Multi-source sensing fusion and target information optimization method in complex intersection environment

By upgrading and optimizing the hardware and software of AI cameras and millimeter wave radars, combined with multi-source data fusion technology, a multi-objective joint optimization model is established, and the accuracy and delay problems of traditional perception equipment in vehicle-road collaborative safety warning system are solved, and the accuracy and delay problems of traditional perception equipment in vehicle-road collaborative safety warning system are achieved. The improvement of pedestrian behavior analysis and safety warning in complex traffic scenarios has been improved, and the accuracy of traffic environment perception and the efficiency of early warning system have been improved.

CN119089136APending Publication Date: 2024-12-06NANTONG TIANCHENG OPTOELECTRONICS TECH CO LTD +4
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411121201.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

Traditional perception devices have low perceptual accuracy and transmission time delay in vehicle-road collaborative safety warning system, resulting in insufficient information redundancy and transmission time delay, and are unable to effectively monitor and early warning of pedestrian behavior in complex traffic scenarios.

Method used

Using general-purpose AI cameras and millimeter-wave radar as main equipment, hardware and software upgrade and optimization are carried out, multi-source data fusion is combined with the detection results of AI cameras, multi-objective joint optimization model is established, and deep reinforcement learning is used to obtain the optimal path. Side target information is transmitted to the vehicle and the cloud through the RSU.

Benefits of technology

Improve the system's confidence in safety warning of roadside traffic targets, realize pedestrian behavior analysis and multi-target tracking, reduce information transmission delay, make up for the shortcomings of existing equipment in target classification and pedestrian behavior analysis, and improve Traffic environment perception accuracy and early warning system efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089136A_ABST
    Figure CN119089136A_ABST
Patent Text Reader

Abstract

The invention discloses a multi-source sensing fusion and target information optimization method in a complex intersection environment. The method comprises the following steps: step 1, determining sensing equipment: selecting a universal AI camera and a millimeter wave radar as main equipment for road side target sensing; according to the method, multi-source data fusion is beneficial for improving the credibility of the system for safety early warning between traffic targets in different directions of the roadside; effective application of pedestrian behavior analysis, including multi-target tracking analysis and pedestrian abnormal behavior analysis such as wandering and behavior mutation, on key intersections or road sections of national and provincial roads is facilitated, and more refined scene application is met; the intersection target information is optimized, the optimal roadside target information required by vehicle-road cooperation is acquired and transmitted to the vehicle, the early warning unit and the cloud through the RSU, and the information transmission time delay and resource waste are reduced while the requirement of the system for national and provincial road environment perception is met.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of vehicle-road collaboration technology, and specifically to a multi-source perception fusion and target information optimization method in a complex intersection environment. Background Art

[0002] With the rapid development of the economy, the number of cars in cities has increased rapidly, and urban traffic problems have become increasingly serious. In areas with complex urban land use and concentrated residential, office, entertainment, and hospital spaces, traffic demand is often greater than traffic supply, that is, travel is congested, chaotic, and inefficient. At the same time, complex traffic scenarios have various forms of traffic organization, including various types of deformed intersections, five-way or more intersections, and double intersections within a short distance. Complex traffic scenarios have complex traffic supply and diverse traffic contradictions due to road conditions that are different from T-intersections or cross intersections.

[0003] As an important technology that runs through multiple fields of new infrastructure, vehicle-road collaboration is not only an important bridge to realize smart transportation and build a strong transportation country, but also can help planners and builders to rationally plan roads, improve traffic efficiency, and ultimately allow people to enjoy convenient, safe and efficient travel services. Under the general framework of vehicle-road collaboration, more attention is currently paid to how to enable vehicles to obtain more accurate and real-time traffic data to meet the development needs of intelligent vehicles, while there is a lack of more targeted research on the perception and specific behavior analysis of roadside pedestrians. National and provincial highways are important channels connecting major transportation distribution centers such as provinces, cities, and counties. In order to enable the national and provincial highway safety warning system based on vehicle-road collaboration to more effectively reflect the "prevention first" traffic safety management strategy and achieve more comprehensive monitoring, warning, guidance and control of traffic behavior, it is necessary to further improve the traffic targets in the warning system, especially the perception of pedestrians. At present, there are many types of target perception devices for vehicles and pedestrians on the market, and the complex environment of intersections is perceived through the establishment of perception devices.

