, presentation content for rendering VR presentation 915 can be downloaded and stored on some headset 925 implementations.
The foregoing operational principles are easily applied to other common VR applications such as gaming and socially enabled experiences. In gaming for example, game play is conducted by users while wearing a VR/AR presentation device. The game is displayed to a user (“gamer”) by the VR/AR presentation device. The visual-inertial sensor 200 equipped VR/AR presentation device conducts orientation tracking, translation movement of the gamer that can be reflected to the game displayed by the VR/AR presentation device, enabling the gamer to move freely within a certain area of surrounding space. A visual-inertial sensor 200 equipped VR/AR presentation device provides mobile VR games with such “inside-out” e.g., integrated within the VR/AR presentation device, tracking thereby obviating the need for some kind of external position sensors such as are required by conventional VR/AR headsets. Yet further, if the gamer uses the VR/AR presentation device frequently in the same area, e.g., the gamer's living room, the visual-inertial sensor 200 can smartly re-use or update an existing (e.g., previously created by the visual-inertial sensor 200) map of this space to provide position tracking.
In socially enabled experiences for example, a celebrity/sports star/singer (creator) performs live in front of a stereo video camera that provides a live video stream as content to fans. Fans wearing a VR/AR presentation device equipped with the visual-inertial sensor 200 (attached or embedded) can consume this live video feed as a VR presentation. When the fans get excited/sad/happy, they move about just as they would during an in person live experience. The visual-inertial sensor 200 performs position tracking, and provides input to the VR/AR presentation device enabling adjusting the live video feed accordingly based on the position tracking. Accordingly, the fans and creator can “interact” with the help of the disclosed technology.
Headsets 1005, 1025 can be communicatively coupled to a source of content for presenting AR presentation 1015 to the wearer(s). While depicted as a wireless connection to a hotspot coupled to a network gateway that further couples to a source of content for AR presentation 1015, headsets 1005, 1025 can be coupled to presentation content using any of a wide variety of mechanisms, wired or wireless. Further, presentation content for rendering AR presentation 1015 can be downloaded and stored on some headsets 1005, 1025 implementations.
The foregoing operational principles are easily applied to other common AR applications such as gaming, socially enabled experiences and industrial applications. Like VR gaming, in AR gaming, game play is conducted by users while wearing a VR/AR presentation device. Gamers can move freely about the room(s) while the visual-inertial sensor 200 equipped VR/AR presentation device tracks the gamer' position. The VR presentation device manages positioning virtual things in the real world environment because of the accurate mapping of the real world environment provided by the visual-inertial sensor 200. Accordingly, the AR gamer can participate in virtual battles in their own living room, accouter their kitchen with virtual overlays of 18th century kitchen items and so forth. Each virtual item can be positioned accurately and realistically into the gamer's real world using the position and localization information provided by the virtual-inertial sensor 200.
In socially enabled experiences for example, video chat applications for VR can enable a user to conduct a video chat with their best friend Tom, and a virtual Tom (displayed using an AR goggles, VR headset, projected by robot or other presentation rendering devices) appears in a position of the user's house. The user can walk around virtual Tom, change the presentation—making him larger or smaller or even changing the rendering altogether clothing him in a hat and suit typically worn by Napoleon.
Many other applications of virtual-inertial sensor 200 equipped VR/AR presentation devices exist. Virtual shopping trips, remote real estate showings, bicycle rides captured and replayed during indoor cycling sessions, virtual flight instruction, indoor/outdoor position tracking for hikes, bike rides, motorcycle travel and so forth.
In each of the scenarios listed above, the VR/AR presentation device utilizes the technology disclosed herein in order to track its own location and to recognize the objects that it encounters. Also, since the VR/AR applications include many complex tasks, each with real-time constraints, it is beneficial that the sensing be done rapidly to accelerate the perception pipeline. In implementations, this can be achieved by offloading some computation from the main processor to the visual-inertial sensor 200. In addition, since many VR/AR presentation devices can be wearable—or at least portable—the VR/AR presentation device carries limited storage capacity battery, energy consumption is a design point. In implementations, some computational tasks are off loaded from the main processor to the visual-inertial sensor 200 to reduce power consumption, thereby enabling implementations to achieve overall energy efficiency. Cost is an issue in VR/AR applications, since lowering the cost of the VR/AR presentation device makes VR/AR affordable to more customers. Hence cost can be another factor for sensor and guidance system design. In implementations, one low-cost grayscale sensor is used for localization tasks, and one colored sensor for recognition tasks. This design point enables these implementations to significantly reduce the cost over a stereo colored sensor designs without sacrificing performance.
Processes
The method described in this section and other sections of the technology disclosed can include one or more of the following features and/or features described in connection with additional methods disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this method can readily be combined with sets of base features identified as implementations such as system overview, tracking, mapping, robot implementation, VR/AR implementation, etc.
Process 1100 continues at action 1120 where inertial measurements from a multi-axis inertial measuring unit (IMU) capable of providing measurement of at least acceleration are buffered. The inertial measuring unit can be collocated with the visual sensor or can be disposed at an offset thereto.
At action 1130, the sets of image data are received at a visual inertial control unit.
At action 1140, the sets of sensor data from the multi-axis IMU are received at a visual inertial control unit.
At action 1150, the inertial data received is timestamped. The time stamping can be performed by an inertial measurement engine.
At action 1160, inertial readouts in the timestamped inertial data are corrected for bias.
At action 1170, the inertial readouts are scaled using a stored scale factor the corrected inertial data.
At action 1180, the scaled inertial data are corrected for misalignment in the IMU to form localization data.
At action 1190, imaging undistortion is performed on the sets of image data.
At action 1195, the localization data and the undistorted sets of image data are provided across a communications interface to a host controlling a mobile platform.
Other implementations of the method described in this section can include a non-transitory computer readable storage medium storing instructions executable by a processor to perform any of the methods described above. Yet another implementation of the method described in this section can include a system including memory and one or more processors operable to execute instructions, stored in the memory, to perform any of the methods described above.
The method described in this section and other sections of the technology disclosed can include one or more of the following features and/or features described in connection with additional methods disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this method can readily be combined with sets of base features identified as implementations such as system overview, tracking, mapping, robot implementation, VR/AR implementation, etc.
Process 1200 continues at action 1220 where while waiting for a new frame, between successive camera frames, the initial pose is updated using inertial data from the multi-axis IMU, to generate a propagated pose.
At action 1230, drift between the propagated pose, based on the inertial data, and an actual perspective of a new pose, is corrected using the new frame captured by the camera. This action can include a number of actions as well:
At action 1240, using the propagated pose, an overlap between the successive camera frames is estimated to reduce computation requirements, and the new frame is correlated with a previous frame by 2D comparison of the successive camera frames, beginning with the estimated overlap.
At action 1250, at least some feature points within a field of view of the propagated pose are retrieved from a 3D map using the propagated pose.
At action 1260, new features are extracted from the new frame.
At action 1270, the extracted new features are matched to the retrieved feature points based on (1) reuse of matched features from the previous frame and (2) matching of features in the new frame with reprojected feature positions from the 3D map onto a 2D view from a perspective of the propagated pose, producing a list of matching features.
At action 1280, a visually corrected pose is calculated using positions of the matching features in the list of matching feature to determine a perspective from which the new frame was viewed by the camera.
At action 1290, responsive to requests for location of the mobile unit and/or the perspective of the camera, data based on one or both of the propagated pose, based on the inertial data, and the visually corrected pose is provided.
Other implementations of the method described in this section can include a non-transitory computer readable storage medium storing instructions executable by a processor to perform any of the methods described above. Yet another implementation of the method described in this section can include a system including memory and one or more processors operable to execute instructions, stored in the memory, to perform any of the methods described above.
The method described in this section and other sections of the technology disclosed can include one or more of the following features and/or features described in connection with additional methods disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this method can readily be combined with sets of base features identified as implementations such as system overview, tracking, mapping, robot implementation, VR/AR implementation, etc.
Process 1300 continues at action 1320 where the keyrigs are reviewed to select keyrig content to include in a point cloud of features, based upon comparisons of keyrig content with content of other selected keyrigs subject to one or more intelligent thresholds. As used herein, intelligent thresholds includes a threshold that is adaptively adjusted based upon device movement, sensor readings, situational or environmental variables (e.g., low light, fog, bright light, and so forth) or combinations thereof.
At action 1330, for selected keyrigs, (a) new feature points in the keyrig are triangulated using feature points of keyrigs previously added to the point cloud of features to obtain feature points in a coordinate system of the device, and (b) coordinates of the feature points in the point cloud of features are aligned to a coordinate system having a z-axis aligned with gravity.
At action 1340, a multilayered hybrid point grid is created from the feature points selected for the point cloud of features, using at least one layer of a multilayered 2D occupancy grid by the following:
At action 1350, a 2D occupancy grid corresponding to one selected from a plurality of x-y layers covering the feature points in the point cloud of features is initialized.
At action 1360, at least one layer of the occupancy grid is populated with points from the point cloud of features within a height range using ray tracing from an observed location of a point in the keyrig aligned to a corresponding point in the occupancy grid and a location of a corresponding point reprojected on the layer of the occupancy grid.
At action 1370, cells along a ray between the aligned observed point and the corresponding point reprojected on the layer are found and the found cells are marked as empty.
At action 1380, responsive to receiving a command to travel to a location, the occupancy grid is used to plan a path of travel to a location commanded and contemporaneously the descriptive point cloud is used while traveling the planned path to avoid colliding with obstructions.
Other implementations of the method described in this section can include a non-transitory computer readable storage medium storing instructions executable by a processor to perform any of the methods described above. Yet another implementation of the method described in this section can include a system including memory and one or more processors operable to execute instructions, stored in the memory, to perform any of the methods described above.
