A regional multi-target tracking system based on edge collaboration

Abstract

The present invention relates to the field of computer vision technology, and in particular to a regional multi-target tracking system based on edge collaboration, including multiple sensor terminals and a central server, and when deploying a target recognition model, a distributed reasoning method of segmentation learning is used to divide the target recognition model into three parts, wherein the first network structure is deployed in a micro-solution processor of the sensor terminal, the second network structure is adaptively deployed in the micro-solution processor of the sensor terminal and the central server, and the third network structure is deployed in the central server. On the basis of image feature matching, a space and time series mapping method is added to perform target matching on targets collected by different sensor terminals, and enhance target recognition of small targets. The present invention reduces data transmission between the central server and the sensor terminal, relieves network pressure, improves the flexibility, scalability and real-time performance of system deployment, and further improves the accuracy of target tracking.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a regional multi-target tracking system based on edge collaboration. Background Art

[0002] Target tracking is a low-level problem in the field of computer vision. It is the basis for advanced tasks such as target classification, behavior understanding, and identity re-identification. Research on target tracking algorithms is of great practical significance.

[0003] Multi-target tracking is a branch of the target tracking problem. It can automatically track multiple targets in a video and assign an identifier to each target. Multi-target tracking is widely used in intelligent monitoring, behavior analysis, autonomous driving and other fields. In particular, when tracking targets in a wide area, multiple monitoring devices are often used to achieve target recognition and tracking in a larger range.

[0004] At present, the conventional multi-sensor target collaborative tracking system transmits all monitoring images to the central server for collaborative processing. The central server needs to be configured with higher bandwidth and computing power, resulting in high system deployment costs. When the system needs to be expanded, bandwidth and computing power bottlenecks are prone to occur, resulting in a decrease in collaborative performance. And because the monitoring areas of different cameras overlap, the same target will appear in one or more monitoring images. The common method currently used for target matching between multiple cameras is still limited to image feature matching algorithms, which distinguish or connect through individual significant image features. When multiple similar targets appear, the use of image feature matching algorithms will cause the tracking performance of the system to drop sharply. Summary of the invention

[0005] To this end, the technical problem to be solved by the present invention is to overcome the problems in the prior art that the multi-sensor target collaborative tracking system has high bandwidth and computing power requirements for the central server, is difficult to expand, and has low target matching accuracy.

[0006] In order to solve the above technical problems, the present invention provides a regional multi-target tracking system based on edge collaboration, comprising:

[0007] Multiple sensor terminals, including micro-resolver processors, for collecting image data of the target;

[0008] The central server is used to receive the image data of the targets from multiple sensor terminals and perform target recognition and target matching on the targets collected by different sensor terminals;

[0009] Among them, when deploying the target recognition model, the distributed reasoning method of segmentation learning is used to divide the target recognition model into three parts: a first network structure, a second network structure and a third network structure; the first network structure is deployed in the micro-solving processor of the sensor terminal, and is used to perform underlying feature extraction on the image data of the target collected by the sensor terminal; the second network structure is adaptively deployed in the micro-solving processor of the sensor terminal and the central server according to the remaining computing power of the sensor terminal, and is used for feature extraction and information transmission; the third network structure is deployed in the central server, and is used to perform feature analysis and semantic analysis on the features extracted by the sensor terminal to complete target recognition.

[0010] Preferably, the second network structure is adaptively deployed on the micro-processor and the central server of the sensor terminal according to the remaining computing power of the sensor terminal, and includes:

[0011] The second network structure is adaptively loaded or unloaded on the sensor terminal or the central server according to the remaining computing power of the sensor terminal: when the remaining computing power of the sensor terminal is less than the preset computing power threshold, the sensor terminal adaptively unloads the second network structure, and the central server loads the second network structure; when the remaining computing power of the sensor terminal is greater than or equal to the preset computing power threshold, the sensor terminal adaptively loads the second network structure, and the central server unloads the second network structure.

[0012] Preferably, the target recognition model of the central server is a YOLOv5 target recognition model.

