Historical target identification method and system, end-edge collaborative bounding box filtering method

Abstract

The application discloses a historical target identification method and system and an end-edge collaborative bounding box filtering method. ReID When facing mass video data, firstly, the time-space transfer characteristics of a track are analyzed to calculate the re-identification value of each frame of each camera. The higher the re-identification value is, the greater the probability of target appearance is. The greater first N frames are taken to perform re-identification, so as to save computing resources. Therefore, as many targets as possible can be identified for the frame pictures with higher re-identification value under limited computing resources. When a new target needs to be searched from an online video picture for analysis, the bounding box of the historical target which has been identified can be deleted, and the bounding box does not need to be transmitted to a computing center for subsequent analysis, so that the bandwidth resources required for transmission and the computing resources of the computing center are saved.

Description

Technical Field

[0001] This invention belongs to the field of video analysis technology, and more specifically, relates to a historical target recognition method and system, and an end-edge collaborative bounding box filtering method. Background Technology

[0002] Currently, CCTV cameras are being deployed more and more widely, with numerous cameras installed in major traffic arteries, public gathering places, shopping malls, hospitals, and other locations with complex security situations. Video data is transmitted to processing centers via dedicated video networks and the internet, where deep neural networks are used for analysis to classify public events and identify targets. For example, when it is necessary to identify new targets from video footage for behavioral analysis, directly analyzing the massive amounts of video data without processing would consume significant computing resources and be extremely time-consuming. Summary of the Invention

[0003] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a historical target recognition method and system, and an end-edge collaborative bounding box filtering method, which aims to solve the technical problem that video target recognition and analysis occupies too many computing resources and is time-consuming.

[0004] To achieve the above objectives, according to one aspect of the present invention, a historical target identification method is provided, characterized by comprising:

[0005] Acquire historical video data of a camera cluster and establish a target trajectory set between different cameras, the target trajectory set including the transfer trajectory of the target between two different cameras;

[0006] Based on the target trajectory set, construct the transition probability matrix and the first arrival time probability distribution matrix between cameras, where the elements in the transition probability matrix... The first-arrival time probability distribution matrix represents the probability that target r directly moves from camera i to camera j. Let r represent the probability that target r moves directly from camera i to camera j for the first time after k frames;

[0007] The re-identification value for each frame from each camera is calculated as follows:

[0008]

[0009] Let t1 represent the re-identification value of camera j in frame p, r represent the target index, R represent the number of targets, and t1 represent the number of targets. r This indicates the frame number of the target r before it leaves camera i before it first moves directly to camera j;

[0010] The re-identification values ​​of each frame from each camera in each time slot are sorted, and the top N with the largest values ​​are selected.ReID The frames are re-identified to obtain historical target bounding boxes.

[0011] In one embodiment, each element in the transition probability matrix The calculation formula is:

[0012]

[0013] n ij Let g be the number of transfer trajectories for target r to move directly from camera i to camera j, and g represent the camera number, g = 0, 1, 2 ... i-1, i+1, ... m, where m represents the number of cameras;

[0014] Elements in the first arrival time probability distribution matrix The calculation formula is:

[0015]

[0016] This represents the number of times target r is directly transferred from camera i to camera j after k frames, where l represents the number of frames.

[0017] In one embodiment, establishing the target trajectory set includes:

[0018] The system matches the transfer trajectories in the historical video data with the existing transfer trajectories in the current target trajectory set. If the same transfer trajectory exists, the count of the current transfer trajectory is incremented. If the same transfer trajectory does not exist, the current transfer trajectory is added to the target trajectory set and the count of the current transfer trajectory is recorded as 1.

[0019] According to another aspect of the present invention, an end-edge collaborative bounding box filtering method is provided, comprising:

[0020] Identify historical target bounding boxes using any of the methods described above;

[0021] The system acquires each frame of each camera in each time slot online and uploads it to the processing center. It identifies the current target bounding box, compares the current target bounding box of the same camera and the same frame with the historical bounding boxes, removes bounding boxes that are similar to the historical bounding boxes, and then performs target analysis based on the remaining bounding boxes.

