A multi-task-aware traffic control network method for high-speed roads

By using an improved CSPDarkNet architecture and a multi-task network model with a feature fusion layer, the computational complexity and energy efficiency issues of intelligent transportation on highways were solved, achieving real-time accuracy and real-time performance of perception, and alleviating traffic congestion.

CN118015564BActive Publication Date: 2024-11-26DONGGUAN UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410179917.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-02-17
Publication Date
2024-11-26
Estimated Expiration
2044-02-17

AI Technical Summary

Technical Problem

In high-speed sections, the computational complexity and energy efficiency issues of intelligent transportation lead to insufficient accuracy and real-time performance in real-time perception, which fails to effectively alleviate traffic congestion.

Method used

By adopting an improved CSPDarkNet architecture and a multi-task network model with a feature fusion layer, combined with lightweight technology, it is deployed on edge monitoring devices to achieve real-time vehicle trajectory guidance and traffic control.

Benefits of technology

It improves the network's data feature extraction and multi-task processing capabilities, ensuring the accuracy and real-time nature of real-time perception, and effectively alleviating traffic congestion.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118015564B_ABST
    Figure CN118015564B_ABST
Patent Text Reader

Abstract

The invention discloses a multi-task perception traffic control network method that can be used for a highway section, comprising the following steps: step S1: setting a plurality of highway section monitoring points; step S2: building a multi-task network model, the multi-task network model consisting of a shared encoder and three decoders; the shared encoder consists of a backbone network and a feature fusion layer, the backbone network adopts a CSPDarkNet architecture in which a C3f module is replaced by a CSP module, and the feature fusion layer includes an FPN architecture and an SPPF module; step S3: applying a multi-task learning loss function to the training of the multi-task network model; after the multi-task network model is trained, the NNCF algorithm is used to make the multi-task network model lightweight; step S4: deploying the multi-task network model in an edge monitoring device of the highway section; step S5: when congestion occurs on the highway section, the edge monitoring device guides the vehicle to change lanes in advance and enter a queue sequence; the above method can ensure that the edge monitoring device of the highway section has real accuracy and real-time performance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of traffic control networks for expressway sections, and specifically relates to a multi-task perception traffic control network method that can be used for expressway sections. Background Art

[0002] In recent years, the rapid economic development has brought about a surge in vehicles in daily life. At the same time, traffic congestion has also occurred on highways in various places. The prosperity of the transportation industry has also put forward higher requirements for road construction and traffic control. Against this background, intelligent transportation came into being.

[0003] However, there are the following problems when applying intelligent transportation on highways:

[0004] 1. Embedded devices used to calculate and identify information collected by image sensors in intelligent transportation need to consider factors such as computing overhead and resource usage. If two or more independent models are used, it will have a negative impact on the system's computational complexity and energy efficiency.

[0005] 2. Due to the existence of the first defect, the reasoning speed will decrease, making it difficult for edge devices deployed on highways to ensure the accuracy and real-time performance of real-time perception, ultimately making it impossible for intelligent transportation to alleviate traffic congestion on highways. Summary of the invention

[0006] The purpose of the present invention is to overcome the above-mentioned defects in the prior art and provide a multi-task perception traffic control network method that can be used in high-speed sections, which solves the following problems:

[0007] First, the multi-task network model of the present invention replaces the C3f module with the improved CSPDarkNet architecture of the CSP module as the backbone network, and the feature fusion layer is used for feature data extraction and fusion, and processes the drivable area segmentation, vehicle target detection segmentation and lane line detection segmentation tasks respectively, which greatly improves the network's ability to extract and fuse data features and multi-task processing capabilities, thereby reducing the computational complexity and improving the reasoning speed.

[0008] Second, the present invention directly applies the trained lightweight multi-task network model to the edge monitoring equipment of the highway section, which can ensure the real-time accuracy and real-time performance of real-time perception, and can be effectively applied to the vehicle trajectory guidance on the highway section, avoiding the butterfly effect caused by temporary lane changes, and effectively alleviating traffic congestion on the highway section.

