A target vehicle tracking method based on an improved Deepsort algorithm

Abstract

The application discloses a target vehicle tracking method based on an improved Deepsort algorithm, which comprises the following steps: 1, image dataset construction; 2, target detection of target vehicles in images is carried out through YOLOv5s, and target vehicle detection frame position information is acquired; 3, target vehicle prediction frame position information is acquired through a Deepsort target tracking algorithm; 4, a vehicle re-identification model is constructed to extract target vehicle appearance features; 5, cascade matching is carried out on the detection frame and the prediction frame of the target vehicle; 5, intersection-over-union matching is carried out on the detection frame and the prediction frame of the target vehicle; 6, distance-intersection-over-union matching is carried out on the detection frame and the prediction frame of the target vehicle. The application can effectively capture appearance feature information of target vehicles and can more effectively deal with the problem of tracking target loss caused by factors such as vehicle occlusion and overlap in a complex traffic environment, so as to improve the accuracy and robustness of target vehicle tracking, improve traffic efficiency, and ensure the safety of vehicle driving.

Description

Technical Field

[0001] This invention belongs to the field of intelligent connected vehicle perception technology, specifically a target vehicle tracking method based on an improved Deepsort algorithm. Background Technology

[0002] With the significant increase in China's car ownership, higher demands have been placed on road traffic efficiency and safety standards. The rapid advancement of artificial intelligence technology has brought innovative solutions to the field of intelligent connected vehicle perception technology. The combination of deep learning and computer vision technology offers new possibilities for accurate vehicle identification and tracking in complex traffic scenarios. The application of automation technology aims to achieve real-time vehicle positioning and monitoring of abnormal driving behavior, which is crucial for the development of intelligent transportation systems. However, in some specific complex traffic scenarios, factors such as vehicle overlap, occlusion, high-speed movement, and environmental changes often lead to the loss of tracking targets. Furthermore, although existing tracking algorithms have shown effectiveness in some aspects, their robustness and stability still need improvement. Specifically, traditional convolutional neural networks have limitations in the accuracy of vehicle feature extraction, which limits the ability to continuously track the same target vehicle during vehicle tracking. Therefore, existing technologies may not be able to ensure continuous identification and tracking of target vehicles during the tracking process, which is particularly evident in complex environments. Summary of the Invention

[0003] This invention aims to address the shortcomings of existing technologies by proposing a target vehicle tracking method based on an improved Deepsort algorithm. This method aims to improve the tracking accuracy and robustness of vehicles in complex traffic scenarios, thereby enhancing the efficiency and safety of transportation.

[0004] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0005] The present invention provides a target vehicle tracking method based on an improved Deepsort algorithm, characterized in that it is applied to a mixed road driving environment that includes connected autonomous vehicles and traditional manually driven vehicles. The target vehicle tracking method includes the following steps.

[0006] Step 1: Constructing the image dataset:

[0007] Step 1.1: Use a camera to collect video data of the target vehicle to be tracked in a daytime scene in the highway weaving area and perform frame extraction processing to obtain a video image sequence set. Among them, Image i Let i represent the i-th video frame, and I represent the total number of video frames.

[0008] Step 1.2: Use the LabelImg annotation tool to annotate the location and category of the target vehicles in the video image sequence set F, obtaining an image dataset Img containing annotation information, denoted as... in, Image represents the i-th frame. i The marked location information of the nth target vehicle, and Image represents the i-th frame. i The width and height of the bounding box containing the nth target vehicle. Image of the i-th frame i The coordinates of the center point of the bounding box containing the nth target vehicle; N represents the total number of target vehicles;

[0009] Step 2: Input Img into the YOLOv5s object detection model to process the i-th frame image Image. i Perform object detection to obtain the i-th frame image. i Location information of the detection box corresponding to the nth target vehicle in, Image represents the i-th frame. i The width, height, and depth of the detection frame for the nth target vehicle. Image represents the i-th frame. i The coordinates of the center point of the detection frame for the nth target vehicle are denoted as .

