CN117078718A - Multi-target vehicle tracking method in expressway scene based on deep SORT - Google Patents

Abstract

The invention discloses a method for tracking multiple target vehicles in a expressway scene based on deep SORT, which comprises the following steps: firstly, constructing a data set for evaluating performance of a multi-target vehicle tracking model in an expressway scene; secondly, constructing a Faster-RCNN-VDHS model and a YOLOV5s-VDHS model for vehicle detection in the expressway scene; thirdly, motion prediction and state estimation, namely modeling and estimating the possible arrival position of a target in a future frame by counting the motion behaviors and parameters of the target object in the previous frame; then, constructing a DNFM-RDS model, a ResNet50-VRHS model, a DenseNet121-VRHS model and a SheffleNetV 2-VRHS model for vehicle re-identification, extracting motion characteristics and apparent characteristics of a target detection frame, and constructing a cost matrix according to the similarity between the characteristics; and finally, performing target vehicle association matching. The invention has the beneficial effects that: the intelligent high-speed system can effectively track the multi-target vehicles in the expressway scene, and improve the monitoring efficiency of the expressway system, so that the development of the intelligent high-speed system is promoted.

Description

Multi-target vehicle tracking method in highway scenes based on DeepSORT

Technical field

The invention belongs to the research field of smart high-speed and smart perception, and specifically relates to a DeepSORT-based multi-target vehicle tracking method in a highway scene.

Background technique

The rapid development of artificial intelligence provides new solutions and approaches for real-time monitoring in highway scenarios. How to use machine learning theory, deep learning methods and computer vision technology to accurately detect and track vehicles traveling on highways based on highway surveillance videos, and achieve automated identification, positioning, traffic statistics and abnormal behavior of highway vehicles Detection and providing scientific decision-making for the traffic management department is an important task in the development of intelligent transportation systems.

Due to the characteristics of open-air scenes on highways, such as significant changes in light during different time periods such as day, evening, and night, serious monitoring contamination, and most passing vehicles traveling at speeds above 80 kilometers per hour, video images are prone to blur, resulting in high-speed The accuracy of vehicle detection and tracking in highway scenes is reduced. The multi-target vehicle tracking technology in the highway scene based on DeepSORT in the present invention can effectively improve the efficiency of highway system monitoring and promote the development of smart high-speed systems.

Contents of the invention

Purpose of the invention: In order to overcome the deficiencies in the existing technology, provide a multi-target vehicle tracking method in highway scenes based on DeepSORT, and construct FFRV-RV-DM, YV-RV-DM, FRV- for highway asphalt pavement disease detection. DV-DM, YV-DV-DM, FRV-SV-DM, YV-SV-DM, FRV-DR-DM, YV-DR-DM network models have undergone model training, parameter optimization and model comparison, and can effectively It tracks multi-target vehicles in highway scenes, has good MOTA, MOTP and average detection time, and can provide technical support for highway system monitoring.

Technical solution: In order to achieve the above objectives, the present invention provides a multi-target vehicle tracking method in a highway scene based on DeepSORT, which includes the following steps:

S1: Construction of the data set. Construct a data set for evaluating the performance of the multi-target vehicle tracking model in highway scenes, including three scenes: daytime, evening, and night;

S2: Target detection, construct the Faster-RCNN-VDHS model and YOLOV5s-VDHS model for vehicle detection in highway scenes, which are used to detect moving targets and obtain image information;

S3: Motion prediction and state estimation. By counting the motion behavior and parameters of the target object in previous frames, we can model and estimate the possible position that the target will reach in future frames;

S4: Feature extraction and similarity calculation, construct the DNFM-RDS model, ResNet50-VRHS model, DenseNet121-VRHS model and ShuffleNetV2-VRHS model for vehicle re-identification, which are used to extract the motion features and appearance features of the target detection frame. Add distance calculation through motion information and appearance information, and build a cost matrix based on the similarity between features;

S5: Target vehicle association matching, convert the target association problem into the problem of solving the optimal matching, and perform optimal matching of each target between frames;

S6: Model training and optimization of parameters, preferably FFRV-RV-DM, YV-RV-DM, FRV-DV-DM, YV-DV-DM, FRV-SV-DM, YV-SV-DM, FRV-DR-DM , YV-DR-DM model performs multi-target vehicle tracking in highway scenes.