[0009] To achieve the above object, the present invention provides a multi-task perception traffic control network method that can be used in high-speed road sections, which includes the following steps:

[0010] Step S1: acquiring real-time traffic flow video data by setting a number of monitoring points on the expressway;

[0011] Step S2: Building a multi-task network model, which is an image detection and segmentation network consisting of a shared encoder and three decoders;

[0012] The shared encoder consists of a backbone network and a feature fusion layer. The backbone network adopts the CSPDarkNet architecture in which the CSP module is replaced by the C3f module. The feature fusion layer includes the FPN architecture and the SPPF module integrated into the FPN architecture. The three decoders are the vehicle target detection decoder, the lane line detection decoder, and the drivable area segmentation decoder.

[0013] Among them, after the real-time traffic flow video data is input into the shared encoder, the backbone network extracts the feature data of different scales of the real-time traffic flow video data, processes the drivable area segmentation task in advance and outputs it to the drivable area segmentation decoder, and the feature fusion layer receives the feature data of different scales of the real-time traffic flow video data, processes the vehicle target detection segmentation and lane line detection segmentation tasks respectively, and outputs them to the vehicle target detection decoder and lane line detection decoder respectively;

[0014] Step S3: training the multi-task network model;

[0015] And construct a multi-task learning loss function, and apply the multi-task learning loss function to the training of the multi-task network model;

[0016] After the multi-task network model is trained, the NNCF algorithm is used to make the trained multi-task network model lightweight;

[0017] Step S4: deploying the trained lightweight multi-task network model in the edge monitoring equipment of the highway section, wherein vehicle target detection and lane line detection assist in searching the optimal control route, and drivable area segmentation determines the safety of vehicle distance on the highway section;

[0018] Step S5: The edge monitoring equipment on the highway section performs real-time route control. When congestion occurs on the highway section, the edge monitoring equipment sends the control route data to the Internet of Vehicles or map software in advance, and guides the vehicle to change lanes in advance and enter the queue sequence.

[0019] Preferably, in step S2, the convolution module of the CSPDarkNet architecture of the backbone network is composed of a Conv convolution layer, a BatchNorm batch normalization layer and an activation function; the convolution kernel of the Conv convolution layer is set to 3*3.

[0020] Preferably, in step S2, the C3f module is an improved module for improving the C2f module in YOLOv8. The C3f module divides the input into n branches and processes them through n Bottleneck modules respectively. At the same time, the 3*3 convolution kernel in the Bottleneck module is split into 3*1 and 1*3 convolution kernels for processing, and finally all branches are spliced ​​with the residual branch, wherein the parameter n is a parameter that can be adjusted manually.

[0021] Preferably, in step S2, the SPPF module of the feature fusion layer is used to expand the receptive field, generate and fuse feature information of different scales, the FPN architecture of the feature fusion layer performs feature fusion at different semantic levels, and the SPPF module and the FPN architecture are connected in series, and the two are connected in series so that the generated features ultimately contain information of multiple scales and multiple semantic levels.

[0022] Preferably, in step S2, the drivable area segmentation decoder and the lane line detection decoder are the same.

[0023] Preferably, the feature fusion layer also includes two layers of upsampling. The FPN architecture of the feature fusion layer performs an upsampling operation after receiving the feature map, cancels the convolution operation before the upsampling operation, uses the C3f module for processing after upsampling, and splices the feature maps output at different stages to complete the feature fusion.

[0024] Preferably, in step S2, the vehicle target detection decoder adopts a PAN structure, which is composed of a C3f module, a convolution layer, and a Concat layer. The FPN architecture of the feature fusion layer and the PAN structure form an FPN-PAN structure. The FPN-PAN structure realizes bidirectional transmission of shallow feature maps and deep feature maps, and finally outputs the results through several target detection heads connected to the vehicle target detection decoder.

[0025] Preferably, then, the output features of the FPN architecture of the feature fusion layer are directly transmitted to the lane detection decoder, and the drivable area segmentation decoder directly obtains features from the position between the back of the SPPF module of the feature fusion layer and the front of the FPN architecture.

