Vehicle tracking method and device, electronic equipment and medium

Abstract

The invention provides a vehicle tracking method and device, electronic equipment and a medium, and the method comprises the steps: collecting a traffic stream video of a monitoring region in real time through a multi-path network camera, and disassembling the traffic stream video into video frame images, thereby obtaining a traffic stream image data set, detecting whether a target vehicle exists in each image in the traffic flow image data set by adopting a vehicle multi-target tracking model to obtain a target vehicle tracking result, generating a target vehicle trajectory based on the tracking result, and calculating the tracking confidence of the target vehicle trajectory, and judging whether the tracking confidence of the target vehicle trajectory is lower than a preset tracking confidence threshold, and if the tracking confidence of the target vehicle trajectory is not lower than the preset tracking confidence threshold, uploading and storing the target vehicle trajectory. According to the method, the vehicle multi-target tracking model based on attention and dynamic mechanism optimization is adopted to track the vehicle, so that the real-time performance and the operation speed are kept while high accuracy is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and Internet of Things, and in particular to a vehicle tracking method, a vehicle tracking device, an electronic device and a computer-readable medium. Background Art

[0002] Vehicle tracking technology plays a key role in modern traffic management, intelligent transportation systems, urban monitoring, and autonomous driving. With the rapid progress of global urbanization, the volume of road traffic is growing, which makes real-time vehicle tracking and management particularly important. Whether it is to maintain smooth traffic, improve road safety, or achieve more efficient resource allocation, accurate and timely vehicle tracking information is indispensable.

[0003] The challenges of vehicle tracking come not only from the complex environment on the road, such as occlusion by other vehicles, different light conditions and changeable weather conditions, but also from the differences in size, shape and color of vehicle types. In addition, with the development of 5G and IoT technologies, traffic multi-target tracking systems need to be able to process large amounts of data and respond in real time, which also brings new challenges to the design and optimization of the system.

[0004] Among the existing vehicle tracking technologies, radar and environmental sensor-based methods have performed well in certain specific applications, but they are often limited by equipment cost, installation location, and maintenance issues. In contrast, video-based vehicle tracking methods, especially those that combine deep learning and computer vision techniques, offer great potential for providing high accuracy and wide applicability. However, in order to meet the real-time requirements, it is necessary to further optimize the algorithm performance and reduce the computational burden. With the rapid development of computer vision technology, video-based vehicle tracking has become a hot topic of research. Some advanced methods use convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to track vehicles in real time. These methods usually rely on a large amount of labeled data for training and can achieve high tracking accuracy in various environments. For example, some existing technologies use object detection frameworks such as YOLO and Faster R-CNN, combined with tracking algorithms such as SORT or DeepSORT, to perform real-time vehicle tracking. These methods have shown good performance on public datasets such as MOT16 or UA-DETRAC. However, although these techniques have achieved impressive results in certain scenarios, they still suffer from the following shortcomings, such as data dependency: many deep learning methods have a strong dependency on large amounts of high-quality labeled data, which can be a limitation in real-world applications; computational complexity: deep learning models, especially those complex models designed to achieve high accuracy, may require high-performance computing resources, which may not be practical on edge devices; dynamic environment adaptability: existing tracking methods may encounter difficulties in environments with rapid changes or complex lighting conditions; occlusion handling: tracking models may lose track or produce false detections in the event of multi-vehicle occlusion or interaction. Summary of the invention

[0005] In view of the above problems, the present invention is proposed to provide a vehicle tracking method and a corresponding vehicle tracking device, an electronic device and a computer-readable medium that overcome the above problems or at least partially solve the above problems.

[0006] The present invention discloses a vehicle tracking method, which comprises:

[0007] Collect traffic video of the monitoring area in real time through multiple network cameras;

[0008] Decomposing the traffic video of the monitoring area into video frame images to obtain a traffic image dataset;

[0009] A vehicle multi-target tracking model is used to detect whether there is a target vehicle in each image in the traffic image dataset, obtain a target vehicle tracking result for each image, and generate a target vehicle trajectory based on the tracking result of each image;

[0010] Calculate the tracking confidence of the target vehicle trajectory;

[0011] Determining whether the tracking confidence of the target vehicle trajectory is lower than a preset tracking confidence threshold;

[0012] If the tracking confidence of the target vehicle trajectory is not lower than a preset tracking confidence threshold, the target vehicle trajectory is uploaded and stored.

[0013] Optionally, the method further comprises:

[0014] Counting the vehicle flow in the traffic video and counting the tunnel passenger flow at each time point;

[0015] Determining whether the vehicle flow exceeds a preset vehicle flow threshold and determining whether the tunnel passenger flow at each time node exceeds a preset tunnel passenger flow threshold;

[0016] If the vehicle flow exceeds the preset vehicle flow threshold and / or the tunnel passenger flow at each time node exceeds the preset tunnel passenger flow threshold, the vehicle flow, the tunnel passenger flow at each time node, and the time distribution of the vehicle flow exceeding the threshold and / or the time distribution of the tunnel passenger flow exceeding the threshold are uploaded and stored, and an information pop-up alarm and an audible and visual alarm are issued, and the vehicle flow image data set is locally saved to iteratively upgrade the vehicle multi-target tracking model.

[0017] Optionally, the vehicle multi-target tracking model includes a detection branch, a dynamic re-identification branch, and a motion appearance prediction model; the vehicle multi-target tracking model is used to detect whether there is a target vehicle in each image in the traffic image dataset, obtain a target vehicle tracking result for each image, and generate a target vehicle trajectory based on the tracking result of each image, including:

[0018] The DLA-34 backbone network and the cross attention module of the detection branch are used to extract features and detect the position of each image in the traffic image dataset, and the position feature map of the target vehicle in each image is obtained;

[0019] The heatmap detection head, box size detection head, and center offset detection head of the detection branch are used to detect the position feature map of the target vehicle in each image to obtain the detection position of the target vehicle in each image.