Further, in the step S1, the daytime, evening, and night scene surveillance videos of a certain section of the Beijing-Shanghai Expressway are used. The FPS is 25 and the resolution is 1920×1080. The lightweight video labeling tool darklabel 2.4 is used to convert the above video stream Import it into the annotation tool, assign different IDs to vehicles according to the format of the MOT16 Challenge data set, and construct a vehicle tracking data set in the highway scene.

Further, in step S2, the Faster-RCNN-VDHS model and YOLOV5s-VDHS model for vehicle detection in highway scenes are constructed, the input frames of the video are analyzed, the moving targets are identified and detected, and the corresponding images of the detected targets are created. Trajectory, initialize the motion variables of the next stage.

Further, Kalman filtering is used as the motion model in step S3. After Kalman filtering is modeled according to the vehicle bounding box in the t-1th frame, the prediction frame of the target in the tth frame can be obtained; then, through the target The detection method uses the detection frame of the t-th frame as the observation value, combines the prediction frame and the detection frame, and obtains the optimal estimate of all target detection frames in the t-th frame of the current video; finally, the optimal estimate is used to obtain the t+1th frame. The new predicted value of the target in the frame, through continuous iteration, finally approaches the actual value of the target. An 8-dimensional space is used to describe the state of the vehicle trajectory in the highway scene at a certain moment. The expression formula is:

Among them, (u, v) is the center coordinate of the vehicle detection frame, r is the aspect ratio, h is the height of the frame, (x·,y·,r·,h·) is the speed of (u,v,r,h) information.

Furthermore, the step S4 includes the acquisition of motion features and apparent features of the target detection frame. The Mahalanobis distance based on the motion features of the vehicle position can provide information that is more effective for short-term prediction. The calculation formula is as follows:

Among them, d ⁽¹⁾ (i, j) is the matching degree between the j-th detection frame and the i-th trajectory, d _j is the position coordinate information of the j-th detection frame, and y _j is the relationship between the current frame trajectory and the next frame. Prediction, _Pi is the covariance matrix of the observation space at the current moment predicted by Kalman filter.

Furthermore, step S4 includes the acquisition of motion features and apparent features of the target detection frame. The cosine distance based on the apparent information features of the vehicle plays an important role in recovering the target ID after the target has been occluded for a long time. The calculation formula is as follows:

Among them, the restriction condition of r _j is ||r _j ||=1, R _i is the feature vector library that stores the entire tracking trajectory, is one of the feature vectors in the feature vector library of the tracking trajectory.

Further, the target vehicle correlation matching in step S5 includes the Hungarian algorithm, cascade matching, and IOU matching. The target correlation problem is converted into a problem of optimal matching, and each target between frames is optimally matched.

Further, in step S6, ResNet50-VRHS, DenseNet121-VRHS, ShuffleNetV2-VRHS, and DNFM-RDS are constructed as re-identification networks for feature extraction and similarity calculation, combined with YOLOV5s-VDHS and Faster-RCNN-VDHS models, including FFRV -RV-DM, YV-RV-DM, FRV-DV-DM, YV-DV-DM, FRV-SV-DM, YV-SV-DM, FRV-DR-DM, YV-DR-DM8 models, built Vehicle tracking model in highway scenes based on deep learning.

Further, in step S6, based on the constructed vehicle tracking data set in the highway scene, the verification set data feedback is provided through the loss function curve, and hyperparameters such as the number of model iterations, learning rate, and attenuation weight are adjusted in sequence to obtain three networks respectively. The MOTA, MOTP and average detection time of the model are comprehensively compared to select the optimal multi-target vehicle tracking method in the highway scene.