[0022] In one embodiment, the method further includes: analyzing the video block division method of each frame of video from the same camera and the video encoding quality level of each video block based on historical data;

[0023] The system acquires and uploads frames from each camera in each time slot to the processing center online, including:

[0024] Based on historical data, the video blocks of each frame are divided and the video encoding quality level of each video block is determined. The current online video frame is then encoded and uploaded to the processing center.

[0025] In one embodiment, determining the video encoding quality level of each video block in the video frame based on historical data includes:

[0026] The video feed of each camera is divided into multiple video units. For each camera, a subset of target trajectories between different video units is constructed. The subset of target trajectories includes the transfer trajectory of the target between two different video units in the same camera.

[0027] Based on a subset of the target trajectory, construct a transition probability submatrix and a first-arrival time probability distribution submatrix between different video units within the same camera. The elements S in the transition probability submatrix... r ′ ab The element T in the first-arrival time probability distribution submatrix represents the probability that target r directly transitions from video unit a to video unit b. r ′ ab (k) represents the probability that target r is directly transferred from video unit a to video unit b for the first time after k frames;

[0028] The re-identification value of each video unit in the same camera is calculated as follows:

[0029]

[0030] This represents the re-identification value of video unit b in frame p within the same camera, where r represents the target sequence number, R represents the number of targets, and t2 r This indicates the frame number in which target r leaves video unit a before it is first directly transferred to video unit b;

[0031] Based on the re-identification value of each video unit within the same camera, the corresponding video frame is divided into multiple video blocks, and the re-identification value of each video unit within each video block is evenly distributed.

[0032] The video coding quality level of each video block is determined based on the average re-identification value of each video block in the same video frame.

[0033] In one embodiment, each element S in the transition probability submatrix r ′ ab The calculation formula is:

[0034]

[0035] n abLet h be the number of transfer trajectories for target r to move directly from video unit a to video unit b, where h represents the video unit number, h = 0, 1, 2 ... a-1, a+1, ... f, and f represents the number of video units.

[0036] The formula for calculating each element in the first arrival time probability distribution matrix is ​​as follows:

[0037]

[0038] This represents the number of times target r is directly transferred from video unit a to video unit b after k frames, where l represents the number of frames.

[0039] In one embodiment, the corresponding video frame is divided into multiple video blocks based on the re-identification value of each video unit within the same camera, including:

[0040] First, divide the video frame into multiple video tiles of the same size, with each video tile serving as an initial video block;

[0041] Determine whether the repartition value distribution of video units in each video block is uniform. For video blocks with uneven distribution, continue to divide them using a quadtree partitioning method until the maximum number of partitions is reached or until the repartition value distribution of video units in the current video block is uniform.

[0042] In one embodiment, comparing the current target bounding box with historical bounding boxes includes:

[0043] Calculate the size of the intersection region between the two bounding boxes in the corresponding frame. When the intersection region is greater than or equal to a set threshold, the current target bounding box is considered to be similar to the historical bounding box.

[0044] According to another embodiment of the present invention, a historical target identification system is provided, comprising:

[0045] The trajectory collection unit is used to acquire historical video data of the camera cluster and establish a target trajectory set between different cameras, the target trajectory set including the transfer trajectory of the target between two different cameras;

[0046] The transition probability matrix construction unit is used to construct the transition probability matrix between cameras based on the target trajectory set, where the elements in the transition probability matrix... This represents the probability that target r moves directly from camera i to camera j;

[0047] The first arrival time probability distribution matrix construction unit is used to construct the first arrival time probability distribution matrix based on the target trajectory set. The elements in the first arrival time probability distribution matrix... Let r represent the probability that target r moves directly from camera i to camera j for the first time after k frames;

