Patent Application
20220391796
The present disclosure relates generally to a warehouse or distribution environment, and more specifically to improving the efficiency of warehouse management by identifying and documenting areas of greatest risk.
In a distribution system, order fulfillment is a key process in managing the supply chain. It includes generating, filling, delivering and servicing customer orders. A typical order fulfillment process includes various sub-processes such as receiving order, picking an order, packing an order, and shipping the order. Receiving refers to the acceptance and storage of incoming inventory at a fulfillment center. When the fulfillment center receives the inventory, the items may be stored in dedicated warehouse locations, such as pallets. A pallet is a portable, rigid platform that is flat and can carry the load. In the picking sub-process, the picking team receives a packing slip with the items, quantities, and storage locations at the facility to collect the ordered products from their respective pallets.
Also, two features influence the operational efficiency of a warehouse or distribution centre. These aspects relate to the dynamic nature of the warehouse environment, and the performance of human operators during a pallet handling/order-picking process. In view of the above, there is a need for addressing the problem of order fulfillment efficiency in a warehouse distribution system, and enabling better operational management by redesigning package handling routes, and optimisation of package handling procedures during order fulfilment.
In an aspect of the present disclosure, there is provided a system for identifying and managing areas of risk in a warehouse environment. The system may include one or more video sensors configured to capture one or more video streams thereof, to generate one or more monitored zones, and one or more uncovered zones in the warehouse environment, based on the Field of View of the one or more video sensors. The system may further include a central processing unit communicatively coupled to the one or more video sensors. The central processing unit includes a raw risk information collection unit configured to store information captured by the one or more video sensors, and a processing and aggregating unit configured to process and aggregate the one or more video streams to produce risk identification information associated with an Operator Route traversed by a warehouse operator while performing a warehouse operation, wherein the risk identification information includes at least one risk zone, and corresponding risk type, and risk level, wherein a risk zone is an area in the warehouse environment that corresponds to one or more risk instances. The system may further include a risk map generation unit configured to generate a Warehouse Risk Map based on the risk identification information, wherein the Warehouse Risk Map is generated by superimposing an identified risk zone on a warehouse map. The system may further include a risk map updating unit for updating the Warehouse Risk Map in real-time when at least one of the risk type, risk level, and risk zone changes for at least one risk instance recorded on the Warehouse Risk Map.
In another aspect of the present disclosure, there is provided a method for identifying and managing areas of risk in a warehouse environment. The method includes capturing one or more video streams thereof, to generate one or more monitored zones, and one or more uncovered zones in the warehouse environment, based on the Field of View of the one or more video sensors. The method may further include storing information captured by the one or more video sensors. The method may further include processing and aggregating the one or more video streams to produce risk identification information associated with an Operator Route traversed by a warehouse operator while performing a warehouse operation, wherein the risk identification information includes at least one risk zone, and corresponding risk type, and risk level, wherein a risk zone is an area in the warehouse environment that corresponds to one or more risk instances. The method may further include generating a Warehouse Risk Map based on the risk identification information, wherein the Warehouse Risk Map is generated by superimposing an identified risk zone on a warehouse map. The method may further include updating the Warehouse Risk Map in real-time when at least one of the risk type, risk level, and risk zone changes for at least one risk instance recorded on the Warehouse Risk Map.
In yet another aspect of the present disclosure, there is provided a computer programmable product for identifying and managing areas of risk in a warehouse environment, the computer programmable product comprising a set of instructions. The set of instructions when executed by a processor causes the processor to capture one or more video streams thereof, to generate one or more monitored zones, and one or more uncovered zones in the warehouse environment, based on the Field of View of the one or more video sensors, store information captured by the one or more video sensors, process and aggregate the one or more video streams to produce risk identification information associated with an Operator Route traversed by a warehouse operator while performing a warehouse operation, wherein the risk identification information includes at least one risk zone, and corresponding risk type, and risk level, wherein a risk zone is an area in the warehouse environment that corresponds to one or more risk instances, generate a Warehouse Risk Map based on the risk identification information, wherein the Warehouse Risk Map is generated by superimposing an identified risk zone on a warehouse map, and update the Warehouse Risk Map in real-time when at least one of the risk type, risk level, and risk zone changes for at least one risk instance recorded on the Warehouse Risk Map.
Various embodiments of the present disclosure perform analysis of known and observed potentially changing environmental and human risk factors to generate and update a spatially defined risk map in a warehouse environment. By relating risk factor information to spatial information, the present disclosure allows causative correlations to be drawn between observed performance variables and specific locations within the warehouse environment or areas proximal thereto. The risk map may be used to detect and identify current and future potential performance impacting problems that include, but are not limited to, rack areas of less accessibility for order pickers, for example, where items are stacked at the back of the rack space, or stacked too high in the rack space, spillage areas, poorly illuminated areas, areas where products of awkward size of shape are more likely to be stacked, or stacked badly, areas where order pickers are more likely to slow down, and areas of greater security risk. Also, the risk map is updated frequently and potentially in real-time to enable speedy adaptation to rapidly changing risk factors, to minimise the damaging effects of rapidly evolving scenarios. Thus, insights obtained from the risk map may be used to improve the warehouse environment design, to increase the operational efficiency and to implement automatic detectors that are able to trigger alarms when an incident happens.
It will be appreciated that features of the present disclosure are susceptible to being combined in various combinations without departing from the scope of the present disclosure as defined by the appended claims.
The summary above, as well as the following detailed description of illustrative embodiments, is better understood when read in conjunction with the appended drawings. For the purpose of illustrating the present disclosure, exemplary constructions of the disclosure are shown in the drawings. However, the present disclosure is not limited to specific methods and instrumentalities disclosed herein. Moreover, those in the art will understand that the drawings are not to scale. Wherever possible, like elements have been indicated by identical numbers.
In the accompanying drawings, an underlined number is employed to represent an item over which the underlined number is positioned or an item to which the underlined number is adjacent. A non-underlined number relates to an item identified by a line linking the non-underlined number to the item. When a number is non-underlined and accompanied by an associated arrow, the non-underlined number is used to identify a general item at which the arrow is pointing.
The following detailed description illustrates embodiments of the present disclosure and ways in which they can be implemented. Although the best mode of carrying out the present disclosure has been disclosed, those skilled in the art would recognize that other embodiments for carrying out or practicing the present disclosure are also possible.
The warehouse environment 100 includes first and second storage racks 102a and 102b, and a trolley 103 for transporting goods in the warehouse environment 100. Although, two storage racks are shown herein, it would be apparent to one of skill in the art, that the warehouse environment 100 may include more than two racks and trolley.
The warehouse environment 100 may further include first and second video sensors 104a and 104b fixedly mounted over the first and second racks 102a and 102b respectively. Example of the video sensors 104a and 104b includes, but is not limited to, video cameras. The first and second video sensors 104a and 104b has a Field of View 106 that corresponds to a spatial volume in which the presence of objects may be detected in the absence of obstructions that would otherwise conceal the object. In the context of the present disclosure, the Field of View 106 also covers an Operator Route, where the Operator Route is defined as the path traversed by a warehouse operator during a task period, and the task period is defined as the time period extending from the moment the operator receives a task list from the supervisor until she/he has finished all the tasks on the task list. It should be noted that a task on the task list may include multiple operations such as a handling, order-filling, pallet-loading/unloading, and rack-filling.
The operational efficiency of the warehouse environment 100 is dependent on the dynamic nature of the warehouse environment 100, and the performance of human operators during a pallet handling/order-picking process. A variety of factors influence the pallet handing/order-picking process. These factors are hereinafter referred to as risks.