Furthermore, the MOTA and MOTP specifically refer to: MOTA is the multi-target tracking accuracy, which is an important indicator to measure the ability of the multi-target tracking method to continuously track the vehicle trajectory. The closer its value is to 1, the better the tracking performance. The calculation formula is as follows:

Among them, FN _t is the number of missed target vehicles in the t-th frame, FP _t is the number of falsely-detected target vehicles in the t-th frame, and IDSW _t is the number of target vehicle ID jumps on several vehicle trajectories in the t-th frame. , GT _t is the actual number of target vehicles marked in the t-th frame.

MOTP stands for multi-target tracking accuracy. It is an important indicator for measuring the positioning accuracy of the detector in the multi-target vehicle tracking method. It represents the average measurement between the detection frame of the multi-target vehicle tracking algorithm and the manually labeled GT. The calculation formula is as follows:

Among them, c _t is the number of target frames that successfully match the predicted target frame and GT in the t-th frame, is the intersection ratio between the i-th target vehicle and the real annotation box in the t-th frame. The larger the MOTP value, the better the detector positioning performance.

This invention integrates the re-identification model of ResNet50-VRHS, DenseNet121-VRHS, ShuffleNetV2-VRHS, DNFM-RDS feature extraction and similarity calculation, and the YOLOV5s-VDHS, Faster-RCNN-VDHS detection model, and is constructed for use in highway scenes. Multi-objective vehicle tracking model. Through model training and parameter optimization, it was found that in terms of the ability to continuously track vehicles in highway scenes, the performance of the vehicle tracking model in highway scenes using YOLOV5s-VDHS as the detector and the DeepSORT-MVTHS framework is better than that of Faster-RCNN- VDHS is the detector and the tracking model of the DeepSORT-MVTHS framework; in terms of positioning accuracy of vehicles in the highway scene, the performance of the highway scene vehicle tracking model using Faster-RCNN-VDHS as the detector and the DeepSORT-MVTHS framework is better than Tracking model using YOLOV5s-VDHS as detector and DeepSORT-MVTHS framework.

The beneficial effects of the present invention are: the constructed multi-target vehicle tracking method in the highway scene based on DeepSORT can effectively track the multi-target vehicles in the highway scene through model training, parameter optimization and model comparison, and has better performance. MOTA, MOTP and average detection time can provide technical support for highway system monitoring.

Description of drawings

Figure 1 is the flow chart of multi-target tracking.

Figure 2 shows an example of monitoring video data set annotation using the vehicle tracking method in a highway scene.

Figure 3 is the DeepSORT-MVTHS target vehicle association matching flow chart.

Figure 4 is the structure diagram of the YV-RV-DM model.

Figure 5 is the structure diagram of the FRV-SV-DM model.

Figure 6 is a comparison chart of vehicle tracking model evaluation indicators in highway scenes.

Figure 7 is a comparison chart of the image processing time of the vehicle tracking model in the highway scene.

Detailed ways

The present invention will be further elucidated below in conjunction with the accompanying drawings and specific embodiments.

The present invention provides a multi-target vehicle tracking method in a highway scene based on DeepSORT, which specifically includes the following steps:

The overall process of the detection-based tracking method can be divided into the target detection stage, motion prediction stage, feature extraction and similarity calculation stage, and data association stage, as shown in Figure 1.

S1: Construction of the data set. Construct a data set for evaluating the performance of the multi-objective vehicle tracking model in highway scenes, including three scenes: day, evening, and night. The daytime, evening, and night scene surveillance video of a section of the Beijing-Shanghai Expressway is used. The FPS is 25 and the resolution is 1920×1080. The lightweight video annotation tool darklabel 2.4 is used to import the above video stream into the annotation tool. According to The format of the MOT16 Challenge data set assigns different IDs to vehicles to construct a vehicle tracking data set in highway scenes, as shown in Figure 2, including the video frame sequence number, vehicle ID number, and the x position coordinate of the upper left corner vertex of the labeled target frame. , y position coordinate, target box width, target box height;

