Crossroad traffic flow monitoring, management and optimization methods of campus CPS system

Abstract

The present invention discloses a method for monitoring, managing and optimizing traffic flow at an intersection of a campus CPS system, wherein the monitoring and management methods include: S1, dividing the campus intersection into four sub-area systems; S2, each sub-area system includes a camera perception module, an information processing module and a CPS calculation module; S3, deploying the trained SAM‑s target detection model into the CPS calculation module to locate the vehicle and determine the traffic flow; S4, judging whether a campus traffic accident has occurred; S5, when the traffic flow at any two intersections is greater than k_1, a one-way green light is used for N seconds, and when the traffic flow at any intersection is greater than x_1, additional personnel are dispatched to manage the intersection; wherein the optimization method includes: S6, using the monitoring and management methods as described above at the school gate and all intersections, setting up graph nodes and forming a traffic flow network, and using an improved graph neural network to regulate them; purpose: to achieve real-time and regional level traffic optimization and regulation, and alleviate regional congestion during campus peak hours.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation technology, and in particular to a method for monitoring, managing and optimizing traffic flow at an intersection of a campus CPS system. Background Art

[0002] In the traditional traffic management methods on campus, there is no real-time and accurate monitoring and statistics of traffic flow, lack of data support, and it is difficult to make efficient intersection adjustments. Most of the time and fixed points rely on teachers to observe, adjust and manage on site. Therefore, the traditional traffic management methods of the campus are inefficient, labor-intensive for teachers, and not intelligent enough.

[0003] In recent years, with the development of the Internet of Things and intelligent transportation, some campus traffic flow monitoring methods based on computer vision algorithms have emerged. However, most of these methods directly use a single camera to count the traffic flow at each intersection on campus, which makes it difficult to achieve regional-level traffic monitoring management optimization.

[0004] In particular, traffic safety is an important part of campus safety. In the management of intersections on campus, due to the irregularity of students and faculty, the traditional timed traffic light control plus timed and fixed-point manual management is directly adopted, or a single camera is used for traffic statistics and monitoring. These two traffic flow monitoring and management methods are not effective and it is still difficult to solve the traffic flow monitoring and management of campus intersections during peak hours after classes.

[0005] In order to cope with the regional congestion during the peak hours of concentrated get out of class dismissal, there is an urgent need for a traffic flow monitoring and management method for campus intersections to achieve real-time and regional-level traffic optimization and control at campus intersections. Summary of the invention

[0006] The purpose of the present invention is to overcome the above-mentioned defects in the prior art, and to provide a method for monitoring, managing and optimizing traffic flow at intersections of a campus CPS system, so as to realize real-time and regional-level traffic optimization and control at intersections within the campus, by monitoring the vehicle flow of the intersection sub-area system, automatically judging whether a campus traffic accident occurs within the sub-area system, setting a threshold k_1 to avoid excessive one-way intersection flow in the sub-area system, and setting a threshold x_1 to timely conduct manual on-site intersection flow diversion, as well as control means through a series of traffic flow networks formed by connecting the graph nodes at the school gate and all intersections, thereby effectively alleviating the regional congestion problem at campus intersections caused by the peak period of school dismissal.

[0007] To achieve the above object, the present invention is implemented through the following two aspects:

[0008] In a first aspect, the present invention provides a method for monitoring and managing traffic flow at an intersection of a campus CPS system, comprising the following steps:

[0009] S1. The campus intersection is divided into four sub-area systems according to the four directions, and the vehicle relationship, pedestrian relationship, intersection relationship and related physical relationship of the four sub-area systems are formed into information connection relationships, and the campus CPS system is constructed through the information connection relationship of the four sub-area systems;

[0010] S2, each sub-area system includes a camera perception module, an information processing module and a CPS calculation module. The camera perception module obtains the real-time image of the intersection of the sub-area system at a set time, and the information processing module performs image preprocessing on the real-time image of the intersection and transmits it to the CPS calculation module;