[0020] The bottom-up feature fusion module of the dynamic re-identification branch is used to perform feature re-identification on the position feature map of the target vehicle in each image, and the re-identification features of the target vehicle in each image are obtained;

[0021] The motion appearance prediction model is used to output the predicted position of the target vehicle based on the Kalman filter, and the predicted position of the target vehicle is matched with the detection position and re-identification features of the target vehicle in each image through the Hungarian algorithm to obtain the target vehicle tracking result of each image, and the target vehicle trajectory is generated based on the target vehicle tracking result of each image.

[0022] Optionally, the generation of the vehicle multi-target tracking model includes:

[0023] The collected traffic video is disassembled into video frame images and annotated to obtain a vehicle image annotation dataset;

[0024] The vehicle image annotation dataset is divided into a vehicle image training set and a vehicle image test set;

[0025] Construct vehicle multi-target tracking model network;

[0026] The vehicle multi-target tracking model network is trained using the vehicle image training set, and tested using the vehicle image test set to obtain the vehicle multi-target tracking model.

[0027] Optionally, the method further comprises:

[0028] Using the vehicle image training set to train the vehicle multi-target tracking model network to obtain an initial vehicle multi-target tracking model;

[0029] The target detection loss, embedding loss and regression loss are calculated, and the back propagation and parameter update of the model are performed based on the target detection loss, embedding loss and regression loss to obtain the vehicle multi-target tracking model.

[0030] The present invention also discloses a vehicle tracking device, which comprises:

[0031] The vehicle flow video acquisition module is used to collect vehicle flow videos in the monitoring area in real time through multiple network cameras;

[0032] A vehicle flow image data set generation module is used to decompose the vehicle flow video of the monitoring area into video frame images to obtain a vehicle flow image data set;

[0033] A tracking module, used to detect whether there is a target vehicle in each image in the traffic image data set by using a vehicle multi-target tracking model, obtain a target vehicle tracking result for each image, and generate a target vehicle trajectory based on the tracking result of each image;

[0034] A tracking confidence calculation module, used to calculate the tracking confidence of the target vehicle trajectory;

[0035] A tracking confidence judgment module, used to judge whether the tracking confidence of the target vehicle trajectory is lower than a preset tracking confidence threshold;

[0036] The target vehicle trajectory storage module is used to upload and store the target vehicle trajectory if the tracking confidence of the target vehicle trajectory is not lower than a preset tracking confidence threshold.

[0037] Optionally, the device further comprises:

[0038] A traffic statistics module, used to count the vehicle flow in the traffic video and the passenger flow in the tunnel at each time node;

[0039] A flow determination module, used to determine whether the vehicle flow exceeds a preset vehicle flow threshold and whether the tunnel passenger flow at each time node exceeds a preset tunnel passenger flow threshold;

[0040] The traffic upload module is used to upload and store the vehicle traffic, the tunnel passenger flow at each time node, and the time distribution of the vehicle traffic exceeding the threshold and / or the time distribution of the tunnel passenger flow exceeding the threshold if the vehicle traffic exceeds the preset vehicle traffic threshold and / or the tunnel passenger flow at each time node exceeds the preset tunnel passenger flow threshold, and issue an information pop-up alarm and an audible and visual alarm, and locally save the vehicle traffic image data set to iteratively upgrade the vehicle multi-target tracking model.

[0041] Optionally, the vehicle multi-target tracking model includes a detection branch, a dynamic re-identification branch, and a motion appearance prediction model; the tracking module includes:

[0042] The position feature map generation submodule is used to extract features and detect positions of each image in the traffic image dataset using the DLA-34 backbone network of the detection branch and the cross attention module to obtain the position feature map of the target vehicle in each image;

[0043] The position detection submodule is used to detect the position of the target vehicle in each image using the heatmap detection head, box size detection head, and center offset detection head of the detection branch to obtain the detection position of the target vehicle in each image.

[0044] A feature re-identification submodule is used to perform feature re-identification on the position feature map of the target vehicle in each image using a bottom-up feature fusion module of a dynamic re-identification branch to obtain a re-identification feature of the target vehicle in each image;

[0045] The target vehicle is generated according to the submodule, which is used to output the predicted position of the target vehicle based on the Kalman filter using the motion appearance prediction model, and match the predicted position of the target vehicle with the detection position and re-identification features of the target vehicle in each image through the Hungarian algorithm to obtain the target vehicle tracking result of each image, and generate the target vehicle trajectory based on the target vehicle tracking result of each image.

[0046] Optionally, the device further comprises:

[0047] The vehicle image annotation module is used to collect traffic flow videos, disassemble them into video frame images, and annotate them to obtain a vehicle image annotation dataset;

[0048] A vehicle image annotation dataset partitioning module is used to partition the vehicle image annotation dataset into a vehicle image training set and a vehicle image test set;

[0049] Model network construction module, used to build vehicle multi-target tracking model network;

[0050] The model training module is used to train the vehicle multi-target tracking model network using the vehicle image training set and to test the vehicle image test set to obtain the vehicle multi-target tracking model.

[0051] Optionally, the device further comprises:

[0052] A preliminary training module, used to train the vehicle multi-target tracking model network using the vehicle image training set to obtain an initial vehicle multi-target tracking model;

[0053] The model parameter optimization module is used to calculate the target detection loss, embedding loss and regression loss, and perform back propagation and parameter update of the model based on the target detection loss, embedding loss and regression loss to obtain the vehicle multi-target tracking model.

[0054] The present invention also discloses an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0055] The memory is used to store computer programs;

[0056] The processor is used to implement the vehicle tracking method of the present invention when executing the program stored in the memory.

[0057] The present invention also discloses one or more computer-readable media, on which instructions are stored. When executed by one or more processors, the processors are enabled to execute the vehicle tracking method according to the present invention.