[0010] Step 3, The image is input into the Deepsort target tracking model for prediction, resulting in the (i+1)th frame image. i+1 The predicted bounding box location information of the nth target vehicle in the image. in, Representing the (i+1)th frame image (Image) i+1 The width and height of the predicted bounding box for the nth target vehicle in the image. Image represents the (i+1)th frame. i+1 The coordinates of the center point of the predicted bounding box of the nth target vehicle in the image are denoted as .

[0011] Step 4: Construct a vehicle re-identification model, including a residual network and a channel attention module, and process the i-th frame image Image. i After feature extraction and feature encoding of the nth target vehicle, the appearance features of the nth target vehicle are obtained. in, Image represents the i-th frame. i The color characteristics of the nth target vehicle. Image represents the i-th frame. i Texture features of the nth target vehicle. Image represents the i-th frame. i Shape characteristics of the nth target vehicle;

[0012] Step 5: For the i-th frame image Image i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The predicted bounding boxes are cascaded for matching. If the match is successful, proceed to step 8; otherwise, proceed to step 6.

[0013] Step 6: For the i-th frame image Image i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The predicted bounding boxes of the nth target vehicle are matched by intersection-union comparison. If the match is successful, proceed to step 8; otherwise, proceed to step 7.

[0014] Step 7: For the i-th frame image Image i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The predicted bounding box of the nth target vehicle is matched with distance-intersection-union ratio. If the match is successful, proceed to step 8; otherwise, it means that the nth target vehicle is lost, the tracking fails, and return to step 2 to track other target vehicles.

[0015] Step 8: After assigning i+1 to i, return to step 3 and execute sequentially until i>I or tracing fails.

[0016] The target vehicle tracking method based on the improved Deepsort algorithm described in this invention is also characterized in that step 5 includes:

[0017] Step 5.1: Calculate the image of the i-th frame using equation (1). i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 Euclidean distance between the predicted bounding boxes of the nth target vehicle

[0018]

[0019] Step 5.2: Calculate the image of the i-th frame using equation (2). i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The cosine distance between the prediction boxes of the nth target vehicle

[0020]

[0021] In equation (2), Image represents the (i+1)th frame. i+1 R is the set of appearance features of all target vehicles i+1 The k-th eigenvector in the dataset; k∈N;

[0022] Step 5.3: Use equation (3) to calculate the image of the i-th frame. i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 Matching similarity between the predicted bounding boxes of the nth target vehicle

[0023]

[0024] In equation (3), λ represents the weighting factor, and 0 < λ < 1;

[0025] Step 5.4, if Then the image of the i-th frame (Image) i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The predicted bounding boxes of the nth target vehicle in the image were successfully matched, and the Image in the (i+1)th frame was also matched. i+1 Predicted bounding box location information for the nth target vehicle Perform correction to obtain the Image in the (i+1)th frame. i+1 Location information of the detection frame for the nth target vehicle Then proceed to step 8; where, Represents the Image in the (i+1)th frame. i+1 The width, height, and depth of the detection frame for the nth target vehicle. Represents the Image in the (i+1)th frame. i+1 The coordinates of the center point of the detection box for the nth target vehicle are denoted as .

[0026] like This indicates a matching failure, proceed to step 6; where, Th α The threshold value for cascading matching is set.

[0027] Step 6 includes:

[0028] Step 6.1: Calculate the image of the i-th frame using equation (4). i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 Cross-Union Loss between the predicted bounding boxes of the nth target vehicle

[0029]

[0030] In equation (4), Image represents the i-th frame. i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The area of ​​the intersection of the predicted bounding boxes of the nth target vehicle. Image represents the i-th frame. i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The total coverage area between the predicted bounding boxes of the nth target vehicle;

[0031] Step 6.2, if This indicates a successful match, and the Image in the (i+1)th frame is selected. i+1 Predicted bounding box location information for the nth target vehicle Perform correction to obtain the Image in the (i+1)th frame. i+1 Location information of the detection frame for the nth target vehicle And proceed to step 8;

[0032] like If the match fails, proceed to step 7; where, Th β The set crossover-union ratio matching threshold.