S2: Construct the Faster-RCNN-VDHS model and YOLOV5s-VDHS model for vehicle detection in highway scenes, analyze the input frames of the video, identify and detect moving targets, create trajectories corresponding to the detected targets, and initialize the next stage motion variables;

S3: Kalman filter is used as the motion model. After Kalman filter is modeled according to the vehicle bounding box in the t-1th frame, the prediction frame of the target in the tth frame can be obtained; then, the tth frame is calculated using the target detection method. The detection frame is used as the observation value, and the prediction frame and the detection frame are combined to obtain the optimal estimate of all target detection frames in the tth frame of the current video. Finally, the optimal estimated value is used to obtain a new predicted value of the target in frame t+1. Through continuous iteration, the actual value of the target is finally approximated. An 8-dimensional space is used to describe the vehicle trajectory in the highway scene at a certain point. The state at the moment is represented by the formula:

Among them, (u, v) is the center coordinate of the vehicle detection frame, r is the aspect ratio, h is the height of the frame, (x·,y·,r·,h·) is the speed of (u,v,r,h) information;

The prediction equation calculation formula of the state vector is as follows, which describes the transfer process of the vehicle motion state from the t-1 frame to the t-th frame in the highway scene. F is the state transition matrix. During the vehicle motion state transfer process, it is also necessary to The uncertainty of the vehicle's motion state is described. Kalman filter assumes that the state variables in the system all obey Gaussian distribution, so each variable can be described by mean and variance.

X＝FX _t-1

The following formula makes a preliminary estimate of the vehicle motion state and the uncertainty of the vehicle motion state. P is the covariance matrix, which is used to describe the uncertainty of the vehicle motion state. The larger the value in P, the greater the uncertainty of the vehicle motion state. The higher, Q is the system noise matrix, which is used to describe the noise in vehicle motion state prediction.

P＝FP _t-1 F _T +Q

S4: Including the acquisition of motion features and appearance features of the target detection frame. The Mahalanobis distance based on the motion features of the vehicle position can provide more effective information for short-term prediction. The calculation formula is as follows:

Among them, d ⁽¹⁾ (i, j) is the matching degree between the j-th detection frame and the i-th trajectory, d _j is the position coordinate information of the j-th detection frame, and y _j is the relationship between the current frame trajectory and the next frame. Prediction, _Pi is the covariance matrix of the observation space at the current moment predicted by Kalman filter.

The cosine distance based on the apparent information characteristics of the vehicle plays an important role in recovering the target ID after the target has been occluded for a long time. The calculation formula is as follows:

Among them, the restriction condition of r _j is ||r _j ||=1, R _i is the feature vector library that stores the entire tracking trajectory, is one of the feature vectors in the feature vector library of the tracking trajectory.

After completing the calculation of the matching degree between the target detection vehicle and the trajectory in the high-speed middle road scene, the gate control calculation needs to be performed through the threshold matrix before the cost matrix can be calculated to complete the subsequent data association. The calculation method of the threshold matrix is as shown in the following formula.

in, It is an indicator of the Mahalanobis distance, used to compare the Mahalanobis distance with the chi-square distribution threshold. Only when the Mahalanobis distance is less than the threshold t ⁽¹⁾ , the matching is successful;/> It is an indicator of cosine distance. Only when the cosine distance is less than the preset hyperparameter t ^(2), the matching is successful;/> It is a joint indicator, and only when its value is 1 is the initial match considered successful.