[0026] Preferably, in step S3, the multi-task learning loss function L all The specific calculation formula is as follows:

[0027] L all =β1L det +β2L da-seg +β3L ll-seg

[0028] Among them, L detIt is represented as the vehicle target detection task, and β1 is represented as L det The corresponding weight coefficient, L da-seg It is represented as the drivable area segmentation task, and β2 is represented by L da-seg The corresponding weight coefficient, L ll-seg It is represented as the lane detection task, and β3 is represented as L ll-seg The corresponding weight coefficient;

[0029] Among them, the target detection task L det The specific calculation formula is as follows:

[0030] L det =α1L class +α2L obj +α3L box

[0031] Among them, L class It is represented as the focal loss function of the penalty term for classification, and α1 is represented as L class The corresponding weight coefficient, L obj It is expressed as the focal loss function for the predicted confidence, and α2 is expressed as L obj The corresponding weight coefficient, L box It is expressed as CIoU loss function, α3 is expressed as L box The corresponding weight coefficient.

[0032] Preferably, in step S3, the bdd100k dataset is used to train the multi-task network model;

[0033] In step S3, the NNCF algorithm is used to make the trained multi-task network model lightweight, and the lightweight optimization steps include:

[0034] Step S31: Establish a quantitative data set Dataset;

[0035] Step S32: Run the NNCF algorithm to optimize the multi-task network model;

[0036] Step S33: Use the serialization function in openvino to obtain the OpenVINO IR model, which is a lightweight model of the multi-task network model in different formats.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. The present invention is provided with a multi-task network model, in which the CSPDarkNet architecture adopted by the backbone network is integrated with the C3f module, and the FPN architecture adopted by the feature fusion layer is integrated with the SPPF module, which greatly improves the network's ability to extract fusion data features;

[0039] 2. According to the difficulty of the segmentation task, the drivable area segmentation task is connected before the FPN architecture, and the vehicle target detection segmentation and lane line detection are connected after the FPN architecture to further improve the accuracy of each task and the reasoning speed of the network;

[0040] 3. The multi-task learning loss function and NNCF algorithm are applied to the multi-task network model respectively to achieve high precision and fast reasoning speed of the multi-task network model. At the same time, the multi-task network model obtains the minimum precision loss after quantization, so that the multi-task network model can be easily deployed on the edge monitoring equipment of the highway section.

[0041] 4. In summary, after the multi-task network model is deployed on the edge monitoring equipment of the highway section, it can ensure the accuracy and real-time performance of real-time perception, effectively guide vehicles to change lanes, thereby avoiding the butterfly effect caused by frequent lane changes or serpentine lane changes during peak periods, and effectively alleviate traffic congestion. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] 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.

[0043] Figure 1 It is a module principle diagram of a backbone network of a multi-task perception flow control network method that can be used in a highway section provided by an embodiment of the present invention;

[0044] Figure 2 It is a module schematic diagram of a C2f module provided by the prior art;

[0045] Figure 3 It is a module schematic diagram of a C2f module of a multi-task perception traffic control network method that can be used in a highway section provided by an embodiment of the present invention;

[0046] Figure 4 It is a module principle diagram of an SPPF module of a multi-task perception flow control network method that can be used for a highway section provided by an embodiment of the present invention;

[0047] Figure 5It is a module principle diagram of a multi-task network model of a multi-task perception flow control network method that can be used for a highway section provided by an embodiment of the present invention;

[0048] Figure 6 It is a flow diagram of a multi-task-aware traffic control network method that can be used in expressway sections, provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0049] 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.

[0050] See also Figure 6 This embodiment provides a multi-task perception traffic control network method that can be used in high-speed sections, which includes the following steps:

[0051] Step S1: acquiring real-time traffic flow video data by setting a number of monitoring points on the expressway;

[0052] Step S2: Build a multi-task network model. The structure of the multi-task network model is as follows: Figure 5 As shown, the multi-task network model is an image detection and segmentation network consisting of a shared encoder and three decoders;

[0053] The shared encoder consists of a backbone network and a feature fusion layer. The backbone network adopts an improved CSPDarkNet architecture that replaces the CSP module with a C3f module. The feature fusion layer includes an FPN architecture and an SPPF module integrated into the FPN architecture. Under the conditions of three-task processing, the shared encoder can share information between different tasks and effectively establish connections between tasks. At the same time, the tasks can influence each other and act as a regular term to prevent overfitting of each task.