The incidence of specific types of risks may be monitored in different locations of the warehouse environment 100, according to parameters such as the time/date of the risk incidents or the identity of the operator or the forklift truck etc. The video sensors 104a and 104b may provide more detailed information regarding an operator or the type of handled packages involved in a given risk incident. This may assist warehouse managers in detecting and identifying patterns in risk incidents, for example a warehouse operator A may be more likely to spill items from a pallet close to the first rack 102a, thereby enabling the warehouse managers to undertake appropriate remedial action. The remedial actions may include, but not limited to, improving the lighting close to a rack where lot of risk incidents occur, increasing the spacing between racks or between racks and walls, providing additional training to particular warehouse operators about lifting or stacking items into racks or onto pallets, changing policy regarding the stacking of heavy or large items on different (higher/lower) rack spaces etc.
The individual risks may be expressed as risk instances. A risk instance comprises the following attributes: the classification of the risk, the one or more zones in the warehouse environment 100 where the relevant risk could happen (thereby enabling localization of the risk instance), and the risk level (the probability of the risk occurring in the or each relevant zone). For brevity, the one or more zones in the warehouse environment 100 where a risk could happen may be referred to henceforth as risk zones.
The system 200 is connected to the first and second video sensors 104a and 104b through a wired or wireless communication network (not shown) to process video streams recorded by the video sensors 104a and 104b.
The system 200 includes a central processing unit (CPU) 201, an operation panel 203, and a memory 205. The CPU 201 is a processor, computer, microcontroller, or other circuitry that controls the operations of various components such as the operation panel 203, and the memory 205. The CPU 201 may execute software, firmware, and/or other instructions, for example, that are stored on a volatile or non-volatile memory, such as the memory 205, or otherwise provided to the CPU 201. The CPU 201 may be connected to the operation panel 203, and the memory 205, through wired or wireless connections, such as one or more system buses, cables, or other interfaces. In an embodiment of the present disclosure, the CPU 201 may include a custom Graphic processing unit (GPU) server software to provide real-time object detection and prediction, for all cameras on a local network.
The operation panel 203 may be a user interface and may take the form of a physical keypad or touchscreen. The operation panel 203 may receive inputs from one or more users relating to selected functions, preferences, and/or authentication, and may provide and/or receive inputs visually and/or audibly.
The memory 205, in addition to storing instructions and/or data for use by the CPU 201, may also include user information associated with one or more users. For example, the user information may include authentication information (e.g. username/pas sword pairs), user preferences, and other user-specific information. The CPU 201 may access this data to assist in providing control functions (e.g. transmitting and/or receiving one or more control signals) related to operation of the operation panel 203, and the memory 205.
In an embodiment of the present disclosure, the CPU 201 includes a raw risk information collection unit 202 for receiving information captured by the video sensors 104a and 104b and storing the information in the storage unit 210, and a processing and aggregating unit 204 configured to process and aggregate video streams to detect the activation by a warehouse operator of one or more trigger conditions associated with one or more risk instances. On detection of the activation of the trigger condition, the processing and aggregating unit 204 is configured to identify and document the attributes of each risk instance.
In the context of the present disclosure, risks may be broadly grouped into two classes, namely predefined risks and heuristic risks. The predefined risks are well-known risks, that may be pre-defined by a management team of the warehouse environment. By contrast, heuristic risks are to be discovered and learned by observation of the warehouse environment. Predefined risks may include risks arising from heavy packages, as heavy packages may cause injuries when they are manipulated by operators. Another example of a predefined risk includes risks arising from fragile packages, as incorrect handling of fragile packages may cause stock and financial loss. While a predefined risk may be established by the management team, the location of occurrences of the said predefined risk may vary with time owing to the dynamic nature of the warehouse environment. For example, the location of heavy and awkwardly-shaped packages on storage racks may change over time.
Localization of a given risk instance may be expressed with different granularities. In particular, whereas a coarse risk localization may rely on identifiers of the racks in the warehouse environment, a fine-grained risk localization may provide more precise location information.
In an embodiment of the present disclosure, the risk level includes two components, namely, recent risk level Precent and global risk level Pglobal. Precent expresses the number of risk incidents that recently occurred in a risk zone as a fraction of the total number of operations undertaken in the risk zone. Pglobal expresses the total number of occurrences of risk incidents in the risk zone since the establishment of the warehouse, as a fraction of the total number of operations undertaken during that time period in that risk zone. Precent and Pglobal respectively contribute 75% and 25% to the overall risk level computation.
More specifically,
L
magnitude,magnitudeϵ[1,10],
magnitude=round(10·(3·Precent+Pglobal)/4),
P
recent=count(incidentst,tϵΔT)/count(operationst),tϵΔT,
P
global=count(incidentst)/count(operationst),tϵ[−∞ . . . now],
where:
The central processing unit 201 further includes a risk map generation unit 206 for generating a Warehouse Risk Map 210 (as shown in
The Warehouse Risk Map 210 is used to optimize the spatial deployment of video cameras in the warehouse environment so that their collective Field of View cover all the locations associated with each risk instance.
The central processing unit 201 includes a risk map updating unit 208 for updating the Warehouse Risk Map 210 according to a set of one or more of a set of pre-defined triggers (i.e. when there is a change in at least one of the risk types, risk levels, or risk zones) each of which is stored in the storage unit 210 and specifically linked with a given risk type. For example, when heavy packages are moved to another rack, the location of the risk associated with each heavy package changes to the new rack. Similarly, if the heavy packages are replaced with fragile ones, the type of risk changes for that risk instance. This allows fine customization of the moment when an update is necessary for the Warehouse Risk Map 210. For efficiency, not every risk incident occurrence causes an update to the Warehouse Risk Map 210. Additionally, the system settings for risk types and corresponding triggers may be periodically re-configured by the warehouse managers.
In an embodiment of the present disclosure, the risk map updating unit 208 is configured to automatically detect the occurrence of one or more risk incidents, and mark their location on the Warehouse Risk Map 210 to thereby illustrate the risk instances. However, since the location associated with a risk instance may vary with time, the Warehouse Risk Map 210 may be dynamically updated based on a risk-specific trigger to reflect these variations.
In the example of risk incidents arising from heavy packages, the location of such risk incidents may be ascertained from an inventory list of the warehouse environment 100. Thus, a rule for updating the trigger for the corresponding risk instances could be “Update the Warehouse Risk Map 210 every time the inventory list changes”. Similarly, for risk incidents arising from fragile packages, the location of such risk incidents may be ascertained through the detection of damaged packages during order-picking. For example, the occurrence of such risk incidents may be detected by a Package Integrity Check AI (PICAI) component (not shown) of the processing and aggregating unit 204. Thus, a rule for updating the trigger for this risk instance could be “Update the Warehouse Risk Map 210 every time the PICAI detects a damaged package”.
The PICAI determines package integrity status by processing video data captured by the video sensors 104a and 104b. More specifically, the PICAI comprises a trained deep neural network classifier (not shown) adapted to process a video stream from a video camera positioned to monitor the warehouse environment where packages are manipulated. The PICAI classifier may implement an architecture such as a visual geometry group (VGG) or a residual neural network (Resnet), and may be trained with a set of images labelled into two classes, namely damaged and non-damaged packages.
The second warehouse environment 400 includes first through sixth Monitored Zones (MZi) 402a till 402f (hereinafter collectively referred to as Monitored Zones 402) monitored by corresponding video sensors 404a till 404f with respective Fields of View. A Monitored Zone is substantially rectangular in shape, and its area is limited by the Field of View of the corresponding monitoring video sensor (i.e. video camera).
The second warehouse environment 400 includes first through seventh Uncovered Zones (UZj) 406a till 406g (hereinafter collectively referred to as Uncovered Zones 406) which the video sensors 404a till 404f are unable to monitor. An Uncovered Zone (UZj) jϵ[1 . . . M], where M is equal to the total number of such Uncovered Zones, may be an aperture (if any) between two consecutive Monitored Zones, or an aperture between a Monitored Zone and a proximal wall of the warehouse. Each successive Uncovered Zone is conferred with a unique identifier, for example, an index j incrementing from 1 according to the requirements of the warehouse management.