S5: Target vehicle correlation matching includes Hungarian algorithm, cascade matching, and IOU matching. The target correlation problem is converted into the problem of solving the optimal matching, and the optimal matching of each target between frames is performed, as shown in Figure 3. First, target detection is performed on the candidate image to obtain the candidate vehicle detection frame, and the tracking trajectory predicted as a definite state by the Kalman filter is cascade matched with the detected target frame. The results obtained are divided into three situations: successfully matched trajectory MT (MatchedTracks, MT), unmatched detection UT (Unmatched Tracks, UT), unmatched tracks UD (UnmatchedDetections, UD). In the case of a successful match, that is, the predicted bounding box matches the vehicle detection bounding box of the current frame, only Kalman filtering is needed to complete the trajectory update. For the case of unmatched detections and unmatched trajectories, the intersection-union ratio needs to be calculated for IOU matching. The tracking trajectory predicted by the Kalman filter to be in an uncertain state and the detected target frame are directly matched by IOU, and its cost matrix is calculated as the input of the Hungarian algorithm to achieve matching association. After IOU matching, for successful matching, the Kalman filter is used to update the previous trajectory; for unmatched trajectories, if the trajectory is in an unconfirmed state at this time, the trajectory will be deleted directly, otherwise mismatching is required It enters the deletion state after reaching a certain number of times; in the case of unmatched detection, that is, a new target appears in the current frame, but the new target does not exist in the predicted bounding box. In this case, a new target needs to be initialized for the target. traces of.

S6-1: Construct ResNet50-VRHS, DenseNet121-VRHS, ShuffleNetV2-VRHS, and DNFM-RDS as re-identification networks for feature extraction and similarity calculation, combined with YOLOV5s-VDHS and Faster-RCNN-VDHS models, including FFRV-RV-DM , YV-RV-DM, FRV-DV-DM, YV-DV-DM, FRV-SV-DM, YV-SV-DM, FRV-DR-DM, YV-DR-DM8 models to build a deep learning-based Vehicle tracking model in highway scenes.

Taking the YV-RV-DM and FRV-SV-DM models as examples, the structure of the constructed YV-RV-DM model is shown in Figure 4. First, the highway surveillance video frame sequence is sequentially passed through the Focus module in the YOLOV5s-VDHS model, After processing by the CBL standard convolution module, CSP1_X module, SPP spatial pyramid pooling module, non-maximum suppression module, and NMS module, the position of the highway monitoring target vehicle is located; secondly, Kalman filtering is used to predict the position of the vehicle in the current frame based on the vehicle trajectory. 8-dimensional state vector. At the same time, the vehicle detection frame passes through the Conv1_x, Conv2_x, Conv3_x, Conv4_x, Conv5_x modules and 7×7 average pooling in the ResNet50-VRHS model to output a 1×1×2048-dimensional vehicle before the fully connected layer. Appearance feature vector, and calculate the minimum cosine distance between it and the last 100 successfully associated feature sets of each vehicle trajectory, thereby establishing a cost matrix; the constructed FRV-SV-DM model structure is shown in Figure 5, which combines the highway surveillance video The frame is input to the Faster-RCNN-VDHS network. After being processed by 13 convolutional layers, 4 max pooling layers, and 3 fully connected layers in VGG-16, a 1×1×1000 global feature is output, and passed through the RPN network, ROI The pooling and classification module locates the position of the target vehicle in the highway scene; secondly, Kalman filtering is used to predict the 8-dimensional state vector of the vehicle in the current frame based on the vehicle trajectory. At the same time, the vehicle detection frame passes through the convolution layer in the ShuffleNetV2-VRHS model , the most pooling layer, Stage2 module, Stage3 module, and Stage4 module output the 1×1×1024-dimensional vehicle appearance feature vector before the fully connected layer, and calculate the minimum of the 100 most recent successfully associated feature sets for each vehicle trajectory. cosine distance to create a cost matrix

S6-2: Based on the constructed vehicle tracking data set in the highway scene, the verification set data feedback is provided through the loss function curve, and hyperparameters such as the number of model iterations, learning rate, and attenuation weight are adjusted in sequence to obtain the MOTA, MOTP and average detection time are shown in Table 1. The MOTA and MOTP specifically refer to: MOTA is the multi-target tracking accuracy, which is an important indicator to measure the ability of the multi-target tracking method to continuously track the vehicle trajectory. The closer its value is to 1, the better the tracking performance. The calculation formula is as follows:

Among them, FN _t is the number of missed target vehicles in the t-th frame, FP _t is the number of falsely-detected target vehicles in the t-th frame, and IDSW _t is the number of target vehicle ID jumps on several vehicle trajectories in the t-th frame. , GT _t is the actual number of target vehicles marked in the t-th frame.