[0011] S3, create a campus traffic data set and train the SAM-s target detection model, deploy the trained SAM-s target detection model to the CPS computing module, transfer the preprocessed image of step S2 to the SAM-s target detection model for calculation and obtain the segmentation result map, use the softmax algorithm in the segmentation result map to calculate the probabilities of pedestrians and vehicles respectively, and obtain the segmentation result map that distinguishes pedestrians and vehicles according to different colors, calculate the number of pedestrians and vehicles respectively according to the segmentation result maps of different colors, and use the feature pyramid structure to extract different image feature details, locate the vehicles in the sub-area system and determine the traffic volume;

[0012] S4, according to the different colors of the segmentation result map in step S3, check whether there is a part where vehicles and pedestrians overlap to determine whether a campus traffic accident occurs;

[0013] S5. According to the number of pedestrians and vehicles calculated by the CPS calculation module in step S3, different quantity weights are set for pedestrians and vehicles respectively and the traffic flow at the intersection of each sub-area system is generated. The traffic flow at the intersection of the four sub-area systems is compared, and thresholds k_1 and x_1 are set respectively, wherein the threshold k_1 is the difference in traffic flow at the intersection between any two sub-area systems. When the difference in traffic flow at the intersection between any two sub-area systems is greater than k_1, one-way green light control is performed for N seconds according to the direction of maximum traffic flow, wherein N is a natural number greater than 1, and x_1 is the maximum traffic flow limit at the intersection of the sub-area system. When the traffic flow value at the intersection of any sub-area system is greater than x_1, it is uploaded to the dispatch center through the network, and additional personnel are dispatched to perform on-site intersection traffic management.

[0014] Preferably, in step S2, the camera sensing module includes a plurality of camera sensing devices arranged at different positions of the sub-area system.

[0015] Preferably, in step S2, the real-time image of the intersection of the camera perception module is transmitted to the information processing module through the 5G network, and the pre-processed image of the information processing module is transmitted to the SAM-s target detection model of the CPS calculation module through the 5G network.

[0016] Preferably, in step S2, the specific steps of image preprocessing of the information processing module are as follows: performing image denoising operation on the received real-time image of the intersection, wherein the noise interference in the real-time image is removed by a median filtering method.

[0017] Preferably, in step S3, the specific steps of preparing a campus traffic data set and training the SAM-s target detection model are as follows:

[0018] S310, obtaining the road videos and road pictures of the campus intersection that are actually collected and include vehicle targets and pedestrian targets, and making them into a campus traffic dataset;

[0019] S311, use labelme software to perform smearing and labeling operations on vehicle targets and pedestrian targets in the campus traffic dataset;

[0020] S312, the number of convolutional layers of the vgg16 network of the SAM-s target detection model is set to 16, the number of categories is 2, the batch size is 4, and the pixel-level softmax cross entropy loss function is used;

[0021] S313. The training data of step S311 is applied to the vgg16 network of the SAM-s target detection model to obtain a trained SAM-s target detection model.

[0022] Preferably, in step S3, the preprocessed image of step S2 is transferred to the SAM-s target detection model for calculation and the specific calculation steps for obtaining the segmentation result map are as follows: the preprocessed image is input into the SAM-s target detection model, and passes through the encoding channel composed of the vgg16 network of the SAM-s target detection model. The 16 convolution blocks gradually reduce the size of the feature map of the preprocessed image, and increase the channels, and extract semantic features of different scales. Then, after passing through the decoder, the decoder part first upsamples the encoder output, and then combines the upsampled features with the encoder features of the corresponding level to achieve cross-scale feature fusion. Finally, the convolution layer is continued to extract the fused features. The number of convolution kernels in the last layer corresponds to the number of categories, and a segmentation result map consistent with the input size is output.

[0023] Preferably, in step S3, the softmax algorithm is used to calculate the probabilities of pedestrians and vehicles in the segmentation result graph, and the probability calculation formula for pedestrians and vehicles is:

[0024]

[0025] Among them, c is the vehicle type and pedestrian category index, x is the sample feature vector of the target detected in the input current picture, is the weight proportion vector of category c, is the weight vector of all categories, b c is the bias term for category c, b j is the bias term for all categories, and P(c|x) is the probability that the region image is of category c.