The NERD component 408 is communicatively coupled to the set of video sensors (404a till 404f in
In an embodiment of the present disclosure, the NERD component 408 includes a stream buffer 410 for receiving and buffering video streams from the video sensors (404a till 404f in
In an embodiment of the present disclosure, the first through kth detectors 412a till 412k are configured to process the video streams from video sensors (404a till 404f in
The inference unit 414 is configured to learn “normal” operational parameters expressed as time spent by an operator in a given zone of the warehouse, and to identify abnormalities suggestive of the occurrence of a new risk type, for example, excessive time spent by an operator in the said zone.
Referring to
In an embodiment, the NERD component 408 comprises a Trajectory Computation Unit 420, a Human pose estimation unit 422, a Merchandise Classification Unit 424, and a Pattern analysis unit 426. The trajectory computation unit 420 is communicably coupled with the Human Pose Estimation Unit 422 and the Merchandise Classification Unit 424, and each of which is communicably coupled in turn with a Pattern Analysis Unit 426. Each of these will be described in more detail below.
In an embodiment, the Trajectory Computation Unit 420 comprises a Detector Unit 430, a Cropper Unit 432, a video Knowledge Distillation Unit 434 (hereinafter referred to as VKD unit 434), a QR Detector 437, a Kalman Filter Unit 438, a Matcher Unit 440, a Previous State Database 442 and a Tracking Database 444.
In an embodiment, the Detector Unit 430 is communicatively coupled with one or more video sensors (for example, the first and second video sensors 104a and 104b) (not shown) installed at one or more locations in the warehouse. The video footage obtained from the video sensor (not shown) comprises a plurality of successively captured video frames, wherein p is the number of video frames in the captured video footage. A given video frame Fr(τ+iΔt)ϵn×m is captured by a video camera at time instant (also known as sampling time) τ+iΔt wherein r is the time at which capture of the video footage starts and Δt is the time interval (also known as the sampling interval) between the capture of a first video frame and the capture of the next video frame. Using this notation, the video footage captured by a video camera can be described as VIDϵ
n×(p×m)=[Fr(τ), Fr(τ+Δt), Fr(τ+2Δt) . . . Fr(τ+pΔt)].
Similarly, in the event video footage is captured from a plurality of video sensors, individual video frames captured by q>1 video sensors at a given sampling time (τ+iΔt) can be concatenated, so that the video footage captured by the collective body of video sensors can be described as:
VIDϵ
(p×m)×(n×q)=[[Fr0(τ),Fr1(τ) . . . Frq(τ)]T,[Fr0(τ+Δt),Fr1(τ+Δt) . . . Frq(τ+Δt)]T, . . . ,[Fr0(τ+pΔt),Fr1(τ+pΔt) . . . Frq(τ+pΔt)]T] (1)
For brevity, a video frame formed by concatenating a plurality of video frames each of which was captured at the same sampling time (for example, [Fr0(τ), Fr1(τ) . . . Frg(τ)]T) will be referred to henceforth as a “Concatenated Video Frame”. Similarly, individual video frames concatenated within a Concatenated Video Frame will be referred to henceforth as “Concatenate Members”. For clarity, in the following discussions, a current sampling time tk is given by tk=τ+NΔt, where N<n. A previous sampling time tp is a sampling time that precedes the current sampling time tk and is given by tp=τ+DΔt where 0<D<N. A current video frame Fr(tk) is a video frame captured at a current sampling time tk. A previous video frame Fr(tp) is a video frame captured at a previous sampling time tp.
In an embodiment, the Detector Unit 430 comprises an object detector algorithm adapted to receive a video frame or a Concatenated Video Frame and to detect therein the presence of a person, pallet and package. For brevity, persons, pallets and packages visible in a video frame will be referred to henceforth as “Risk Subjects”.
For further clarity, a currently detected Risk Subject is a Risk Subject that has been detected by the Detector Unit 430 in a current video frame Fr(tk). A previously detected Risk Subject is a Risk Subject that has been detected in a previous video frame Fr(tp). A previous detection of a Risk Subject is the detection of the Risk Subject by the Detector Unit 430 in a previous video frame Fr(tp). A current detection of a Risk Subject is the detection of the Risk Subject by the Detector Unit 430 in the current video frame Fr(tk). Further, a most recent previous detection of a Risk Subject is a one of a one or more previous detections of a given Risk Subject by the Detector Unit 430 at a previous sampling time that is closest to the current sampling time, or in other words, at a given current time tk, a most recent previous detection of a Risk Subject is the last previous detection of the Risk Subject in the previous video frames.
In an embodiment, the object detector algorithm is further configured to classify a detected Risk Subject as either a person, pallet or package, and to determine the location(s) of the Risk Subject(s); and to determine the location of the detected Risk Subject in the video frame or Concatenated Video Frame. The location of a detected Risk Subject is represented by the co-ordinates of a bounding box which is configured to enclose the Risk Subject. The co-ordinates of a bounding box are established with respect to the co-ordinate system of the video frame or Concatenated Video Frame. In particular, the object detector algorithm is adapted to receive individual successively captured video frames Fr(i+iΔt) from the video footage VID; and to process each video frame Fr(i) to produce one or more details of a set of bounding boxes B(τ)=[b1(τ), b2(τ) . . . bi(τ))]T i≤NRS(τ), where NRS(τ) is the number of Risk Subjects detected and identified in the video frame Fr(τ) and bi(τ) is the bounding box encompassing an ith Risk Subject. The details of each bounding box bi(τ) comprise four variables, namely [x,y], h and w, where [x,y] is the co-ordinates of the upper left corner of the bounding box relative to the upper left corner of the video frame (whose coordinates are [0,0]); and h,w are the height and width of the bounding box respectively.
Thus, the output from the Detector Unit 430 is one or more Detection Measurement vectors each of which comprises the co-ordinates of a bounding box enclosing a Risk Subject detected in the received video frame, together with a classification label for each such detected Risk Subject. For brevity, the co-ordinates of a bounding box enclosing a Risk Subject detected in the received video frame will be referred to henceforth as a Detection Measurement vector. The Detector Unit 430 is communicatively coupled with the Kalman Filter Unit 438 and the Cropper Unit 432 to transmit thereto the Detection Measurement vector(s) calculated from the received video frame (Fr(τ)).
To this end, the object detector algorithm comprises a deep neural network whose architecture is substantially based on the EfficientDet (as described in M. Tan, R. Pang and Q. V. Le, EfficientDet: Scalable and Efficient Object Detection, 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, Wash., USA, 2020, pp. 10778-10787). The architecture of the deep neural network of the object detector algorithm may also be based on YOLOv4 (as described in A Bochkovskiy, C-Y Wang and H-Y M Liao, 2020 arXiv: 2004.10934). However, the person skilled in the art will acknowledge that the above-mentioned neural network and architecture are provided for illustrative purposes only. In particular, the skilled person will understand that the preferred embodiment is not limited to these deep neural network architectures. On the contrary, the preferred embodiment is operable with any object detector architecture and/or training algorithm which is suitable for the detection and localization of people and objects such as pallets and packages in video footage. For example, the object detector algorithm could comprise a YOLOv5 architecture with an S or L architecture, or any other single-shot detector such as SSD, RetinaNet etc.
The goal of training the object detector is to cause it to establish an internal representation of a Risk Subject, wherein the internal representation allows the object detector to recognize a Risk Subject in subsequently received video footage. To meet this aim, the dataset used to train the object detector consists of video footage of a variety of scenarios recorded in a variety of different warehouses. The video footage, which will henceforth be referred to as the Training Dataset is assembled with the aim of providing robust, class-balanced information about Risk Subjects derived from different views of a Risk Subject obtained from different viewing angles. The members of the Training Dataset are selected to create sufficient diversity to overcome the challenges to subsequent Risk Subjects recognition posed by variations in illumination conditions, perspective changes, a cluttered background and most importantly intra-class variation. In most instances, images of a given scenario are acquired from multiple cameras, thereby providing multiple viewpoints of the scenario.