[0013] Preferably, the first network structure includes:

[0014] The sensor terminal inputs the collected image data into the Focus module, the first convolutional layer, the first 3Bott l eNeckCSP module, the second convolutional layer, the first 9Bott l eNeckCSP module in the forward propagation direction, and outputs the first-level feature map;

[0015] The second network structure includes:

[0016] The first-level feature map is sequentially input into the third convolutional layer and the second 9Bott l eNeckCSP module along the forward propagation direction, and the second-level feature map is output;

[0017] The second-level feature map is sequentially input into the fourth convolutional layer and the SPP module along the forward propagation direction, and the third-level feature map is output;

[0018] The third-level feature map is input into the fifth convolutional layer in the bottom-up feature fusion unit, and the third sampling feature map is output;

[0019] Connecting the third sampling feature map with the second level feature map after upsampling, and outputting a third fused feature map;

[0020] The third network structure includes:

[0021] The third fused feature map is sequentially input into the first spatial attention module and the second 3Bott l eNeckCSP module in the bottom-up feature fusion unit along the forward propagation direction, and the third enhanced feature map is output;

[0022] Input the third enhanced feature map to the sixth convolutional layer in the bottom-up feature fusion unit, and output the second sampled feature map;

[0023] Connecting the second sampling feature map with the first level feature map after upsampling, and outputting a second fused feature map;

[0024] The second fused feature map is sequentially input into the second spatial attention module and the third 3Bott l eNeckCSP module in the bottom-up feature fusion unit along the forward propagation direction, and the second enhanced feature map is output, and the second enhanced feature map is input into the first detection head as the first target fused feature map;

[0025] The second enhanced feature map is input into the seventh convolutional layer in the top-down feature fusion unit, and its output is connected with the second sampling feature map and then input into the fourth 3Bott l eNeckCSP module in the top-down feature fusion unit to output the second target fusion feature map; the second target fusion feature map is input into the second detection head;

[0026] The second fused feature map is input into the eighth convolutional layer in the top-down feature fusion unit, and its output is connected with the third sampling feature map and then input into the fifth 3Bott l eNeckCSP module in the top-down feature fusion unit, which outputs the third target fused feature map and is output to the third detection head.

[0027] Preferably, the central server performs target recognition on targets collected by different sensor terminals, including:

[0028] Through the spatial topological structure and coordination mechanism, the different distances and angles of each target identified by different sensor terminals are mapped to the same spatial coordinate system. According to the coordinate value of each target in the spatial coordinate system, a globally unique target number is generated for each target, and the large target information, medium target information and small target information collected by different sensor terminals for the same target are bound.

[0029] The spatial position of the target is calibrated with the recognition results of the large target information and the medium target information to generate an attention heat map. When the target is recognized based on the small target information, the attention heat map of the large target information and the medium target information is input into the spatial attention module of the third network structure to enhance the characteristics of the small target information.

[0030] Preferably, the structure of the spatial attention module includes:

[0031] The attention heat map is normalized and multiplied with the features output by the upper network to obtain the enhanced features as the input features of the lower network.

[0032] Preferably, the central server performs target matching on targets collected by different sensor terminals, including:

[0033] The central server obtains the target recognition confidence output by the target recognition model based on the image features, obtains the spatial topological similarity between two targets based on the spatial features, and obtains the temporal topological similarity between two targets based on the temporal features. Then, according to the target recognition confidence, spatial topological similarity and temporal topological similarity, the similarity weights of each target in each sensor terminal and all targets in other sensor terminals are calculated respectively, and a many-to-many target mapping between multiple sensor terminals is obtained;

[0034] The many-to-many target mapping between multiple sensor terminals is organized into an asymmetric weighted bipartite graph, and the obtained adjacency matrix is ​​a non-square matrix. The non-square matrix is ​​first padded with zeros to obtain a square matrix, and then the Hungarian algorithm is used to calculate and solve it, and the matching combination with the highest weight is taken as the global optimal matching.