[0026] Preferably, in step S3, after extracting different image feature details using a feature pyramid structure, target detection frames are generated for different vehicles, and the number of vehicle target detection frames is calculated to accurately locate the vehicles in the sub-area system and determine the traffic volume.

[0027] Preferably, in step S5, when the difference in traffic flow at the intersection of any two sub-area systems is less than or equal to k_1, the intersections of the sub-area systems perform green light control for M seconds in turn, where M is a natural number greater than 1.

[0028] In a second aspect, the present invention provides a method for optimizing traffic flow at an intersection of a campus CPS system, further comprising the following steps:

[0029] S6. The school gate and all the crossroads in the campus adopt the crossroad traffic flow monitoring and management method of the campus CPS system as described in the first aspect above. At the same time, graph nodes are set up at the school gate and all the crossroads to form a series of traffic flow networks, and a spatiotemporal fusion graph neural network based on deep learning is used to regulate the series of traffic flow networks. The graph nodes and real-time images of the intersections at the school gate and all the crossroads are input into the spatiotemporal fusion graph neural network based on deep learning. The spatiotemporal fusion graph neural network based on deep learning adopts a gated TCN module, uses Silu function to process spatial information, and uses sigmoid function to process practical information. Finally, the intersection flow information of the graph node at each moment is integrated with the intersection road space information between the graph nodes. The calculation formula is:

[0030] H=Silu(W1·X+b1)⊙sigmoid(W2·X+b2)

[0031] Among them, W1 and W2 are learning weights, b1 and b2 are learning biases, X is input data, and H is output data. At the same time, the intersection flow information of the input graph node passes through the graph convolution module parallel to the gated TCN module, and the intersection flow information of the graph node is integrated with the real-time observation information to realize the prediction of the traffic flow at each intersection in the next time period, and combine the intersection flow management of each intersection to prevent peak congestion at local intersections.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. The present invention obtains a segmentation result map through an improved SAM-s target detection model based on deep learning. The softmax algorithm is used in the segmentation result map to calculate the probabilities of pedestrians and vehicles respectively, which can distinguish and detect pedestrians and vehicles more quickly and accurately, and ensure the real-time and accuracy of subsequent vehicle positioning, traffic flow monitoring and sub-area system intersection traffic management.

[0034] 2. The present invention uses an improved spatiotemporal fusion graph neural network based on deep learning to effectively regulate the series traffic flow network formed by the graph nodes at the school gate and all intersections. By obtaining the intersection traffic information of each graph node, the deep learning method is used to predict future congestion, which is more accurate than traditional methods. It can also avoid peak congestion at local intersections through early regulation.

[0035] 3. A method for monitoring, managing and optimizing intersection traffic flow of a campus CPS system of the present invention, on the first hand, realizes real-time and regional-level traffic optimization and control of intersections in the campus by monitoring the vehicle flow of the intersection sub-area system, automatically determining whether a campus traffic accident occurs in the sub-area system, setting a threshold k_1 to avoid excessive one-way intersection traffic in the sub-area system, and setting a threshold x_1 to timely conduct manual on-site intersection traffic diversion, while greatly improving the vehicle detection accuracy and speed. On the second hand, the series traffic flow network formed by the graph nodes at the school gate and all intersections is regulated by an improved spatiotemporal fusion graph neural network based on deep learning, so as to realize real-time and regional-level traffic optimization and control of the entire intersection in the campus. Both the first and second aspects effectively alleviate the regional congestion problem at campus intersections caused by the peak period of school dismissal. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1 It is a flow diagram of a method for monitoring, managing and optimizing crossroad traffic flow in a campus CPS system provided by the present invention. DETAILED DESCRIPTION

[0038] The technical scheme in this embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in this embodiment of the present invention. Obviously, the described embodiment is one embodiment of the present invention, not all embodiments of the present invention. Based on this embodiment of the present invention, all other embodiments of the present invention obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] Combination Figure 1 As shown, an embodiment of the present invention provides a method for monitoring, managing and optimizing crossroad traffic flow of a campus CPS system, wherein the following steps S1-S5 are a method for monitoring and managing crossroad traffic flow of a campus CPS system, and the following steps S1-S6 are a method for optimizing crossroad traffic flow of a campus CPS system, and steps S1-S6 are specifically as follows:

[0040] S1. The campus intersection is divided into four sub-area systems according to the four directions, and the vehicle relationship, pedestrian relationship, intersection relationship and related physical relationship of the four sub-area systems are connected to form an information connection relationship, and the campus CPS system is built through the information connection relationship of the four sub-area systems (the full name of the campus CPS system is Campus Cyber-Physical Systems. The campus CPS system is a multi-dimensional complex system that integrates computing, network and physical environment. Through the organic integration and deep collaboration of 3C (Computation, Communication, Control) technologies, it realizes real-time perception, dynamic control and information services of large-scale engineering systems).

[0041] S2. Each sub-area system includes a camera perception module, an information processing module and a CPS calculation module. The camera perception module obtains the real-time image of the intersection of the sub-area system at a set time (for example, 5 seconds). The information processing module performs image preprocessing on the real-time image of the intersection and transmits it to the CPS calculation module.

[0042] The camera sensing module includes a number of camera sensing devices arranged at different positions of the sub-area system. The distribution of different positions is based on the previous vehicle flow and pedestrian trends at the intersection. Through a number of camera sensing devices, real-time images of intersections at different positions are obtained, so that the road flow information of the sub-area system can be fully obtained.

[0043] The real-time images of the intersection from the camera perception module are transmitted to the information processing module through the 5G network, and the pre-processed images from the information processing module are transmitted to the SAM-s target detection model of the CPS computing module through the 5G network.

[0044] Among them, the specific steps of image preprocessing in the information processing module are as follows: perform image denoising operation on the received real-time picture of the intersection, wherein the median filtering method is used to remove noise interference in the real-time picture, thereby reducing the number of pixels that need to be processed and analyzed by the subsequent target detection network, thereby reducing the amount of calculation and improving processing efficiency, and improving the accuracy of subsequent processing. After the information processing module completes the image preprocessing, the processed picture is transmitted to the CPS computing module via the 5G network.

[0045] S3. Prepare a campus traffic data set and train the SAM-s target detection model (the full name of the SAM-s target detection model is Segment Anything Model-s target detection model), deploy the trained SAM-s target detection model to the CPS computing module, transfer the pre-processed image of step S2 to the SAM-s target detection model for calculation and obtain the segmentation result map, use the softmax algorithm in the segmentation result map to calculate the probabilities of pedestrians and vehicles respectively, and obtain the segmentation result map that distinguishes pedestrians and vehicles according to different colors, and calculate the number of pedestrians and vehicles respectively according to the segmentation result maps of different colors, and use the feature pyramid structure of the SAM-s target detection model to extract different image feature details. The advantage of this processing is that, on the basis of retaining the original size of the image, the image features are extracted to make the image segmentation more accurate, so as to accurately locate all vehicles in the sub-area system and determine the traffic volume;

[0046] The specific steps for creating a campus traffic dataset and training the SAM-s target detection model are as follows:

[0047] S310, obtaining the road video (using real video) and road pictures (using small batches of high-definition pictures) of the campus intersection that are actually collected and include vehicle targets and pedestrian targets, and making them into a campus traffic data set. The advantage of using the actually collected campus traffic data set is that it is easier for the SAM-s target detection model to identify and distinguish campus vehicles and campus pedestrians;

[0048] S311, use labelme software to perform smearing and labeling operations on vehicle targets and pedestrian targets in the campus traffic dataset;

[0049] S312, the number of convolutional layers of the vgg16 network of the SAM-s target detection model is set to 16, the number of categories is 2, the batch size is 4, and the pixel-level softmax cross entropy loss function is used;

[0050] S313, the training data of step S311 is used for the vgg16 network of the SAM-s target detection model, so as to obtain a trained SAM-s target detection model;