[0004] However, traditional sensing devices have the following shortcomings: Traditional sensing devices are set up in the vehicle-road collaborative safety warning system. Due to the low perception accuracy of the sensing devices and the time delay in transmission, the vehicle-road collaborative safety warning system has deficiencies in information redundancy, transmission delay and other aspects. Summary of the invention

[0005] The purpose of the present invention is to provide a multi-source perception fusion and target information optimization method in a complex environment of an intersection, so as to solve the problem that the traditional perception equipment proposed in the above background technology is set up in the vehicle-road collaborative safety warning system. Due to the low perception accuracy of the perception equipment and the time delay in transmission, the vehicle-road collaborative safety warning system has insufficient processing of information redundancy, transmission delay and other aspects.

[0006] To achieve the above object, the present invention provides the following technical solution: a method for multi-source perception fusion and target information optimization in a complex environment of an intersection, comprising the following steps:

[0007] Step 1: Determine the perception equipment: Select general AI cameras and millimeter-wave radars as the main equipment for roadside object perception;

[0008] Step 2: Upgrade the sensing equipment: perform hardware upgrade and software optimization on the camera and radar module respectively;

[0009] Step 3: Multi-source data fusion: After the perception equipment is upgraded, it combines the perception information of other roadside perception equipment, mainly millimeter-wave radar, and uses the KM algorithm to correlate and fuse the detection results of the AI ​​camera on the roadside to obtain the fused and optimized roadside target information and scene analysis;

[0010] Step 4: Information optimization: For the roadside target information obtained, a multi-objective joint optimization model is established, and the policy gradient descent algorithm is used to update the neural network parameters to obtain the optimal behavior strategy of deep reinforcement learning to obtain the optimal roadside target information required for vehicle-road collaboration;

[0011] Step 5: Information transmission: The optimized information of the complex environment of the intersection is transmitted to the vehicle, early warning unit and cloud through RSU.

[0012] As a preferred technical solution of the present invention, the hardware upgrade and optimization in step 2 specifically includes selecting chips for designing the camera's built-in controller and edge controller. For some application scenarios, starlight-level AI cameras are used to improve the effectiveness of target perception in severe weather conditions. With the help of algorithm optimization, the camera module and radar module have the advantages of lightweight and miniaturization, and are directly embedded in the roadside equipment for use, thereby improving the level of integration of the roadside equipment.

[0013] As a preferred technical solution of the present invention, the software upgrade optimization in the step 2 is specifically based on the YOLOv5 target detection framework, using MobileNetV2 to replace the backbone network of the camera module, and replacing the ordinary convolution of the PAN structure with a deep separable convolution, which greatly reduces the number of parameters of the network model and supports the lightweight and miniaturized design of the camera module; introducing an attention mechanism, strengthening the ability of the network structure to extract key features from feature information, enhancing model performance, and improving the algorithm speed and target classification accuracy of the camera. The YOLOv5 directly detects and locates the target from the original image, and the MobileNetV2 is composed of a series of stacked InvertedResidual Blocks, each of which includes the following three layers: a 1x1 dimensionality-enhancing layer that increases the number of output channels to expand the feature map, a deep separable convolution layer: reducing the computational cost by applying spatial deep separable convolutions, which includes 3x3 deep separable convolutions and batch normalization operations, and a linear bottleneck layer: this is a 1x1 convolution layer that reduces the number of output channels and adds a linear activation function.

[0014] As a preferred technical solution of the present invention, the specific steps of using the KM algorithm in step 3 to associate and fuse with the detection results of the AI ​​camera also on the roadside are as follows: S1, image initialization: initialize the weights of all edges in the graph perceived by the perception device to 0, and construct an initial staggered tree T = (X, Y), where X and Y represent two vertex sets respectively; S2, determine equal subgraphs: count each vertex x of T in the image initialized in the perception device, if there is a path that satisfies lx + ly = W, where lx is the sum of the weights of all the tree edges of x in T, and ly is x is the sum of the weights of all out-of-tree edges in T, W is the weight of x in the original graph, then add x to the equal subgraph; S3, find unmatched nodes: find unmatched nodes in the equal subgraph and mark them as unmatched nodes; S4, find augmented paths: if there are unmatched nodes, select a starting node from it and find a path consisting of edges that alternate with it, so that the last edge of the path is an edge associated with the unmatched node; S5, modify the label: for each vertex x in the equal subgraph, select a vertex y associated with it from the matched nodes and calculate d=min(l x +l y -w), where w is the weight of the edge between x and y. In order to make the augmenting path better, subtract d from the weight of the in-tree edges of all x nodes in T, and add d to the weight of the out-tree edges of all y nodes in T; S6, update the interleaved tree: adjust all the edges in the augmenting path and all the non-tree edges in T to obtain a new interleaved tree.