[0048] The re-identification value calculation unit is used to calculate the re-identification value of each frame from each camera:

[0049]

[0050] Let t1 represent the re-identification value of camera j in frame p, r represent the target index, R represent the number of targets, and t1 represent the number of targets. r This indicates the frame number of the target r before it leaves camera i before it first moves directly to camera j;

[0051] The historical target bounding box recognition unit is used to sort the re-identification values ​​of each frame from each camera in each time slot, and select the top N with the largest values. ReID The frames are re-identified to obtain historical target bounding boxes.

[0052] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0053] This invention first proposes a historical target recognition method, which can quickly identify targets from massive amounts of historical video data. The method first analyzes the target's transfer trajectory between different cameras within a large amount of historical video data. Because target transfer exhibits a certain spatiotemporal correlation in specific locations, meaning there are statistical patterns in target transfer between cameras, this invention constructs a transfer probability matrix and a first-arrival time probability distribution matrix by analyzing the transfer trajectories of a large number of trajectories. It then calculates the re-identification value of each camera in each frame. A higher re-identification value indicates a more frequent appearance of the target in that frame; in other words, a higher re-identification value indicates a greater probability of the target appearing in that frame. Then, based on the calculated re-identification values, the re-identification values ​​of each frame from each camera are sorted in each time slot. The number of frames N that can be analyzed is set according to the computing power of the processing center. ReID Only the top N elements with higher recognition value are selected. ReID Frames are analyzed and the first N frames are identified. ReID The bounding boxes of targets appearing in the frames are used as historical target bounding boxes appearing in a massive amount of historical video footage. This invention does not directly identify targets from a massive amount of video footage, but selects a few frames with the highest re-identification value for identification. Although some image recognition may be lost, it saves computing resources and can identify as many targets as possible from frames with high re-identification value with limited computing resources.

[0054] The second aspect of the proposed end-edge collaborative bounding box filtering method is that when it is necessary to search for new targets in online video frames for analysis, it first identifies all target bounding boxes in the online video frames, and then deletes the identified historical target bounding boxes without transmitting them to the computing center for subsequent analysis, such as behavior detection. In other words, the same target does not need to be analyzed multiple times in the computing center, thereby saving the bandwidth resources required for transmission and the computing resources of the computing center. Attached Figure Description

[0055] Figure 1 This is a diagram of a monitoring system architecture as shown in one embodiment;

[0056] Figure 2 This is a flowchart of the steps of a historical target recognition method according to one embodiment;

[0057] Figure 3 This is a flowchart of the steps of an embodiment of an end-edge collaborative bounding box filtering method;

[0058] Figure 4 This is a flowchart of the steps in a video coding quality level classification method according to one embodiment;

[0059] Figure 5 This is a structural block diagram of a historical target recognition system according to one embodiment. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0061] To make it easier to understand, let's first introduce the typical architecture of a monitoring system, such as... Figure 1 As shown, cameras distributed in different areas capture videos, encode the video footage, and transmit it to the edge center. The edge center performs preprocessing before transmitting it to the cloud for further analysis. Data transmission from the cameras to the edge center and from the edge center to the cloud both utilize limited bandwidth resources.

[0062] like Figure 2 The diagram shows a flowchart of a historical target recognition method according to an embodiment, which mainly includes the following steps:

[0063] Step S110: Obtain historical video data of the camera cluster and establish a target trajectory set between different cameras. The target trajectory set includes the transfer trajectory of the target between two different cameras.

[0064] In certain scenarios, cameras are fixedly installed at specific locations. This camera cluster monitors targets in that environment and uploads the data to the cloud for analysis. After a period of time, the cloud stores a large amount of video data. Based on this historical video data, a set of target trajectories is constructed.