Prior to their use in the Training Dataset, the video footage is processed to remove video frames/images that are very similar. Similarly, some members of the Training Dataset are also used to train the VKD Unit 434. The members of the Training Dataset may also be subjected to further data augmentation techniques to increase the diversity thereof and thereby increase the robustness of the eventual trained object detector model. Specifically, the images/video frames are resized to a standard size wherein the size is selected to balance the advantages of more precise details in the video frame/image against the cost of more computationally expensive network architectures required to process the video frame/image. Similarly, all of the images/video frames are re-scaled to a value in the interval [−1, 1], so that no features of an image/video frame have significantly larger values than the other features.
In a further pre-processing step, individual images/video frames in the video footage of the Training Dataset are provided with one more bounding boxes, wherein each such bounding box is arranged to enclose a Risk Subjects visible in the image/video frame. The extent of occlusion of the view of a Risk Subjects in an image/video frame is assessed. Those Risk Subjects whose view in an image/video frame is more than 70% un-occluded are labelled with the class of the Risk Subjects (wherein the class label is selected from the set comprising person, pallet and package). Individual images/video frames are further provided with a unique identifier which is used, as will be described later, for the training of the VKD Unit 434.
In an embodiment, the Kalman Filter Unit 438 is also communicatively coupled with a Previous State Database 442. The Previous State Database 442 comprises a plurality of Previous State vectors psj, j≤NPV, each of which is derived from the most recent observation of a Risk Subject detected in a previously captured video frame. Specifically, a Previous State vector psj of a jth Risk Subject is denoted by psj=[ϕ; u, v, s, r, u′, v′, s′, r′]T where:
In an embodiment, the Previous State Database 442 is initially populated with Previous State vectors derived by the Detector Unit 430 from the first video frame Fr(τ0) received (at sampling time τ0), wherein NRS(τ0) is the total number of Risk Subjects observed in the first video frame and the first derivative terms (u′, v′, s′ and r′) of each of the initialised Previous State vectors are set to a value of zero.
In an embodiment, the Kalman Filter Unit 438 is adapted to receive a Detection Measurement vector from the Detector Unit 430, and configured to retrieve the Previous State vectors from the Previous State Database 442. The Kalman Filter Unit 438 is further adapted to estimate candidate dynamics of the Risk Subject enclosed by the bounding box whose details are contained in the Detection Measurement vector based on the estimated dynamics of previously detected Risk Subjects (represented by the Previous State vectors retrieved from the Previous State Database 442). For brevity, the estimated dynamics of a currently detected Risk Subject based on the Previous State vector (of a previously detected Risk Subject), will be referred to henceforth as the Predicted State vector of the currently detected Risk Subject. Thus, using this nomenclature, for a given detected Risk Subject in a current video frame, the Kalman Filter Unit 438 is adapted to calculate one or more candidate Predicted State vectors. The Kalman Filter Unit 438 is communicatively coupled with the Matcher Unit 440 to transmit thereto the candidate Predicted State vector(s) and the Actual Measurement vector of the currently detected Risk Subject.
In an embodiment, the Cropper Unit 432 is adapted to receive a video frame (Fr(τ)) and receive one or more Detection Measurement vectors from the Detector Unit 430. The Cropper Unit 432 is further adapted to crop the received video frame (Fr(τ)) to the region(s) enclosed by the bounding box(es) specified in the Detection Measurement vectors. For brevity, a cropped region that is enclosed by a bounding box, will be referred to henceforth as a Cropped Region. The Cropper Unit 432 is further adapted to transmit the Cropped Region(s) to the VKD Unit 434. While the Cropper Unit 432 is described herein as being a separate component to the Detector Unit 430, the skilled person will understand that the Cropper Unit 432 and the Detector Unit 430 could also be integral components.
In an embodiment, the VKD Unit 434 comprises a Views Knowledge Distillation (VKD) network (as described in Porrello A., Bergamini L. and Calderara S., Robust Re-identification by Multiple View Knowledge Distillation, Computer Vision, ECCV 2020, Springer International Publishing, European Conference on Computer Vision, Glasgow, August 2020). The VKD unit 434 further comprises a pre-trained Teacher Network 435 communicatively coupled with a Student Network 436. The Teacher Network 435 and the Student Network 436 have substantially matching architectures, namely a ResNet-101 convolutional neural network (as described in He K., Zhang X., Ren S. and Sun J. “Deep Residual Learning for Image Recognition”, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, N V, 2016, pp. 770-778) with a bottleneck attention module (as described in Park, J., Woo, S., Lee, J., Kweon, I. S.: “BAM: bottleneck attention module” in British Machine Vision Conference (BMVC) 2018). The skilled person will understand that the above network architectures are provided for example only. In particular, the skilled person will understand that the preferred embodiment is in no way limited to the above-mentioned network architectures. Instead, the preferred embodiment is operable with any network architecture capable of forming an internal representation of a Risk Subject based on one or more of its physical appearance attributes, for example, the network employed in the VKD Unit 434 could be a ResNet-34, ResNet-50, DenseNet-121 or a MobileNet.
In an embodiment, prior to operation of the Trajectory Computation Unit 420, the Teacher Network 435 is trained on a selected plurality of video frames; and the Student Network 436 is trained from the Teacher Network 435 in a self-distillation mode as described below. In this way, the Teacher Network 435 and the Student Network 436 are trained to establish an internal representation of the appearance of a Risk Subject sufficient to permit subsequent identification of the Risk Subject should it appear in further captured video frames.
The Teacher Network 435 and the Student Network 436 are respectively trained using a first subset and a second subset of a gallery comprising a plurality of Concatenated Video Frames. Thus, the gallery comprises a plurality of scenes viewed from different viewpoints by a plurality of video sensors. In at least some of the scenes, one or more classes of Risk Subject are visible. These scenes mirror those used to establish the Training Dataset for the object detector algorithm of Detector Unit 430. Hence at least some of the members of the Training Dataset may be used as members of the gallery.
The first subset (Tr_SS1) comprises a first number (X) of Concatenated Video Frames from the gallery, as shown below:
Tr_SS1ϵ(p×m)×(n×X)=[[Fr0(τ),Fr1(τ) . . . FrX(τ)]T,[Fr0(τ+Δt),Fr1(τ+Δt) . . . FrX(τ+Δt)]T, . . . ,[Fr0(τ+pΔt),Fr1(τ+pΔt) . . . FrX(τ+pΔt)]T]
The second subset (Tr_SS2) comprises a second number (Y) of Concatenated Video Frames from the gallery, wherein Y<X, as shown below.
Tr_SS2ϵ(p×m)×(n×Y)=[[Fr0(τ),Fr1(τ) . . . FrY(τ)]T,[Fr0(τ+Δt),Fr1(τ+Δt) . . . FrY(τ+Δt)]T, . . . ,[Fr0(τ+pΔt),Fr1(τ+pΔt) . . . FrY(τ+pΔt)]T]
Thus, the first and second subset comprise the same images from the gallery. However, the first subset and the second subset differ according to the number of Concatenate Members in their respective Concatenated Video Frames, wherein the first subset comprises Concatenated Video Frames with a larger number of Concatenate Members than the Concatenated Video Frames of the second subset.
The gallery further comprises:
Accordingly, the first subset is accompanied with the details of the bounding box(es) enclosing each Risk Subject detected in a video frame of the first subset and identifiers of the Risk Subjects. Similarly, the second subset is accompanied with the details of the bounding box(es) enclosing each Risk Subject detected in a video frame of the second subset and identifiers of the Risk Subjects.
The goal of training the Teacher Network 435 is to establish an internal representation which enables it to subsequently recognize a Risk Subject visible in a Concatenated Video Frame based on the Risk Subject's physical appearance attributes. The Teacher Network 435 expresses its establishment of an internal representation of a Risk Subject's appearance by returning a ranked list of identifiers for the Risk Subject, the said ranked list comprising identifiers selected by the Teacher Network 435 from the first subset. The performance of the training process is assessed by computing how many times the correct identifier for a Risk Subject visible in a Concatenated Video Frame is among the first pre-defined number of identifiers returned by the Teacher Network 435 in response to that Concatenated Video Frame.