[0015] As a preferred technical solution of the present invention, the scene analysis in step three includes pedestrian behavior analysis, vehicle behavior analysis and vehicle-pedestrian differentiation. The pedestrian behavior analysis specifically analyzes the pedestrian behavior at the intersection. The vehicle behavior analysis analyzes the vehicle behavior at the intersection in terms of acceleration, deceleration, lane change, direction change and parking. The vehicle-pedestrian differentiation distinguishes between vehicles and pedestrians at the intersection. The pedestrian behavior analysis is provided with multi-target tracking analysis and pedestrian abnormal behavior analysis. The multi-target tracking analysis tracks and analyzes multiple pedestrians at the intersection. The pedestrian abnormal behavior analysis analyzes and warns pedestrians of wandering or sudden changes in behavior.

[0016] As a preferred technical solution of the present invention, the goal of establishing the multi-objective joint optimization model in step 4 is to balance the system's average task processing delay, average vehicle energy consumption, perception accuracy and perception range.

[0017] As a preferred technical solution of the present invention, in step four, deep reinforcement learning optimizes the target information after the association and fusion of the intersection area.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] 1. Multi-source data fusion is conducive to improving the credibility of the system's safety warnings between different traffic targets on the roadside; it is conducive to the effective application of pedestrian behavior analysis at key intersections or sections of national and provincial roads, including multi-target tracking analysis, pedestrian abnormal behavior analysis such as wandering and behavioral mutations, to meet more refined scenario applications;

[0020] 2. Optimize intersection target information, obtain the optimal roadside target information required for vehicle-road collaboration, and transmit it to vehicles, warning units and the cloud through RSU, which not only meets the system's needs for environmental perception of national and provincial roads, but also reduces information transmission delay and resource waste;

[0021] 3. Optimize the software and hardware of existing target perception equipment, and combine it with multi-source data fusion technology to make up for the deficiencies of existing roadside perception equipment on national and provincial roads in target classification, pedestrian behavior analysis, and scenario applicability; at the same time, with the help of target information optimization technology, obtain the optimal roadside target information required by the system to make up for the deficiencies in information redundancy and transmission delay in the current national and provincial road vehicle-road collaborative safety warning system. Under the premise of rational use of information resources, further reduce the communication delay between the perception unit and the warning unit and the vehicle, so as to more effectively prevent human-related traffic conflicts between national and provincial roads and their road traffic targets, and provide reliable front-end technical support for subsequent research on vehicle-road collaborative information service models and deployment of more reasonable roadside warning solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a flow chart of the present invention;

[0023] Figure 2 This is a schematic diagram of the roadside target data fusion principle of the present invention;

[0024] Figure 3 This is a flow chart of multi-source data fusion of the present invention;

[0025] Figure 4 A schematic diagram of the architecture of the scenario analysis of the present invention;

[0026] Figure 5 A data optimization model diagram for the present invention;

[0027] Figure 6 Schematic diagram of the architecture of the multi-objective joint optimization model of the present invention. DETAILED DESCRIPTION

[0028] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0029] See also Figure 1-6 The present invention provides a method for multi-source perception fusion and target information optimization in a complex environment of an intersection, comprising the following steps:

[0030] Step 1: Determine the perception equipment: Select general AI cameras and millimeter-wave radars as the main equipment for roadside object perception;

[0031] Step 2: Upgrade the sensing equipment: perform hardware upgrade and software optimization on the camera and radar module respectively;

[0032] Step 3: Multi-source data fusion: After the perception equipment is upgraded, it combines the perception information of other roadside perception equipment, mainly millimeter-wave radar, and uses the KM algorithm to correlate and fuse the detection results of the AI ​​camera on the roadside to obtain the fused and optimized roadside target information and scene analysis;

[0033] Step 4: Information optimization: For the roadside target information obtained, a multi-objective joint optimization model is established, and the policy gradient descent algorithm is used to update the neural network parameters to obtain the optimal behavior strategy of deep reinforcement learning to obtain the optimal roadside target information required for vehicle-road collaboration;

[0034] Step 5: Information transmission: The optimized information of the complex environment of the intersection is transmitted to the vehicle, early warning unit and cloud through RSU.