[0033] Step 7 includes:

[0034] Step 7.1: Calculate the image of the i-th frame using equation (5). i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 Distance between predicted bounding boxes of the nth target vehicle - Intersection over Union (IoU) loss

[0035]

[0036] In equation (5), Image represents the i-th frame. i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The diagonal distance between the minimum enclosed regions between the predicted bounding boxes of the nth target vehicle;

[0037] Step 7.2, if This indicates a successful match, and the Image in the (i+1)th frame is selected. i+1 Predicted bounding box location information for the nth target vehicle Perform correction to obtain the Image in the (i+1)th frame. i+1Location information of the detection frame for the nth target vehicle And proceed to step 8;

[0038] This indicates a matching failure, meaning the nth target vehicle was lost. γ This indicates the set distance-intersection-union ratio (DIU) matching threshold.

[0039] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the target vehicle tracking method, and the processor is configured to execute the program stored in the memory.

[0040] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the target vehicle tracking method.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] 1. This invention constructs a novel vehicle re-identification model that uses a residual network with better feature extraction performance as the feature extractor. This network structure is deeper and has stronger feature extraction capabilities, enabling it to more effectively capture the appearance feature information of the target vehicle. Simultaneously, the introduction of a channel attention mechanism allows the network to focus more intently on the target region, improving the accuracy and robustness of target vehicle tracking.

[0043] 2. This invention reconstructs the matching stage of the Deepsort algorithm by adding distance-intersection-union ratio matching during the association matching process between the detection box and the prediction box of the target vehicle. This can effectively address the problem of target loss caused by factors such as vehicle occlusion and overlap in complex traffic environments. Attached Figure Description

[0044] Figure 1 This is the overall flowchart of the present invention;

[0045] Figure 2 This is a diagram of the residual network structure.

[0046] Figure 3 This is a diagram of the module structure within the channel.

[0047] Figure 4 This is a structural diagram of the vehicle re-identification model. Detailed Implementation

[0048] In this example, a target vehicle tracking method based on an improved Deepsort algorithm is applied to a mixed driving environment involving both connected autonomous vehicles and traditional manually driven vehicles. Figure 1As shown, the target vehicle tracking method includes the following steps;

[0049] Step 1: Constructing the image dataset:

[0050] Step 1.1: Use a camera to collect video data of the target vehicle to be tracked in a daytime scene in the highway weaving area and perform frame extraction processing to obtain a video image sequence set. Among them, Image i Let i represent the i-th video frame, and I represent the total number of video frames.

[0051] Step 1.2: Use the LabelImg annotation tool to annotate the location and category of the target vehicles in the video image sequence set F, obtaining an image dataset containing annotation information, denoted as . in, Image represents the i-th frame. i The marked location information of the nth target vehicle, and Image represents the i-th frame. i The width and height of the bounding box containing the nth target vehicle. Image of the i-th frame i The coordinates of the center point of the bounding box containing the nth target vehicle; N represents the total number of target vehicles.

[0052] Step 2: Input Img into the YOLOv5s object detection model to process the i-th frame image Image. i Perform object detection to obtain the i-th frame image. i Location information of the detection box corresponding to the nth target vehicle in, Image represents the i-th frame. i The width, height, and depth of the detection frame for the nth target vehicle. Image represents the i-th frame. i The coordinates of the center point of the detection frame for the nth target vehicle are denoted as .

[0053] Step 3, The image is input into the Deepsort target tracking model, where it is predicted using Kalman filtering to obtain the (i+1)th frame image. i+1 The predicted bounding box location information of the nth target vehicle in the image. in, Representing the (i+1)th frame image (Image) i+1 The width and height of the predicted bounding box for the nth target vehicle. Image represents the (i+1)th frame. i+1The coordinates of the center point of the predicted bounding box of the nth target vehicle in the image are denoted as .

[0054] In this embodiment, Kalman filtering includes a prediction phase and an update phase;

[0055] In the Kalman filter prediction stage, the i-th frame image Image i Target vehicle detection frame location information As input, the image of frame i+1 is obtained. i+1 The predicted bounding box location information of the nth target vehicle in the image.