We describe a system and various implementations for updating a position of a mobile unit that includes a camera with distance calculation and multi-axis inertial measuring unit (IMU).
Some additional implementations and features include:
Some particular implementations and features are described in the following discussion.
In one implementation, described is a system including a mobile platform having disposed on it at least one camera; a multi-axis inertial measuring unit (IMU); and an interface to a host. The host can include one or more processors coupled to a non-transitory memory, the memory loaded with computer instructions to updating a position of the mobile platform that includes the camera with distance calculation and multi-axis IMU. When executed on the processors, the instructions implement various actions, for example, receiving a location of the mobile unit and perspective, including view direction, of the camera, referred to collectively as an initial pose is performed by the system.
While waiting for a new frame, between successive camera frames, the system updates the initial pose using inertial data from the multi-axis IMU to generate a propagated pose.
The system corrects drift between the propagated pose, based on the inertial data, and an actual perspective of a new pose, using the new frame captured by the camera, by:
using the propagated pose, estimating an overlap between the successive camera frames to reduce computation requirements, correlating the new frame with a previous frame by 2D comparison of the successive camera frames, beginning with the estimated overlap; retrieving at least some feature points within a field of view of the propagated pose from a 3D map using the propagated pose; extracting new features from the new frame; matching the extracted new features to the retrieved feature points based on (1) reuse of matched features from the previous frame and (2) matching of features in the new frame with reprojected feature positions from the 3D map onto a 2D view from a perspective of the propagated pose, producing a list of matching features; and calculating a visually corrected pose using positions of the matching features in the list of matching feature to determine a perspective from which the new frame was viewed by the camera.
Responsive to requests for location of the mobile unit and/or the perspective of the camera, the system provides data based on one or both of the propagated pose, based on the inertial data, and the visually corrected pose.
Other implementations include methods including performing the actions of the system, non-transitory machine readable storage media storing program logic implementing such methods, substituents and components thereof, and devices incorporating any or combinations of the foregoing.
The terms and expressions employed herein are used as terms and expressions of description and not of limitation, and there is no intention, in the use of such terms and expressions, of excluding any equivalents of the features shown and described or portions thereof. In addition, having described certain implementations of the technology disclosed, it will be apparent to those of ordinary skill in the art that other implementations incorporating the concepts disclosed herein can be used without departing from the spirit and scope of the technology disclosed. Accordingly, the described implementations are to be considered in all respects as only illustrative and not restrictive.
The following detailed description is made with reference to the figures. Sample implementations are described to illustrate the technology disclosed, not to limit its scope, which is defined by the claims. Those of ordinary skill in the art will recognize a variety of equivalent variations on the description that follows.
The discussion is organized as follows. First, a high-level description of one implementation will be discussed at an architectural level. Next, apparatuses and processes used by some implementations to efficiently process image and data from the auxiliary sensors and build a 3D map are discussed. The references to specific examples are intended to be illustrative of the approaches disclosed herein rather than limiting.
This document describes 3D map building techniques employed by an autonomous mobile unit by using visual sensory data from a multi-ocular system that provides a 360-degrees view. In one embodiment, the multi-ocular sensor includes four interfaces to couple with four cameras, each camera providing a field of view of at least 200 degrees. The four cameras are configured into two back-to-back pairs, whereby each pair of cameras can be disposed of horizontally or vertically with respect to the other pair of cameras. As each camera has a field of view of at least 200 degrees, each back-to-back pair of cameras provides a 360-degrees view. Therefore two back-to-back pairs of cameras will provide two 360-degrees views. By comparing these two 360-degrees views, the relative depth information of objects around the autonomous mobile unit can be obtained in the form of a disparity map, which encodes the difference in horizontal coordinates of corresponding image points. The values in this disparity map are inversely proportional to the scene depth at the corresponding pixel location. As used herein, the term “quadocular” will be used to describe a multi-ocular system that includes four cameras, each providing a field of view of at least 200 degrees. In other embodiments, the multi-ocular sensor may include more than one camera, each camera providing a field of view of more than 180 degrees. Such sensors can be used to identify and localize absolute or relative position, create maps through simultaneous localization and mapping (SLAM) or structure from motion (SfM), and detect stationary and non-stationary obstacles.
When an autonomous mobile unit moves, images by the quadocular system can be utilized to build a 3D map. Therefore, the pair of 360-degrees frames can be compared by extracting and matching key features in the frames. One main limitation of this process is the inaccurate estimation of the absolute depth scale of the movement of the mobile unit. Indeed, even if camera pose estimation and scene reconstruction are carried out accurately, the error in the estimated depth of key features by stereo matching grows quadratically with depth or distance of the key features. In other words, the inaccuracy of the depth of key features in the far range is much higher than that in the near range. While the depth accuracy in the far range is unusably bad, the depth accuracy in the near range is significant. Some approaches suggest solving the issue via object detection by matching the scene with a pre-defined set of 3D models, so to recover the initial depth scale based on the estimated object size, which nevertheless fails in the absence of known shapes in the scene.
Data from one or more auxiliary sensors can be used to improve the pose accuracy and depth accuracy. In one embodiment, the auxiliary sensor can be a multi-axis inertial measurement unit (IMU), which measures and reports the mobile unit's linear acceleration using one or more accelerometers and rotational rate using one or more gyroscopes. In one embodiment, the auxiliary sensor can be a global positioning system (GPS) receiver, which provides the autonomous mobile unit's geolocation at a given time. In one embodiment, the mobile unit can have an IMU and a GPS receiver as its two auxiliary sensors. As used herein, the term “quadocular-auxiliary sensor” will be used to describe any multi-ocular sensors that include one or more auxiliary sensors.
The cameras or the imaging sensors in the multi-ocular system can be RGB or grayscale. Using low-end imaging sensors to construct a sensor, e.g., cameras having a resolution of 1280×720, obviates the cost of high-end image sensors. Wider field of view can be achieved by ultra wide-angle fisheye lenses that produce strong visual distortion intended to create a wide panoramic or hemispherical image.
Autonomous units such as autonomous vehicles, delivery robots and telepresence robot (moving around a remote environment automatically) can benefit from fast, accurate and reliable positional awareness. For example, for an autonomous unit to follow a road, it needs to know its location on the road, where it has been previously and where it is planning on going. For the autonomous vehicle to stay in a particular lane, it needs to know the location of the lane markers. When an obstacle is detected in the planned path, the planned route needs to be modified by the autonomous unit to avoid the obstacle and continue its way to its destination. In general, highways tend to be more predictable and orderly, with road surfaces typically well maintained and lanes well-marked. In contrast, residential or urban driving environments feature a much higher degree of unpredictability with many generic objects, inconsistent lane markings, and elaborate traffic flow patterns. For an autonomous unit to stay in a lane, the localization requirements are in the order of decimeters.
The map described herein is a three-dimensional (3D) map used for navigation. In one implementation, a 3D map is created having an accuracy in a range of 10 centimeters to 20 centimeters. A further implementation provides a 3D map having an accuracy in a range of up to 10 cm. A still further implementation provides a 3D map having an accuracy of up to 10 cm. One measurement of accuracy is a difference between a location of an object depiction on the 3D map and an actual location of the object in space. Embodiments of the present invention provide improved methods for mapping in autonomous units, allowing rich, dynamic information to be downloaded rapidly and efficiently from a map server. In one implementation, the 3D maps are stored on the server and are available to be downloaded by various autonomous units. In one implementation, the 3D map is constructed at a map server with visual, inertial and geolocation data provided by one or more autonomous units. In another implementation, 3D maps are built by autonomous units with quadocular-auxiliary sensors and then stored at a map server. In another implementation, autonomous units with quadocular-auxiliary sensors and the map server share the workload of building the 3D map. Implementations can provide improved map generation. Some implementations can provide generating maps with better accuracy. Some implementations can provide improved guidance of electro-mechanical systems.
The 3D maps are constructed in a hierarchy, wherein each step corresponds to a different level or type of feature on the 3D map. In one implementation, the construction of 3D maps has three distinct steps:
In one implementation, the autonomous unit's path of travel can be planned using a proto-roadmap. The proto-roadmap consists of an existing digital map used for navigation in autonomous and non-autonomous vehicles and is distributed to the autonomous units by a map server. Digital maps usually contain accurate representations of a particular area, detailing major and minor road arteries, points of interests and service locations. Some existing digital maps may have sub-meter level accuracy. Some existing digital maps may have lane-level accuracy. As used herein, the term “sub-meter level accuracy” will be used to describe accuracy within one meter. Example digital maps include Google Maps™, MapQuest, Bing Maps™, Rand McNally™, Here™, Waze™, OpenStreetMap™, Navmii™, Sygic™, Scout™, MAPS.ME™ and CoPilot™.
In one implementation, the map server or the autonomous units may not have enough semantic information to identify one or more objects on the 3D map, and thereby, not include the unidentified object in the final 3D map. Each constructed 3D map may store a time of day or weather when the map was created. The master 3D map in the map server may include one or more 3D maps of the same location, each 3D map constructed with visual information captured at various times of the day or at various weather conditions.
The 3D map can be available in one of the following data formats: Extended Markup Language (XML), AutoCAD™ DXF, Cartesian coordinate system (XYZ), Digital line graph (DLG), Esri™ TIN, Geography Markup Language (GML), GeoJSON, GeoMedia, Keyhole Markup Language (KML), MapInfo™ TAB, National Transfer Format (NTF), Spatialite, Shapefile, Simple Features, SOSI, Spatial Data File, and Topologically Integrated Geographic Encoding and Referencing Vector Product Format (TIGER VPF). Those of ordinary skill in the art will recognize a variety of equivalent variations.