[0035] Preferably, the calculation formula for the similarity weight between every two targets is:

[0036]

[0037] Among them, w is the similarity weight, and its value range is [0,1]; Indicates whether the two targets are of the same type. If they are, the value is 1, and if they are not, the value is 0. α is the product of the recognition confidence of the two targets, v is the spatial topological similarity, and I is the time series similarity.

[0038] The time series similarity is the ratio of the time when two targets are successfully matched to the target survival time in the historical matching records of the two targets.

[0039] Preferably, the method for calculating the spatial topological similarity is:

[0040] The 360-degree range around each target is divided into intervals of 10 degrees, and the distances of other similar targets closest to the target in the interval are obtained to obtain the spatial topological vector V of the target. The elements in the spatial topological vector V represent the distances of adjacent targets in each interval direction.

[0041] The cyclic shift similarity is calculated for the spatial topological vectors V of the two targets, and the maximum similarity is taken as the spatial topological similarity.

[0042] Preferably, the micro-resolver processor is mounted on a NVIDIA Jetson board.

[0043] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0044] The regional multi-target tracking system based on edge collaboration described in the present invention adopts a distributed reasoning method of segmentation learning to divide the target recognition model into three parts, migrates part of the computing resources to the sensor terminal, performs edge data processing nearby, removes redundant information, reduces a large amount of data transmission between the central server and the sensor terminal, reduces the demand for bandwidth, alleviates network pressure, and can better support distributed applications, thereby improving the flexibility, scalability and real-time performance of system deployment.

[0045] The regional multi-target tracking system based on edge collaboration described in the present invention adds space and time series mapping methods on the basis of image feature matching, performs target matching on targets collected by different sensor terminals, realizes target tracking linkage matching under multiple sensor terminals, and further improves the accuracy of target tracking between multiple sensor terminals. In addition, the present invention adds a spatial attention module to the target recognition model, uses the spatial attention heat map of large and medium targets to provide prior information of the area where the small targets are located, and can improve the attention and feature extraction capabilities of small target recognition to specific areas, improve the sensitivity and accuracy of small target recognition, expand the recognition distance of the sensor terminal to the target, and enable the system to have target tracking capabilities in a larger area. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below according to specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0047] Figure 1 It is a schematic diagram of a dynamic segmentation structure of a regional multi-target tracking system based on edge collaboration of the present invention;

[0048] Figure 2 It is a schematic diagram of the target recognition model segmenting the first network structure, the second network structure diagram and the third network structure;

[0049] Figure 3 It is a flowchart of small target enhanced recognition combining spatial topological structure and synergy mechanism;

[0050] Figure 4 It is the structural diagram of the spatial attention module;

[0051] Figure 5 It is a schematic diagram of the evolution from the target space topology structure to the space topology vector;

[0052] Figure 6 is an asymmetric weighted bipartite graph matching two sensor terminals. DETAILED DESCRIPTION

[0053] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0054] Reference Figure 1 As shown, the present invention provides a regional multi-target tracking system based on edge collaboration, comprising:

[0055] Multiple sensor terminals, including micro-resolver processors, for collecting image data of the target;

[0056] The central server is used to receive image data of targets from multiple sensor terminals and perform target recognition and target matching on targets collected by different sensor terminals.

[0057] In order to reduce the computing pressure of the central server, the present invention adds a micro-computing processor to the sensor terminal, so that the sensor terminal also has a certain data feature extraction capability. When deploying the target recognition model, the distributed reasoning method of segmentation learning is used to divide the target recognition model into three parts: the first network structure, the second network structure and the third network structure.

[0058] The first network structure is deployed in the micro-resolving processor of the sensor terminal, and is used to extract underlying features of the image data of the target collected by the sensor terminal, such as texture, color, and edge feature information.

[0059] The second network structure is adaptively deployed on the micro-processor and central server of the sensor terminal according to the remaining computing power of the sensor terminal for feature extraction and information transmission, and specifically includes:

[0060] The second network structure is adaptively loaded or unloaded on the sensor terminal or the central server according to the remaining computing power of the sensor terminal: when the remaining computing power of the sensor terminal is less than the preset computing power threshold, the sensor terminal adaptively unloads the second network structure, and the central server loads the second network structure; when the remaining computing power of the sensor terminal is greater than or equal to the preset computing power threshold, the sensor terminal adaptively loads the second network structure, and the central server unloads the second network structure.