[0065] Specifically, the target trajectory set can record the transfer trajectories and trajectory information of various targets in a list format, including the number of trajectories and the time of occurrence. Understandably, different targets can have the same transfer trajectory, and the same target can have different transfer trajectories. When constructing the set, the transfer trajectories in historical video data are matched with the existing transfer trajectories in the current target trajectory set. If a matching transfer trajectory exists, the count of the current transfer trajectory is incremented. If no matching transfer trajectory exists, it is considered a newly appearing transfer trajectory and added to the target trajectory set, with the count of the new transfer trajectory recorded as 1. For example, when target A moves from camera 1 to camera 2, and target B also moves from camera 1 to camera 2, the number of trajectories for the target moving from camera 1 to camera 2 is 2. If the time it takes for a target to move from one camera to another exceeds a set duration t... exit If a target leaves a camera and does not reappear, it is considered to have left the coverage area of ​​the camera cluster.

[0066] Step S120: Construct the transition probability matrix and the first arrival time probability distribution matrix between cameras based on the target trajectory set, wherein the elements in the transition probability matrix... The first-arrival time probability distribution matrix represents the probability that target r directly moves from camera i to camera j. Let r represent the probability that target r is directly transferred from camera i to camera j for the first time after k frames.

[0067] In this step, two key matrices are constructed: the transition probability matrix and the first arrival time probability distribution matrix.

[0068] Specifically, a spatial transfer model of targets between cameras is established, the transfer probability of targets in different camera frames in historical video is calculated, and a transfer probability matrix is ​​formed. Specifically, the transfer probability refers to the one-step transfer probability, let... Let represent the probability that target r moves directly from camera i to camera j. Assume there are m cameras, and target r may move from camera i to any camera g other than camera i, where g = 0, 1, 2, ..., i-1, i+1, ..., m, where g = 0 indicates that the target has traveled through camera t. exit S still did not appear in the footage from other camera clusters. i0 This indicates that after the target leaves the camera i, it travels through t... exitThe probability of detecting the target in other camera cluster images remains unchanged. i = 1, 2, ..., m, j = 0, 1, 2, ..., i-1, i+1, ..., m. Therefore, the transition probability matrix is ​​an m × m dimensional probability transition matrix, where each element S... ij The calculation formula is:

[0069]

[0070] n ij Let r be the number of transfer trajectories from camera i to camera j.

[0071] Simultaneously, a time-transfer model of targets between cameras was established; and the first-arrival time distribution of targets in different camera shots in historical videos was statistically analyzed. Let represent the probability that target r directly moves from camera i to camera j for the first time after k frames. Since the target may linger after entering camera j, and the target appears in multiple consecutive frames of camera j, the time of the first direct transfer to camera j is the earliest time of such a transfer. The time taken for the target to first enter camera j after leaving camera i varies. By iterating through all cameras, an m×m dimensional first-arrival time distribution is formed for each frame, where each element... The calculation formula is:

[0072]

[0073] Let represent the number of times target r directly moves from camera i to camera j for the first time after k frames, and l represent the total number of frames that target r experiences when it directly moves from camera i to camera j for the first time.

[0074] Step S130: Calculate the re-identification value of each frame from each camera. Let t1 represent the re-identification value of camera j in frame p, r represent the target index, R represent the number of targets, and t1 represent the number of targets. r This indicates the frame number of the target r before it leaves camera i before it first moves directly to camera j.

[0075] Among them, p-t1 r Let r be the number of frames that target r takes to move directly from camera i to camera j for the first time.

[0076] By traversing all cameras and frame numbers in step S130, the re-identification value of each camera in each frame can be calculated. The higher the re-identification value, the more likely the target is to appear.

[0077] Step S140: Sort the re-identification values ​​of each frame from each camera in each time slot, and select the top N with the largest values. ReIDThe frames are re-identified to obtain historical target bounding boxes.