Examples of systems, apparatus, and methods according to the disclosed implementations are described in an autonomous vehicle context. The technology disclosed relates to improving utilization of computing resources such as computational power and memory use during processing of image and data from auxiliary sensors inside a single input-multiple data (SIMD) architecture. The technology disclosed can be implemented in the context of any computer-implemented system including a reduced instruction set (RISC) system, emulated hardware environment, or the like. Moreover, this technology can be implemented using two or more separate and distinct computer-implemented systems that cooperate and communicate with one another. This technology can be implemented in numerous ways, including as a process, a method, an apparatus, a system, a device, a computer-readable medium such as a computer-readable storage medium that stores computer readable instructions or computer program code, or as a computer program product comprising a computer usable medium having a computer readable program code embodied therein. The technology disclosed can be implemented in the context of any computer-implemented system like a NEON ARM VFP9-S processor, an ARM core processor, an INTEL SSE SIMD processor or a compatible processor implementation.
The system 1400 includes an autonomous vehicle 1405 with a quadocular-auxiliary sensor 1410, an autonomous vehicle 1406 with a quadocular-auxiliary sensor 1412, an autonomous vehicle 1408 with a quadocular-auxiliary sensor 1414, a delivery robot 1416 with a quadocular-auxiliary sensor 1420, a delivery robot 1418 with a quadocular-auxiliary sensor 1422, the network(s) 1404, and a map hosting server system 1402. For clarity sake, only three autonomous vehicles and two delivery robots are shown to be connected to the map hosting server system 1402 through the network(s) 1404. However, any number of autonomous units with quadocular-auxiliary sensors can be connected to the map hosting server system 1402 through the network(s) 1404.
The interconnection of the elements of system 1400 will now be described. Network(s) 1404 couples the autonomous vehicles 1405, 1406, 1408, delivery robots 1416, 1418, and the map hosting server system 1402, all in communication with each other (indicated by solid double-arrowed lines). The actual communication path can be point-to-point over public and/or private networks. The communications can occur over a variety of networks, e.g., private networks, VPN, MPLS circuit, or Internet, and can use appropriate application programming interfaces (APIs) and data interchange formats, e.g., Representational State Transfer (REST), JavaScript™ Object Notation (JSON), Extensible Markup Language (XML), Simple Object Access Protocol (SOAP), Java Message Service (JMS), and/or Java Platform Module System. All of the communications can be encrypted. The communication is generally over a network such as the LAN (local area network), WAN (wide area network), telephone network (Public Switched Telephone Network (PSTN), Session Initiation Protocol (SIP), wireless network, point-to-point network, star network, token ring network, hub network, Internet, inclusive of the mobile Internet, via protocols such as EDGE, 3G, 4G LTE, Wi-Fi, and WiMAX. Additionally, a variety of authorization and authentication techniques, such as username/password, Open Authorization (OAuth), Kerberos, SecureID, digital certificates and more, can be used to secure the communications.
Quadocular-Auxiliary Sensor
An IMU 1510 provides raw sensor data for agent localization pipeline, which consumes IMU data at a high frequency at least 200 Hz to generate agent positional information in real-time. In an implementation, the localization pipeline combines information from IMU 1510 which runs at relatively high frequency to provide frequent updates of less accurate information, and cameras 1502, 1504, 1506, 1508 which run at a relatively lower frequency, e.g., 30 Hz in one implementation, to provide more accurate information with less frequency.
A GPS provider 1512 provides a global position estimate at, e.g., 10 Hz in one implementation. While GPS is a relatively accurate localization sensor, its update rate may be too slow to provide real-time updates. IMU 1510 provides a faster update with less accurate results. The IMU 1510 propagates the autonomous mobile unit's position periodically, such as for example every 5 milliseconds in one implementation, but the error accumulates as time progresses. Therefore, in one implementation, every 100 milliseconds, a GPS update is received which helps correct the IMU error. Data from the GPS receiver implementations has an accuracy of about one meter. Furthermore, the GPS signal can experience multipath problems, meaning that the signal may bounce off buildings and introduce more noise. A GPS receiver also benefits from an unobstructed view of the sky and thus is less effective in closed environments.
The control unit 1514 controls the sensors (IMU 1510, GPS receiver 1512 and cameras 1502, 1504, 1506, 1508), performs pre-computation in order to accelerate the localization pipeline, and packages raw data 1516 to be sent to a host.
The control unit 1514 enables the quadocular-auxiliary sensor 1500 to interact with a host. The host (not shown in
Control Unit
Referring to
Dust, dirt or scratches on the camera lenses of the quadocular-auxiliary sensor block incoming light before it reaches the visual sensors, and appears as dark spots in images. The presence of these artifacts will result in an image different than the scene in front of the cameras, and affect detection of key features. Image correction of artifacts by artifact removal processor 1803 can be used to recreate missing pixel data. Various algorithms used for image correction by artifact removal processor 1803 will be readily apparent to those skilled in the art.
The image undistortion processor 1804 corrects distortion in the image data in the captured frames. The image distortion is generally referred to an optical aberration that deforms and bends physically straight lines and makes them appear curvy in images. Optical distortion occurs as a result of optical design. To achieve reliable computer vision results, image undistortion processor 1804 can undistort the image before further processing is performed. This can be achieved by using a lookup table of the size of the input image and performing a remapping operation to undistort the whole image.
In cases when the remaining portions of the processing pipeline do not require the whole image, but only the feature points within the image, the feature undistortion processor 1808 perform a feature undistortion operation on the CU. In detail, this operation runs after the feature extraction stage and undistorts each feature point.
The processor feature detection engine 1806 performs feature detection upon image frames using Shi-Tomasi. Features are “interesting” parts of an image. The Shi-Tomasi feature detection includes methods that aim at computing abstractions of image information and making local decisions at every image point whether there is an image feature of a given type at that point or not. The resulting features will be subsets of the image domain, often in the form of isolated points. Some implementations perform the feature detection on the control unit 1514 to relieve the host from performing such tasks, and to accelerate the feature detection process. A type of feature can be (A) two small Eigenvalues: the feature is not interesting; (B) one small, one big value: the feature is likely an edge; (C) two big values: the feature is likely a corner; and (D) other type of features. Accordingly, in an implementation, processing includes: (a) action 1: calculate Eigenvalue of the intensity value of each pixel and its surrounding pixels, and determine (i) whether the feature is of interest; and (ii) for features of interest, a type of feature; and (b) action 2: refine by applying non-maximum suppression, applying spatial binning, applying heuristics or applying other types of refinement.
The feature description engine 1810 performs feature description on detected features. The feature description includes methods to identify each detected points in an image uniquely. Feature description can be used to compare and match feature points between different images. Some implementations perform the feature description on the control unit 1514 to relieve the host from performing such tasks, and to accelerate the feature description process.
One implementation of feature description engine 1810 uses a SIMD-accelerated ORB descriptor to describe features. The description of a feature can be used for matching purposes and describe a feature's uniqueness. The ORB descriptor approach was selected for its relative rotational invariance and immunity to Gaussian image noise. One example of an ORB feature detector and binary descriptor can be found at “ORB feature detector and binary descriptor,” http:slashslash scikit-image.org/docs/dev/auto_examples/plot_orbdothtml (last accessed Aug. 17, 2016). For further information on ORB Descriptor, reference may be had to Ethan Rublee, et al., “ORB: an efficient alternative to SIFT or SURF,” which is incorporated herein by reference for all purposes.
In some implementations, the feature detection engine 1806 and the feature description engine 1810 can be combined, as demonstrated in Convolutional Neural Networks (CNNs). A convolution neural network is a type of deep neural network that is widely used in object recognition tasks. A general CNN evaluation pipeline usually consists of the following layers: (i) the convolution layer uses different filters to extract different features from the input image with each filter containing a set of “learnable” parameters that will be derived from the training stage; (ii) the activation layer decides whether to activate the target neuron or not; (iii) the pooling layer reduces the spatial size of the representation to reduce the number of parameters and consequently the computation in the network; and (iv) the fully connected layer connects all neurons to all activations in the previous layer. CNNs have proven very useful in areas such as feature recognition and classification in images from visual sensors. CNNs have managed to identify faces, objects and traffic signs for autonomous robots and vehicles. CNNs are data-centric and require heavy computation. In the last few years, increases in available storage and computation capabilities have enabled CNNs to achieve success in supervised perception tasks of key features in images from visual sensors. After training for days or even weeks on a large data set, a CNN can be capable of in real-time perception of key features in images. For example, CNNs have achieved top results with ImageNet Large Scale Visual Recognition Challenge (ILSVRC) dataset, which contains 1-2 million images in over 1000 categories.
In autonomous vehicles, CNNs can be used to perform lane and vehicle detection while running at frame rates required for a real-time system. A labeled data set is required to represent all possible driving environments (rain, snow, night, day, etc.), and scenarios (pedestrian crossing the road, a car in front, mailbox by the side of the road, etc.). Various versions of CNNs can be used to implement the technology disclosed herein, e.g., AlexNet, Overfeat CNN, R-CNN, etc.
The optical flow feature correspondence processor 1812 performs 2D feature correspondence generation for the features. The feature correspondence computation is used to identify the feature points that appear on both the left and the right cameras. Once feature correspondence is identified for any two feature points, triangulation can be applied to the feature points to derive the depth of the point in space. This depth information is employed by processes later in the localization pipeline. Some implementations perform the feature correspondence generation on the control unit 1514 to relieve the host from performing such tasks, and to accelerate the feature correspondence generation.