The goal of training the Student Network 436 is to use the content of the second subset together with aspects of the internal representation formed by the Teacher Network 435, to enable the Student Network 436 to form its own internal representation of a Risk Subject's physical appearance attributes; thereby allowing the Student Network 436 to subsequently recognize a Risk Subject visible in a video frame based on the Risk Subject's physical appearance attributes. To this end, the training procedure for the Student Network 436 employs a cost function comprising a weighted sum of a triplet loss term and a classification loss term, a knowledge distillation loss and an L2 distance term.
The goal of training the Student Network 436 is to use the content of the second subset together with aspects of the internal representation formed by the Teacher Network 435, to enable the Student Network 436 to form its own internal representation of a Risk Subject's physical appearance attributes; thereby allowing the Student Network 436 to subsequently recognize a Risk Subject visible in a video frame based on the Risk Subject's physical appearance attributes. To this end, the training procedure for the Student Network 436 employs a cost function comprising a weighted sum of a triplet loss term and a classification loss term, a knowledge distillation loss and an L2 distance term.
In an embodiment, I knowledge distillation loss is a cross entropy loss term expressing the difference between the identifier returned by the Teacher Network 435 in response to a Concatenated Video Frame and the identifier returned by the Student Network 436 in response to a Concatenated Video Frame comprising a subset of video frames from the Concatenated Video Frame given as input to the Teacher Network 435. Thus, the second cost function is formulated to cause the Student Network 436 to output a Detected Appearance vector that closely approximate the appearance vector outputted by the Teacher Network 435. Since the Teacher Network 435 is trained on Concatenated Video Frame comprising a larger number of Concatenate Members, the Teacher Network 435 will establish appearance vectors containing more information. The second cost function causes the additional information to be distilled into the Detected Appearance vectors outputted by the Student Network 436, even though the Student Network 436 does not receive as rich an input as the Teacher Network 435. The second cost function further comprises an L2 distance term expressing the distance between the internal representation formed in the Teacher Network 435 and that formed in the Student Network 436.
Prior to their use in the gallery, images are processed to remove those of similar appearance. The resulting images are further pre-processed by resizing, padding, random cropping, random horizontal flipping and normalization. Similarly, the images are subjected to a random erasing operation in which some of the pixels in the image are automatically erased. This is useful for simulating occlusion, so that the Tracking System becomes more robust to occlusion.
Once suitably trained and cross-validated, the Student Network 436 is configured to process a Cropped Region received from the Cropper Unit 432 to produce a set of Detected Appearance vectors A(τ)=[α1(τ), α2(τ) . . . αj(τ))]T i≤NRS(i) relating to the Risk Subjects appearing in the Cropped Region. A Detected Appearance vector αi(τ), i≤NRS(τ) (wherein ∥αi(τ)∥=1) is formed from the activation states of the neurons in the Student Network 436. Thus, a Detected Appearance vector α1(τ) comprises the appearance attributes of a given Risk Subject as internally represented by the Student Network 436. The Student Network 436 is further adapted to transmit the set of Detected Appearance vectors A(τ) to the Matcher Unit 440.
The Matcher Unit 440 is communicatively coupled with the Kalman Filter Unit 438; and is configured to receive therefrom a set of candidate Predicted State vector(s) and the Actual Measurement vector of a currently detected Risk Subject. The Matcher Unit 440 is further communicatively coupled with the VKD Unit 434; and is configured to receive therefrom a set of Detected Appearance vectors A(τ)=[α1(τ), α2(τ) . . . αi(τ))]T i≤NRS(i) of each and every Risk Subject detected in a given video frame Fr(τ) respectively.
The Matcher Unit 440 is also communicatively coupled with the Tracking Database 444. The Tracking Database 444 comprises a Tracking matrix TRϵN
N
N
To address the complexity posed by the timing of individual Previous Appearance vectors in different Tracklet vectors Trj(τ), for simplicity a universal index k will be used henceforth to refer to individual Previous Appearance vectors Tαk in a given Tracklet vector, (wherein Trj(τ)={PαkϵN
In an embodiment, the Tracking Database 444 is initially populated with Detected Appearance vectors αj(1) j≤NRS(τ0) calculated by the Student Network 436 in response to the first video frame F(τ0) received (at sampling time τ0). Thus, the Tracking Database 444 is an appearance-based counterpart for the dynamics/state-based Previous State Database 442. Indeed, since the Tracking Database 444 and the Previous State Database 442 are both populated according to the order in which Risk Subjects are detected in a monitored area, the ordering of the Tracklet vectors Trj(ϕ), j≤NPV in the Tracking Database 444 matches that of the Previous State vectors psj(ϕ), j≤NPV in the Previous State Database 442.
In an embodiment, the Matcher Unit 440 comprises a Motion Cost Unit 448, an Appearance Cost Unit 450 and, an Intersection over Union (IoU) Unit 452, all of which are communicatively coupled with a Hungarian Algorithm Unit 454. The Hungarian Algorithm Unit 454 is further communicatively coupled with an Update Unit 456, wherein the Update Unit 456 is itself communicatively coupled with the Previous State Database 442 and the Tracking Database 444.
In an embodiment, the Motion Cost Unit 448 is adapted to calculate the squared Mahalanobis distance ΔM matrix representing the squared distance (δi,jM) between a given Actual Measurement vector zi(τ) and a Predicted Measurement vector ({circumflex over (m)}j(τ)) calculated from each Predicted State vector (i.e. {circumflex over (m)}j(τ)=Hτ{circumflex over (x)}j(τ)|τ) at a given sampling instant τ.
ΔM=Y(τ)TSMY(τ) (2)
where SM is the covariance matrix of Y(τ)
State estimation uncertainty is addressed by measuring how many standard deviations the Actual Measurement vector zi(τ) is from the Predicted Measurement vector ({circumflex over (m)}j(τ)). Thus, an unlikely association of a given Actual Measurement vector zi(τ) with a given Previous State vector psj(ϕ) can be excluded, by thresholding the Mahalanobis distance ΔM at a 95% confidence interval calculated from the χ2 distribution. Specifically, by implementing this thresholding function (Th(M)), a State Indicator matrix SIϵN
In an embodiment, the Appearance Cost Unit 450 is configured to retrieve from the Tracking Database 444 each of a plurality of Tracklet vectors Trj(τ)ϵN
δi,j,kA=min(1−αi(τ)TPAjk),k≤100 (3)
In an analogous manner to the Motion Cost Unit 448, the Appearance Cost Unit 450 employs a threshold operation on the minimum cosine distance (δi,j,kA) to exclude an unlikely association of the Detected Appearance vector (αj(τ)) of a given Risk Subject with a given Previous Appearance vector TAk in a given Tracklet vector Trj(τ) in the Tracking Database 444. Specifically, by implementing this thresholding function (Th(A)), an Appearance Indicator matrix AIϵN
In an embodiment, the IoU Unit 452 is adapted to receive the Predicted Measurement vectors ({circumflex over (m)}j(τ), j≤NPV) from the Kalman Filter Unit 438; and the bounding box vectors (bj(τ), j≤NRS(τ)) established by the Detector Unit 430 in response to a received video frame Fr(τ). The IoU Unit 452 is further adapted to calculate an intersection over union (IoU) measurement between each bounding box vector bj(τ) and each Predicted Measurement vector {circumflex over (m)}j(τ) and to employ a thresholding operation on the minimum IoU value, to exclude an unlikely association of a bounding box vector bj(τ) calculated from a received video frame Fr(τ) and a predicted bounding box based on predicted system dynamics.