[0035] Step 2 of hardware upgrade and optimization is to select chips for the design of camera built-in controller and edge controller. For some application scenarios, starlight-level AI cameras are used to improve the effectiveness of target perception in severe weather conditions. With the help of algorithm optimization, the camera module and radar module have the advantages of lightweight and miniaturization, and are directly embedded in the roadside equipment for use, thereby improving the level of integration of roadside equipment.

[0036] The software upgrade and optimization in step 2 is specifically based on the YOLOv5 target detection framework, using MobileNetV2 to replace the backbone network of the camera module, and replacing the ordinary convolution of the PAN structure with a deep separable convolution, which greatly reduces the number of parameters of the network model and supports the lightweight and miniaturized design of the camera module; introducing the attention mechanism, strengthening the network structure's ability to extract key features from feature information, enhancing model performance, and improving the camera's algorithm speed and target classification accuracy. YOLOv5 detects and locates targets directly from the original image, and MobileNetV2 consists of a series of stacked Inverted Residual Blocks. Each Inverted Residual Block contains the following three layers: a 1x1 dimensionality-enhancing layer that increases the number of output channels to expand the feature map, a deep separable convolution layer: reduces computational costs by applying spatial deep separable convolutions, and this layer contains 3x3 deep separable convolutions and batch normalization operations, and a linear bottleneck layer: this is a 1x1 convolution layer that reduces the number of output channels and adds a linear activation function.

[0037] In step 3, the specific steps of using the KM algorithm to associate and fuse with the detection results of the AI ​​camera also on the roadside are as follows: S1, image initialization: initialize the weights of all edges in the graph perceived by the perception device to 0, and construct an initial staggered tree T = (X, Y), where X and Y represent two vertex sets respectively; S2, determine equal subgraphs: count each vertex x in T in the image initialized in the perception device, if there is a path that satisfies lx + ly = W, where lx is the sum of the weights of all the tree edges entering x in T, and ly is the weight of all the tree edges leaving x in T. , where W is the weight of x in the original graph, then add x to the equal subgraph; S3, find unmatched nodes: find unmatched nodes in the equal subgraph and mark them as unmatched nodes; S4, find augmented paths: if there are unmatched nodes, select a starting node from them and find a path consisting of edges that alternate with them, so that the last edge of the path is an edge associated with the unmatched node; S5, modify the label: for each vertex x in the equal subgraph, select a vertex y associated with it from the matched nodes and calculate d=min(l x +l y-w), where w is the weight of the edge between x and y. In order to make the augmenting path better, subtract d from the weight of the in-tree edges of all x nodes in T, and add d to the weight of the out-tree edges of all y nodes in T; S6, update the interleaved tree: adjust all the edges in the augmenting path and all the non-tree edges in T to obtain a new interleaved tree.

[0038] The scene analysis in step three includes pedestrian behavior analysis, vehicle behavior analysis and vehicle-pedestrian differentiation. Pedestrian behavior analysis conducts specific analysis on pedestrian behavior at intersections. Vehicle behavior analysis conducts acceleration, deceleration, lane change, direction change and parking analysis on vehicle behavior at intersections. Vehicle-pedestrian differentiation distinguishes vehicles and pedestrians at intersections. Pedestrian behavior analysis includes multi-target tracking analysis and pedestrian abnormal behavior analysis. Multi-target tracking analysis tracks and analyzes multiple pedestrians at intersections. Pedestrian abnormal behavior analysis analyzes and warns pedestrians of wandering or sudden changes in behavior.

[0039] The goal of establishing the multi-objective joint optimization model in step 4 is to balance the system's average task processing delay, average vehicle energy consumption, perception accuracy, and perception range.

[0040] In step 4, deep reinforcement learning optimizes the target information after association and fusion in the intersection area.

[0041] In the present invention, general AI cameras and millimeter-wave radars are selected as the main equipment for roadside target perception; the hardware and software upgrades and optimizations are performed on the camera and radar modules respectively; after the perception equipment is upgraded, the perception information of other roadside perception equipment mainly based on millimeter-wave radar is combined, and its detection results are correlated and fused with the detection results of the AI ​​camera also on the roadside using the KM algorithm to obtain the fused and optimized roadside target information and scene analysis; for the obtained roadside target information, a multi-objective joint optimization model is established, and the policy gradient descent algorithm is used to update the neural network parameters to obtain the optimal behavior strategy of deep reinforcement learning, so as to obtain the optimal roadside target information required for vehicle-road collaboration; the obtained optimization information of the complex environment of the intersection is transmitted to the vehicle, the early warning unit and the cloud through the RSU.