[0054] The three decoders are vehicle target detection decoder, lane line detection decoder and drivable area segmentation decoder;

[0055] Among them, after the real-time traffic flow video data is input into the shared encoder, the improved CSPDarkNet architecture of the backbone network extracts the feature data of different scales of the real-time traffic flow video data, processes the drivable area segmentation task in advance and outputs it to the drivable area segmentation decoder. Since the drivable area segmentation task is simple to process, processing this task in advance is beneficial to improving the reasoning speed of the network method. The feature fusion layer receives the feature data of different scales of the real-time traffic flow video data, processes the vehicle target detection segmentation and lane line detection segmentation tasks respectively, and outputs them to the vehicle target detection decoder and lane line detection decoder respectively. The feature fusion layer is beneficial to improving the accuracy of the vehicle target detection segmentation and lane line detection segmentation tasks;

[0056] The backbone network structure is as follows: Figure 1 As shown, the backbone network performs multi-scale feature data extraction, and the different scale features extracted by the backbone network output three task branches. The CSPDarkNet architecture of the backbone network can not only solve the problem of gradient duplication in the optimization process, but also provide feature transfer and reuse. After replacing the BottleneckCSP module with the C3f module, the CSPDarkNet architecture of the backbone network includes a convolution module and a C3f module.

[0057] Specifically, the convolution module of the CSPDarkNet architecture of the backbone network consists of a Conv convolution layer, a BatchNorm batch normalization layer and an activation function; the convolution layer uses a fixed 3*3 convolution kernel to slide on the feature map, and obtains new feature pixel values ​​by multiplying and adding corresponding pixel values; the batch normalization layer performs batch processing on the output feature map of the convolution layer, converting the distribution of the feature map into a normal distribution; the activation function is the SiLU function.

[0058] In addition, the structures of C2f module and C3f module are as follows: Figure 2 , 3 As shown in the figure, the C3f module is an improved module for the C2f module in YOLOv8. The C3f module divides the input into n branches and processes them through n Bottleneck modules respectively. It is designed with reference to the residual module and the idea of ​​ELAN, which can enable the network to obtain richer gradient flow information and improve the ability to integrate data features, further improving the network detection speed. At the same time, the 3*3 convolution kernel in the Bottlenec module is split into 3*1 and 1*3 convolution kernels for processing, effectively reducing the number of parameters. Finally, all branches are spliced ​​with the residual branch, where the parameter n is a parameter that can be adjusted manually.

[0059] In traffic scenes, the camera’s shooting distance is relatively far, and the detection targets in the image are usually at different scales. Therefore, the SPPF module improves the multi-task network model’s ability to extract multi-scale features for different detection targets, which can improve the detection accuracy of the network model. Specifically, the improved SPPF module of the feature fusion layer is as follows: Figure 4 As shown in the figure, it is used to expand the receptive field, generate and fuse feature information of different scales. The SPPF module enhances the learning ability of the model and can effectively avoid image distortion caused by cropping and scaling operations of the image area. At the same time, it solves the problem of convolutional neural network extracting repeated features related to the graph, greatly improves the speed of generating candidate boxes, and saves computing costs, thereby greatly improving the computing and reasoning speed of the network.

[0060] The FPN architecture of the feature fusion layer performs feature fusion at different semantic levels. The SPPF module and the FPN architecture are connected in series. After the two are connected in series, the generated features finally contain information of multiple scales and multiple semantic levels.

[0061] Among them, in step S2, the vehicle flow video data is generated by a shared encoder to generate a feature map (feature data of different scales is a feature map), and the feature map is respectively passed to the drivable area segmentation task, vehicle target detection segmentation and lane line detection segmentation task. The drivable area segmentation decoder and the lane line detection decoder are the same. The feature fusion layer also includes two layers of upsampling. The FPN architecture of the feature fusion layer performs an upsampling operation after receiving the feature map. The convolution operation is canceled before the upsampling operation. After upsampling, the C3f module is used for processing, and the feature maps output at different stages are spliced ​​to complete the feature fusion.