One optical flow feature correspondence processor 1812 implementation employs optical flow methods to calculate the motion between two image frames, taken at times t and t+Δt at each voxel position. One such method, called a differential method, is based on local Taylor series approximations of the image signal, using partial derivatives concerning the spatial and temporal coordinates. Accordingly, in an implementation, processing includes: (a) input: last image, current image, a list of detected feature locations from the last image; (b) output: a list of locations of the last image's detected features' in the current image; (c) assumption: brightness consistency, image changes by and only by motion; (d) action 1: predict the output locations by either just assuming there is no motion, so the current locations are the same as last frame's feature locations, OR use information retrieved from the auxiliary sensors to predict the locations; and (e) action 2: refine the pixel coordinate for each feature point by searching around the predicted location, matching patches, and using matching score to determine the refined position. Accordingly, the technology disclosed can provide implementations with the ability to gain increased performance by using information from the auxiliary to narrow the search and save time.
Referring again to
The time stamping processor 2004 of an inertial component 1612 time stamps each set of inertial measurement data that the control unit 1514 receives from the IMU sensor 1510 data, in order to assure that the quadocular-auxiliary sensor 1500 maintains a temporally accurate stream of sensor data. Such rigorous attention to maintaining the integrity of the sensor data stream enables implementations to provide agent localization that works reliably. Time-stamping raw data by the quadocular-auxiliary sensor obviates the need for complex synchronization tasks.
The bias correction processor 2006 corrects IMU data read out from the time-stamped inertial data. Due to manufacturing imperfections, IMU sensors usually have bias problems such that its measurements contain errors. A bias error, if not removed from the measurement, is integrated twice as part of the mechanization process. In this case, a constant bias (error) in acceleration becomes a linear error in velocity and a quadratic error in position. A constant bias in attitude rate (gyro) becomes a quadratic error in velocity and a cubic error in position. The bias can be derived from the offline factory sensor calibration stage. This calibration information in control unit 1514 to perform bias correction task on the control unit 1514.
The scale correction processor 2008 applies stored scale factor information to the corrected inertial data. Scale factor error is the relationship between input and output. If the input is 100%, the expected output is 100%. The actual output is the result of a linear effect, where the output is proportional to the input but scaled. For example, if the input is 10 m/s2, but there is a 2% scale factor error, the output measurement is 10.2 m/s2. The scale factor can be derived from the offline factory sensor calibration stage. This calibration information in control unit 1514 to perform scale correction task on the control unit 1514.
The misalignment correction processor 2010 corrects misalignments of sensory elements of the inertial measurement sensor. There is a 3-axis gyroscope, and a 3-axis accelerometer mounted orthogonally to each other. The mountings, however, have errors and so are not perfectly at 90 degrees. This leads to a correlation between sensors. For example, in one configuration, one axis is pointed perfectly up, and the IMU is level. The accelerometer on this axis measures gravity. If the other two axes were perfectly orthogonal, they do not measure any of the effects of gravity. If there is a non-orthogonality, the other axes also measure gravity, leading to a correlation in the measurements. The effect of non-orthogonality occurs within sensor sets (between accelerometers or gyroscopes), between sensor sets or between the sensor sets and the enclosure (package misalignment). Careful manufacturing, as well as factory calibration, can help minimize this error source. Continuous estimation and correction during system operation is also an approach used to minimize this effect. Package misalignment (between the IMU 1510 and the enclosure) can be removed by performing a bore-sighting estimation to determine the offset between the IMU 1510 measurement frame and the sensor (objective) frame. The misalignment numbers can be derived from the offline factory sensor calibration stage. This calibration information in control unit 1514 to perform misalignment correction task on the control unit 1514.
In some implementations, the IMU 1510 and the cameras 1502, 1504, 1506 and 1508 do not reside at the same physical location; there is a distance between the IMU 1510 and the cameras 1502, 1504, 1506 and 1508. Accordingly, in order to enable later processes in the localization pipeline to treat the IMU 1510 and the cameras 1502, 1504, 1506 and 1508 as being co-located, on implementation determines a transformation matrix between the IMU 1510 and the cameras 1502, 1504, 1506 and 1508, which can be achieved from an offline production or post-production calibration stage. In control unit 1514, this transformation matrix is stored locally and applied to the IMU data. This technique enables later processes to be able to treat the IMU 1510 and the cameras 1502, 1504, 1506 and 1508 to be co-located
In an embodiment, 3D maps for navigation are built in the control unit 1514.
The sparse 3D map builder 2203 builds a sparse 3D mapping with visual data from the quadocular-auxiliary sensor. The sparse 3D mapping can be used by the autonomous unit to localize itself and have positional awareness continuously, e.g., the sensor system can determine where it is in the real world.
The 3D map builder 2204 builds a 3D map from the sparse 3D mapping created by the sparse 3D map builder 2203 with visual data from the quadocular-auxiliary sensor.
The semantic view builder 2205 identifies objects in the visual data with the help of the feature description engine 1810 or a combined feature detection engine 1806 and feature description engine 1810 (Convolutional Neural Networks). The semantic information is used to classify the objects in the visual data as moving or non-moving.
Referring now to
The computing environment may also include other removable/non-removable, volatile/nonvolatile computer storage media. For example, a hard disk drive may read or write to non-removable, nonvolatile magnetic media. A magnetic disk drive may read from or write to a removable, nonvolatile magnetic disk, and an optical disk drive may read from or write to a removable, nonvolatile optical disk such as a CD-ROM or other optical media. Other removable/non-removable, volatile/nonvolatile computer storage media that can be used in the exemplary operating environment include, but are not limited to, magnetic tape cassettes, flash memory cards, digital versatile disks, digital video tape, solid state RAM, solid state ROM, and the like. The storage media are typically connected to the system bus through a removable or non-removable memory interface.
In an embodiment, the processor 2332 is a NEON ARM processor implementing a single input-multiple data (SIMD) architecture as a reduced instruction set computer (RISC) architecture. Depending on implementation, however, processor 2332 can alternatively be a realized using a specific purpose microcontroller, peripheral integrated circuit element, a CSIC (customer-specific integrated circuit), an ASIC (application-specific integrated circuit), a logic circuit, a digital signal processor, a programmable logic device such as an FPGA (field-programmable gate array), a PLD (programmable logic device), a PLA (programmable logic array), an RFID processor, smart chip, or any other device or arrangement of devices that are capable of implementing the actions of the processes of the technology disclosed.
Communications interface 2342 can include hardware and/or software that enables communication between quadocular-auxiliary positioning system 2300 and other systems controlling or enabling customer hardware and applications (hereinafter, a “host system” or “host”) such as for example, a robot or other guided mobile platform, an autonomous vehicle, or the like (not shown in
Instructions defining FASAT process 2344 are stored in memory 2334, and these instructions, when executed, perform analysis on image frames captured by the cameras 1502, 1504, 1506 and 1508, inertial data captured by the IMU 1510 and geolocation data from the GPS receiver 1512 connected to quadocular-auxiliary positioning system 2300. In one implementation, FASAT process 2344 includes various logical processes, such as a feature extractor 2352 that receives a raw image and determines a salient points' representation of objects in the image thereby representing the geometry understanding of the objects from a machine's perspective view. In some implementations, feature extractor 2352 analyzes images (e.g., image frames captured via cameras 1502, 1504, 1506 and 1508) to detect edges of an object therein and/or other information about the object's location. A sensor fusion tracking process 2354 uses feature extraction results, inertial data from the IMU 1510 and geolocation data from the GPS receiver 1512 to generate pose accurately and rapidly. A smart interaction map 2356 enables using a known map of obstructions to localize the sensor. The map is built using mapping functionality of mapping process 2392, which is described in further detail herein below. A Re-localizer process 2358 recovers device positional awareness when the device has lost track of device position. A system diagnostic and response (SDAR) 2366 manages of current localizing state of the device and provide response strategy.
A mapping process 2392 generates a 3D map that maps the surrounding and objects recognized by the feature extractor 2352 in the ground perspective view and spatial perspective view of the 3D maps.
In some implementations, other processing 2394 analyzes audio or ultrasonic signals (e.g., audio signals captured via sonar or audio sensors comprising non-tactile input 2382) to localize objects and obstructions by, for example, time distance of arrival, multilateralism or the like (“Multilateration is a navigation technique based on the measurement of the difference in distance to two or more stations at known locations that broadcast signals at known times. See Wikipedia, at <http:slashslash en.wikipediadotorg/w/index.php?title=Multilateration&oldid=523281858>, on Nov. 16, 2012, 06:07 UTC). Audio signals place the object on a known surface, and the strength and variation of the signals can be used to detect object's presence. If both audio and image information is simultaneously available, both types of information can be analyzed and reconciled to produce a more detailed and/or accurate path analysis.
In some implementations, other processing 2394 determines paths to track and predict device movements in space based upon the 3D maps generated by mapping process 2392. One or more applications 2396 can be loaded into memory 2334 (or otherwise made available to processor 2332) to augment or customize functioning of the quadocular sensors thereby enabling the system 2300 to function as a platform. Successive camera images are analyzed at the pixel level to extract object movements and velocities. In some implementations, presentation interface 2378 includes a video feed integrator provides integration of live video feed from the cameras 1502, 1504, 1506 and 1508, and one or more virtual objects. Video feed integrator governs the processing of video information from disparate types of cameras. For example, information received from pixels that provide quadochromatic imaging and from pixels that provide color imaging (e.g., RGB) can be separated by the integrator and processed differently. Information from one type of sensor can be used to enhance, correct, and/or corroborate information from another type of sensor. Information from one type of sensor can be favored in some types of situational or environmental conditions (e.g., low light, fog, bright light, and so forth). The device can select between providing presentation output based upon one or the other types of image information, either automatically or by receiving a selection from the user.
