A vehicle multi-target tracking method suitable for highway tunnel monitoring scenarios

Abstract

The present invention discloses a vehicle multi-target tracking method suitable for highway tunnel monitoring scenarios. The tracking method of the present invention takes the actual highway tunnel environment as the starting point, and on the basis of vehicle target detection, skillfully uses motion features and trajectory features. In order to adapt to different state trajectories, a targeted prediction algorithm is designed, and the similarity matching method is improved. With the help of a hierarchical cascade matching strategy, accurate vehicle tracking is successfully achieved. The method of the present invention effectively reduces the vehicle target matching error caused by factors such as target occlusion and target loss, which not only improves the accuracy of tracking, but also meets the real-time requirements.

Description

Technical Field

[0001] The invention belongs to the field of intelligent transportation, and in particular relates to a vehicle multi-target tracking method suitable for highway tunnel monitoring scenarios. Background Art

[0002] As an indispensable part of the urban transportation network, the safety and smoothness of highway tunnels are crucial to ensure the seamless transportation. With the continuous increase in traffic volume and the increase in vehicle speed, the demand for monitoring and management of highway tunnels is becoming increasingly urgent. In this context, vehicle multi-target tracking technology, as a key technology, has received extensive attention and application. In the highway tunnel monitoring scenario, this technology has the ability to provide accurate, real-time and comprehensive vehicle tracking and management, thus showing a wide range of application prospects.

[0003] However, existing vehicle tracking methods face some challenges in highway tunnel scenarios. First, for multi-target tracking, existing methods may face performance bottlenecks and find it difficult to accurately track multiple targets at the same time. Second, there is a large volume of traffic and high speeds in highway tunnels, and vehicles frequently occlude each other and move rapidly. This poses a severe test to the accuracy and stability of the tracking algorithm, and can easily lead to problems such as missed and false detection of targets. For example, BoYin et al. designed a vehicle re-identification method that combines multiple features, which improves the accuracy of vehicle re-identification from the perspective of feature extraction. However, this method consumes a lot of computing resources and time, affecting the efficiency and practicality of the model.

[0004] Therefore, it is necessary to conduct in-depth research on a multi-target tracking solution that adapts to the characteristics of highway tunnels based on actual conditions, so as to reduce vehicle target matching errors caused by target occlusion and loss. At the same time, it is also necessary to cope with the special needs of high-speed vehicles in highway tunnels and reduce the occurrence of problems such as missed detection of tracking models and ID switching. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide a vehicle multi-target tracking method suitable for highway tunnel monitoring scenes. The present invention aims to solve the problems of large vehicle target matching error, poor tracking accuracy and low real-time performance in the existing methods.

[0006] To achieve the above object, the present invention provides a vehicle multi-target tracking method suitable for highway tunnel monitoring scenarios, comprising the following steps:

[0007] S1. Use the target detection model to detect the vehicle target of the current frame, and classify the vehicle target detection results and vehicle target trajectories;

[0008] The vehicle target detection results are divided into: high confidence detection results D highAnd low confidence detection results D low ;

[0009] The vehicle target trajectory is divided into: confirmed trajectory T confirmed and unconfirmed trajectory T unconfirmed ;

[0010] Confirmed track T confirmed Divided into: Confirmed tracks T in the activated state confirmed and the confirmed trajectory T of the lost state confirmed ;

[0011] S2. The high confidence detection result D in step S1 high and all confirmed trajectories T confirmed Matching is performed to obtain the successfully matched association pair 1, the unmatched high confidence detection result and the unmatched trajectory;

[0012] S3. The low confidence detection result D in step S1 is low Matching with the trajectory that is in the activated state and not successfully matched in step S2, so as to obtain the successfully matched association pair 2, the unsuccessfully matched high confidence detection result and the unsuccessfully matched trajectory;

[0013] S4. Combine the unmatched high confidence detection results in step S2 and the unconfirmed trajectory T in step S1 unconfirmed Matching is performed to obtain successfully matched association pairs 3, unmatched high-confidence detection results and unmatched trajectories;

[0014] S5. Update the matching results of steps S2 to S4, repeat steps S1 to S4 and the update of step S5, that is, construct a multi-target vehicle tracking model in the tunnel;

[0015] S6. Create a highway tunnel vehicle target tracking dataset, and use the dataset together with the target detection model to train a multi-target tracking model for vehicles in the tunnel.