[0062] Among them, the vehicle target detection decoder adopts the PAN structure, which is composed of a C3f module, a convolution layer, and a Concat layer. The FPN architecture of the feature fusion layer and the PAN structure form an FPN-PAN structure. The FPN-PAN structure realizes the bidirectional transmission of shallow feature maps and deep feature maps. (The deep feature map has strong semantic features and weaker position information, and the shallow feature map consists of weaker semantic features and position information. The FPN-PAN structure can enrich the semantic features, reduce the loss of semantic information, and improve the accuracy. Then, the grid of each multi-scale feature map will be assigned to different aspect ratio anchor frames, and the detection head connected to each decoder will predict the position offset, aspect ratio, and the possibility of each category and the confidence of the prediction.) Finally, the results are output by several target detection heads connected to the vehicle target detection decoder; then, the output features of the FPN architecture of the feature fusion layer (resolution is [W / 8, H / 8, 256]) are directly transmitted to the lane detection decoder, and the drivable area segmentation decoder directly obtains features from the position between the back of the SPPF module of the feature fusion layer and the front of the FPN architecture.

[0063] Among them, the vehicle target detection decoder, lane line detection decoder and drivable area segmentation decoder all have upsampling layers, which can reduce a certain amount of calculation. The lane line detection segmentation task branch and the drivable area segmentation task branch both use C3f module feature extraction and three upsampling processes to restore the feature map of scale (W / 8, H / 8, 256) to the scale of (W, H, 2). Each pixel in the [W, H, 2] image has the probability of being a driving area or a lane line. (The driving area probability is the probability of the driving area relative to the background area outside the driving area, and the lane line probability is the probability of the lane line area relative to the background area outside the lane line area). Therefore, the above three decoders not only have high-precision output, but also have very fast inference speed. At the same time, upsampling using the neighboring method instead of deconvolution can greatly reduce computing resources.

[0064] Step S3: Train the multi-task network model.

[0065] Among them, after the multi-task network model is built, the required data set is produced. The data set should contain the data sets required for different tasks, and have data under different conditions, such as sunny days, cloudy days, foggy days, evenings and other factors, so as to improve the generalization of the network model. The data set is composed of the bdd100k data set. Bdd100k is one of the largest and most diverse traffic scene data sets to date. In the data set, road object detection is to mark 2D bounding boxes on 100,000 pictures of buses, traffic lights, traffic signs, people, bicycles, trucks, motorcycles, cars, trains and passengers; instance segmentation is used to explore pixel-level and rich instance-level annotations, with more than 10,000 related images; lane markings are multiple lane markings on 100,000 driving guide pictures. Therefore, using the bdd100k data set to train the multi-task network model can improve the generalization of the multi-task network model during training;

[0066] A multi-task learning loss function is constructed and applied to the training of the multi-task network model. The multi-task learning loss function is used to adjust the model parameters when training the multi-task network model to obtain better detection accuracy.

[0067] Among them, the multi-task learning loss function L all The specific calculation formula is as follows:

[0068] L all =β1L det +β2L da-seg +β3L ll-seg

[0069] Among them, L detIt is represented as the vehicle target detection task, and β1 is represented as L det The corresponding weight coefficient, L da-seg It is represented as the drivable area segmentation task, and β2 is represented by L da-seg The corresponding weight coefficient, L ll-seg It is represented as the lane detection task, and β3 is represented as L ll-seg The corresponding weight coefficient;

[0070] Among them, the target detection task L det The specific calculation formula is as follows:

[0071] L det =α1L class +α2L obj +α3L box

[0072] Among them, L class It is represented as the focal loss function of the penalty term for classification, and α1 is represented as L class The corresponding weight coefficient, L obj It is expressed as the focal loss function for the predicted confidence, and α2 is expressed as L obj Corresponding weight coefficients, focal loss loss function is used to reduce the loss function of simple tasks, allowing the network to focus more on difficult instances, L box It is expressed as CIoU loss function, α3 is expressed as L box Corresponding weight coefficients, the CIoU loss function can take into account the distance, overlap, scale similarity and aspect ratio between the predicted box and the true box.

[0073] For the drivable area segmentation task L da-seg and lane detection task L ll-seg , their structures are the same, and their loss functions both use the cross entropy loss function containing Logits, aiming to minimize the classification error between the network output pixels and the target; however, the lane line detection task L ll-seg The IoU loss function should also be added because it is particularly effective for predicting sparse lane lines.

[0074] In the above loss function α1α2α3β1β2β3 can be adjusted to balance all parts of the total loss.