Presentation interface 2378, audio presentation 2379, non-tactile input 2382, and communications interface 2342 can be used to facilitate user interaction via a quadocular sensor with quadocular-auxiliary positioning system 2300. These components can be of highly customized design, generally conventional design or combinations thereof as desired to provide any type of user interaction. In some implementations, results of analyzing captured images using inertial measuring unit 1510, GPS receiver 1512 and cameras 1502, 1504, 1506, 1508 and FASAT program 2344 can be interpreted as representing objects and obstacles in 3D space. For example, a robot equipped with quadocular-auxiliary sensor can perform path planning and/or obstacle avoidance across a surface that has been analyzed using FASAT program 2344, and the results of this analysis can be interpreted as an occupancy map by some other program executing on processor 2332 (e.g., a motion planner, localization and tracking process, or other application). Smart interaction map 2356 may use the representation of space built by mapping 2392 to plan a path for a robot or mobile platform through space, e.g., to improve localization and tracking of the robot or platform through space.
It will be appreciated that the quadocular-auxiliary positioning system 2300 is illustrative and that variations and modifications are possible. Quadocular-auxiliary positioning systems can be implemented in a variety of form factors, including “cloud” computing systems of servers and networks, desktop systems, laptop systems, tablets, smartphones or personal digital assistants, and so on. A particular implementation may include other functionality not described herein for clarity sake. In some implementations, one or more cameras and two or more microphones may be built into the quadocular-auxiliary system 2300 or may be supplied as separate components. Further, an image or audio analyzer can be implemented using only a subset of a quadocular-auxiliary positioning system 2300 components (e.g., as a processor executing program code, an ASIC, or a fixed function digital signal processor, with suitable I/O interfaces to receive image data and output analysis results).
While quadocular-auxiliary positioning system 2300 is described herein with reference to particular blocks, it is to be understood that the blocks are defined for convenience of description and are not intended to imply a particular physical arrangement of component parts. Further, the blocks need not correspond to physically distinct components. To the extent that physically distinct components are used, connections between components (e.g., for data communication) can be wired and/or wireless as desired. Thus, for example, execution of feature extractor 2352 by processor 2332 can cause processor 2332 to operate IMU 1510, GPS receiver 1512 and cameras 1502, 1504, 1506, 1508 to capture images and/or audio signals of an object traveling across and in contact with a surface to detect its entrance by analyzing the image and/or audio data.
Tracking for Autonomous Units with Quadocular-Auxiliary Sensors
Tracking refers to capabilities rendered by system hardware and functional processes that enable a controlled device (robot, mobile platform, autonomous vehicles, or other hardware) to localize itself and have positional awareness continuously, e.g., the sensor system can determine where it is in the real world.
Architecture for Tracking System
Feature Extraction
A feature extractor (FE) 2402 represents the geometric understanding of an environment from an autonomous mobile unit's perspective view. Feature extractor 2402 receives raw image information from control unit 1514 and provides a propagated pose of the autonomous mobile unit to a sensor fusion tracker (SFT) 2412. One implementation of a feature extractor 2402 performs a variety of actions including image preprocessing, feature detection, and feature descriptor preparation. Image processing tasks also include performing artifact removal and Gaussian blur and gamma correction on raw image data.
Feature Detection: Optical Flow
Optical flow gives the 2D-2D correspondence between previous image and a current image. Feature extractor 2402 generates a hierarchy of levels; each level generation is aligned with optical flow needs, and therefore need only be computed once.
Processed images are stored in a multiple hierarchical patch and/or undistorted patch.
In an example implementation, as levels in the hierarchy increase, the more blur exists in the image. At each level in the hierarchy, the size of the image is reduced to ¼ of the size of the image from which it is derived, e.g., the size of the image in the previous level. For example, if an image on the first (“zero level”) has a size 640×480, then the size of the corresponding image on the next higher level (“first level”) has a size 320×240, and a next corresponding image at the next higher level (“second level”) has a size 160×120. If the same feature point appears at more than one level, then that feature point is determined by the system to be a strong feature point. In the foregoing example, a scale factor of 2 is used. However, any scale factor can be selected, e.g., a scale factor of 1.6 provides the original 640×480 image with a next level image having a resolution of 400×300.
One feature extractor 2402 implementation employs optical flow methods to calculate the motion between two image frames, taken at times t and t+Δt at each voxel position. One such method, called a differential method, is based on local Taylor series approximations of the image signal, using partial derivatives with respect to the spatial and temporal coordinates. Accordingly, in an implementation, processing includes: (a) input: last image, current image, a list of detected feature locations from the last image; (b) output: a list of locations of the last image's detected features' in the current image; (c) assumption: brightness consistency, image changes by and only by motion; (d) action 1: predict the output locations by either just assuming there is no motion, so the current locations are the same as last frame's feature locations, OR use information retrieved from one or more auxiliary sensors to predict the locations; and (e) action 2: refine the pixel coordinate for each feature point by searching around the predicted location, matching patches, and using a matching score to determine the refined position. Accordingly, the technology disclosed can provide implementations with the ability to gain increased performance by using information from one or more auxiliary sensors to narrow the search and save time.
Feature Detection:
One implementation of feature extraction processor 2402 uses NEON-accelerated Shi-Tomasi feature detection with spatial binning in order to obtain features from regions to cover as much as possible for the full image. The feature extractor 2402 uses data from one or more auxiliary sensors and the pose information received from sensor fusion tracker 2412 in order to dynamically decide the regions to track and the parameters to use. Features are “interesting” parts of an image. A type of feature can be (A) two small Eigenvalues: the feature is not interesting; (B) one small, one big value: the feature is likely an edge; (C) two big values: the feature is likely a corner; and (D) other type of features. Accordingly, in an implementation, processing includes: (a) action 1: calculate Eigenvalue of the intensity value of each pixel and its surrounding pixels, and determine (i) whether the feature is of interest; and (ii) for features of interest, a type of feature; and (b) action 2: refine by applying non-maximum suppression, applying spatial binning, applying heuristics or applying other types of refinement.
Feature Descriptor:
One implementation of feature extractor 2402 uses a SIMD-accelerated ORB descriptor to describe features. The description of a feature can be used for matching purposes and describe a feature's uniqueness. The ORB descriptor approach was selected for its relative rotational invariance and immunity to Gaussian image noise. One example of an ORB feature detector and binary descriptor can be found at “ORB feature detector and binary descriptor,” http:slashslashscikit-image.org/docs/dev/auto_examples/plot_orbdothtml (last accessed Aug. 17, 2016). For further information on ORB Descriptor, reference may be had to Ethan Rublee, et al., “ORB: an efficient alternative to SIFT or SURF,” which is incorporated herein by reference for all purposes.
Bag-of-Words
One implementation of feature extractor 2402 performs bag-of-words extraction, determining from one image a byte array, sparse histogram over vocabulary. Accordingly, in an implementation, Bag-of-words processing includes using a list of feature descriptors in one image to extract Bag-of-words feature for one image. For example, in an image having 400 features with 400 corresponding descriptors, and a bag-of-words feature with the size of 256 bits. The resulting Bag-of-words is a 256-bit array, and one image only needs/has one 256 bit array. Bag-of-words is useful in performing relocalization as described herein below in further detail. Bag-of-words can indicate among A, B, C images, an image D has a matching score individually D-A, D-B, D-C. Say for example that D-B has the highest score, and then the camera taking image D is closest to the camera pose used when taking image B, not A or C.
One implementation of feature extractor 2402 adaptively adjusts a threshold that is applied to a number of features needed in order for the system to keep track of a moving object. Such intelligent thresholds include a threshold that is adaptively adjusted based on device movement, sensor readings, situational or environmental variables (e.g., low light, fog, bright light, and so forth) or combinations thereof. Once a sufficient number of features (e.g., above the threshold) are identified, the system will continue to track an object. Otherwise, the system does not track the object. Feature extractor 2402 will adaptively adjust the threshold based on device movement and other sensor levels.
SFT: Generating Pose Information from Feature Extraction Results and Data from Auxiliary Sensors
Sensor fusion based tracker (SFT) 2412 determines how to smartly use feature extraction results and data from one or more auxiliary sensors (IMU 1510, GPS receiver 1512) to generate pose accurately and rapidly. Sensor fusion based tracker 2412 performs initialization for the system state, feature extractor initial state, interaction with map state, and other states. Sensor fusion based tracker 2412 has logic to decide how to initialize the state including use how many images to skip, and any other initialization steps for specific hardware.
In an implementation, sensor fusion based tracker 2412 performs buffer management for the auxiliary sensors. Timestamps of the data from the auxiliary sensors are corrected and aligned. Additionally, data information from the auxiliary sensors that have not been processed yet is maintained in a buffer.
Generating Pose Information for a Quadocular-Auxiliary Sensor where an IMU and a GPS Receiver are Used as the Auxiliary Sensors:
Propagation: In a quadocular-auxiliary sensor with two auxiliary sensors, an intermediate pose is estimated using a first propagated pose from data from one auxiliary sensor together with a second propagated pose from visual information. A final pose is estimated by calibrating the intermediate pose with a third propagated pose from the second auxiliary sensor. Those of ordinary skill in the art will recognize a variety of equivalent variations. In an implementation, an IMU can be the first auxiliary sensor while a GPS receiver can be a secondary auxiliary sensor.
Implementations can employ extended Kalman filtering (EKF) to extrapolate the final pose of the autonomous unit.
The method described in this section and other sections of the technology disclosed can include one or more of the following features and/or features described in connection with additional methods disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this method can readily be combined with sets of base features identified as implementations such as system overview, tracking, mapping, robot implementation, autonomous vehicle, etc.