[0051] The specific calculation steps of transferring the preprocessed image of step S2 to the SAM-s target detection model for calculation and obtaining the segmentation result map are as follows: input the preprocessed image into the SAM-s target detection model, pass through the encoding channel composed of the vgg16 network of the SAM-s target detection model, and use 16 convolution blocks to gradually reduce the feature map size of the preprocessed image and increase the channels, extracting semantic features of different scales, and then pass through the decoder. The decoder part first upsamples the encoder output, and then combines the upsampled features with the encoder features of the corresponding level to achieve cross-scale feature fusion, and finally continue to extract fusion features through the convolution layer. The number of convolution kernels in the last layer corresponds to the number of categories, and the segmentation result map consistent with the input size is output. Furthermore, a residual connection is added to the volume base layer of the generated mask before decoding, so that the generated segmented image is more accurate and the network runs faster.

[0052] Among them, the softmax algorithm is used in the segmentation result map to calculate the probability of pedestrians and vehicles respectively. The probability calculation formula for pedestrians and vehicles is:

[0053]

[0054] Among them, c is the vehicle type and pedestrian category index, x is the sample feature vector of the target detected in the input current picture, is the weight proportion vector of category c, is the weight vector of all categories, b c is the bias term for category c, b j is the bias term for all categories, and P(c|x) is the probability that the region image is of category c.

[0055] Among them, after using the feature pyramid structure to extract different image feature details, target detection frames are generated for different vehicles, and the number of vehicle target detection frames is automatically calculated to achieve accurate positioning of all vehicles in the sub-area system and determine the traffic flow.

[0056] S4. The result graph is obtained according to the different colors of the segmentation in step S3. Whether there is any overlap between vehicles and pedestrians in the dense area segmented by color can be used to determine whether a campus traffic accident occurs.

[0057] S5. According to the number of pedestrians and vehicles calculated by the CPS calculation module in step S3, different weights are set for pedestrians and vehicles respectively, and the traffic flow of each sub-region system intersection is generated (for example, the weight of each car is set to 1, the weight of electric car is set to 0.6, and the weight of pedestrian is set to 0.3 through the calculated number of vehicles. The sub-region system intersection traffic flow = the number of cars * 1 + the number of electric cars * 0.6 + the number of pedestrians * 0.3, and the sub-region system intersection traffic flow is obtained by weighting;), compare the sizes of the intersection traffic flows of the four sub-region systems, and set thresholds k_1 and x_1 respectively, where the threshold k_1 is the difference in traffic flow at the intersection of any two sub-region systems. When the difference in traffic flow at the intersection of any two sub-region systems is greater than k_1, proceed according to the direction of maximum traffic flow. The green light for one-way traffic is controlled for N seconds, where N is a natural number greater than 1 (specifically, N can be preferably 20 seconds, and 20 seconds is the optimal green light control time). When the difference in traffic flow at the intersection of any two sub-area systems is less than or equal to k_1, the intersections of each sub-area system take turns to control green light traffic for M seconds, where M is a natural number greater than 1 (specifically, M can be preferably 30 seconds); x_1 is the maximum traffic flow limit at the intersection of the sub-area system. When the traffic flow value at the intersection of any sub-area system is greater than x_1, the sub-area system is deemed to have abnormal traffic behavior, such as vehicle collision, etc. At this time, the information is uploaded to the dispatch center through the network in a timely manner, and personnel are immediately dispatched to conduct on-site intersection traffic management, handle abnormal traffic conditions, and manually evacuate people and vehicles in a timely manner.

[0058] S6. The school gate and all the crossroads in the campus adopt the crossroad traffic flow monitoring and management method of the campus CPS system as described in steps S1-S5 above. At the same time, graph nodes are set up at the school gate and all the crossroads to form a series of traffic flow networks, and an improved spatiotemporal fusion graph neural network based on deep learning is used to regulate the series of traffic flow networks. The graph nodes and real-time images of the intersections at the school gate and all the crossroads are input into the improved spatiotemporal fusion graph neural network based on deep learning. The improved spatiotemporal fusion graph neural network based on deep learning adopts a gated TCN module, uses Silu function to process spatial information, and uses sigmoid function to process practical information. Finally, the intersection flow information of the graph node at each moment is integrated with the intersection road space information between the graph nodes. The calculation formula is:

[0059] H=Silu(W1·X+b1)⊙sigmoid(W2·X+b2)

[0060] Among them, W1 and W2 are learning weights, which can be updated according to the spatial information of the intersection roads between the graph nodes, B1 and b2 are learning biases, which can be updated according to the spatial information of the intersection roads between the graph nodes, X is the input data, and H is the output data. At the same time, the intersection flow information of the input graph node passes through the graph convolution module in parallel with the gated TCN module, and the intersection flow information of the graph node is integrated with the real-time observation information to realize the prediction of the traffic flow at each intersection in the next time period, and the intersection flow management of each intersection is combined to realize the real-time and overall regional level traffic optimization and control at the school gate and all intersections to prevent peak congestion at local intersections. (That is, to prevent peak congestion at individual intersections, and fail to obtain timely evacuation of the previous or next intersection)

[0061] Furthermore, this improved deep learning-based spatiotemporal fusion graph neural network combines the gated TCN module with the parallel processing of the graph convolutional neural network, so that the processing of relatively few graph nodes is faster and more accurate.

[0062] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. 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 method for monitoring and managing traffic flow at an intersection of a campus CPS system, characterized in that: The following steps are involved: S1. The campus intersection is divided into four sub-area systems according to the four directions, and the vehicle relationship, pedestrian relationship, intersection relationship and related physical relationship of the four sub-area systems are formed into information connection relationships, and the campus CPS system is constructed through the information connection relationship of the four sub-area systems; S2, each sub-area system includes a camera perception module, an information processing module and a CPS calculation module. The camera perception module obtains the real-time image of the intersection of the sub-area system at a set time, and the information processing module performs image preprocessing on the real-time image of the intersection and transmits it to the CPS calculation module; S3, create a campus traffic data set and train the SAM-s target detection model, deploy the trained SAM-s target detection model to the CPS computing module, transfer the preprocessed image of step S2 to the SAM-s target detection model for calculation and obtain the segmentation result map, use the softmax algorithm in the segmentation result map to calculate the probabilities of pedestrians and vehicles respectively, and obtain the segmentation result map that distinguishes pedestrians and vehicles according to different colors, calculate the number of pedestrians and vehicles respectively according to the segmentation result maps of different colors, and use the feature pyramid structure to extract different image feature details, locate the vehicles in the sub-area system and determine the traffic volume; In step S3, the specific steps of making a campus traffic dataset and training the SAM-s target detection model are as follows: S310, obtaining the road videos and road pictures of the campus intersection that are actually collected and include vehicle targets and pedestrian targets, and making them into a campus traffic dataset; S311, use labelme software to perform smearing and labeling operations on vehicle targets and pedestrian targets in the campus traffic dataset; S312, the number of convolutional layers of the vgg16 network of the SAM-s target detection model is set to 16, the number of categories is 2, the batch size is 4, and the pixel-level softmax cross entropy loss function is used; S313, the training data of step S311 is used for the vgg16 network of the SAM-s target detection model, so as to obtain a trained SAM-s target detection model; In step S3, the specific calculation steps of transferring the preprocessed image of step S2 to the SAM-s target detection model for calculation and obtaining the segmentation result map are as follows: the preprocessed image is input into the SAM-s target detection model, and passes through the encoding channel composed of the vgg16 network of the SAM-s target detection model. The feature map size of the preprocessed image is gradually reduced by 16 convolution blocks, and the channels are enlarged to extract semantic features of different scales. Then, it passes through the decoder. The decoder part first upsamples the encoder output, and then combines the upsampled features with the encoder features of the corresponding level to achieve cross-scale feature fusion. Finally, the convolution layer is continued to extract the fusion features. The number of convolution kernels in the last layer corresponds to the number of categories, and the segmentation result map consistent with the input size is output; In step S3, the softmax algorithm is used to calculate the probabilities of pedestrians and vehicles in the segmentation result map. The probability calculation formula for pedestrians and vehicles is: Among them, c is the vehicle type and pedestrian category index, x is the sample feature vector of the target detected in the input current picture, is the weight proportion vector of category c, is the weight vector of all categories, b c is the bias term for category c, b j is the bias term for all categories, P(c|x) is the probability that the region image is of category c; S4, according to the different colors of the segmentation result map in step S3, check whether there is a part where vehicles and pedestrians overlap to determine whether a campus traffic accident occurs; S5. According to the number of pedestrians and vehicles calculated by the CPS calculation module in step S3, different quantity weights are set for pedestrians and vehicles respectively and the traffic flow at the intersection of each sub-area system is generated. The traffic flow at the intersection of the four sub-area systems is compared, and thresholds k_1 and x_1 are set respectively, wherein the threshold k_1 is the difference in traffic flow at the intersection between any two sub-area systems. When the difference in traffic flow at the intersection between any two sub-area systems is greater than k_1, one-way green light control is performed for N seconds according to the direction of maximum traffic flow, wherein N is a natural number greater than 1, and x_1 is the maximum traffic flow limit at the intersection of the sub-area system. When the traffic flow value at the intersection of any sub-area system is greater than x_1, it is uploaded to the dispatch center through the network, and additional personnel are dispatched to perform on-site intersection traffic management.