MOTP stands for multi-target tracking accuracy. It is an important indicator for measuring the positioning accuracy of the detector in the multi-target vehicle tracking method. It represents the average measurement between the detection frame of the multi-target vehicle tracking algorithm and the manually labeled GT. The calculation formula is as follows:

Among them, c _t is the number of target frames that successfully match the predicted target frame and GT in the t-th frame, is the intersection ratio between the i-th target vehicle and the real annotation box in the t-th frame. The larger the MOTP value, the better the detector positioning performance.

Table 1 Eight model evaluation index tables

As shown in Figure 6, RV-RV-DM, YV-RV-DM, FRV-DV-DM, YV-DV-DM, FRV-SV-DM, YV-SV-DM, FRV-DR-DM, YV -The eight DR-DM models all have good results in vehicle tracking on the highway scene surveillance video data set, and their MOTA and MOTP indicators both reach about 80%. The YV-RV-DM model using YOLOV5s-VDHS as the detector and ResNet50-VRHS as the re-identification network has a MOTA of 82.1% and a MOTP of 80.5%; using Faster-RCNN-VDHS as the detector and ResNet50-VRHS as the re-identification network, the MOTA is 82.1% and the MOTP is 80.5%. The FRV-RV-DM model of the recognition network has a MOTA of 79.7% and a MOTP of 85.3%; the YV-DV-DM model using YOLOV5s-VDHS as the detector and DenseNet121-VRHS as the re-identification network has a MOTA of 82.6% and a MOTP of 82.6%. is 80.8%; the MOTA of the FRV-DV-DM model using Faster-RCNN-VDHS detector and DenseNet121-VRHS as the re-identification network is 80.5% and MOTP is 85.4%; using YOLOV5s-VDHS as the detector and ShuffleNetV2- The MOTA of the YV-SV-DM model with VRHS as the re-identification network is 81.9%, and the MOTP is 80.3%; the MOTA of the FRV-SV-DM model with Faster-RCNN-VDHS as the detector and ShuffleNetV2-VRHS as the re-identification network is 79.4%, and the MOTP is 85.0%; the YV-DR-DM model using YOLOV5s-VDHS as the detector and DNFM-RDS as the re-identification network has a MOTA of 83.2% and a MOTP of 81.1%; using Faster-RCNN-VDHS for detection The MOTA of the FRV-DR-DM model using DNFM-RDS as the re-identification network is 81.0%, and the MOTP is 85.8%. In terms of the ability to continuously track vehicles, if the same re-identification network is used, the MOTA indicators of the vehicle tracking model in the highway scene using YOLOV5s-VDHS as the detector and the DeepSORT-MVTHS framework are both higher than those using Faster-RCNN-VDHS. The tracking model of the DeepSORT-MVTHS framework of the detector; in terms of positioning accuracy, the MOTP indicators of the highway scene vehicle tracking model of the DeepSORT-MVTHS framework with Faster-RCNN-VDHS as the detector are higher than those with YOLOV5s-VDHS as the detector. And the tracking models of the DeepSORT-MVTHS framework FRV-RV-DM, YV-RV-DM, FRV-DV-DM, YV-DV-DM, FRV-SV-DM, YV-SV-DM, FRV-DR-DM, The average processing time of each frame of the eight models of YV-DR-DM is shown in Figure 7. The average processing time of each frame of the YV-RV-DM model is 27ms; the average processing time of each frame of the FRV-RV-DM model is 55ms; the average processing time of each frame of the YV-DV-DM model is 31ms; FRV- The average processing time of each frame of the DV-DM model is 62ms; the average processing time of each frame of the YV-SV-DM model is 22ms; the average processing time of each frame of the FRV-SV-DM model is 50ms; YV-DR- The average processing time of each frame of the DM model is 81ms; the average processing time of each frame of the FRV-DR-DM model is 108ms.