[0058] The present invention includes the following advantages:

[0059] The vehicle tracking method of the present invention collects the traffic video of the monitoring area in real time through a multi-channel network camera, disassembles the traffic video of the monitoring area into video frame images to obtain a traffic image data set, uses a vehicle multi-target tracking model to detect whether each image in the traffic image data set has a target vehicle to obtain the target vehicle tracking result of each image, and generates the target vehicle trajectory based on the tracking result of each image, calculates the tracking confidence of the target vehicle trajectory, and determines whether the tracking confidence of the target vehicle trajectory is lower than a preset tracking confidence threshold. If the tracking confidence of the target vehicle trajectory is not lower than the preset tracking confidence threshold, the target vehicle trajectory is uploaded and stored. The present invention uses a vehicle multi-target tracking model optimized based on attention and dynamic mechanisms to track vehicles, ensuring that real-time performance and computing speed are maintained while maintaining high accuracy. This optimization shows its superiority in a tunnel environment, especially in complex light, occlusion and other specific factors, and greatly improves the real-time tracking accuracy of the model. And according to the set tracking confidence threshold, high-confidence tracking results are screened out to ensure data accuracy. The vehicle multi-target tracking model realizes target detection and data association in a single network, greatly simplifies the calculation process, and thus reduces the calculation complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is a flow chart of the steps of a vehicle tracking method provided by an embodiment of the present invention;

[0061] Figure 2 This is a workflow diagram for real-time traffic flow statistics in a tunnel provided by an embodiment of the present invention;

[0062] Figure 3 It is a framework diagram of a vehicle multi-target tracking model provided by an embodiment of the present invention;

[0063] Figure 4 is a workflow diagram of a vehicle multi-target tracking model provided by an embodiment of the present invention;

[0064] Figure 5 is a workflow diagram of a traffic multi-target tracking system provided by an embodiment of the present invention;

[0065] Figure 6 is a structural block diagram of a vehicle tracking device provided by an embodiment of the present invention;

[0066] Figure 7 is a block diagram of an electronic device provided by an embodiment of the present invention;

[0067] Figure 8 It is a schematic diagram of a computer-readable medium provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0069] Reference Figure 1 , shows a flow chart of the steps of a vehicle tracking method provided in an embodiment of the present invention, which may specifically include the following steps:

[0070] Step 101, collecting traffic video of the monitoring area in real time through multiple network cameras;

[0071] Step 102, decomposing the traffic video of the monitoring area into video frame images to obtain a traffic image dataset;

[0072] Step 103, using a vehicle multi-target tracking model to detect whether there is a target vehicle in each image in the traffic image data set, obtaining a target vehicle tracking result for each image, and generating a target vehicle trajectory based on the tracking result of each image;

[0073] Step 104, calculating the tracking confidence of the target vehicle trajectory;

[0074] Step 105, determining whether the tracking confidence of the target vehicle trajectory is lower than a preset tracking confidence threshold;

[0075] Step 106: If the tracking confidence of the target vehicle trajectory is not lower than a preset tracking confidence threshold, the target vehicle trajectory is uploaded and stored.

[0076] In an embodiment of the present invention, first, the traffic video data of the monitored area can be collected in real time through multiple network cameras. These video data are then disassembled into continuous video frame images to form a traffic image data set. Next, a pre-deployed vehicle multi-target tracking model based on attention and dynamic mechanism optimization is used to perform in-depth analysis on each frame of the image to detect whether there is a target vehicle therein. The vehicle multi-target tracking model of the present invention, with its superior real-time performance and computing speed, can maintain high-accuracy vehicle detection in complex environments such as tunnels, even in the face of challenges such as changing light conditions and vehicle occlusion. After the position of the target vehicle is detected in each image, the position change of the vehicle in each frame of the image is continuously tracked through data association, thereby generating a complete trajectory of the target vehicle. The vehicle multi-target tracking model realizes target detection and data association in a single network, greatly simplifies the calculation process, and thus reduces the computational complexity.

[0077] In order to ensure the reliability of the tracking results, the present invention introduces the calculation of tracking confidence, which can be obtained based on a comprehensive evaluation of the continuity and stability of the vehicle trajectory and the accuracy of the model in identifying the target vehicle's features. Subsequently, the tracking confidence of each target vehicle trajectory is compared with the preset tracking confidence threshold. This threshold can be flexibly set according to actual needs to balance the accuracy of tracking. If the tracking confidence of a certain trajectory reaches the threshold, it is regarded as a valid tracking result, and the trajectory data can be uploaded and stored; conversely, if the tracking confidence is insufficient, the corresponding trajectory will be regarded as unreliable and discarded, thereby effectively filtering out low-quality tracking data and ensuring the accuracy and usefulness of the data. Users can personalize the tracking confidence according to different places to adapt to applications in different scenarios and provide more flexible scalability.

[0078] In addition, the present invention also provides a traffic multi-target tracking system, the vehicle tracking method of the present invention is applied to the traffic multi-target tracking system, and the traffic multi-target tracking system may include the following parts: a video acquisition module, a central processing module, a cloud server and a user end. The video acquisition module is connected to the central processing module, the central processing module is connected to the cloud server, and the cloud server is connected to the user end. The video acquisition module uses a network camera to capture the video image of the monitoring area, and transmits the video image to the central processing module; the central processing module builds an environment to deploy a vehicle multi-target tracking model, and after receiving the video image data, the central processing module will complete the tracking of each vehicle through real-time reasoning of the vehicle multi-target tracking model. The model first detects whether the received image information has a target vehicle, if not, it returns to re-collect, if yes, it completes the tracking task through data association, and judges whether the vehicle meets the tracking condition through the set tracking confidence threshold; if the preset tracking confidence is not reached, the detected target trajectory is discarded, and if the preset tracking confidence is reached, it is uploaded and stored. By deploying collection equipment on both sides of the tunnel entrance and exit to realize distributed architecture and data intercommunication, statistics are collected on passenger flow statistics in real-time fixed monitoring areas, tunnel passenger flow statistics within time nodes, time distribution of flow exceeding thresholds, and other data. At the same time, relevant data are uploaded to the cloud server, and the cloud server sends the data to the user end. The user can receive and view real-time information through the mobile phone APP, completing the basic construction of the vehicle tracking method of the present invention in the system application.