[0075] After the multi-task network model is trained, the NNCF algorithm is used to make the trained multi-task network model lightweight;

[0076] Among them, the lightweight optimization steps include:

[0077] Step S31: Establish a quantitative data set Dataset;

[0078] Step S32: Run the NNCF algorithm to optimize the multi-task network model;

[0079] Step S33: Use the serialization function in openvino to obtain the OpenVINO IR model. The OpenVINO IR model is a lightweight model of a multi-task network model in different formats. The accuracy of the OpenVINO IR model has not decreased significantly compared to before optimization, ensuring that the model has high accuracy and real-time performance when deployed on embedded devices.

[0080] Step S4: deploy the trained lightweight multi-task network model in the edge monitoring equipment of the highway section, wherein vehicle target detection and lane line detection are used to assist in the search for the best control route, and the drivable area segmentation is used to determine the distance between vehicles on the highway section while ensuring the safety of the control route. The image detection and segmentation network of the multi-task network model can control the implementation of multiple tasks at the minimum deployment cost while ensuring the high accuracy and real-time performance of the network. The edge monitoring equipment of the highway section is a long straight road intersection monitoring camera or an intersection camera;

[0081] Step S5: After the multi-task network model is processed, the edge monitoring device of the highway section performs real-time route control. When congestion occurs on the highway section, the edge monitoring device sends the control route data to the Internet of Vehicles or map software in advance, and guides the vehicle to change lanes in advance and enter the queue sequence, thereby avoiding the butterfly effect caused by frequent lane changes or serpentine lane changes during peak periods and alleviating traffic congestion.

[0082] This network takes into account the problems of computing resources, precision loss, and computing reasoning speed when deploying embedded devices. The present invention improves the backbone network, integrates the C3f module to enrich the gradient flow information, improves the ability to integrate data features, introduces the fast pyramid pooling SPPF module to improve the reasoning speed, designs decoders for different tasks and multi-task learning functions, and designs a network architecture of a shared encoder according to the difficulty of different tasks. Through multi-task learning loss function learning, the detection accuracy of the network model is improved, and the NNCF algorithm is used to realize the lightweight of the multi-task network model, ensuring that the accuracy loss of the model is minimized when it is deployed.

[0083] In summary, this multi-task network model can ensure accuracy and real-time performance without affecting the detection speed of the model, while greatly reducing the computational overhead and resource usage. It is a good solution for deployment on embedded devices in intelligent transportation.

[0084] In addition, the multi-task network model can be deployed on long straight intersection monitoring cameras or intersection cameras. The purpose is to guide the vehicle to change lanes through the network's multi-task perception before the headway is too small, thereby avoiding the butterfly effect caused by frequent lane changes or serpentine lane changes during peak periods and alleviating traffic congestion.