[0056] In the Kalman filter update phase, in this example, the i-th frame image Image i The detection bounding box of the target vehicle and the image of the (i+1)th frame. i+1 After the predicted bounding box of the nth target vehicle is successfully matched, the predicted bounding box position information obtained in the prediction stage is further corrected and updated to obtain the corrected i+1th frame image. i+1 Location information of the detection frame for the nth target vehicle

[0057] Step 4: Construct the vehicle re-identification model. The network structure diagram of the vehicle re-identification model is shown below. Figure 2 As shown, it includes: a residual network and a channel attention module, and applies the i-th frame image Image. i After feature extraction and feature encoding of the nth target vehicle, the appearance features of the nth target vehicle are obtained. in, Image represents the i-th frame. i The color characteristics of the nth target vehicle. Image represents the i-th frame. i Texture features of the nth target vehicle. Image represents the i-th frame. i The shape characteristics of the nth target vehicle.

[0058] like Figure 3 As shown, the residual network ResNet34 consists of a 7x7 convolutional layer, a max pooling layer, 16 consecutively connected residual blocks, an average pooling layer, and a fully connected layer. The residual block consists of two 3x3 convolutional layers, two batch normalization layers, and a ReLU activation function.

[0059] like Figure 4As shown, the channel attention module includes: one average pooling layer and two fully connected layers, one multiplier, one ReLU activation function, and one Sigmoid activation function. The first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the Sigmoid activation function.

[0060] The channel attention module is embedded in the ResNet34 network in a series manner to form a vehicle re-identification model.

[0061] Step 5: For the i-th frame image Image i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The predicted bounding boxes are cascaded for matching. If the match is successful, proceed to step 8; otherwise, proceed to step 6.

[0062] Step 5.1: Calculate the image of the i-th frame using equation (1). i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 Euclidean distance between the predicted bounding boxes of the nth target vehicle

[0063]

[0064] Step 5.2: Calculate the image of the i-th frame using equation (2). i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The cosine distance between the prediction boxes of the nth target vehicle

[0065]

[0066] In equation (2), Image represents the (i+1)th frame. i+1 R is the set of appearance features of all target vehicles i+1 The k-th eigenvector in the vector; k∈N.

[0067] Step 5.3: Use equation (3) to calculate the image of the i-th frame. i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 Matching similarity between the predicted bounding boxes of the nth target vehicle

[0068]

[0069] In equation (3), λ represents the weighting factor, and 0 < λ < 1;

[0070] Step 5.4, if Then the image of the i-th frame (Image) i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The predicted bounding boxes of the nth target vehicle in the image were successfully matched, and the Kalman filter was used to process the Image in the (i+1)th frame. i+1 Predicted bounding box location information for the nth target vehicle Perform correction to obtain the Image in the (i+1)th frame. i+1 Location information of the detection frame for the nth target vehicle Then proceed to step 8; where, Represents the Image in the (i+1)th frame. i+1 The width, height, and depth of the detection frame for the nth target vehicle. Represents the Image in the (i+1)th frame. i+1 The coordinates of the center point of the detection frame for the nth target vehicle are denoted as .

[0071] like This indicates a matching failure, proceed to step 6; where, Th α The threshold value for cascading matching is set.

[0072] Step 6: For the i-th frame image Image i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The predicted bounding boxes of the nth target vehicle are matched by intersection-union comparison. If the match is successful, proceed to step 8; otherwise, proceed to step 7.

[0073] Step 6.1: Calculate the image of the i-th frame using equation (4). i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 Cross-Union Loss between the predicted bounding boxes of the nth target vehicle

[0074]

[0075] In equation (4), Image represents the i-th frame. i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The area of ​​the intersection of the predicted bounding boxes of the nth target vehicle. Image represents the i-th frame. i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1The total coverage area between the predicted bounding boxes of the nth target vehicle.

[0076] Step 6.2, if This indicates a successful match, and the Image in the (i+1)th frame is then processed using Kalman filtering. i+1 Predicted bounding box location information for the nth target vehicle Perform correction to obtain the Image in the (i+1)th frame. i+1 Location information of the detection frame for the nth target vehicle And proceed to step 8;

[0077] like If the match fails, proceed to step 7; where, Th β The set crossover-union ratio matching threshold.