[0078] Specifically, each time slot t slot There will be t from m cameras slot ×FrameRate×M frames, where FrameRate represents the frame rate captured by the camera. The number of analyzable frames N is pre-set based on the computing resources of the computing center. ReID The frames within the time slot are sorted according to their re-identification value, and the top N frames are selected. ReID The frame is re-identified to identify the historical target bounding box. Specifically, the frame number to which the detected bounding box belongs and the position of the bounding box within the frame can be recorded in a temporary list.

[0079] Through steps S110-S140 above, based on the analysis of massive video data, the statistical trajectory of the target's movement between cameras is obtained, and the re-identification value of each frame of each camera is calculated. Only the top N frames with the largest re-identification values ​​are selected. ReID The analysis is performed on each frame. This invention does not directly identify targets from a massive amount of video footage, but rather selects a few frames with the highest re-identification value for identification. Although some image recognition may be lost, it saves computing resources and can identify as many targets as possible from frames with high re-identification value with limited computing resources.

[0080] Based on the above, the present invention further proposes an end-edge collaborative bounding box filtering method, such as... Figure 3 As shown, the method mainly includes:

[0081] Step S100: Obtain historical target bounding boxes.

[0082] Step S100 involves obtaining historical target bounding boxes based on the method described above.

[0083] Step S200: Acquire each frame image of each camera in each time slot online and upload it to the processing center, identify the current target bounding box, compare the current target bounding box of the same camera and the same frame with the historical bounding box, and remove the bounding box that is close to the historical bounding box.

[0084] In one embodiment, the camera captures video footage online, which is then encoded by a video encoder and uploaded to the edge center. The edge center identifies the bounding boxes, removes historical bounding boxes, and then uploads the footage to the cloud for computation and analysis.

[0085] Among them, based on the new camera footage, the bounding box can be obtained using conventional object detection algorithms.

[0086] The video footage is encoded by a video encoder before being transmitted to the edge center. Depending on the encoding method, the bandwidth resources occupied by the transmission vary.

[0087] In one embodiment, the video block division method and video encoding quality level of each video block of the same camera can be analyzed first based on historical video data. Then, the current online video is encoded and uploaded to the edge center based on the video block division and video encoding quality level of each video block obtained from historical data.

[0088] Specifically, such as Figure 4 As shown, the method for dividing video blocks and the video encoding quality level of each video block are determined based on historical video data analysis of the same camera. This mainly includes the following steps:

[0089] Step S210: Divide the video frame of each camera into multiple video units, and construct a target trajectory subset between different video units for each camera. The target trajectory subset includes the transfer trajectory of the target between two different video units in the same camera.

[0090] For example, a single frame from a camera is 1280*720 pixels, which can be divided into multiple video units with a minimum size of 16*16 pixels, and each video unit is 16*16 pixels in size.

[0091] Within the same camera, the target may also have a transfer trajectory, moving from one video unit to another. The transfer relationship between different video units of the target within the same camera is analogous to the transfer relationship between the target and different cameras, and a subset of the target trajectory is constructed for each camera.

[0092] Step S220: Construct a transition probability submatrix and a first-arrival time probability distribution submatrix between different video units within the same camera based on the target trajectory subset, wherein the elements S in the transition probability submatrix r ′ ab The element T in the first-arrival time probability distribution submatrix represents the probability that target r directly transitions from video unit a to video unit b. r ′ ab (k) represents the probability that target r is directly transferred from video unit a to video unit b for the first time after k frames.

[0093] Specifically, each element S in the transition probability submatrix r ′ ab The calculation formula is:

[0094]

[0095] n abLet h be the number of transfer trajectories for target r to move directly from video unit a to video unit b, where h represents the video unit number, h = 0, 1, 2 ... a-1, a+1, ... f, and f represents the number of video units.

[0096] The formula for calculating each element in the first arrival time probability distribution matrix is ​​as follows:

[0097]

[0098] This represents the number of times target r is directly transferred from video unit a to video unit b after k frames, where l represents the number of frames.