[0061] The flexible deployment of the second network structure ensures the joint model reasoning between the sensor terminal and the central server on the one hand, and takes into account the feature extraction of different examples by the sensor terminal on the other hand, so that the central server can adapt to more types of sensor terminals.

[0062] The third network structure is deployed on the central server and is used to perform feature analysis and semantic analysis on the features extracted by the sensor terminals.

[0063] Preferably, the present invention uses NVIDIA Jetson onboard as a micro-solving processor of the sensor terminal. The micro-solving processor adopts a graphics processing unit (GPU) architecture with a peak computing power of 1.33TFLOPS. It runs deep neural networks on sensor terminals with a power consumption as low as 7.5W and obtains higher accuracy. It is suitable for scenarios where bandwidth and latency problems need to be solved. The micro-solving processor meets the distributed deployment requirements of segmentation learning. Without losing recognition accuracy, it can make full use of the computing power of the edge, reduce network bandwidth and central server computing power requirements, and meet the application requirements of multi-channel synchronous solving, low latency, and simple expansion.

[0064] Preferably, the target recognition model is a YOLOv5 target recognition model, which not only has a fast reasoning speed but also has a relatively high detection accuracy, and can meet the accuracy requirements of target recognition. Figure 2 As shown, the segmentation methods of the YOLOv5 target recognition model specifically include:

[0065] The first network structure includes:

[0066] The sensor terminal inputs the collected image data into the Focus module, the first convolutional layer, the first 3Bott l eNeckCSP module, the second convolutional layer, the first 9Bott l eNeckCSP module in the forward propagation direction, and outputs the first-level feature map;

[0067] The second network structure includes:

[0068] The first-level feature map is sequentially input into the third convolutional layer and the second 9Bott l eNeckCSP module along the forward propagation direction, and the second-level feature map is output;

[0069] The second-level feature map is sequentially input into the fourth convolutional layer and the SPP module along the forward propagation direction, and the third-level feature map is output;

[0070] The third-level feature map is input into the fifth convolutional layer in the bottom-up feature fusion unit, and the third sampling feature map is output;

[0071] Connecting the third sampling feature map with the second level feature map after upsampling, and outputting a third fused feature map;

[0072] The third network structure includes:

[0073] The third fused feature map is sequentially input into the first spatial attention module and the second 3Bott l eNeckCSP module in the bottom-up feature fusion unit along the forward propagation direction, and the third enhanced feature map is output;

[0074] Input the third enhanced feature map to the sixth convolutional layer in the bottom-up feature fusion unit, and output the second sampled feature map;

[0075] Connecting the second sampling feature map with the first level feature map after upsampling, and outputting a second fused feature map;

[0076] The second fused feature map is sequentially input into the second spatial attention module and the third 3Bott l eNeckCSP module in the bottom-up feature fusion unit along the forward propagation direction, and the second enhanced feature map is output, and the second enhanced feature map is input into the first detection head as the first target fused feature map;

[0077] The second enhanced feature map is input into the seventh convolutional layer in the top-down feature fusion unit, and its output is connected with the second sampling feature map and then input into the fourth 3Bott l eNeckCSP module in the top-down feature fusion unit to output the second target fusion feature map; the second target fusion feature map is input into the second detection head;

[0078] The second fused feature map is input into the eighth convolutional layer in the top-down feature fusion unit, and its output is connected with the third sampling feature map and then input into the fifth 3Bott l eNeckCSP module in the top-down feature fusion unit, which outputs the third target fused feature map and is output to the third detection head.

[0079] To sum up, in the distributed reasoning method based on segmentation learning, the central server does not need to obtain the original data. Instead, each sensor terminal extracts features from the original image data, removes redundant information, and obtains the feature data, which is then transmitted to the central server. Without losing recognition accuracy, the bandwidth pressure of network transmission is greatly reduced, and the computing power shortage of the central server is alleviated. Therefore, the present invention has significant advantages for target tracking scenarios with high real-time requirements, and can simultaneously meet the application requirements of multi-channel synchronous solution, low latency, and simple expansion.