[0042] In the present invention, general AI cameras and millimeter-wave radars are selected as the main equipment for roadside target perception, and the general AI cameras are upgraded and optimized with corresponding software and hardware to meet the specific scenarios of abnormal pedestrian behavior detection, customized coverage of collision risk areas at non-signal-controlled intersections, and all-weather application requirements. In addition, they still have certain advantages in terms of construction cost and popularity. The comparison of mainstream target perception devices is shown in the following table:

[0043]

[0044] In terms of hardware upgrade and optimization: domestic chips are preferred for the design of camera built-in controllers and edge controllers to reduce hardware costs, which has a positive role in promoting the wide application of the project across the country; for some application scenarios, starlight-level AI cameras are used to improve the effectiveness of target perception in severe weather conditions; with the help of algorithm optimization methods, the camera module and radar module have the advantages of lightweight and miniaturization, and are directly embedded in the roadside equipment for use, which improves the level of integration of roadside equipment and makes the application more flexible; in terms of software upgrade and optimization: based on the YOLOv5 target detection framework, MobileNetV2 is used to replace the backbone network of the camera module, and the PAN structure is replaced by the PAN structure. The ordinary convolution of the structure is replaced with a depth-wise separable convolution, which greatly reduces the number of parameters of the network model and supports the lightweight and miniaturized design of the camera module; the attention mechanism is introduced to strengthen the ability of the network structure to extract key features from feature information, enhance model performance, and improve the algorithm speed and target classification accuracy of the camera to meet the classification requirements for pedestrians, ordinary motor vehicles, and large motor vehicle traffic targets in some scenarios; among them, pedestrian traffic targets include walking pedestrians and non-motor vehicle cyclists; for application scenarios that only need to perceive moving targets, millimeter-wave radar is used to make up for the shortcomings of general AI cameras in bad weather conditions and improve the perception accuracy of roadside targets; for pedestrian crossing safety warnings, In scenarios where speed limit reminders need to be issued based on specific classification targets, the optimized and upgraded AI camera can be used to effectively distinguish between pedestrians and vehicles, and even analyze the specific behavior of pedestrians, so as to implement the project construction purpose of "refined scenarios corresponding to refined equipment". The perception information of other roadside perception devices, mainly millimeter-wave radars, is combined, and its detection results are correlated and fused with the detection results of AI cameras on the roadside using the KM algorithm to obtain the fused and optimized roadside target information. On the one hand, it is conducive to improving the system's credibility in safety warnings between traffic targets in different directions on the roadside; on the other hand, it is conducive to the analysis of pedestrian behavior at key intersections or sections of national and provincial roads. It can effectively apply multi-target tracking and analysis, pedestrian abnormal behavior analysis such as wandering and behavioral mutation to meet more refined scenario applications, establish a multi-objective joint optimization model with the goal of balancing the system's average task processing delay, average vehicle energy consumption, perception accuracy and perception range, and use the policy gradient descent algorithm to update the neural network parameters to obtain the optimal behavior strategy for deep reinforcement learning, and further optimize the target information after correlation and fusion in the intersection area to obtain the optimal roadside target information required for vehicle-road collaboration, and transmit it to vehicles, early warning units and the cloud through RSU, which can meet the system's perception needs for national and provincial road environments while reducing information transmission delays and resource waste.

[0045] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-source perception fusion and target information optimization method under complex intersection environments, characterized in that: The following steps are involved: Step 1: Determine the perception equipment: Select general AI cameras and millimeter-wave radars as the main equipment for roadside object perception; Step 2: Upgrade the sensing equipment: perform hardware upgrade and software optimization on the camera and radar module respectively; Step 3: Multi-source data fusion: After the perception equipment is upgraded, it combines the perception information of other roadside perception equipment, mainly millimeter-wave radar, and uses the KM algorithm to correlate and fuse the detection results of the AI ​​camera on the roadside to obtain the fused and optimized roadside target information and scene analysis; Step 4: Information optimization: For the roadside target information obtained, a multi-objective joint optimization model is established, and the policy gradient descent algorithm is used to update the neural network parameters to obtain the optimal behavior strategy of deep reinforcement learning to obtain the optimal roadside target information required for vehicle-road collaboration; Step 5: Information transmission: The optimized information of the complex environment of the intersection is transmitted to the vehicle, early warning unit and the cloud through the RSU.