At action 2510, a first propagated pose is estimated from the inertial data from the IMU. Using iterated Extended Kalman Filter (EKF) framework, a filter state Xt at time t is parameterized as Xt=[qGBt pBtG vBtG ba bg], where qGBt is the unit quaternion describing the rotation from the global frame {G} of the autonomous unit to the IMU body frame {B}, pBtG and vBtG are the IMU position and velocity with respect to {G}, and finally ba and bg are gyroscope and accelerometer biases. When a new IMU measurement (angular velocity wt and linear acceleration at) is received, the filter state Xt is propagated by the kinematic equations as follows:
R
G
Bt=(RBtG)T=(RBt-lGRBtBt-1)T Eq. (1)
R
Bt
Bt-1
=f((wt−bg)Δt) Eq. (2)
ΔwBt=(RBtG(at−ba)+gG)Δt Eq. (3)
v
{dot over (B)}t
=v
Bt-1
+Δv
Bt Eq. (4)
p
{dot over (B)}t
=p
Bt-1
+v
Bt-1
Δt+1/2ΔvBtΔt Eq. (5)
where f( ) is the function converting axis-angle rotation to rotation matrix, ΔvBt is the velocity increment, RGBt is the rotation matrix of the quaternion qGBt and gG represents the gravity correction in the global coordinate system. Propagation by the IMU can be performed in a number of ways: (i) the propagation happens right after an inertial measurement taken by the IMU arrives; or (ii) delayed propagation happens in a batch right after an image taken by the visual sensors arrives but before the update procedure described below is performed.
In some implementations, a motion sensor keeps a window of readings from the accelerometer of the IMU. The motion sensor computes the variance of the acceleration in the window of accelerator readings and determines whether the mobile unit is moving or static, based on the calculated variance. A variance greater than a predetermined threshold variance indicates the mobile unit is moving, and propagation is performed by using the EKF method. A variance smaller than a predetermined threshold variance indicates the mobile unit is static and propagation is performed by using the EKF method assuming zero velocity. The predetermined threshold variance can be estimated by various techniques as readily apparent to those skilled in the art.
At action 2520, a second propagated pose is when visual information from the cameras in the quadocular-auxiliary sensor is received. A set of correspondences between 2D image features In a single observation implementation, based on the current image observation, the information needed for propagation, e.g., features, poses, map points, etc. is prepared. Then 2D-3D correspondence information for the optical flow tracked 2D features is obtained by directly using 2D-2D correspondences from optical flow tracking results. For the remaining 3D map points, smartly search over 3D with adaptive parameters by: (i) if the remaining number of points is small, e.g., below a threshold, perform a linear search, otherwise, (ii) if the remaining number of points is fairly big, e.g., above a threshold, perform a log(n) search with the use of a kd-tree or octree. Next, perform the update, and generate pose for the current frame using only “good outstanding” points. Alternatively, perform the update twice: first only with high weights “good outstanding” points, then with all the map points. As used herein, the term “good outstanding” points denotes high-quality points, e.g., those points having smaller reprojection error and/or being tracked many times in a good state, which can be selected by application of some heuristics.
At action 2530, the first propagated pose is fused with the second propagated fuse to estimate an intermediate pose. Residual for each correspondence is the re-projection error of each 3D map point PG in the global coordinate.
(vu)=Å(RBCi(RBtG(PG−pBtG))+pBci) Eq. (6)
where (RBCi, pBCi) is the transform from the IMU to the coordinate of the i-th camera obtained through offline calibration, and RGBt is the rotation matrix of the quaternion qGBt. Å is the perspective projection of a 3D point to the image coordinate (vu). The actual EKF residual is constructed by concatenating residuals from all the measurements.
At action 2540, a second propagated pose is when visual information from the cameras in the quadocular-auxiliary sensor is received. The GPS receiver supplies the control unit the new geolocation of the autonomous mobile unit. The GPS receiver can provide position, velocity, and covariance information in North-east down (NED) format, or Doppler, pseudo-ranges, and covariance information in Earth-centered, Earth-fixed (ECEF) format. Propagation by the GPS receiver can be performed right after a new geolocation arrives from the satellite.
At action 2550, the third propagated pose is calibrated with the intermediate pose to estimate the final pose. Final pose consists of constant velocity propagation, SLAM update, and geolocation update. Calibration is done offline by solving the orientation alignment known as Wahba's problem (G. Wahba, “A Least Squares Estimate of Spacecraft Attitude,” SIAM Review, 1965, 7(3), pp. 409). [p v ψ] is sued to present the 9×1 vector containing position, velocity, and orientation (axis angle representation) of the autonomous unit. Equations 7-9 represent constant velocity propagation, SLAM update, and geolocation update respectively. μ is the zero mean Gaussian random noise with covariance; y is the zero mean Gaussian measurement noise with covariance. RGSB represents the rotation matrix from the GPS receiver to IMU body frame.
Other implementations of the method described in this section can include a non-transitory computer-readable storage medium storing instructions executable by a processor to perform any of the methods described above. Yet another implementation of the method described in this section can include a system including one or more memory and one or more processors operable to execute instructions, stored in the memory, to perform any of the methods described above.
SIM: Smart Map Interaction and Localize
Smart interaction with map (SIM) processor 2414 enables smart interaction by a device being guided (i.e., receiving positional awareness information from the quadocular-auxiliary sensor system) with a known map in order to localize itself. Accordingly, one implementation of a smart interaction map processor 2414 processing includes: (a) Perform a query of the map based on a need and a current status. Some query components included by specific implementations are map points (latitude, longitude, uncertainty, average reprojection error, etc.), keyrigs' poses, and occupancy grid (one layer or multiple layers). As used herein, the term keyrig denotes a set of image data received. In some implementations, the sets of image data include feature points and pose information. The pose information includes a location of the mobile device and view of the camera that captured the image data. The sets of image data are referred to collectively as keyrigs; (b) If the map as a whole is not big, e.g., meets a particular threshold size, return the entire map; (c) Otherwise, if the map is big, e.g., meets or exceeds a particular threshold size and covers large areas, e.g., spatial area meets or exceeds a particular threshold, smartly grab a local region of map based on the device's current position; (d) If some area of the map is known to have bad quality, don't use it; and (e) Adaptively down-sample map based on system performance according to an adaptive parameter. For example: (i) if the 3D feature-based tracking process has been accurate, e.g., below a threshold number of errors or invocations of relocalization processing to relocalize a device that has lost its place, then reduce the sample size to conserve resources, otherwise, (ii) if the device is getting lost frequently, e.g., meets or exceeds a threshold, then increase the sample size to provide better tracking.
SDAR: Manage Localization State and Response Strategy
System diagnostics and response (SDAR) processor 2403 performs system-level management of the current localizing state of a quadocular-auxiliary sensor and provides response strategy. In one implementation, the system diagnostics and response processor 2403 monitors tracked map points and corresponding reprojection error, monitors the longevity of untracked map points, and suggests actions to mapping module, monitors device speed/acceleration/jitter/drift, monitors map “activity” (i.e., is it updated regularly?), performs newly added map contribute enough to the overall mapping (i.e., so it's not doing useless hovering?), monitors keyrig topology (i.e., stretch enough or too much?), suggests when and where the device should add keyrig, and monitors feature extractor pipeline
Relocalize a Lost Device
Re-locator 2422 recovers a device that has lost track of its positional from this state. Once re-locator 2422 returns the lost device back on track, the device can resume normal interacting with smart interaction map processor 2414, and/or can contribute to building maps, using the techniques described herein below.
The physical hardware component of network interfaces are sometimes referred to as network interface cards (NICs), although they need not be in the form of cards: for instance they could be in the form of integrated circuits (ICs) and connectors fitted directly onto a motherboard, or in the form of macrocells fabricated on a single integrated circuit chip with other components of the computer system.
User interface input devices 2622 may include a keyboard, pointing devices such as a mouse, trackball, touchpad, or graphics tablet, a scanner, a touchscreen incorporated into the display, audio input devices such as voice recognition systems, microphones, and other types of input devices. In general, use of the term “input device” is intended to include all possible types of devices and ways to input information into map hosting server system 1402 or onto computer network 2618.
User interface output devices 2620 may include a display subsystem, a printer, a fax machine, or non-visual displays such as audio output devices. The display subsystem may include a cathode ray tube (CRT), a flat panel device such as a liquid crystal display (LCD), a projection device, or some other mechanism for creating a visible image. The display subsystem may also provide a nonvisual display such as via audio output devices. In general, use of the term “output device” is intended to include all possible types of devices and ways to output information from map hosting server system 1402 to the user or another machine or computer system.
Storage subsystem 2624 stores the basic programming and data constructs that provide the functionality of certain embodiments of the present invention. For example, the various modules implementing the functionality of certain embodiments of the invention may be stored in storage subsystem 2624 such as the image processor 2626, the map builder 2628, the proto-roadmap storage 2632 and the master 3D map storage 2634. The software modules the image processor 2626 and the map builder 2628 are generally executed by processor subsystem 2614.
Storage subsystem 2624 typically includes a number of memories including a main random access memory (RAM) for storage of instructions and data during program execution, a read-only memory (ROM) in which fixed instructions are stored, and a file storage subsystem that provides persistent storage for program and data files, and may include a hard disk drive, a floppy disk drive along with associated removable media, a CD ROM drive, an optical drive, or removable media cartridges. The map storages 2632, 2634 and modules implementing the functionality of certain embodiments of the invention may have been provided on a computer-readable medium such as one or more CD-ROMs and may be stored by file storage subsystem. The storage subsystem 2624 contains, among other things, computer instructions which, when executed by the processor subsystem 2614, cause the map hosting server system to operate or perform functions as described herein. As used herein, processes and software that are said to run in or on “the map server” or “the map hosting server system,” execute on the processor subsystem 2614 in response to computer instructions and data in the storage subsystem 2624 including any other local or remote storage for such instructions and data.
Bus subsystem 2612 provides a mechanism for letting the various components and subsystems of map hosting server system 1402 communicate with each other as intended. Although bus subsystem 2612 is shown schematically as a single bus, alternative embodiments of the bus subsystem may use multiple busses.