[0078] Step 7: For the i-th frame image Image i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The predicted bounding box of the nth target vehicle is matched with distance-intersection-union ratio. If the match is successful, step 8 is executed. Otherwise, the nth target vehicle is lost, the tracking fails, and the process returns to step 2 to track other target vehicles.

[0079] Step 7.1: Calculate the image of the i-th frame using equation (5). i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 Distance between predicted bounding boxes of the nth target vehicle - Intersection over Union (IoU) loss

[0080]

[0081] In equation (5), Image represents the i-th frame. i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The diagonal distance between the minimum enclosed regions of the predicted bounding boxes of the nth target vehicle.

[0082] Step 7.2, if This indicates a successful match, and the Image in the (i+1)th frame is then processed using Kalman filtering. i+1 Predicted bounding box location information for the nth target vehicle Perform correction to obtain the Image in the (i+1)th frame. i+1 Location information of the detection frame for the nth target vehicle And proceed to step 8;

[0083] This indicates that the match failed, the nth target vehicle was lost, and the tracking failed. Where, Th γ This represents the set distance-intersection-union ratio (DIU) matching threshold.

[0084] Step 8: After assigning i+1 to i, return to step 3 and execute sequentially until i>I or tracing fails.

[0085] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0086] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0087] In summary, this invention reconstructs the vehicle re-identification model and the association matching stage of the detection box and prediction box of the target vehicle. It can effectively capture the appearance feature information of the target vehicle and more effectively deal with the problem of target loss due to vehicle occlusion, overlap and other factors in complex traffic environments. This can improve the accuracy and robustness of target vehicle tracking, improve traffic efficiency, and ensure the safety of vehicle driving. It is of great significance for the further development of intelligent transportation systems.

Claims

1. A target vehicle tracking method based on an improved Deepsort algorithm, characterized in that, The target vehicle tracking method is applied to mixed road driving environments that include connected autonomous vehicles and traditional manually driven vehicles, and includes the following steps; Step 1: Constructing the image dataset: Step 1.1: Use a camera to collect video data of the target vehicle to be tracked in a daytime scene in the highway weaving area and perform frame extraction processing to obtain a video image sequence set. Among them, Image i Let i represent the i-th video frame, and I represent the total number of video frames. Step 1.2: Use the LabelImg annotation tool to annotate the location and category of the target vehicles in the video image sequence set F, obtaining an image dataset Img containing annotation information, denoted as... in, Image represents the i-th frame. i The marked location information of the nth target vehicle, and Image represents the i-th frame. i The width and height of the bounding box containing the nth target vehicle. Image of the i-th frame i The coordinates of the center point of the bounding box containing the nth target vehicle; N represents the total number of target vehicles; Step 2: Input Img into the YOLOv5s object detection model to process the i-th frame image Image. i Perform object detection to obtain the i-th frame image. i Location information of the detection box corresponding to the nth target vehicle in, Image represents the i-th frame. i The width, height, and depth of the detection frame for the nth target vehicle. Image represents the i-th frame. i The coordinates of the center point of the detection frame for the nth target vehicle are denoted as . Step 3, The image is input into the Deepsort target tracking model for prediction, resulting in the (i+1)th frame image. i+1 The predicted bounding box location information of the nth target vehicle in the image. in, Representing the (i+1)th frame image (Image) i+1 The width and height of the predicted bounding box for the nth target vehicle in the image. Image represents the (i+1)th frame. i+1 The coordinates of the center point of the predicted bounding box of the nth target vehicle in the image are denoted as . Step 4: Construct a vehicle re-identification model, including a residual network and a channel attention module, and process the i-th frame image Image. i After feature extraction and feature encoding of the nth target vehicle, the appearance features of the nth target vehicle are obtained. in, Image represents the i-th frame. i The color characteristics of the nth target vehicle. Image represents the i-th frame. i Texture features of the nth target vehicle. Image represents the i-th frame. i Shape characteristics of the nth target vehicle; Step 5: For the i-th frame image Image i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The predicted bounding boxes are cascaded for matching. If the match is successful, proceed to step 8; otherwise, proceed to step 6. Step 6: For the i-th frame image Image i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The predicted bounding boxes of the nth target vehicle are matched by intersection-union comparison. If the match is successful, proceed to step 8; otherwise, proceed to step 7. Step 7: For the i-th frame image Image i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The predicted bounding box of the nth target vehicle is matched with distance-intersection-union ratio. If the match is successful, proceed to step 8; otherwise, it means that the nth target vehicle is lost, the tracking fails, and return to step 2 to track other target vehicles. Step 8: After assigning i+1 to i, return to step 3 and execute sequentially until i>I or tracing fails.