2. The method for multi-source perception fusion and target information optimization under complex intersection environment according to claim 1 is characterized by: The hardware upgrade and optimization in step 2 specifically involves selecting chips for the design of camera built-in controllers and edge controllers. For some application scenarios, starlight-level AI cameras are used to improve the effectiveness of target perception in severe weather conditions. With the help of algorithm optimization, the camera module and radar module have the advantages of lightweight and miniaturization, and can be directly embedded in the roadside equipment for use, thereby improving the level of integration of the roadside equipment.

3. The method for multi-source perception fusion and target information optimization under complex intersection environment according to claim 1 is characterized by: The software upgrade optimization in the step 2 is specifically based on the YOLOv5 target detection framework, using MobileNetV2 to replace the backbone network of the camera module, and replacing the ordinary convolution of the PAN structure with a deep separable convolution, which greatly reduces the number of parameters of the network model and supports the lightweight and miniaturized design of the camera module; introducing the attention mechanism, strengthening the ability of the network structure to extract key features in feature information, enhancing the model performance, and improving the algorithm speed and target classification accuracy of the camera. The YOLOv5 directly detects and locates the target from the original image, and the MobileNetV2 is composed of a series of stacked InvertedResidualBlocks, each of which contains the following three layers: a 1x1 dimensionality-enhancing layer that increases the number of output channels to expand the feature map, a deep separable convolution layer: reduces the computational cost by applying spatial deep separable convolutions, and the layer contains 3x3 deep separable convolutions and batch normalization operations, and a linear bottleneck layer: this is a 1x1 convolution layer that reduces the number of output channels and adds a linear activation function.

4. The method for multi-source perception fusion and target information optimization under complex intersection environment according to claim 1 is characterized by: The specific steps of using the KM algorithm in step 3 to associate and fuse the detection results of the AI ​​camera also on the roadside are as follows: S1, image initialization: initialize the weights of all edges in the graph perceived by the perception device to 0, and construct an initial staggered tree T = (X, Y), where X and Y represent two vertex sets respectively; S2, determine equal subgraphs: count each vertex x of T in the image initialized in the perception device, if there is a line that satisfies lx + ly = W, where lx is the sum of the weights of all the tree edges entering x in T, and ly is the sum of the weights of all the tree edges exiting x in T. S3. Find unmatched nodes: Find unmatched nodes in the equal subgraph and mark them as unmatched nodes; S4. Find augmented paths: If there are unmatched nodes, select a starting node from them and find a path consisting of edges that alternate with them, so that the last edge of the path is an edge associated with the unmatched node; S5. Modify the label: For each vertex x in the equal subgraph, select a vertex y associated with it from the matched nodes and calculate d=min(l x +l y -w), where w is the weight of the edge between x and y. To make the augmenting path better, subtract d from the weight of the incoming tree edge of all x nodes in T, and add d to the weight of the outgoing tree edge of all y nodes in T; S6. Update the interleaved tree: adjust all edges in the augmented path and all non-tree edges in T to obtain a new interleaved tree.

5. The method for multi-source perception fusion and target information optimization under complex intersection environment according to claim 1 is characterized by: The scene analysis in step three includes pedestrian behavior analysis, vehicle behavior analysis and vehicle-pedestrian differentiation. The pedestrian behavior analysis specifically analyzes the pedestrian behavior at the intersection. The vehicle behavior analysis analyzes the vehicle behavior at the intersection in terms of acceleration, deceleration, lane change, direction change and parking. The vehicle-pedestrian differentiation distinguishes between vehicles and pedestrians at the intersection. The pedestrian behavior analysis is provided with multi-target tracking analysis and pedestrian abnormal behavior analysis. The multi-target tracking analysis tracks and analyzes multiple pedestrians at the intersection. The pedestrian abnormal behavior analysis analyzes and warns pedestrians of wandering or sudden changes in behavior.

6. The multi-source perception fusion and target information optimization method under complex intersection environment according to claim 1 is characterized by: The goal of establishing the multi-objective joint optimization model in step 4 is to balance the system's average task processing delay, average vehicle energy consumption, perception accuracy and perception range.

7. The method for multi-source perception fusion and target information optimization under complex intersection environment according to claim 1 is characterized by: In the step 4, deep reinforcement learning optimizes the target information after the intersection area is associated and fused.