[0080] In the application environment of multiple sensor terminals, the larger the range monitored by each sensor terminal, the higher the spatial overlap rate of the sensor terminals, so there will be multiple angles of snapshot information for the same target. The central server is of great significance to improving the accuracy of target recognition and target tracking through the fusion and judgment of multi-angle target information. However, the perception process of a large-scale space will face the difficulty of identifying small targets. Since small targets usually have fewer pixels and insignificant feature expressions, the recognition model has low recognition accuracy for distant targets.

[0081] Reference Figure 3 As shown, in order to improve the recognition accuracy of small targets, the present invention combines the spatial topological structure and the coordination mechanism in the application environment of multiple sensor terminals. Through the different distances and different angles of the same target recognized by the multiple sensor terminals, the central server maps the targets recognized by different sensor terminals to the same spatial coordinate system, generates a globally unique target number and associates them. Different sensor terminals collect targets with the same target number, obtain target information of different sizes, calibrate the spatial position of the target with the recognition results of large and medium targets, generate an attention heat map, and associate the sensor terminals that collect small targets. When using small targets to identify targets, the third network structure adds a spatial attention module to the original structure of the YOLOv5 target recognition model, and strengthens the characteristics of small target information based on the attention heat map generated by large and medium targets.

[0082] Specifically, refer to Figure 4 As shown, the spatial attention module structure includes: the attention heat map is normalized and multiplied with the features output by the upper network to obtain the enhanced features as the input features of the lower network.

[0083] In the process of target recognition, adding a spatial attention heat map of the area where small targets appear can make the recognition model focus on the prior information of the area where the small targets are located, and weight the features during the feature transfer process, so as to focus on effective features. By increasing the attention and feature extraction capabilities of the target recognition model to a specific area, the present invention improves the sensitivity and accuracy of small target recognition, expands the recognition distance of the sensor terminal to the target, and enables the system to have the ability to track targets in a larger area.

[0084] After the target is identified, the central server performs target matching to associate and track the target between different frames. Traditional target matching schemes match the target image features collected by the sensor terminal based on their similarity. However, in the application environment of multiple sensor terminals, the image features, size features, and brightness features of the same target in different sensor terminals are significantly different, and it is impossible to match targets between multiple sensor terminals based solely on the similarity of image features.

[0085] In order to solve the above problems, the present invention adds a space and time series mapping method based on image feature matching, which specifically includes:

[0086] The central server obtains the target recognition confidence output by the target recognition model based on image features, obtains the spatial topological similarity between two targets based on spatial features, and obtains the temporal topological similarity between two targets based on temporal features. Then, based on the target recognition confidence, spatial topological similarity, and temporal topological similarity, the central server calculates the similarity weights of each target in each sensor terminal and all targets in other sensor terminals. The calculation formula for the similarity weight between each two targets is:

[0087]

[0088] Among them, w is the similarity weight, and its value range is [0,1]; It indicates whether the two targets are of the same type. If they are, the value is 1, and if they are not, the value is 0. α is the product of the recognition confidence of the two targets, v is the spatial topological similarity, and I is the time series similarity.

[0089] The time series similarity is the ratio of the time when two targets are successfully matched to the target survival time in the historical matching records of the two targets.

[0090] Reference Figure 5 As shown, the method for calculating the spatial topological similarity specifically includes: dividing the 360-degree range around each target into intervals of 10 degrees, obtaining the distances of other similar targets closest to the target within the interval, and obtaining a 1×36 spatial topological vector V of the target, wherein the elements in the spatial topological vector V represent the distances of adjacent targets in each interval direction; performing cyclic shift similarity calculation on the spatial topological vectors V of the two targets, and taking the maximum similarity as the spatial topological similarity. The method for calculating the spatial topological similarity has the advantage of similar rotation invariance, and can offset the target space rotation changes caused by inconsistent shooting angles of different sensor terminals.