[0099] Step S230: Calculate the re-identification value of each video unit in the same camera. This represents the re-identification value of video unit b in frame p within the same camera, where r represents the target sequence number, R represents the number of targets, and t2 r This indicates the frame number of the target r before it leaves video unit a and is first directly transferred to video unit b.

[0100] Similar to the method of calculating the re-identification value of each frame of different cameras, this embodiment calculates the re-identification value of different video units within each frame of each camera. The higher the re-identification value, the more important the corresponding video unit area.

[0101] Step S240: Divide the corresponding video frame into multiple video blocks according to the re-identification value of each video unit in the same camera, and the re-identification value of each video unit in each video block is evenly distributed.

[0102] In one embodiment, the video frame can be first divided into multiple video tiles of the same size, for example, into 4*3 tiles. The 4*3 division is based on the consideration that most current camera video frame aspect ratios are 16:9 or 4:3, and a 4*3 division ensures that the size of each video tile is strictly equal. Each video tile serves as an initial video block. It is then determined whether the re-division value distribution of video units within each video block is uniform. For video blocks with uneven distribution, a quadtree partitioning method is used to continue partitioning until the maximum number of partitions is reached or until the re-division value distribution of video units within the current video block is uniform. Specifically, based on the re-identification value within a single camera frame captured in real time, a quadtree partitioning method is used to perform deep quadtree partitioning on each video tile. The video tile is used as the starting point for video block partitioning (i.e., the video tile is the largest video block). Under certain conditions, the video block is uniformly divided in a 2*2 format. Specifically, for each video block, the re-partitioning value of its internal video units is queried. If the re-partitioning value of the video units within the video block is very unevenly distributed, the video block is quadtree partitioned. At the same time, in order to prevent infinite iteration of block partitioning and ensure the robustness of the block partitioning strategy, the quadtree partitioning depth is limited to within 3. That is, when a video block has been partitioned to 3 levels, the partitioning of the video block stops.

[0103] Step S250: Determine the video coding quality level of each video block based on the average re-identification value of each video block in the same video frame.

[0104] By calculating the average re-identification value of each video block, important and unimportant regions within the same frame are distinguished. A higher re-identification value indicates a greater likelihood of the target appearing in that video block, requiring improved encoding quality. Conversely, a lower re-identification value indicates a lower probability of the target appearing in that video block, allowing for a reduction in encoding quality to conserve bandwidth. This encoding method prioritizes encoding regions with high re-identification value within limited bandwidth resources, thereby improving recognition accuracy.

[0105] After determining the target bounding box of the new video frame, historical bounding boxes are first discarded. Since historical bounding boxes have already been identified and analyzed, there is no need to analyze them again, thus saving cloud computing resources. For example, for a fixed office location, if three staff members A, B, and C have already been identified and their behavior analyzed in the first frame of camera 1 in the first time slot, then in the new video frame, it is unnecessary to continue analyzing targets A, B, and C; only the newly appearing target D needs to be analyzed, thereby saving computing resources. Furthermore, after discarding historical targets at the edge center, the number of bounding boxes uploaded to the cloud from the edge center is reduced, decreasing the bandwidth occupied and thus saving bandwidth resources as well.

[0106] In one specific embodiment, when performing bounding box comparison, the IoU of the bounding boxes in the corresponding frame can be calculated. IoU represents the size of the intersection region between two bounding boxes. The condition for identifying two bounding boxes as the same target bounding box is IoU(box) = size. i ,box j )≥α thresh ,,α thresh This is the set value.

[0107] Step S300: Perform target analysis based on the remaining bounding box.

[0108] By using the above methods for online target identification and analysis, the number of targets that need to be analyzed for re-identification can be greatly reduced, thus saving a significant amount of computing and bandwidth resources.