Map hosting server system 1402 itself can be of varying types including a personal computer, a portable computer, a workstation, a computer terminal, a network computer, a television, a mainframe, a server farm, or any other data processing system or user device. Due to the ever-changing nature of computers and networks, the description of map hosting server system 1402 depicted in
In addition, while the present invention has been described in the context of a fully functioning data processing system, those of ordinary skill in the art will appreciate that the processes herein are capable of being distributed in the form of a computer-readable medium of instructions and data and that the invention applies equally regardless of the particular type of signal bearing media actually used to carry out the distribution. As used herein, a computer-readable medium is one on which information can be stored and read by a map hosting server system. Examples include a floppy disk, a hard disk drive, a RAM, a CD, a DVD, flash memory, a USB drive, and so on. The computer-readable medium may store information in coded formats that are decoded for actual use in a particular data processing system. A single computer readable medium, as the term is used herein, may also include more than one physical item, such as a plurality of CD ROMs or a plurality of segments of RAM, or a combination of several different kinds of media. As used herein, the term does not include mere time-varying signals in which the information is encoded in the way the signal varies over time.
The feature detection processor 2706 performs feature detection upon image frames using Shi-Tomasi. Features are “interesting” parts of an image. The Shi-Tomasi feature detection includes methods that aim at computing abstractions of image information and making local decisions at every image point whether there is an image feature of a given type at that point or not. The resulting features will be subsets of the image domain, often in the form of isolated points. A type of feature can be (A) two small Eigenvalues: the feature is not interesting; (B) one small, one big value: the feature is likely an edge; (C) two big values: the feature is likely a corner; and (D) other type of features. Accordingly, in an implementation, processing includes: (a) action 1: calculate Eigenvalue of the intensity value of each pixel and its surrounding pixels, and determine (i) whether the feature is of interest; and (ii) for features of interest, a type of feature; and (b) action 2: refine by applying non-maximum suppression, applying spatial binning, applying heuristics or applying other types of refinement. The feature detection processor 2706 is used to detect 2D feature points of road markings in the ground perspective view layer in the 3D map. The feature detection processor 2706 also detects 2D feature points of objects in the spatial perspective view layer in the 3D map.
The feature description processor 2710 performs feature description on detected features. The feature description includes methods to identify each detected points in an image uniquely. Feature description can be used to compare and match feature points between different images. One implementation of feature description engine 2710 uses a SIMD-accelerated ORB descriptor to describe features. The description of a feature can be used for matching purposes and describe a feature's uniqueness. The ORB descriptor approach was selected for its relative rotational invariance and immunity to Gaussian image noise. One example of an ORB feature detector and binary descriptor can be found at “ORB feature detector and binary descriptor,” http:slashslashscikit-imagedotorg/docs/dev/auto_examples/plot_orb.html (last accessed Aug. 17, 2016). For further information on ORB Descriptor, reference may be had to Ethan Rublee, et al., “ORB: an efficient alternative to SIFT or SURF,” which is incorporated herein by reference for all purposes.
In some implementations, the feature detection processor 2706 and the feature description processor 2710 can be combined, as demonstrated in Convolutional Neural Networks (CNNs). A convolution neural network is a type of deep neural network that is widely used in object recognition tasks. A general CNN evaluation pipeline usually consists of the following layers: (i) the convolution layer uses different filters to extract different features from the input image with each filter containing a set of “learnable” parameters that will be derived from the training stage; (ii) the activation layer decides whether to activate the target neuron or not; (iii) the pooling layer reduces the spatial size of the representation to reduce the number of parameters and consequently the computation in the network; and (iv) the fully connected layer connects all neurons to all activations in the previous layer. CNNs have proven very useful in areas such as feature recognition and classification in images from visual sensors. CNNs have managed to identify faces, objects and traffic signs for autonomous robots and vehicles. CNNs are data-centric and require heavy computation. In the last few years, increases in available storage and computation capabilities have enabled CNNs to achieve success in supervised perception tasks of key features in images from visual sensors. After training for days or even weeks on a large data set, a CNN can be capable of in real-time perception of key features in images. For example, CNNs have achieved top results with ImageNet Large Scale Visual Recognition Challenge (ILSVRC) dataset, which contains 1-2 million images in over 2300 categories. In autonomous vehicles, CNNs can be used to perform lane and vehicle detection while running at frame rates required for a real-time system. A labeled data set is required to represent all possible driving environments (rain, snow, night, day, etc.), and scenarios (pedestrian crossing the road, a car in front, mailbox by the side of the road, etc.). CNNS can be used to determine semantic information of road markings in the ground perspective view layer and objects in the spatial perspective view layer in the 3D map. Various versions of CNNs can be used to implement the technology disclosed herein, e.g., AlexNet, Overfeat CNN, R-CNN, etc.
The sparse 3D map builder 2803 builds a sparse 3D mapping with visual data from the quadocular-auxiliary sensor. The sparse 3D mapping can be used by the autonomous unit to localize itself and have positional awareness continuously, e.g., the sensor system can determine where it is in the real world.
The 3D map builder 2804 builds a 3D map from the sparse 3D mapping created by the sparse 3D map builder 2803 with visual data from the quadocular-auxiliary sensor.
The semantic view builder 2805 identifies objects in the visual data from the quadocular-auxiliary sensor with the help of the feature description engine 2710 or a combined feature detection engine 2706 and feature description engine 2710 (Convolutional Neural Networks). The semantic information is used to classify the objects in visual data from the quadocular-auxiliary sensors moving or non-moving.
The map merger 2806 mergers one or more newly created 3D maps with the master multi-map in storage 2634. The map updater 2807 updates certain parts of the master 3D map in storage 2634 with information from one or more newly created 3D maps. The master multi-map is provided via a network(s) 1404 to one or more autonomous for guidance at a future time. The proto-roadmaps stored in storage 2632 are provided via the network(s) 1404 to one or more autonomous for guidance in locations for which 3D maps are not available.
In implementations, geometric information obtained by a 3D feature-based tracking process is represented using a descriptive point cloud representation. The descriptive point cloud is a set of sparse 3D points, where each point encodes a 3D geometric location, an uncertainty of the geometric location, and a set of 2D location plus appearance-based descriptors, each observed by a keyrig. A keyrig contains (i) a set of 360-degrees images; (ii) a timestamp for when the set of 360-degrees images in the keyrig is captured; (iii) a pose (i.e. the transformation from the quadocular-auxiliary sensor's coordinate to the map coordinate at the current time) and its uncertainty; and (iv) a sequence of readings from the auxiliary sensors starting from a first temporal point before the timestamp to a second temporal point after the timestamp. Some implementations further include a flag to specify whether the quadocular-auxiliary sensor is moving or static. This flag can be set from information obtained from the IMU readings.
Construction of a 3D Map
With the help of the captured keyrigs by autonomous units with quadocular-auxiliary sensors, a 3D map can be built for navigation with an accuracy in a range of 5 centimeters to 10 centimeters. In one implementation, the 3D map is built at the map server after the map server receives keyrigs from one or more autonomous units. The newly created map is stored or merged with an existing map at the map server. In another implementation, 3D maps are built by autonomous units with quadocular-auxiliary sensors and sent to the map server to be stored.
The 3D map building technology disclosed will be illustrated with reference to a set of keyrigs captured by an autonomous vehicle 2905 with a quadocular-auxiliary sensor in
In some implementations, a convolution neural network is employed to determine the feature points from the image frames and obtain feature description. Optical flow techniques are used to determine 2D correspondences in the images, enabling matching together features in different images. In some implementations, Shi-Tomasi feature detection is employed to determine the 2D feature points from the image frames. The 2D features points from the first 360-degrees image and the 360-degrees second image will be triangulated to determine the depth of the feature point. A location of feature points will be generated for each of the feature points detected. The map entry will include a global position, including longitude and latitude, and feature descriptors of the. The map entry might further include a time of day (e.g., daytime, nighttime, etc.) and/or the weather condition (e.g., sunny, cloudy, rainy, etc.) when the keyrig was captured.
With the help of the sparse 3D mapping of object feature points, the quadocular-auxiliary sensor-equipped autonomous vehicle 2905 can circumnavigate obstacles and self-localize within the urban environment 3000. This enables the autonomous vehicle 2905 to navigate the urban driving environment 3000 without hitting any obstacles.
After the construction of the sparse 3D mapping, a 3D map 3300 is created from the sparse 3D mapping 3200 to add more feature points.
Map Construction by Autonomous Units
The method described in this section and other sections of the technology disclosed can include one or more of the following features and/or features described in connection with additional methods disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this method can readily be combined with sets of base features identified as implementations such as system overview of quadocular-auxiliary sensors and the map server, 3D map construction, etc.
Process 3700 continues at action 3720 where the map server classifies the objects in the set of 360 degrees images as moving objects and non-moving objects with the help of the semantic perspective view builder 2805 in map builder 2628.
At action 3730, the map server selects a subset of keyrig from the set of keyrigs and builds a sparse 3D mapping of the non-moving objects with the help of sparse 3D map builder 2803 in map builder 2628.
At action 3740, the map server produces a 3D map from the sparse 3D mapping to with the help of 3D map builder 2804 in map builder 2628.
At action 3750, the map server provides the 3D map via a communication link to one or more additional autonomous units for guidance at a future time.
Other implementations of the method described in this section can include a non-transitory computer-readable storage medium storing instructions executable by a processor to perform any of the methods described above. Yet another implementation of the method described in this section can include a system including a memory and one or more processors operable to execute instructions, stored in the memory, to perform any of the methods described above.
The method described in this section and other sections of the technology disclosed can include one or more of the following features and/or features described in connection with additional methods disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this method can readily be combined with sets of base features identified as implementations such as system overview of quadocular-auxiliary sensors and the map server, 3D map construction, etc.