In an embodiment, the Hungarian Algorithm Unit 454 is adapted to receive the minimum cosine distance (δi,j,kA) from the Appearance Cost Unit 450; and squared Mahalanobis distance (δi,jM) from the Motion Cost Unit 448. The Hungarian Algorithm Unit 454 is adapted to calculate a weighted sum of the minimum cosine distance (δi,j,kA) and the squared Mahalanobis distance (δi,jM), using a weighting variable A which is initially set to a pre-defined value and later tuned as appropriate for the relevant use case.
c
i,j=λδi,jM+(1−λ)δi,j,kA (4)
In an embodiment, the Hungarian Algorithm Unit 454 is further adapted to populate an Association matrix with values formed from the product of the corresponding binary variables of the State Indicator matrix SIϵN
N
In the event a recently detected ith Risk Subject cannot be matched to a jth Tracklet vector Trj(τ), any Tracklet vector Trj(τ) that has not been matched with a Risk Subject over the previous pre-defined number of sample instants are selected, to form a set of historically unmatched Tracklet vectors UTrj(τ). The Hungarian Algorithm Unit 454 is then adapted to implement a further iteration thereof to determine assignments of unmatched recently detected ith Risk Subjects to each of the historically unmatched Tracklet vectors UTrj(τ).
In this process, the Hungarian Algorithm Unit 454 is adapted to sort the unmatched Tracklet vectors UTrj(τ) in ascending order according to their age. Specifically, the unmatched Tracklet vectors UTrj(τ) are ordered according to the elapsed time (qΔt) between a current sampling instant (τ) and the sampling instant of the most recent Previous Appearance vector Pαk, k≤100 thereof. As will be recalled, the most recent Previous Appearance vector of a Tracklet vector represents the last (or most recent) time at which a Risk Subject corresponding with the Tracklet vector was observed. For brevity, the number of samples between the current sampling instant and the sampling instant of the most recent Previous Appearance vector Pαk of an unmatched Tracklet vector UTrj(τ) will be referred to henceforth as the age of the unmatched Tracklet vector UTrj(τ). In other words, an unmatched Tracklet vector UTrj(τ) whose most recent Previous Appearance vector Pαk is one sample prior to the current sampling instant will be referred to as an unmatched Tracklet vector UTrj(τ) of age one sample. Similarly, an unmatched Tracklet vector UTrj(τ) whose most recent Previous Appearance vector Pαk is two samples prior to the current sampling instant will be referred to as having an age of two samples, and so forth.
In an embodiment, the Hungarian Algorithm is implemented to determine assignments of a recently detected ith Risk Subject to each jth unmatched Tracklet vector UTrj(τ) in order of increasing age of the unmatched Tracklet vector UTrj(τ). In particular, the Hungarian Algorithm Unit 454 is adapted to select each of the unmatched Tracklet vectors UTrj(τ) of age one sample and attempt to find an assignment of the recently detected ith Risk Subject therewith.
In an embodiment, if in the event a match is not identified between the recently detected ith Risk Subject and the unmatched Tracklet vectors UTrj(τ) of age one sample, the Hungarian Algorithm Unit 454 is adapted to select each of the unmatched Tracklet vectors UTrj(τ) whose age is two samples and attempt to find an assignment of the recently detected ith Risk Subject therewith. In the event a match is not identified, the Hungarian Algorithm Unit 454 is adapted to select in turn each of the unmatched Tracklet vectors UTrj(τ) whose age is three samples and attempt to find an assignment of the recently detected ith Risk Subject therewith. This process is repeated for a pre-determined number (Amax) number of ages of the unmatched Tracklet vectors UTrj(τ). The resulting distance between an unmatched Tracklet vector UTrj(τ)) and the detected ith Risk Subject will be the smallest distance computed between the detected ith Risk Subject and each Previous Appearance vector in that Tracklet vector UTrj(τ)).
A given iteration of this process will not override an existing matching, as an unmatched Tracklet vector UTrj(τ) under consideration during the iteration will have a different age to the unmatched Tracklet vectors UTrj(τ) considered during a previous iteration. Furthermore, any recently detected ith Risk Subjects that have been matched during a given iteration will be excluded from consideration during subsequent iteration. This approach is premised on the belief that unmatched Tracklet vectors UTrj(τ) of least age are likely to be more similar to a given recently detected ith Risk Subject than older unmatched Tracklet vectors UTrj(τ).
In an embodiment, the Hungarian Algorithm Unit 454 is further adapted to receive the intersection over union measurements from the IoU Unit 452 and to use the intersection over union measurements to determine assignments of unmatched recently detected ith Risk Subjects to each of the unmatched Tracklet vectors UTrj(τ) of age 1 sample.
In an embodiment, the Hungarian Algorithm Unit 454 is adapted to transmit to the Update Unit 456 a set of first matching indices and second matching indices, being the indices i and j of the matching recently detected Risk Subjects and Tracklet vectors.
In an embodiment, the Update Unit 456 is adapted to transmit to the Previous State Database 442, Actual Measurement vectors zi(τ) together with different instructions depending on whether the index of a given Actual Measurement vector zi(τ) matches a first matching index. Specifically, if an index of a given Actual Measurement vector zi(τ) matches a first matching index, the instructions transmitted by the Update Unit 456 comprise an instruction to activate the Kalman Filter Unit 438 to compute a new Predicted State Vector {circumflex over (x)}(γ)|γ-1 using the matching Previous State Vector. The new Predicted State Vector is projected in the Actual Measurement space and the difference between the projected prediction and Actual Measurement is computed having the goal in updating the Previous State vector. The instructions further provide that the Previous State vector psj(ϕ) whose index matches the second matching index is to be updated with the given Actual Measurement vector zi(τ) (and the first derivative components (u′, v′, s′ and r′) of the Previous State vector psj(ϕ) be updated with those of the new Predicted State vector {circumflex over (x)}(γ)|γ-1). By contrast, in the event an index of a given Actual Measurement vector zi(τ) does not match a first matching index, the instructions transmitted by the Update Unit 456 comprise an instruction to use the Actual Measurement vector zi(τ) to add a new Previous State vector psj(ϕ) to the Previous State Database 442. Specifically, the new Previous State vector psj(ϕ) is denoted by psj(ϕ)=[zi(τ), u′, v′, s′, r′]T where the first derivative terms (u′, v's′ and r′) are initialised to a value of zero.
In an embodiment, the Update Unit 456 is adapted to transmit to the Tracking Database 444, each of a set of Detected Appearance vector A(τ)=[α1(τ), α2(τ) . . .
On receipt of the instructions, the Previous State Database 442 and the Tracking Database 444 are also adapted to review the age of its Previous State vectors psj(τ) and corresponding Tracklet vectors Trj(τ). The age of a Tracklet vector Trj(τ) is denoted as the elapsed time (qΔt) between a current sampling instant (τ) and the sampling instant of the first Tracking Appearance vector Tot′ of the Tracklet vector. In the event the age of a Tracklet vector Trj(τ) exceeds a pre-defined number of sampling intervals, the Previous State Database 442 and the Tracking Database 444 are adapted to delete the Tracklet vector Trj(τ) and corresponding previous State vectors psj(τ). In this way, the Previous State Database 442 and the Tracking Database 444 are cleansed of records of Risk Subjects that have left the warehouse, to prevent the accumulation of unnecessary records therein and thereby control the storage demands of the preferred embodiment.
In an embodiment, the QR Detector 437 implements a QR detection algorithm. The purpose of the QR Detector 437 is to permit the identification of warehouse staff based on the presence of the Quick Response (QR) code on a tag worn on the person's uniform.
In one embodiment the QR Detector 437 is implemented using a neural network based on the Yolo_v5 architecture. The skilled person will acknowledge that this network architecture is provided for illustration purposes only. In particular, the skilled person will acknowledge that the preferred embodiment is not limited to the use of this network architecture. On the contrary, the preferred embodiment is operable with any network architecture which enables the detection and recognition of a QR present in an image. For example, the preferred embodiment is operable with any other single-shot detector such as RetinaNet.