[0085] 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 multi-task perception traffic control network method that can be used in high-speed sections, characterized in that: It includes the following steps: Step S1: acquiring real-time traffic flow video data by setting a number of monitoring points on the expressway; Step S2: Building a multi-task network model, which is an image detection and segmentation network consisting of a shared encoder and three decoders; The shared encoder consists of a backbone network and a feature fusion layer. The backbone network adopts the CSPDarkNet architecture in which the CSP module is replaced by the C3f module. After replacing the BottleneckCSP module with the C3f module, the CSPDarkNet architecture of the backbone network includes a convolution module and a C3f module. The feature fusion layer includes an FPN architecture and an SPPF module integrated into the FPN architecture. The three decoders are a vehicle target detection decoder, a lane line detection decoder, and a drivable area segmentation decoder. Among them, after the real-time traffic flow video data is input into the shared encoder, the backbone network extracts the feature data of different scales of the real-time traffic flow video data, processes the drivable area segmentation task in advance and outputs it to the drivable area segmentation decoder, and the feature fusion layer receives the feature data of different scales of the real-time traffic flow video data, processes the vehicle target detection segmentation and lane line detection segmentation tasks respectively, and outputs them to the vehicle target detection decoder and lane line detection decoder respectively; In step S2, the C3f module is an improved module of the C2f module in YOLOv8. The C3f module divides the input into n branches and processes them through n Bottleneck modules respectively. At the same time, the 3*3 convolution kernel in the Bottleneck module is split into 3*1 and 1*3 convolution kernels for processing. Finally, all branches are spliced ​​with the residual branch. The parameter n is a parameter that can be adjusted manually. In step S2, the SPPF module of the feature fusion layer is used to expand the receptive field, generate and fuse feature information of different scales, and the FPN architecture of the feature fusion layer performs feature fusion at different semantic levels. The SPPF module and the FPN architecture are connected in series, and the two are connected in series so that the generated features finally contain information of multiple scales and multiple semantic levels; Among them, the feature fusion layer also includes two layers of upsampling. The FPN architecture of the feature fusion layer performs upsampling after receiving the feature map, cancels the convolution operation before the upsampling operation, uses the C3f module for processing after upsampling, and splices the feature maps output at different stages to complete feature fusion; Among them, in step S2, the vehicle target detection decoder adopts a PAN structure, which is composed of a C3f module, a convolution layer, and a Concat layer. The FPN architecture of the feature fusion layer and the PAN structure form an FPN-PAN structure. The FPN-PAN structure realizes the bidirectional transmission of shallow feature maps and deep feature maps, and finally outputs the results through several target detection heads connected to the vehicle target detection decoder; Next, the output features of the FPN architecture of the feature fusion layer are directly transmitted to the lane detection decoder, and the drivable area segmentation decoder directly obtains features from the position between the back of the SPPF module of the feature fusion layer and the front of the FPN architecture; Step S3: training the multi-task network model; And construct a multi-task learning loss function, and apply the multi-task learning loss function to the training of the multi-task network model; After the multi-task network model is trained, the NNCF algorithm is used to make the trained multi-task network model lightweight; Step S4: deploying the trained lightweight multi-task network model in the edge monitoring equipment of the highway section, wherein vehicle target detection and lane line detection assist in searching the optimal control route, and drivable area segmentation determines the safety of vehicle distance on the highway section; Step S5: The edge monitoring equipment on the highway section performs real-time route control. When congestion occurs on the highway section, the edge monitoring equipment sends the control route data to the Internet of Vehicles or map software in advance, and guides the vehicle to change lanes in advance and enter the queue sequence.

2. According to claim 1, a multi-task perception traffic control network method that can be used in high-speed sections is characterized in that: In step S2, the convolution module of the CSPDarkNet architecture of the backbone network consists of a Conv convolution layer, a BatchNorm batch normalization layer and an activation function; the convolution kernel of the Conv convolution layer is set to 3*3.

3. According to claim 1, a multi-task perception traffic control network method that can be used in high-speed sections is characterized in that: In step S2, the drivable area segmentation decoder and the lane line detection decoder are the same.

4. According to claim 1, a multi-task perception traffic control network method applicable to high-speed sections is characterized in that: In step S3, the multi-task learning loss function L all The specific calculation formula is as follows: L all =β1L det +β2L da-seg +β3L ll-seg Among them, L det It is represented as the vehicle target detection task, and β1 is represented as L det The corresponding weight coefficient, L da-seg It is represented as the drivable area segmentation task, and β2 is represented by L da-seg The corresponding weight coefficient, L ll-seg It is represented as the lane detection task, and β3 is represented as L ll-seg The corresponding weight coefficient; Among them, the target detection task L det The specific calculation formula is as follows: L det =α1L class +α2L obj +α3L box Among them, L class It is represented as the focal loss function of the penalty term for classification, and α1 is represented as L class The corresponding weight coefficient, L obj It is expressed as the focal loss function for the predicted confidence, and α2 is expressed as L obj The corresponding weight coefficient, L box It is expressed as CIoU loss function, α3 is expressed as L box The corresponding weight coefficient.

5. The multi-task perception traffic control network method applicable to high-speed sections according to claim 1 is characterized in that: In step S3, the multi-task network model is trained using the bdd100k dataset; In step S3, the NNCF algorithm is used to make the trained multi-task network model lightweight, and the lightweight optimization steps include: Step S31: Establish a quantitative data set Dataset; Step S32: Run the NNCF algorithm to optimize the multi-task network model; Step S33: Use the serialization function in openvino to obtain the OpenVINO IR model, which is a lightweight model of the multi-task network model in different formats.