[0091] According to the similarity weight between each target and all targets in other sensor terminals, a many-to-many target mapping between multiple sensor terminals is obtained. Each target is associated with multiple targets in other sensor terminals. The present invention organizes the many-to-many target mapping between multiple sensor terminals into an asymmetric weighted bipartite graph.

[0092] Asymmetric weighted bipartite graph reference for matching two sensor terminals Figure 6 As shown, x1, x2, x3 and x4 are targets collected by one sensor terminal, and y1, y2, y3, y4, y5 and y6 are targets collected by another sensor terminal.

[0093] Since the number of targets captured by each sensor terminal is inconsistent, the adjacency matrix obtained based on the asymmetric weighted bipartite graph is a non-square matrix and cannot be directly solved by the Hungarian algorithm. Therefore, the non-square matrix is ​​first padded with zeros to obtain a square matrix, and then the maximum matching method of the asymmetric weighted bipartite graph is calculated by the Hungarian algorithm, and the matching combination with the highest weight is taken as the global optimal matching. This target matching method has a certain matching anti-disturbance capability, realizes the regional target tracking linkage function under multiple sensor terminals, and further improves the accuracy of target tracking between multiple sensor terminals.

[0094] Obviously, the above embodiments are merely examples for the purpose of clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A regional multi-target tracking system based on edge collaboration, characterized in that: include: Multiple sensor terminals, including micro-resolvers, for collecting image data of the target; The central server is used to receive the image data of the targets from multiple sensor terminals and perform target recognition and target matching on the targets collected by different sensor terminals; Among them, when deploying the target recognition model, the distributed reasoning method of segmentation learning is used to divide the target recognition model into three parts: a first network structure, a second network structure and a third network structure; the first network structure is deployed in the micro-resolving processor of the sensor terminal, and is used to perform underlying feature extraction on the image data of the target collected by the sensor terminal; the second network structure is adaptively deployed in the micro-resolving processor of the sensor terminal and the central server according to the remaining computing power of the sensor terminal, and is used for feature extraction and information transmission; the third network structure is deployed in the central server, and is used to perform feature analysis and semantic analysis on the features extracted by the sensor terminal to complete target recognition; The central server identifies the targets collected by different sensor terminals, including: Through the spatial topological structure and coordination mechanism, the different distances and angles of each target identified by different sensor terminals are mapped to the same spatial coordinate system. According to the coordinate value of each target in the spatial coordinate system, a globally unique target number is generated for each target, and the large target information, medium target information and small target information collected by different sensor terminals for the same target are bound. For the same target, the spatial position of the target is calibrated using the recognition results of the large target information and the medium target information of the target to generate an attention heat map; when target recognition is performed based on the small target information of the target, the attention heat map of the large target information and the medium target information of the target is input into the spatial attention module of the third network structure to enhance the features of the small target information of the target.

2. According to the edge collaboration-based regional multi-target tracking system of claim 1, it is characterized in that: The second network structure is adaptively deployed on the micro-processor and the central server of the sensor terminal according to the remaining computing power of the sensor terminal, and includes: The second network structure is adaptively loaded or unloaded on the sensor terminal or the central server according to the remaining computing power of the sensor terminal: when the remaining computing power of the sensor terminal is less than the preset computing power threshold, the sensor terminal adaptively unloads the second network structure, and the central server loads the second network structure; when the remaining computing power of the sensor terminal is greater than or equal to the preset computing power threshold, the sensor terminal adaptively loads the second network structure, and the central server unloads the second network structure.

3. The regional multi-target tracking system based on edge collaboration according to claim 1 is characterized in that: The target recognition model of the central server is the YOLOv5 target recognition model.