[0079] In one embodiment of the present invention, the method further comprises:

[0080] Counting the vehicle flow in the traffic video and counting the tunnel passenger flow at each time point;

[0081] Determining whether the vehicle flow exceeds a preset vehicle flow threshold and determining whether the tunnel passenger flow at each time node exceeds a preset tunnel passenger flow threshold;

[0082] If the vehicle flow exceeds the preset vehicle flow threshold and / or the tunnel passenger flow at each time node exceeds the preset tunnel passenger flow threshold, the vehicle flow, the tunnel passenger flow at each time node, and the time distribution of the vehicle flow exceeding the threshold and / or the time distribution of the tunnel passenger flow exceeding the threshold are uploaded and stored, and an information pop-up alarm and an audible and visual alarm are issued, and the vehicle flow image data set is locally saved to iteratively upgrade the vehicle multi-target tracking model.

[0083] In addition to setting a tracking confidence threshold to improve tracking accuracy, the present invention can also set a flow threshold for flow alarm. Specifically, the central processing module includes an information storage unit, an information processing unit, and an alarm unit. The information storage unit stores relevant programs, algorithm models, and early warning information; the information processing unit processes the image data of the monitoring area, and calls the algorithm model of the information storage unit to detect and track vehicles; the alarm unit includes an information pop-up box alarm of the user-side APP and an external sound and light alarm.

[0084] The information processing unit counts the vehicle flow in the traffic video and the tunnel passenger flow at each time node, determines whether the vehicle flow exceeds the preset vehicle flow threshold and determines whether the tunnel passenger flow at each time node exceeds the preset tunnel passenger flow threshold. If the vehicle flow exceeds the preset vehicle flow threshold and / or the tunnel passenger flow at each time node exceeds the preset tunnel passenger flow threshold, then upload relevant information including vehicle flow, tunnel passenger flow at the time node, and time distribution of flow exceeding the threshold to the cloud server, wherein the continuous frame image information of the vehicle video stored locally is used for iterative upgrade of the algorithm model. The image information stored in the cloud server is further pushed to the user end in the background, and the user can view the real-time vehicle flow, tunnel passenger flow at the time node, and time distribution of flow exceeding the threshold through the mobile phone APP.

[0085] In the implementation of traffic flow statistics in the tunnel, the information processing unit uses a tracking method to perform real-time passenger flow statistics on both lanes after the vehicles enter and exit the area line of the two-way lanes on both sides of the tunnel entrance and exit, writes them into the cache, and then calculates them separately and returns the results to the system's quantity statistics interface. Figure 2 Specifically, the video stream data of devices A and B on both sides of the tunnel entrance and exit are collected first, and the number of devices A and B entering and exiting is counted in real time based on algorithm reasoning. in , A out ), B in and out quantity (B in , B out) , real-time calculation results in tunnel flow = (A in -B out )+(B in -A out ), determine whether it exceeds the set threshold. If it exceeds the threshold, the central processing module sends an alarm signal and records the information.

[0086] In one embodiment of the present invention, a vehicle multi-target tracking model includes a detection branch, a dynamic re-identification branch, and a motion appearance prediction model; the vehicle multi-target tracking model is used to detect whether there is a target vehicle in each image in the traffic image data set, obtain a target vehicle tracking result for each image, and generate a target vehicle trajectory based on the tracking result of each image, including:

[0087] The DLA-34 backbone network and the cross attention module of the detection branch are used to extract features and detect the position of each image in the traffic image dataset, and the position feature map of the target vehicle in each image is obtained;

[0088] The heatmap detection head, box size detection head, and center offset detection head of the detection branch are used to detect the position feature map of the target vehicle in each image to obtain the detection position of the target vehicle in each image.

[0089] The bottom-up feature fusion module of the dynamic re-identification branch is used to perform feature re-identification on the position feature map of the target vehicle in each image, and the re-identification features of the target vehicle in each image are obtained;

[0090] The motion appearance prediction model is used to output the predicted position of the target vehicle based on the Kalman filter, and the predicted position of the target vehicle is matched with the detection position and re-identification features of the target vehicle in each image through the Hungarian algorithm to obtain the target vehicle tracking result of each image, and the target vehicle trajectory is generated based on the target vehicle tracking result of each image.

[0091] Reference Figure 3 The vehicle multi-target tracking model of the present invention integrates a detection branch, a dynamic re-identification branch, and a motion appearance prediction model. Figure 4 Specifically, the vehicle multi-target tracking model first uses the DLA-34 backbone network and cross attention module of the detection branch to perform refined feature extraction and position positioning on each frame of the traffic monitoring video to ensure that the position feature map of the target vehicle can be accurately captured. Subsequently, the heatmap detection head, box size detection head, and center offset detection head in the detection branch are used to further refine the precise position information of the target vehicle to ensure the accuracy of tracking.

[0092] In order to improve the continuity and robustness of tracking, a dynamic re-identification branch is designed in the vehicle multi-target tracking model. This branch uses a bottom-up feature fusion module to perform deep learning and re-identification processing on the extracted position feature map, generate unique re-identification features of the target vehicle, and effectively distinguish and track different vehicles. Finally, the motion appearance prediction model predicts the position of the target vehicle based on the Kalman filter algorithm, and efficiently matches the predicted position with the detected position and re-identification features through the Hungarian algorithm, thereby achieving continuous and stable tracking of vehicle targets, and constructing accurate target vehicle trajectories based on this, thereby greatly improving the real-time and accuracy of multi-target vehicle tracking.