During training, a Reference Frame is created, wherein the Reference Frame is video frame from video footage captured of the warehouse in the absence of a QR code. In a next step, a training dataset is created by cutting short videos from the raw video footage captured by the video sensors in the warehouse. The short videos contain sequences where a QR code is shown to a video camera. To ensure diversity of feature distribution, video frames are extracted from the short videos using an average hashing algorithm.
Once trained, the Yolo_v5 network is configured to receive a video frame Fr(τ); and in response thereto, to output three vectors, as follows:
On detecting a QR code in a video frame Fr(τ), the QR Detector 437 is configured to crop a corresponding region from the video frame Fr(τ). The cropped region corresponds with that occupied by the bounding box surrounding the QR code with 20 pixels added on each side of the bounding box, to ensure the whole QR code is contained in the cropped region. The QR code in the cropped region is then decoded using a barcode reading tool which may comprise any of Pyzbar, PyQRCode, qrcode and qrtools etc. The output from the barcode reading tool comprises a string of characters decoded from the QR code. The QR Detector 437 is configured to associate the string with the person detected, by the Detector Unit 430, at a position closest to the QR code in the video frame Fr(τ). Thus, the capability of the Trajectory Computation Unit 420 to re-identify a person from one video frame to another is enhanced through its coupling with an identity assigned to the person based on the QR code they present to the video sensors of the warehouse.
The Merchandise Classification Unit 424 is configured to analyse the content of a pallet or a package detected by the Detector Unit 430 in a received video frame Fr(τ). The Merchandise Classification Unit 424 comprises two communicatively coupled modules, namely an instance segmentation Unit and an image retrieval Unit. The instance segmentation Unit performs instance segmentation and the image retrieval Unit uses an image retrieval algorithm to classify cropped bounding boxes of products, that were detected by the instance segmentation Unit.
To address periodic changes in the appearance of products, the image retrieval Unit comprises a first model to detect the presence of a product and a second model to recognize the product using prior knowledge in the form of a product database that can be easily updated. The second model is trained with the classes “pack”, “box” and “vegetables”. However, the classes are extendable to include “small pack”, “medium pack” and “large pack”.
To this end, the Instance segmentation is employed because products may be stacked erratically on a pallet; and a pixel level mask will increase the accuracy of detecting the products. In a preferred embodiment, the instance segmentation Unit employs a transformer-based model (as described by Z Liu, Y, Lin, Y. Cao, H. Yu, Y. Wei, Z. Zhang, S. Lin, and B. Gao, Swin Transformer: Hierarchical Vision Transformer using Shifted Windows, Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 10012-10022). However, the skilled person will acknowledge that this transformer model is provided for illustration purposes only. In particular, the skilled person will acknowledge that the preferred embodiment is not limited to the use of this transformer model. On the contrary, the preferred embodiment is operable with any transformer-based or CNN-based backbone that can be used for instance segmentation.
In an embodiment, the Image Retrieval Unit implements an algorithm for product re-identification using a neural network that learns an embedding representation for each instance of a product contained in a product image database. More specifically, the Image Retrieval Unit compares the visual appearance of a pallet in a received video frame Fr(i) with appearance information of products contained in the warehouse's stock inventory. The appearance information is expressed in images of the products, the said images being stored in the product database. From these images, information about the appearance of a product under a variety of conditions (e.g. different viewing perspectives and rotation angles) can be expressed as embedding vectors formed by an embedding model such as VKD or Siamese Nets. The skilled person will understand that these embedding models are provided for illustrative purposes only. In particular, the skilled person will understand that the preferred embodiment is not limited to these embedding networks. On the contrary, the preferred embodiment is operable with any encoder model, capable of forming embedding vector representations of the appearance of a product, for example classic convolutional neural network (CNN). Several images of each product are used to train the embedding model. The embedding model can also be trained on image(s) of an entire pallet rather than images of each product.
Using their embedding vectors, the Image Retrieval Unit compares a product detected in a received video frame Fr(i) with products in the product database by means of a simple distance metric in the embedding space. In one embodiment the distance metric is a cosine metric or a Euclidean distance metric. The output from the Image Retrieval Unit is a label identifying the product whose image is contained in the product image database and which matches the product detected in a received video frame by the Detector Unit 430 of the Trajectory Computation Unit 420. For brevity, this label will be referred to henceforth as a “product label”.
Returning to
In an embodiment, the Human Pose Estimation Unit 422 is configured to recognize and extract a set of human actions from the received video footage. The actions most likely to be performed by staff in the warehouse comprise squatting, bending, and reaching movements. These actions are recognised through human pose estimation of each person tracked by the Trajectory Computation Unit 420. A human pose is formed by 16 important points of the human skeleton, namely, right/left ankle, right/left knee, right/left hip, pelvis, thorax, upper neck, head top, right/left wrist, right/left shoulder, right/left elbow. The Human Pose Estimation Unit 422 is configured to detect movements of individual body parts and combines successive detected movements to classify the action performed by the ith person. For example, to detect a squatting action, we analyse a person's head and shoulder joints for the presence of at least three repetitive movements in the vertical axis performed within a 3-minute time interval.
Using the co-ordinates of a person received from the Trajectory Computation Unit 420, the Human Pose Estimation Unit 422 is configured to apply single-person pose estimation to determine the pose of that person. To this end, the Human Pose Estimation Unit 422 is configured to use the UniPose neural network architecture that uses historical information to allow a temporal component to be considered and thereby better estimate joints in the presence of blurring or occlusion. The network is trained with frames labelled in the same way as the MPII Dataset but simplified to use 2D joints. The human pose estimated by the network is then post-processed using a time-based sliding window of an interval tp (e.g. tp=5 s) to extract the action corresponding to the person's body movements in the interval. The movements are defined by the trajectory (over each tp interval) of the body joint positions identified by the network. One or more actions (e.g. picking, dropping, bending, reaching, or throwing) corresponding with the detected body joint trajectories is determined in accordance with pre-defined detection strategies (i.e. which link known body part movements to specific known activities).
In an embodiment, the output from the Human Pose Estimation Unit 422 is a series of labels, each of which identifies one of several potential actions that could have been performed by the ith person at time t. In this way, the series of labels identifies the actions performed by the ith person at different points of time during their work in the warehouse.
In an embodiment, the Pattern Analyser 426 is configured to receive action labels from the Human Pose Estimation Unit 422, product labels from the Merchandise Classification Unit 424 and trajectories of detected Risk Subjects from the Trajectory Computation Unit 420. The Pattern Analyser 426 is configured to use internal logic elements to process the above-mentioned received information, to detect patterns of movement and activities undertaken by individual persons, stock items, pallets and equipment in the warehouse, wherein these patterns are consistent with the identified predefined risks and heuristic risks of the warehouse.
In an embodiment of the present disclosure, the Monitored Zones may have a numbering scheme based on identifiers of video sensors positioned to capture video footage in the respective monitored zones. Alternatively, the Monitored Zones may have a fixed numbering scheme (independent of the route taken by a warehouse operator) according to the requirements of the warehouse managers.
In entirety, the NERD component 408 processes the video data captured by the array of video sensors (404a till 404f) to create new heuristic risk types. Using this, a corresponding risk instance may be created based on observations of different process anomalies in each Monitored Zone and/or Uncovered Zone along the Operator Route.