2. According to the method for monitoring and managing traffic flow at intersections of a campus CPS system in claim 1, it is characterized in that: In step S2, the camera sensing module includes a plurality of camera sensing devices arranged at different positions of the sub-area system.

3. The method for monitoring and managing traffic flow at an intersection of a campus CPS system according to claim 1 is characterized in that: In step S2, the real-time intersection image of the camera perception module is transmitted to the information processing module through the 5G network, and the pre-processed image of the information processing module is transmitted to the SAM-s target detection model of the CPS computing module through the 5G network.

4. The method for monitoring and managing traffic flow at an intersection of a campus CPS system according to claim 1 is characterized in that: In step S2, the specific steps of image preprocessing of the information processing module are as follows: performing image denoising operation on the received real-time image of the intersection, wherein the noise interference in the real-time image is removed by a median filtering method.

5. The method for monitoring and managing traffic flow at an intersection of a campus CPS system according to claim 1 is characterized in that: In step S3, after extracting different image feature details using the feature pyramid structure, target detection frames are generated for different vehicles, and the number of vehicle target detection frames is calculated to accurately locate the vehicles in the sub-area system and determine the traffic flow.

6. The method for monitoring and managing traffic flow at an intersection of a campus CPS system according to claim 1 is characterized in that: In step S5, when the difference in traffic volume at the intersections of any two sub-area systems is less than or equal to k_1, the intersections of the sub-area systems take turns to perform green light control for M seconds, where M is a natural number greater than 1.

7. A method for optimizing traffic flow at intersections of a campus CPS system, characterized in that: The following steps are also included: S6. The school gate and all the crossroads in the campus adopt the crossroad traffic flow monitoring and management method of the campus CPS system as described in any one of claims 1-6. At the same time, graph nodes are set up at the school gate and all the crossroads to form a series of traffic flow networks, and a spatiotemporal fusion graph neural network based on deep learning is used to regulate the series of traffic flow networks, wherein the graph nodes and real-time images of the intersections at the school gate and all the crossroads are input into the spatiotemporal fusion graph neural network based on deep learning. The spatiotemporal fusion graph neural network based on deep learning adopts a gated TCN module, uses Silu function to process spatial information, and uses sigmoid function to process time information. Finally, the intersection flow information of the graph node at each moment is integrated with the intersection road space information between the graph nodes. The calculation formula is: H=Silu(W1·X+b1)⊙sigmoid(W2·X+b2) Among them, W1 and W2 are learning weights, b1 and b2 are learning biases, X is input data, and H is output data. At the same time, the intersection flow information of the input graph node passes through the graph convolution module parallel to the gated TCN module, and the intersection flow information of the graph node is integrated with the real-time observation information to realize the prediction of the traffic flow at each intersection in the next time period, and combine the intersection flow management of each intersection to prevent peak congestion at local intersections.

Images