The present invention is a multi-target vehicle tracking method in a highway scene based on DeepSORT. DeepSORT-MVTHS, a multi-target vehicle tracking framework for highway scenes and FRV-RV-DM, YV-RV-DM, FRV-DV-DM, YV-DV-DM, for vehicle tracking in highway scenes, were constructed. FRV-SV-DM, YV-SV-DM, FRV-DR-DM, YV-DR-DM eight models were compared and tested based on the constructed vehicle video surveillance data set in the highway scene. The experimental results show that in the highway scene In terms of the ability to continuously track vehicles in the scene, the performance of the vehicle tracking model in the highway scene using YOLOV5s-VDHS as the detector and the DeepSORT-MVTHS framework is better than that of the tracking model using Faster-RCNN-VDHS as the detector and the DeepSORT-MVTHS framework. , In terms of positioning accuracy of vehicles in highway scenes, the performance of the highway scene vehicle tracking model using Faster-RCNN-VDHS as the detector and the DeepSORT-MVTHS framework is better than using YOLOV5s-VDHS as the detector and the DeepSORT-MVTHS framework. tracking model.

The technical means disclosed in the solution of the present invention are not limited to the technical means disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be pointed out that for those of ordinary skill in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also regarded as the protection scope of the present invention.

Claims (10)