In an example, a risk of excessive time spent by operator in a particular zone of the warehouse may be determined by comparing the time interval spent by an operator in the various Monitored Zones and/or Uncovered Zones along the Operator Route 420, against an expected “normal” time interval spent in the relevant warehouse zone. This risk may indicate the slowing-down of an activity/process undertaken in the warehouse zone. The “normal” time interval spent in the warehouse zone may be estimated as an average of the time intervals spent therein during a past pre-defined number of weeks. Also, the “normal” time interval may be estimated by observing a predefined number of the instances of the process performed in the relevant warehouse zone. Alternatively, the “normal” time interval may be estimated by calculating the average time spent in each Monitored Zone and/or Uncovered Zone along the Operator Route 420 during a pre-defined number (N) of previous days. For this risk type, a rule for updating the trigger could be “Update the Warehouse Risk Map 210 in
In another example, warehouse zones where risk incidents occur frequently, may be discovered by establishing a threshold for the number of process interruptions caused by the occurrence of various uncategorized/unknown incidents in Monitored Zones and/or Uncovered Zones. Such incidents may be reported by a warehouse manager, and may, for example, be caused by overly narrow aisles/spacing between racks, preventing items from being packed securely in the racks, so that packages fall from the rack. For this risk type, a rule for an update trigger could be “Update the Warehouse Risk Map (210 in
Thus, identification of risk areas allows the warehouse managers/operators to quickly take remedial action to address the cause thereof. More importantly, informed decision-making regarding pro-active measures may be taken including redesigning aspects of the warehouse to prevent or minimize the effect of the risk factors. The redesigning aspects may include redefining and/or improving manipulation procedures, redesigning the physical and logistics aspects of the warehouse environment, improving packing/stacking criteria, planning better order pickers routes, implementing enhanced (environmental and operator) monitoring etc.
The computer readable media may include non-transitory computer readable storage media, which may include hard drives, floppy diskettes, optical disks, CD-ROMs, DVDs, read-only memories (ROMs), random access memories (RAMS), EPROMs, EEPROMs, flash memory, magnetic or optical cards, solid-state memory devices, or other types of storage media suitable for storing electronic instructions. In addition, in some implementations, the computer readable media may include a transitory computer readable signal (in compressed or uncompressed form). Examples of computer readable signals, whether modulated using a carrier or not, include but are not limited to, signals that a computer system hosting or running a computer program can be configured to access, including signals downloaded through the Internet or other networks. Finally, the order in which the operations are described is not intended to be construed as a limitation, and any number of the described operations can be combined in any order and/or in parallel to implement the process.
At step 502, each Field of View of one or more video sensors installed in a warehouse environment are used to generate one or more Monitored Zones, and one or more Uncovered Zones therein. The one or more sensors have a Field of View that corresponds to a spatial volume in which the presence of objects may be detected in the absence of obstructions that would otherwise conceal the object. In the context of the present disclosure, the Field of View also covers an Operator Route, where the Operator Route is defined as the path traversed by a warehouse operator during the during a task period, and the task period is defined as the time period extending from the moment the operator receives a task list from the supervisor until she/he has finished all the tasks on the task list. It should be noted that a task on the task list may include multiple operations such as a handling, order-filling, pallet-loading/unloading, and rack-filling. At step 504, information comprising video streams captured by each video sensor is stored.
At step 506, each of the video streams are processed and aggregated to produce information regarding risk instances associated with an Operator Route followed by a warehouse operator while performing a warehouse operation, wherein the risk identification information includes at least one risk zone, and corresponding risk type, and risk level, wherein a risk zone is an area in the warehouse environment that corresponds to one or more risk instances. In an embodiment of the present disclosure, the warehouse operation is selected from at least one of: a handling task, an order filling task, a pallet loading/unloading task, and a rack filling task. A risk is selected from at least one of: a predefined risk arising from a heavy package, a predefined risk arising from a fragile package and a heuristic risk. In an embodiment of the present disclosure, the occurrence of one or more pre-defined risks is detected, and the location of each risk is marked on a Warehouse Risk Map to thereby illustrate corresponding risk instances. In an example, the pre-defined risk includes a risk arising from heavy packages, the location of the said risk is determined from an inventory list, and the corresponding Warehouse Risk Map is updated, each time the inventory list changes.
In an embodiment of the present disclosure, one or more heuristic risks are determined by comparing the time spent by the operator, object handling actions, and the operator's movement pattern with a corresponding pre-defined time spent by the operator, a pre-defined object handling action, and a pre-defined operator movement pattern.
At step 508, a Warehouse Risk Map is generated based on the risk instances information, wherein the Warehouse Risk Map is generated by superimposing an identified risk zone on a two-dimensional map of an observed warehouse environment. The superimposing risk zones are partially overlapped zones (areas) on the map which corresponds to two different risk instances such as first and second racks. The Warehouse Risk Map is used to optimize the spatial deployment of video cameras in the warehouse environment so that their collective Field of View cover all the locations associated with each risk instance.
At step 510, the Warehouse Risk Map is updated in real-time when at least one of the risk type, risk level, and risk zone changes for at least one risk instance recorded on the Warehouse Risk Map. In an embodiment of the present disclosure, a risk level for a risk zone is computed based on probability of a particular risk incident happening at the risk zone, the risk level including two components, a recent risk level, and a global risk level, where the recent risk level expresses a number of risk incidents that recently occurred in the risk zone as a fraction of total number of operations undertaken in the risk zone, and the global risk level expresses a total number of occurrences of risk incidents in the risk zone as a fraction of the total number of operations undertaken.
The computer readable media may include non-transitory computer readable storage media, which may include hard drives, floppy diskettes, optical disks, CD-ROMs, DVDs, read-only memories (ROMs), random access memories (RAMs), EPROMs, EEPROMs, flash memory, magnetic or optical cards, solid-state memory devices, or other types of storage media suitable for storing electronic instructions. In addition, in some implementations, the computer readable media may include a transitory compute: readable signal (in compressed or uncompressed form). Examples of computer readable signals, whether modulated using a carrier or not., include, but are not limited to, signals that a computer system hosting or running a computer program can be configured to access, including signals downloaded through the Internet or other networks. Finally, the order in which the operations are described is not intended to be construed as a limitation, and any number of the described operations can be combined in any order and/or in parallel to implement the process.
In an embodiment, the method may include a first step 602 of obtaining a plurality of video frames over a pre-defined period of time using one or more video sensors. The plurality of video frames may relate to one or more Monitored Zones, and one or more Uncovered Zones therein.
In an embodiment, the method may include a next step 604 of configuring a risk discovery unit, i.e. a New Emerging Risk Discovery (NERD) component 408 to the plurality of video sensors. In this embodiment, the New Emerging Risk Discovery (NERD) component 408 is configured to the central processing unit (CPU) 201, and the CPU 201 is communicably coupled to the plurality of video sensors. The (NERD) component 408 is configured to execute:
The next step 606 of the method, which comprises detecting a current risk subject in the plurality of real time video frames. The current risk subject is detected on the basis of a trigger condition comprising risk instances which may be associated with the current risk subject. The current risk subject is one of a person, pallets and packages visible in the plurality of video frame.
The next step 608 of the method, which comprises determining a location of the current risk subject detected in the obtained plurality of real time video frames. The location comprises location coordinates of the risk subject in the plurality of the video frame. The location coordinates are determined by a bounding box or an area of interest formed in the plurality of video frames. Based on the location of the determined risk subject, the next step 610 of the method is executed, which comprises calculating the detection measurement data of the one or more risk subjects.
The next step 612 of the method, which comprises estimating one or more physical characteristics of the current risk subject. The physical characteristics are related to various pose formed by 16 important points of the human skeleton, namely, right/left ankle, right/left knee, right/left hip, pelvis, thorax, upper neck, head top, right/left wrist, right/left shoulder, right/left elbow. The movement related to the pose are to classify the action formed within a pre-defined period of time.
The next 614 of the method, in which the classified action is received, and along with the determined location, patterns or movement of the current risk subjects are detected.
Modifications to embodiments of the present disclosure described in the foregoing are possible without departing from the scope of the present disclosure as defined by the accompanying claims. Expressions such as “including”, “comprising”, “incorporating”, “consisting of”, “have”, “is” used to describe and claim the present disclosure are intended to be construed in a non-exclusive manner, namely allowing for items, components or elements not explicitly described also to be present. Reference to the singular is also to be construed to relate to the plural.
This application is a Continuation-In-Part of U.S. patent application Ser. No. 16/991,210 filed on 12 Aug. 2020, the entirety of which is incorporated herein.
| Number | Date | Country | |
|---|---|---|---|
| Parent | 16991210 | Aug 2020 | US |
| Child | 17888752 | US |