[0109] Furthermore, this invention also proposes a historical target identification system for analyzing massive amounts of historical data and identifying historical targets under prioritized computing resources. Specifically, such as... Figure 5 As shown, the system includes the following functional units:

[0110] The trajectory collection unit is used to acquire historical video data of the camera cluster and establish a target trajectory set between different cameras. The target trajectory set includes the transfer trajectory of the target between two different cameras.

[0111] The transition probability matrix construction unit is used to construct the transition probability matrix between cameras based on the target trajectory set, where the elements in the transition probability matrix... This represents the probability that target r moves directly from camera i to camera j;

[0112] The first arrival time probability distribution matrix construction unit is used to construct the first arrival time probability distribution matrix based on the target trajectory set. The elements in the first arrival time probability distribution matrix... Let r represent the probability that target r moves directly from camera i to camera j for the first time after k frames;

[0113] The re-identification value calculation unit is used to calculate the re-identification value of each frame from each camera:

[0114]

[0115] Let t1 represent the re-identification value of camera j in frame p, r represent the target index, R represent the number of targets, and t1 represent the number of targets. r This indicates the frame number of the target r before it leaves camera i before it first moves directly to camera j;

[0116] The historical target bounding box recognition unit is used to sort the re-identification values ​​of each frame from each camera in each time slot, and select the top N with the largest values. ReIDThe frames are re-identified to obtain historical target bounding boxes.

[0117] The historical target identification system is used to execute the historical target identification method described above. Each functional unit can only perform the corresponding steps, as detailed above.

[0118] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying historical targets, characterized in that, include: Acquire historical video data of a camera cluster and establish a target trajectory set between different cameras, the target trajectory set including the transfer trajectory of the target between two different cameras; Based on the target trajectory set, construct the transition probability matrix and the first arrival time probability distribution matrix between cameras, where the elements in the transition probability matrix... Indicate target Self-facing camera The probability of direct transfer to camera j, and the elements in the first-arrival time probability distribution matrix. Indicate target Self-facing camera experience The probability that a frame is first directly transferred to camera j; The re-identification value for each frame from each camera is calculated as follows: Indicates that camera j is at the 1st descent. The value of frame re-identification Indicates the target sequence number. Indicates the target quantity. Indicate target First, directly transfer to camera j before leaving camera. Frame number; The re-identification values ​​of each frame from each camera in each time slot are sorted, and the larger ones are selected. The frames are re-identified to obtain historical target bounding boxes.

2. The historical target identification method as described in claim 1, characterized in that, Elements in the transition probability matrix The calculation formula is: For the goal Self-facing camera The number of transfer trajectories directly transferred to camera j. Indicates the camera number. , Indicates the number of cameras; Elements in the first arrival time probability distribution matrix The calculation formula is: Indicate target Self-facing camera experience The number of times a frame is directly transferred to camera j for the first time. Indicates the number of frames.

3. The historical target identification method as described in claim 1, characterized in that, Establishing the target trajectory set includes: The system matches the transfer trajectories in the historical video data with the existing transfer trajectories in the current target trajectory set. If the same transfer trajectory exists, the count of the current transfer trajectory is incremented. If the same transfer trajectory does not exist, the current transfer trajectory is added to the target trajectory set and the count of the current transfer trajectory is recorded as 1.

4. A method for edge-end collaborative bounding box filtering, characterized in that, include: The historical target recognition method according to any one of claims 1 to 3 identifies historical target bounding boxes; The system acquires each frame of each camera in each time slot online and uploads it to the processing center. It identifies the current target bounding box, compares the current target bounding box of the same camera and the same frame with the historical bounding boxes, removes bounding boxes that are similar to the historical bounding boxes, and then performs target analysis based on the remaining bounding boxes.

5. The end-edge collaborative bounding box filtering method as described in claim 4, characterized in that, The method also includes: analyzing the video block division method of each frame of video from the same camera based on historical data and the video encoding quality level of each video block; The system acquires and uploads frames from each camera in each time slot to the processing center online, including: Based on historical data, the video blocks of each frame are divided and the video encoding quality level of each video block is determined. The current online video frame is then encoded and uploaded to the processing center.