4. The regional multi-target tracking system based on edge collaboration according to claim 3 is characterized in that: The first network structure includes: The sensor terminal inputs the collected image data into the Focus module, the first convolutional layer, the first 3BottleNeckCSP module, the second convolutional layer, the first 9BottleNeckCSP module in the forward propagation direction, and outputs the first-level feature map; The second network structure includes: The first-level feature map is sequentially input into the third convolutional layer and the second 9BottleNeckCSP module along the forward propagation direction, and the second-level feature map is output; The second-level feature map is sequentially input into the fourth convolutional layer and the SPP module along the forward propagation direction, and the third-level feature map is output; The third-level feature map is input into the fifth convolutional layer in the bottom-up feature fusion unit, and the third sampling feature map is output; Connecting the third sampling feature map with the second level feature map after upsampling, and outputting a third fused feature map; The third network structure includes: The third fused feature map is sequentially input into the first spatial attention module and the second 3BottleNeckCSP module in the bottom-up feature fusion unit along the forward propagation direction, and the third enhanced feature map is output; Input the third enhanced feature map to the sixth convolutional layer in the bottom-up feature fusion unit, and output the second sampled feature map; Connecting the second sampling feature map with the first level feature map after upsampling, and outputting a second fused feature map; The second fused feature map is sequentially input into the second spatial attention module and the third 3BottleNeckCSP module in the bottom-up feature fusion unit along the forward propagation direction, and the second enhanced feature map is output, and the second enhanced feature map is input into the first detection head as the first target fused feature map; The second enhanced feature map is input into the seventh convolutional layer in the top-down feature fusion unit, and its output is connected with the second sampling feature map and then input into the fourth 3BottleNeckCSP module in the top-down feature fusion unit to output the second target fusion feature map; the second target fusion feature map is input into the second detection head; The second fused feature map is input into the eighth convolutional layer in the top-down feature fusion unit, and its output is connected with the third sampling feature map and then input into the fifth 3BottleNeckCSP module in the top-down feature fusion unit, which outputs the third target fused feature map and is output to the third detection head.

5. The regional multi-target tracking system based on edge collaboration according to claim 1 is characterized in that: The structure of the spatial attention module includes: The attention heat map is normalized and multiplied with the features output by the upper network to obtain the enhanced features as the input features of the lower network.

6. The regional multi-target tracking system based on edge collaboration according to claim 1 is characterized in that: The central server matches the targets collected by different sensor terminals, including: The central server obtains the target recognition confidence output by the target recognition model based on the image features, obtains the spatial topological similarity between two targets based on the spatial features, and obtains the temporal topological similarity between two targets based on the temporal features. Then, according to the target recognition confidence, spatial topological similarity and temporal topological similarity, the similarity weights of each target in each sensor terminal and all targets in other sensor terminals are calculated respectively, and a many-to-many target mapping between multiple sensor terminals is obtained; The many-to-many target mapping between multiple sensor terminals is organized into an asymmetric weighted bipartite graph, and the obtained adjacency matrix is ​​a non-square matrix. The non-square matrix is ​​first padded with zeros to obtain a square matrix, and then the Hungarian algorithm is used to calculate and solve it, and the matching combination with the highest weight is taken as the global optimal matching.

7. The regional multi-target tracking system based on edge collaboration according to claim 6 is characterized in that: The calculation formula for the similarity weight between every two targets is: ； in, is the similarity weight, and its value range is ; Indicates whether the two targets are of the same type. If so, the value is 1; if not, the value is 0. is the product of the recognition confidence of the two targets, is the spatial topological similarity, is the time series similarity; The time series similarity is the ratio of the time when two targets are successfully matched to the target survival time in the historical matching records of the two targets.

8. The regional multi-target tracking system based on edge collaboration according to claim 7 is characterized in that: The calculation method of the spatial topological similarity is: The 360-degree range around each target is divided into intervals of 10 degrees, and the distances of other similar targets closest to the target in the interval are obtained to obtain the spatial topological vector V of the target. The elements in the spatial topological vector V represent the distances of adjacent targets in each interval direction. The cyclic shift similarity is calculated for the spatial topological vectors V of the two targets, and the maximum similarity is taken as the spatial topological similarity.

9. The regional multi-target tracking system based on edge collaboration according to claim 1, characterized in that: The micro-solver processor is an NVIDIA Jetson onboard.

Images