Process 3800 continues at action 3820 the map server generates a second 3D map using the set of keyrigs from the second autonomous unit with the help of sparse 3D map builder 2803, dense 3D map builder 2804 and semantic perspective view builder 2805 in map builder 2628.
At action 3830, the map server merges the second 3D map with the first 3D map with the help of the map merger 2806 in map builder 2628.
Other implementations of the method described in this section can include a non-transitory computer-readable storage medium storing instructions executable by a processor to perform any of the methods described above. Yet another implementation of the method described in this section can include a system including a memory and one or more processors operable to execute instructions, stored in the memory, to perform any of the methods described above.
Map Construction by Map Server
The method described in this section and other sections of the technology disclosed can include one or more of the following features and/or features described in connection with additional methods disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this method can readily be combined with sets of base features identified as implementations such as system overview of quadocular-auxiliary sensors and the map server, 3D map construction, etc.
Process 3900 continues at action 3920 where the autonomous unit classifies the objects in the 360-degrees images as moving objects and non-moving objects with the help of the semantic perspective view builder 2205 in map building component 1628.
At action 3930, the autonomous unit selects a subset of keyrig from the set of keyrigs and builds a sparse 3D mapping of the non-moving objects with the help of sparse 3D builder 2203 in map building component 1628.
At action 3940, the autonomous unit builds a 3D map of the non-moving objects from the sparse 3D mapping with the help of dense 3D map builder 2204 in map building component 1628.
At action 3950, the autonomous unit provides the 3D map via a communications link to the map server.
At action 3960, the map server distributes the 3D map to guide the autonomous unit at a future time and at least one other autonomous unit.
Other implementations of the method described in this section can include a non-transitory computer-readable storage medium storing instructions executable by a processor to perform any of the methods described above. Yet another implementation of the method described in this section can include a system including a memory and one or more processors operable to execute instructions, stored in the memory, to perform any of the methods described above.
The method described in this section and other sections of the technology disclosed can include one or more of the following features and/or features described in connection with additional methods disclosed. In the interest of conciseness, the combinations of features disclosed in this application are not individually enumerated and are not repeated with each base set of features. The reader will understand how features identified in this method can readily be combined with sets of base features identified as implementations such as system overview of quadocular-auxiliary sensors and the map server, 3D map construction, etc.
Process 4000 continues at action 4020 where the autonomous unit generates a first 3D roadmap using the set of keyrigs.
At action 4030, the autonomous unit provides the first 3D map via a communications link to the map server with a second 3D map.
At action 4040, the map server merges the second 3D map with the first 3D map with the help of the map merger 2806 in map builder 2628.
Other implementations of the method described in this section can include a non-transitory computer-readable storage medium storing instructions executable by a processor to perform any of the methods described above. Yet another implementation of the method described in this section can include a system including a memory and one or more processors operable to execute instructions, stored in the memory, to perform any of the methods described above.
Some particular implementations and features are described in the following discussion.
In one implementation, described is a methodology for building 3D maps of roads using information sourced by one or more moving autonomous units. The method includes: receiving a set of keyrigs from a first autonomous unit, each keyrig comprising visual information of the surrounding scenery and a position of the first autonomous unit where the visual information was captured, the position of the autonomous unit generated using combinations of global positioning system, inertial measurement unit, and visual information of the surrounding scenery by the first autonomous unit during travel from a starting point to an end point; classifying at least one of the one or more objects from the visual information of the surrounding scenery from the set of keyrigs into a set of moving objects and a set of non-moving objects; selecting a subset of keyrig from the set of keyrigs and determining a sparse 3D mapping of object feature points taken from visual information of the surrounding scenery from the subset of keyrigs for the set of non-moving objects; building a 3D map of object feature points from the sparse 3D mapping of object feature points using visual information of the surrounding scenery from the set of keyrigs for the set of non-moving objects, wherein the 3D map is accurate within centimeters; and providing the 3D map via a communications link to one or more additional autonomous units to guide the one or more additional autonomous units at a future time.
In some implementations, the method is implemented in a system which includes one or more autonomous units, including a first autonomous unit and a second autonomous unit, each autonomous unit including at least a quad camera visual sensor and at least one selected from a GPS and an IMU; and a map server to build 3D maps of roads using information sourced by the one or more autonomous units. In some implementations, the method is implemented in a system which includes a non-transitory computer-readable medium to store instructions.
In some implementations, the method further requires the starting point of the travel of the first autonomous unit is the end point of the travel of the first autonomous unit. In some implementations, the method further includes at least one 3D map with an accuracy within 10 centimeters. In some implementations, the method further includes at least one 3D map with an accuracy within 5 centimeters. In some implementations, accuracy is determined by a difference between a location of an object depiction on the 3D map and an actual location of the object in space.
In some implementations, the method further includes: receiving from a second autonomous unit a set of keyrigs while hosting a first 3D map, each keyrig is a set of quad images with a pose generated using combinations of global positioning system, inertial measurement unit, and visual information of a scene by the second autonomous unit; generating a second 3D map using the set of keyrigs from the second autonomous unit; and merging the second map 3D with the first map. In some implementations, the first 3D map and the second 3D map both cover common location and further including updating the first 3D map using the second 3D map.
In some implementations, the autonomous unit generates several hundred thousand images during one hour of operation by an autonomous unit. In some implementations, the autonomous unit contemporaneously tracks its position against 3D maps. In some implementations, the method further includes storing a time of day or a weather condition with the 3D map. In some implementations, the method further includes identifying an object feature point to be included in the sparse 3D mapping of object feature points: extracting a first set 2D features of the object from a first 360-degrees image in a keyrig selected from a subset of keyrigs; extracting a second set of 2D features of the object from a second 360-degrees image in the selected keyrig; receiving a position of the autonomous unit when the 360-degrees images were captured including longitude and latitude as input; triangulating the first set of 2D features from the first 360-degrees image and the second set of 2D features from the 360-degrees second image to derive location for feature points of the object relative to the position of the autonomous unit; and generating for at least one feature point of the object, a global position, including longitude, latitude, and height and adding the global position and feature descriptors of the object to the sparse mapping of object feature points. In some implementations, the method further includes identifying an object feature point to be included in the dense 3D mapping of object feature points: extracting a first set 2D features of the object from a first 360-degrees image in a keyrig selected from a set of keyrigs provided by an autonomous unit; extracting a second set of 2D features of the object from a second 360-degrees image in the selected keyrig; receiving a position of the autonomous unit when the 360-degrees images were captured including longitude and latitude as input; triangulating the first set of 2D features from the first 360-degrees image and the second set of 2D features from the 360-degrees second image to derive location for feature points of the object relative to the position of the autonomous unit; and generating for at least one feature point of the object, a global position, including longitude, latitude, and height and adding the global position and feature descriptors of the object to the dense mapping of object feature points.
Also disclosed is an implementation that includes a method for automatically building 3D maps of roads by an autonomous unit that includes at least a quad camera visual sensor and at least one selected from a GPS and an IMU. The method including: in the first autonomous unit, capturing a set of keyrigs, each keyrig comprising visual information of the surrounding scenery and a position of the first autonomous unit where the visual information was captured, the position of the autonomous unit generated using combinations of global positioning system, inertial measurement unit, and visual information of the surrounding scenery by the first autonomous unit during travel from a starting point to an end point; classifying at least one of the one or more objects from the visual information of the surrounding scenery from the set of keyrigs into a set of moving objects and a set of non-moving objects; selecting a subset of keyrig from the set of keyrigs and determining a sparse 3D mapping of object feature points taken from visual information of the surrounding scenery from the subset of keyrigs for the set of non-moving objects; building a 3D map of object feature points from the sparse 3D mapping of object feature points using visual information of the surrounding scenery from the set of keyrigs for the set of non-moving objects, wherein the 3D map is accurate within centimeters; and providing the 3D map via a communications link to a map server that stores and distributes 3D maps to guide the autonomous unit at a future time and at least one other autonomous unit.
The terms and expressions employed herein are used as terms and expressions of description and not of limitation, and there is no intention, in the use of such terms and expressions, of excluding any equivalents of the features shown and described or portions thereof. In addition, having described certain implementations of the technology disclosed, it will be apparent to those of ordinary skill in the art that other implementations incorporating the concepts disclosed herein can be used without departing from the spirit and scope of the technology disclosed. Accordingly, the described implementations are to be considered in all respects as only illustrative and not restrictive.
This application is a continuation of U.S. application Ser. No. 16/550,143, filed Aug. 23, 2019, entitled “VISUAL-INERTIAL POSITIONAL AWARENESS FOR AUTONOMOUS AND NON-AUTONOMOUS TRACKING”, (TRIF 1011-2), which is a continuation of U.S. application Ser. No. 15/942,348, filed Mar. 30, 2018, entitled “VISUAL-INERTIAL POSITIONAL AWARENESS FOR AUTONOMOUS AND NON-AUTONOMOUS TRACKING”, (TRIF 1011-1), which is a continuation-in-part of application of U.S. application Ser. No. 15/250,393, filed Aug. 29, 2016, entitled “VISUAL-INERTIAL POSITIONAL AWARENESS FOR AUTONOMOUS AND NON-AUTONOMOUS TRACKING” (PERC 1001-1). The non-provisional applications are hereby incorporated by reference for all purposes.
Number | Date | Country | |
---|---|---|---|
Parent | 16550143 | Aug 2019 | US |
Child | 17008299 | US | |
Parent | 15942348 | Mar 2018 | US |
Child | 16550143 | US |
Number | Date | Country | |
---|---|---|---|
Parent | 15250393 | Aug 2016 | US |
Child | 15942348 | US |