2. The target vehicle tracking method based on the improved Deepsort algorithm according to claim 1, characterized in that, Step 5 includes: Step 5.1: Calculate the image of the i-th frame using equation (1). i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 Euclidean distance between the predicted bounding boxes of the nth target vehicle Step 5.2: Calculate the image of the i-th frame using equation (2). i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The cosine distance between the prediction boxes of the nth target vehicle In equation (2), Image represents the (i+1)th frame. i+1 R is the set of appearance features of all target vehicles i+1 The k-th eigenvector in the dataset; k∈N; Step 5.3: Use equation (3) to calculate the image of the i-th frame. i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 Matching similarity between the predicted bounding boxes of the nth target vehicle In equation (3), λ represents the weighting factor, and 0 < λ < 1; Step 5.4, if Then the image of the i-th frame (Image) i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The predicted bounding boxes of the nth target vehicle in the image were successfully matched, and the Image in the (i+1)th frame was also matched. i+1 Predicted bounding box location information for the nth target vehicle Perform correction to obtain the Image in the (i+1)th frame. i+1 Location information of the detection frame for the nth target vehicle Then proceed to step 8; where, Represents the Image in the (i+1)th frame. i+1 The width, height, and depth of the detection frame for the nth target vehicle. Represents the Image in the (i+1)th frame. i+1 The coordinates of the center point of the detection box for the nth target vehicle are denoted as . like This indicates a matching failure, proceed to step 6; where, Th α The threshold value for cascading matching is set.

3. The target vehicle tracking method based on the improved Deepsort algorithm according to claim 1, characterized in that, Step 6 includes: Step 6.1: Calculate the image of the i-th frame using equation (4). i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 Cross-Union Loss between the predicted bounding boxes of the nth target vehicle In equation (4), Image represents the i-th frame. i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The area of ​​the intersection of the predicted bounding boxes of the nth target vehicle. Image represents the i-th frame. i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The total coverage area between the predicted bounding boxes of the nth target vehicle; Step 6.2, if This indicates a successful match, and the Image in the (i+1)th frame is selected. i+1 Predicted bounding box location information for the nth target vehicle Perform correction to obtain the Image in the (i+1)th frame. i+1 Location information of the detection frame for the nth target vehicle And proceed to step 8; like If the match fails, proceed to step 7; where, Th β The set crossover-union ratio matching threshold.

4. A target vehicle tracking method based on an improved Deepsort algorithm according to claim 2, characterized in that, Step 7 includes: Step 7.1: Calculate the image of the i-th frame using equation (5). i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 Distance between predicted bounding boxes of the nth target vehicle - Intersection over Union (IoU) loss In equation (5), Image represents the i-th frame. i The detection bounding box of the nth target vehicle and the image of the (i+1)th frame. i+1 The diagonal distance between the minimum enclosed regions between the predicted bounding boxes of the nth target vehicle; Step 7.2, if This indicates a successful match, and the Image in the (i+1)th frame is selected. i+1 Predicted bounding box location information for the nth target vehicle Perform correction to obtain the Image in the (i+1)th frame. i+1 Location information of the detection frame for the nth target vehicle And proceed to step 8; This indicates a matching failure, meaning the nth target vehicle was lost. γ This indicates the set distance-intersection-union ratio (DIU) matching threshold.

5. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing any of the target vehicle tracking methods of claims 1-4, and the processor is configured to execute the programs stored in the memory.

6. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is run by the processor, it performs the steps of the target vehicle tracking method according to any one of claims 1-4.

Images