[0093] In one embodiment of the present invention, the generation of the vehicle multi-target tracking model includes:

[0094] The collected traffic video is disassembled into video frame images and annotated to obtain a vehicle image annotation dataset;

[0095] The vehicle image annotation dataset is divided into a vehicle image training set and a vehicle image test set;

[0096] Construct vehicle multi-target tracking model network;

[0097] The vehicle multi-target tracking model network is trained using the vehicle image training set, and tested using the vehicle image test set to obtain the vehicle multi-target tracking model.

[0098] In one embodiment of the present invention, the method further comprises:

[0099] Using the vehicle image training set to train the vehicle multi-target tracking model network to obtain an initial vehicle multi-target tracking model;

[0100] The target detection loss, embedding loss and regression loss are calculated, and back propagation and parameter update of the model are performed based on the target detection loss, embedding loss and regression loss to obtain the vehicle multi-target tracking model.

[0101] In the embodiment of the present invention, 10 video sequences were collected by using the Internet and field survey experiments. After breaking down the frames, the total number of image data was 17,132. Then the data set was divided into a training set and a test set in a ratio of 9:1.

[0102] The variant DLA-34 designed by the present invention has two changes compared to the original DLA-34: (1) it is decomposed into two different feature fusion methods according to the detection and re-identification branches. The detection branch adopts the original DLA-34 feature fusion method, and the re-identification branch adopts a bottom-up feature fusion method to retain high-level feature map information; (2) a cross attention module (CCA) is added between the first and second layers of the iterative deep aggregation structure and the bottom-up fusion module to weight the feature map to enhance local context information.

[0103] The specific variant DLA-34 network structure uses DLA-34 as the main component module in the backbone network, and encodes the feature map through continuous downsampling. Then, in the detection branch, an iterative deep aggregation structure is used as the neck network, in which the iterative deep aggregation process and the layered deep aggregation process will continue to perform upsampling decoding aggregation operations until the feature map resolution is 1 / 4 of the input image to the second half of the detection branch to complete feature fusion. In the re-identification branch, a bottom-up feature fusion module is used to replace the iterative deep aggregation structure in DLA-34 as the neck network. This structure retains high-level feature map information during the feature fusion process, and outputs a feature map resolution of 1 / 4 of the input image after directly upsampling the 1 / 32 resolution to the second half of the dynamic re-identification branch.

[0104] The detailed calculation process of cross attention is as follows: the input local feature map H is filtered and generated by two 1×1 convolutions and two feature maps, and then the attention map A is further constructed through the calculation of the attraction (Affinity) calculation module. This process uses the average pooling technology to encode the feature map in the horizontal direction, and uses Softmax normalization to process in the vertical direction to achieve effective encoding of each channel. Finally, the pixel information mapped horizontally and vertically is added and fused to output the feature map.

[0105] The calculation formula of the attraction (Affinity) calculation module is: Then a layer is applied on the channel dimension of D to calculate the attention map, and another 1×1 convolution layer is applied to H to generate the vector V for feature adaptation. Finally, the fusion operation of the context global information is performed. The calculation formula of the fusion operation is:

[0106] This model can locate and track multiple targets more accurately and has better robustness, and can achieve continuous tracking in complex scenarios to reduce frequent switching of targets.

[0107] The feature map output by the backbone network is used to comprehensively predict the target position through the heatmap detection head, box size detection head and centeroffset detection head. On this basis, a dedicated re-identification head (Re-ID) is added to extract re-identification features. The original single convolution kernel is reconstructed using the dynamic convolution mechanism (Dynamic Conv), and the high-dimensional shared feature information after the segmented feature fusion and integration of the feature extraction network is adjusted using 256-channel dynamic convolution. Finally, the target re-identification features are extracted according to the target center point. The re-identification features are fused by the convolutional neural network with a dynamic perception mechanism after the re-identification branch output of the feature extraction network. The re-identification branch after reconstruction is called the dynamic re-identification branch (Dynamic Re-ID Embedding, DRE).

[0108] The network of the dynamic re-identification branch (DRE) consists of an upsampling layer and a dynamic convolution layer. First, a 3×3 convolution is performed to extract higher-level features. Then, the global information of the input feature map is compressed by the average pooling sampling layer. Then, a local operation is performed by the dynamic convolution layer. In the dynamic convolution calculation, the attention weights are calculated by applying the squeeze-and-excitation block (SE) to aggregate. Then, the two fully connected layers are sequentially activated by the ReLU activation function and the Softmax normalization to generate the standardized attention weights of the K convolution kernels. Finally, they are added after the distribution calculation of the dynamic perception layer, and then the 128-dimensional dynamic convolution module is finally generated by the fully connected layer. The dynamic re-identification branch dynamically aggregates multiple parallel convolution kernels according to the attention. The attention dynamically adjusts the weight of each convolution kernel according to the input, thereby generating an adaptive dynamic convolution. On the premise of enhancing the association module of the context information to enhance the extraction of the identity feature (identity embedding) of each position, the dynamic embedding network is strengthened to utilize the characteristic that the distance between different targets is greater than the distance between the same targets, and finally the network's discrimination ability is enhanced.

[0109] Data association process: Based on the detector, a motion appearance prediction model is first built, and the position and velocity estimation of the target is continuously updated based on the Kalman filter. At the same time, the detection results and re-identification features output by the detection and re-identification networks are combined and similarity is calculated to associate the actually detected target with the tracker's predicted target. The Hungarian algorithm is used to complete the association matching of the target ID through multiple stages of matching to ensure that the tracker can correctly track the target to its corresponding trajectory.

[0110] Based on the above design, a vehicle multi-target tracking model network including a detection branch, a dynamic re-identification branch, and a motion appearance prediction model was constructed, and then the vehicle multi-target tracking model network was trained using a training set to obtain a vehicle multi-target tracking model.