6. The end-edge collaborative bounding box filtering method as described in claim 5, characterized in that, The video encoding quality level of each video block in the video frame is determined based on historical data, including: The video feed of each camera is divided into multiple video units. For each camera, a subset of target trajectories between different video units is constructed. The subset of target trajectories includes the transfer trajectory of the target between two different video units in the same camera. Based on a subset of the target trajectory, construct a transition probability submatrix and a first-arrival time probability distribution submatrix between different video units within the same camera. The elements in the transition probability submatrix... Indicate target Self-video unit Directly transferred to the video unit The probability of arrival, the elements in the first-arrival time probability distribution submatrix. Indicate target Self-video unit experience The frame is transferred directly to the video unit for the first time. The probability of; The re-identification value of each video unit in the same camera is calculated as follows: Indicates video unit in the same camera exist The value of frame re-identification Indicates the target sequence number. Indicates the target quantity. Indicate target For the first time, the data was directly transferred to the video unit. Previously left the video unit Frame number; Based on the re-identification value of each video unit within the same camera, the corresponding video frame is divided into multiple video blocks, and the re-identification value of each video unit within each video block is evenly distributed. The video coding quality level of each video block is determined based on the average re-identification value of each video block in the same video frame.

7. The edge-edge collaborative bounding box filtering method as described in claim 6, characterized in that, Elements in the transition probability submatrix The calculation formula is: For target r self-video unit Directly transferred to the video unit The number of transfer trajectories, Indicates the video unit number, , Indicates the number of video units; The formula for calculating each element in the first arrival time probability distribution matrix is ​​as follows: The target r is represented by the video unit. experience The frame is directly transferred to the video unit for the first time. The number of times, Indicates the number of frames.

8. The edge-edge collaborative bounding box filtering method as described in claim 7, characterized in that, Based on the re-identification value of each video unit within the same camera, the corresponding video frame is divided into multiple video blocks, including: First, divide the video frame into multiple video tiles of the same size, with each video tile serving as an initial video block; Determine whether the repartition value distribution of video units in each video block is uniform. For video blocks with uneven distribution, continue to divide them using a quadtree partitioning method until the maximum number of partitions is reached or until the repartition value distribution of video units in the current video block is uniform.

9. The edge-to-edge collaborative bounding box filtering method as described in claim 4, characterized in that, Compare the current target bounding box with historical bounding boxes, including: Calculate the size of the intersection region between the two bounding boxes in the corresponding frame. When the intersection region is greater than or equal to a set threshold, the current target bounding box is considered to be similar to the historical bounding box.

10. A historical target identification system, characterized in that, include: The trajectory collection unit is used to acquire historical video data of the camera cluster and establish a target trajectory set between different cameras, the target trajectory set including the transfer trajectory of the target between two different cameras; The transition probability matrix construction unit is used to construct the transition probability matrix between cameras based on the target trajectory set, where the elements in the transition probability matrix... Indicate target Self-facing camera The probability of directly transferring the signal to camera j; The first arrival time probability distribution matrix construction unit is used to construct the first arrival time probability distribution matrix based on the target trajectory set. The elements in the first arrival time probability distribution matrix... Indicate target Self-facing camera experience The probability that a frame is first directly transferred to camera j; The re-identification value calculation unit is used to calculate the re-identification value of each frame from each camera: Indicates that camera j is at the 1st descent. The value of frame re-identification Indicates the target sequence number. Indicates the target quantity. Indicate target First, directly transfer to camera j before leaving camera. Frame number; The historical target bounding box recognition unit is used to sort the re-identification values ​​of each frame from each camera in each time slot, and take the larger value. The frames are re-identified to obtain historical target bounding boxes.

Images