1. The multi-target vehicle tracking method in highway scenes based on DeepSORT is characterized by: including the following steps: S1: Construction of the data set. Construct a data set for evaluating the performance of the multi-target vehicle tracking model in highway scenes, including three scenes: daytime, evening, and night; S2: Target detection, construct the Faster-RCNN-VDHS model and YOLOV5s-VDHS model for vehicle detection in highway scenes, which are used to detect moving targets and obtain image information; S3: Motion prediction and state estimation. By counting the motion behavior and parameters of the target object in previous frames, we can model and estimate the possible position that the target will reach in future frames; S4: Feature extraction and similarity calculation, construct the DNFM-RDS model, ResNet50-VRHS model, DenseNet121-VRHS model and ShuffleNetV2-VRHS model for vehicle re-identification, which are used to extract the motion features and appearance features of the target detection frame. Add distance calculation through motion information and appearance information, and build a cost matrix based on the similarity between features; S5: Target vehicle association matching, convert the target association problem into the problem of solving the optimal matching, and perform optimal matching of each target between frames; S6: Model training and optimization of parameters, preferably FFRV-RV-DM, YV-RV-DM, FRV-DV-DM, YV-DV-DM, FRV-SV-DM, YV-SV-DM, FRV-DR-DM , YV-DR-DM model performs multi-target vehicle tracking in highway scenes. 2. The DeepSORT-based multi-target vehicle tracking method in a highway scene according to claim 1, characterized in that: in step S1, scene monitoring videos of a certain section of a highway during the day, evening and night are used, and the FPS is 25 , the resolution is 1920×1080, using the lightweight video annotation tool darklabel 2.4, import the above video stream into the annotation tool, assign different IDs to the vehicles according to the format of the MOT16 Challenge data set, and construct vehicle tracking data in the highway scene set. 3. The DeepSORT-based multi-target vehicle tracking method in highway scenes according to claim 1, characterized in that: in step S2, a Faster-RCNN-VDHS model and YOLOV5s- are constructed for vehicle detection in highway scenes. The VDHS model analyzes the input frames of the video, identifies and detects moving targets, creates trajectories corresponding to the detected targets, and initializes the motion variables in the next stage. 4. The DeepSORT-based multi-target vehicle tracking method in the highway scene according to claim 1, characterized in that: in step S3, Kalman filtering is used as the motion model, and the Kalman filtering is based on the motion in the t-1th frame. After the vehicle boundary box is modeled, the prediction frame of the target in the t-th frame can be obtained; then, the detection frame of the t-th frame is used as the observation value through the target detection method, and the prediction frame and the detection frame are combined to obtain all the prediction frames of the t-th frame of the current video. The optimal estimated value of the target detection frame; finally, the optimal estimated value is used to obtain the new predicted value of the target in the t+1th frame. Through continuous iteration, the actual value of the target is finally approximated, and an 8-dimensional space is used to describe it. The state of the vehicle trajectory at a certain moment in the highway scene is represented by the formula: Among them, (u, v) is the center coordinate of the vehicle detection frame, r is the aspect ratio, h is the height of the frame, is the speed information of (u, v, r, h). 5. The DeepSORT-based multi-target vehicle tracking method in a highway scene according to claim 1, characterized in that: the step S4 includes the acquisition of motion characteristics and apparent characteristics of the target detection frame, based on the vehicle position motion characteristics The Mahalanobis distance can provide more effective information for short-term prediction. The calculation formula is as follows: Among them, d ⁽¹⁾ (i, j) is the matching degree between the j-th detection frame and the i-th trajectory, d _j is the position coordinate information of the j-th detection frame, and y _j is the relationship between the current frame trajectory and the next frame. Prediction, _Pi is the covariance matrix of the observation space at the current moment predicted by Kalman filter. 6. The DeepSORT-based multi-target vehicle tracking method in a highway scene according to claim 1, characterized in that: the step S4 includes the acquisition of motion characteristics and apparent characteristics of the target detection frame, based on vehicle appearance information. The cosine distance of features plays an important role in recovering the target ID after the target has been occluded for a long time. The calculation formula is as follows: Among them, the restriction condition of r _j is ||r _j ||=1, R _i is the feature vector library that stores the entire tracking trajectory, is one of the feature vectors in the feature vector library of the tracking trajectory. 7. The multi-target vehicle tracking method in highway scenes based on DeepSORT according to claim 1, characterized in that: the target vehicle association matching in step S5 includes Hungarian algorithm, cascade matching, IOU matching, and the target association problem Convert it to the problem of solving the optimal matching, and optimally match each target between frames. 8. The DeepSORT-based multi-target vehicle tracking method in highway scenes according to claim 1, characterized in that: in step S6, ResNet50-VRHS, DenseNet121-VRHS, ShuffleNetV2-VRHS, and DNFM-RDS are constructed as re-identification The network performs feature extraction and similarity calculation, combined with YOLOV5s-VDHS and Faster-RCNN-VDHS models, including FFRV-RV-DM, YV-RV-DM, FRV-DV-DM, YV-DV-DM, FRV-SV- DM, YV-SV-DM, FRV-DR-DM, and YV-DR-DM eight models are used to build a deep learning-based vehicle tracking model in highway scenes. 9. The DeepSORT-based multi-target vehicle tracking method in a highway scene according to claim 1, characterized in that: in step S6, a verification set is provided through a loss function curve based on the constructed vehicle tracking data set in the highway scene. After data feedback, the number of model iterations, learning rate, and attenuation weight were adjusted in sequence to obtain the MOTA, MOTP, and average detection time of the three network models respectively, and the optimal multi-target vehicle tracking method in the highway scene was selected through comprehensive comparison. 10. The DeepSORT-based multi-target vehicle tracking method in a highway scene according to claim 9, characterized in that: the MOTA and MOTP specifically refer to: MOTA is the multi-target tracking accuracy, which measures the impact of the multi-target tracking method on the vehicle. An important indicator of trajectory continuous tracking ability. The closer the value is to 1, the better the tracking performance. The calculation formula is as follows: Among them, FN _t is the number of missed target vehicles in the t-th frame, FP _t is the number of falsely-detected target vehicles in the t-th frame, and IDSW _t is the number of target vehicle ID jumps on several vehicle trajectories in the t-th frame. . MOTP stands for multi-target tracking accuracy. It is an important indicator for measuring the positioning accuracy of the detector in the multi-target vehicle tracking method. It represents the average measurement between the detection frame of the multi-target vehicle tracking algorithm and the manually labeled GT. The calculation formula is as follows: Among them, c _t is the number of target frames that successfully match the predicted target frame and GT in the t-th frame, is the intersection ratio between the i-th target vehicle and the real annotation box in the t-th frame; the larger the MOTP value, the better the detector positioning performance.