[0111] Specifically, the training set is fed into the computing unit of the system to train and fine-tune the vehicle multi-target tracking model network. The backbone network uses a 7x7 convolution kernel in the initial layer, and the remaining convolution layers use a 3x3 convolution kernel with a step size of 1. The image padding uses the same, and the activation function uses ReLU; the pooling layer usually uses a 2x2 kernel with a step size of 2 for downsampling. Feature fusion is achieved by combining upsampling and downsampling with jump connections and IDAUp. The ReID head uses a network based on dynamic convolution to extract feature vectors, with the convolution kernel set to 4 and the convolution network dimension set to 128. Use multi-task loss to set uncertaintyloss to optimize target detection and ReID tasks simultaneously, and set triplet loss to optimize Re-ID features. The initialized vehicle multi-target tracking model network is iteratively trained. The input image size is 1088x608, the initial learning rate is usually set to 1e-4, and the learning rate decay strategy is used to decay to 1e-5 after 40 epochs. The optimizer uses Adam, the weight decay is set to 1e-4, the batch-size is set to 12, and data enhancement techniques such as random cropping, rotation, and color distortion are used. The ImageNet pre-trained model is used, and the iterative training is 60 epochs. The feature map resolution is 272×152.

[0112] Calculate multi-task losses, including target detection loss, embedding loss, and regression loss, use a variant of tripletloss to improve the re-ID embedding branch loss calculation, perform model backpropagation and parameter updates, and optimize the model's detection and tracking performance.

[0113] Finally, the tracking effect of the vehicle multi-target tracking model is verified based on the public test set KITTI. It can be judged whether the vehicle multi-target tracking model has reached the specified accuracy and real-time frame rate. If not, continue to train and optimize the vehicle multi-target tracking model; if it has reached, select the model with the highest average accuracy as the final weight model according to the test set results. And verify the tracking effect of the vehicle multi-target tracking model in a fixed scene. It can be judged whether the vehicle multi-target tracking model has reached the specified accuracy in a specific test environment (such as Jetson Nano embedded devices, tunnels, highway traffic, night scenes, etc.). If not, continue to train and optimize the vehicle multi-target tracking model; if it has reached, select the model with the highest average accuracy as the final weight model according to the test set results.

[0114] The embodiment of the present invention constructs and trains a cross attention module (CCA) between the first and second layers of the iterative deep aggregation structure and the bottom-up fusion module to weight the feature map to enhance local context information. The Re-ID branch uses dynamic convolution to extract the feature map and then learns and improves the network's target re-identification ability. At the same time, the variant triplet loss is used to adjust the optimal solution, thereby further improving the accuracy of the vehicle multi-target tracking model.

[0115] The vehicle multi-target tracking model obtained after training and testing is transplanted into the traffic multi-target tracking system. Figure 5 , the specific processing flow of the traffic multi-target tracking system:

[0116] Data initialization: setting tracking confidence threshold and flow threshold;

[0117] Data collection: The monitoring system collects real-time video streams from vehicles in the high-definition camera area. In tunnel scenarios, collection equipment is deployed in the lanes on both sides of the tunnel entrance and exit;

[0118] Data processing: On the edge computing device, the collected video stream is first tracked in real time through the vehicle multi-target tracking model. The system will count the vehicle flow in the video area in real time, and filter out high-confidence tracking results according to the set tracking confidence threshold to ensure the accuracy of the data. When the vehicle flow exceeds the preset total passenger flow threshold and lasts for a period of time, the system will trigger a notification to remind managers to pay attention to the real-time traffic conditions.

[0119] Data upload: The edge computing device uploads relevant information such as vehicle flow, tunnel passenger flow statistics within a time node, and time distribution of traffic exceeding the threshold to the cloud server. The data center stores and further analyzes the data. At the same time, the cloud server pushes these key data to the mobile application for tunnel management in real time. The continuous frame image information of the vehicle video saved locally is used for iterative upgrades of the algorithm model;

[0120] Data distribution: The image information stored in the cloud server is pushed to the user end. The management personnel can use the mobile phone APP to view the real-time vehicle flow, tunnel passenger flow statistics within the time node, and the distribution of the time when the flow exceeds the threshold. When the flow exceeds the preset threshold, the application will issue a notification to remind the management personnel to take timely measures to ensure on-site traffic prevention and control.

[0121] The vehicle tracking method and traffic multi-target tracking system of the present invention have the following significant beneficial effects: (1) Improved accuracy and real-time performance: A vehicle multi-target tracking model based on attention and dynamic mechanism optimization is adopted to ensure high accuracy while maintaining real-time performance and computing speed. This optimization shows its superiority in tunnel environments, especially in complex light, occlusion and other specific factors, and greatly improves the real-time tracking accuracy of the model. (2) Reduced computational complexity: Combining the joint detection and embedded tracking paradigm, the present model realizes target detection and data association in a single network, greatly simplifies the calculation process, and thus reduces computational complexity. (3) Enhanced model adaptability: The vehicle tracking method of the present invention takes into account various real-scene factors, such as light changes, occlusion and high traffic density, and has stronger adaptability and stability. (4) Real-time remote monitoring: Combined with the Internet of Things architecture, not only local real-time monitoring can be achieved, but also remote data processing and analysis can be achieved. Through mobile APP or other platforms, managers can obtain vehicle tracking and traffic statistics information in real time and respond to various situations in a timely manner. (5) Improved system flexibility and scalability: Users can personalize tracking confidence according to different scenarios, making the system more flexible and scalable. (6) Cost saving: Compared with the tracking system that relies on high-cost radar and environmental sensors, the video image-based tracking system of the present invention has obvious cost advantages while also providing high accuracy and wide applicability.

[0122] It should be noted that, for the sake of simplicity, the method embodiments are described as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present invention are not limited by the order of the actions described, because according to the embodiments of the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0123] Reference Figure 3 , shows a structural block diagram of a vehicle tracking device provided in an embodiment of the present invention, which may specifically include the following modules:

[0124] The vehicle flow video acquisition module is used to collect vehicle flow videos in the monitoring area in real time through multiple network cameras;

[0125] A vehicle flow image data set generation module is used to decompose the vehicle flow video of the monitoring area into video frame images to obtain a vehicle flow image data set;

[0126] A tracking module, used to detect whether there is a target vehicle in each image in the traffic image data set by using a vehicle multi-target tracking model, obtain a target vehicle tracking result for each image, and generate a target vehicle trajectory based on the tracking result of each image;

[0127] A tracking confidence calculation module, used to calculate the tracking confidence of the target vehicle trajectory;

[0128] A tracking confidence judgment module, used to judge whether the tracking confidence of the target vehicle trajectory is lower than a preset tracking confidence threshold;

[0129] The target vehicle trajectory storage module is used to upload and store the target vehicle trajectory if the tracking confidence of the target vehicle trajectory is not lower than a preset tracking confidence threshold.

[0130] Optionally, the device further comprises:

[0131] A traffic statistics module, used to count the vehicle flow in the traffic video and the passenger flow in the tunnel at each time node;

[0132] A flow determination module, used to determine whether the vehicle flow exceeds a preset vehicle flow threshold and whether the tunnel passenger flow at each time node exceeds a preset tunnel passenger flow threshold;

[0133] The traffic upload module is used to upload and store the vehicle traffic, the tunnel passenger flow at each time node, and the time distribution of the vehicle traffic exceeding the threshold and / or the time distribution of the tunnel passenger flow exceeding the threshold if the vehicle traffic exceeds the preset vehicle traffic threshold and / or the tunnel passenger flow at each time node exceeds the preset tunnel passenger flow threshold, and issue an information pop-up alarm and an audible and visual alarm, and locally save the vehicle traffic image data set to iteratively upgrade the vehicle multi-target tracking model.

[0134] Optionally, the vehicle multi-target tracking model includes a detection branch, a dynamic re-identification branch, and a motion appearance prediction model; the tracking module includes:

[0135] The position feature map generation submodule is used to extract features and detect positions of each image in the traffic image dataset using the DLA-34 backbone network of the detection branch and the cross attention module to obtain the position feature map of the target vehicle in each image;

[0136] The position detection submodule is used to detect the position of the target vehicle in each image using the heatmap detection head, box size detection head, and center offset detection head of the detection branch to obtain the detection position of the target vehicle in each image.

[0137] A feature re-identification submodule is used to perform feature re-identification on the position feature map of the target vehicle in each image using a bottom-up feature fusion module of a dynamic re-identification branch to obtain a re-identification feature of the target vehicle in each image;

[0138] The target vehicle is generated according to the submodule, which is used to output the predicted position of the target vehicle based on the Kalman filter using the motion appearance prediction model, and match the predicted position of the target vehicle with the detection position and re-identification features of the target vehicle in each image through the Hungarian algorithm to obtain the target vehicle tracking result of each image, and generate the target vehicle trajectory based on the target vehicle tracking result of each image.

[0139] Optionally, the device further comprises:

[0140] The vehicle image annotation module is used to collect traffic flow videos, disassemble them into video frame images, and annotate them to obtain a vehicle image annotation dataset;

[0141] A vehicle image annotation dataset partitioning module is used to partition the vehicle image annotation dataset into a vehicle image training set and a vehicle image test set;

[0142] Model network construction module, used to build vehicle multi-target tracking model network;

[0143] The model training module is used to train the vehicle multi-target tracking model network using the vehicle image training set and to test the vehicle image test set to obtain the vehicle multi-target tracking model.

[0144] Optionally, the device further comprises:

[0145] A preliminary training module, used to train the vehicle multi-target tracking model network using the vehicle image training set to obtain an initial vehicle multi-target tracking model;

[0146] The model parameter optimization module is used to calculate the target detection loss, embedding loss and regression loss, and perform back propagation and parameter update of the model based on the target detection loss, embedding loss and regression loss to obtain the vehicle multi-target tracking model.

[0147] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0148] In addition, an embodiment of the present invention further provides an electronic device, such as Figure 4 As shown, it includes a processor 401, a communication interface 402, a memory 403 and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404.

[0149] Memory 403, used for storing computer programs;

[0150] The processor 401 is used to implement the vehicle tracking method described in the above embodiment when executing the program stored in the memory 403.

[0151] The communication bus mentioned in the above terminal can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0152] The communication interface is used for communication between the above terminal and other devices.

[0153] The memory may include a random access memory (RAM) or a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0154] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0155] like Figure 5 As shown, in another embodiment provided by the present invention, a computer-readable storage medium 501 is further provided, in which instructions are stored. When the computer-readable storage medium is run on a computer, the computer executes the vehicle tracking method described in the above embodiment.

[0156] In another embodiment of the present invention, a computer program product including instructions is provided. When the computer program product is run on a computer, the computer executes the vehicle tracking method described in the above embodiment.

[0157] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media integrated. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk Solid State Disk (SSD)), etc.

[0158] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0159] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0160] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A vehicle tracking method, characterized in that: The method comprises: Collect traffic video of the monitoring area in real time through multiple network cameras; Decomposing the traffic video of the monitoring area into video frame images to obtain a traffic image dataset; A vehicle multi-target tracking model is used to detect whether there is a target vehicle in each image in the traffic image dataset, obtain a target vehicle tracking result for each image, and generate a target vehicle trajectory based on the tracking result of each image; Calculate the tracking confidence of the target vehicle trajectory; Determining whether the tracking confidence of the target vehicle trajectory is lower than a preset tracking confidence threshold; If the tracking confidence of the target vehicle trajectory is not lower than a preset tracking confidence threshold, the target vehicle trajectory is uploaded and stored.

2. The method according to claim 1, characterized in that: The method further comprises: Counting the vehicle flow in the traffic video and counting the tunnel passenger flow at each time point; Determining whether the vehicle flow exceeds a preset vehicle flow threshold and determining whether the tunnel passenger flow at each time node exceeds a preset tunnel passenger flow threshold; If the vehicle flow exceeds the preset vehicle flow threshold and / or the tunnel passenger flow at each time node exceeds the preset tunnel passenger flow threshold, the vehicle flow, the tunnel passenger flow at each time node, and the time distribution of the vehicle flow exceeding the threshold and / or the time distribution of the tunnel passenger flow exceeding the threshold are uploaded and stored, and an information pop-up alarm and an audible and visual alarm are issued, and the vehicle flow image data set is locally saved to iteratively upgrade the vehicle multi-target tracking model.

3. The method according to claim 1, characterized in that The vehicle multi-target tracking model includes a detection branch, a dynamic re-identification branch, and a motion appearance prediction model; the vehicle multi-target tracking model is used to detect whether there is a target vehicle in each image in the traffic image dataset, obtain the target vehicle tracking result of each image, and generate the target vehicle trajectory based on the tracking result of each image, including: The DLA-34 backbone network and the cross attention module of the detection branch are used to extract features and detect the position of each image in the traffic image dataset, and the position feature map of the target vehicle in each image is obtained; The heatmap detection head, box size detection head, and center offset detection head of the detection branch are used to detect the position feature map of the target vehicle in each image to obtain the detection position of the target vehicle in each image. The bottom-up feature fusion module of the dynamic re-identification branch is used to perform feature re-identification on the position feature map of the target vehicle in each image, and the re-identification features of the target vehicle in each image are obtained; The motion appearance prediction model is used to output the predicted position of the target vehicle based on the Kalman filter, and the predicted position of the target vehicle is matched with the detection position and re-identification features of the target vehicle in each image through the Hungarian algorithm to obtain the target vehicle tracking result of each image, and the target vehicle trajectory is generated based on the target vehicle tracking result of each image.

4. The method according to claim 1, characterized in that: The generation of the vehicle multi-target tracking model includes: The collected traffic video is disassembled into video frame images and annotated to obtain a vehicle image annotation dataset; The vehicle image annotation dataset is divided into a vehicle image training set and a vehicle image test set; Construct vehicle multi-target tracking model network; The vehicle multi-target tracking model network is trained using the vehicle image training set, and tested using the vehicle image test set to obtain the vehicle multi-target tracking model.

5. The method according to claim 4, characterized in that The method further comprises: Using the vehicle image training set to train the vehicle multi-target tracking model network to obtain an initial vehicle multi-target tracking model; The target detection loss, embedding loss and regression loss are calculated, and back propagation and parameter update of the model are performed based on the target detection loss, embedding loss and regression loss to obtain the vehicle multi-target tracking model.

6. A vehicle tracking device, characterized in that: The device comprises: The vehicle flow video acquisition module is used to collect vehicle flow videos in the monitoring area in real time through multiple network cameras; A vehicle flow image data set generation module is used to decompose the vehicle flow video of the monitoring area into video frame images to obtain a vehicle flow image data set; A tracking module, used to detect whether there is a target vehicle in each image in the traffic image data set using a vehicle multi-target tracking model, obtain a target vehicle tracking result for each image, and generate a target vehicle trajectory based on the tracking result of each image; A tracking confidence calculation module, used to calculate the tracking confidence of the target vehicle trajectory; A tracking confidence judgment module, used to judge whether the tracking confidence of the target vehicle trajectory is lower than a preset tracking confidence threshold; The target vehicle trajectory storage module is used to upload and store the target vehicle trajectory if the tracking confidence of the target vehicle trajectory is not lower than a preset tracking confidence threshold.

7. The device according to claim 6, characterized in that The device also includes: A traffic statistics module, used to count the vehicle flow in the traffic video and the passenger flow in the tunnel at each time node; A flow determination module, used to determine whether the vehicle flow exceeds a preset vehicle flow threshold and whether the tunnel passenger flow at each time node exceeds a preset tunnel passenger flow threshold; The traffic upload module is used to upload and store the vehicle traffic, the tunnel passenger flow at each time node, and the time distribution of the vehicle traffic exceeding the threshold and / or the time distribution of the tunnel passenger flow exceeding the threshold if the vehicle traffic exceeds the preset vehicle traffic threshold and / or the tunnel passenger flow at each time node exceeds the preset tunnel passenger flow threshold, and issue an information pop-up alarm and an audible and visual alarm, and locally save the vehicle traffic image data set to iteratively upgrade the vehicle multi-target tracking model.

8. The device according to claim 6, characterized in that The vehicle multi-target tracking model includes a detection branch, a dynamic re-identification branch, and a motion appearance prediction model; the tracking module includes: The position feature map generation submodule is used to extract features and detect positions of each image in the traffic image dataset using the DLA-34 backbone network of the detection branch and the cross attention module to obtain the position feature map of the target vehicle in each image; The position detection submodule is used to detect the position of the target vehicle in each image using the heatmap detection head, box size detection head, and centeroffset detection head of the detection branch to obtain the detection position of the target vehicle in each image. A feature re-identification submodule is used to perform feature re-identification on the position feature map of the target vehicle in each image using a bottom-up feature fusion module of a dynamic re-identification branch to obtain a re-identification feature of the target vehicle in each image; The target vehicle is generated according to the submodule, which is used to output the predicted position of the target vehicle based on the Kalman filter using the motion appearance prediction model, and match the predicted position of the target vehicle with the detection position and re-identification features of the target vehicle in each image through the Hungarian algorithm to obtain the target vehicle tracking result of each image, and generate the target vehicle trajectory based on the target vehicle tracking result of each image.

9. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; The memory is used to store computer programs; The processor is used to implement the vehicle tracking method according to any one of claims 1 to 5 when executing the program stored in the memory.

10. One or more computer-readable media having instructions stored thereon, which, when executed by one or more processors, cause the processors to perform the vehicle tracking method according to any one of claims 1 to 5.

Images