[0016] Furthermore, in step S1, the confidence test result is judged by a low score threshold and a high score threshold;

[0017] Among them, the low score threshold θ L =0.2, high score threshold θ H =0.6;

[0018] The detection score of the vehicle target Score ≥ θ H D is a high confidence detection result high ;

[0019] The detection score of the vehicle target θ L ≤Score<θ HD is a low confidence detection result low .

[0020] Further, the step S2 includes the following sub-steps:

[0021] S2.1 Design prediction algorithm for confirmed trajectory T confirmed Make predictions and obtain predicted values;

[0022] S2.2 Use IoU to calculate the predicted value and high confidence detection result D high The similarity between them generates the distance matrix Dis;

[0023] S2.3 Using the Hungarian algorithm, the distance matrix Dis is used to convert the high confidence detection results D high With confirmed trajectory T confirmed Matching is performed to obtain successfully matched association pairs 1, unmatched high-confidence detection results, and unmatched trajectories.

[0024] Furthermore, in step S2.1, the prediction algorithm has two different prediction branches:

[0025] Prediction branch 1: Confirmed trajectory T confirmed When in the active state, Kalman filtering is used for motion prediction;

[0026] Prediction branch 2: Confirmed trajectory T confirmed When in lost state, check confirmed track T confirmed If the length of the historical trajectory data is greater than 4, the Cubic Spline interpolation method is used for trajectory prediction, otherwise the Kalman filter is used for motion prediction.

[0027] Furthermore, in step S2.2, the calculation expression of the distance matrix Dis is as follows:

[0028] Dis=1-IoU

[0029]

[0030] Where: a is the high confidence detection result D high The bounding box of the confirmed trajectory T confirmed The bounding boxes obtained after prediction; Area(a) and Area(b) represent the areas of a and b respectively.

[0031] Further, the step S3 includes the following sub-steps:

[0032] S3.1 uses CDIoU to calculate the active and unmatched trajectory in step S2 and the low confidence detection result D in step S1 lowThe similarity between them generates the distance matrix Dis CD ;

[0033] S3.2 Using the Hungarian algorithm, the distance matrix Dis CD The low confidence detection result D in step S1 low Matching is performed with the trajectory that is in the activated state and has not been successfully matched in step S2, so as to obtain the successfully matched association pair 2, the unsuccessfully matched low-confidence detection result and the unsuccessfully matched trajectory.

[0034] Furthermore, in step S3.1, the distance matrix Dis CD The calculation expression is as follows:

[0035] Dis CD =1-CDIoU

[0036] CDIoU=IoU+λ(1-diou)

[0037]

[0038] Where: IoU represents the low confidence detection result D in step S1 low and the intersection-over-union ratio of the trajectories that are active and not successfully matched in step S2; λ is the weight coefficient; diou represents the distance metric; DL represents the low confidence detection result D low ; TH represents the bounding box of the trajectory that is activated and not successfully matched in step S2; ||DL-TH||2 represents the sum of the Euclidean distances of the four vertices of the two bounding boxes; MBR represents the minimum matrix surrounding the two bounding boxes DL and TH; Diagonal(MBR) represents the diagonal length of the MBR matrix.

[0039] Further, the step S4 includes the following sub-steps:

[0040] S4.1 Use IoU to calculate the unconfirmed track T in step S1 unconfirmed The distance matrix between the high confidence detection results that were not successfully matched in step S2;

[0041] S4.2 Using the Hungarian algorithm, the distance matrix calculated in step S4.1 is used to match the unmatched high-confidence detection results in step S2 and the unconfirmed trajectories T in step S1. unconfirmed Matching is performed to obtain successfully matched association pairs 3, unmatched high-confidence detection results, and unmatched trajectories.

[0042] Further, the step S5 includes the following sub-steps:

[0043] S5.1 uses Kalman filtering to update the trajectories of association pair 1, association pair 2 and association pair 3 that are successfully matched in steps S2 to S4;

[0044] S5.2 updates the unmatched trajectories in step S3 to a lost state. The lost trajectories can be retained for 60 frames. If the number of frames exceeds 60, the trajectories are updated to a removed state. The unmatched low-confidence detection results in step S3 are deleted.

[0045] S5.3 Update the successfully matched association pair 3 in step S4 to the confirmed trajectory T in the activated state confirmed ; The high confidence detection results that are not matched successfully are updated to new trajectories, that is, unconfirmed trajectories T unconfirmed ; Unmatched trajectories are updated to removed status;

[0046] S5.4 repeats the contents of steps S1 to S4 and steps S5.1 to S5.3.

[0047] Further, the step S6 includes the following sub-steps:

[0048] S6.1 Create a highway tunnel vehicle target tracking dataset;

[0049] Collect video data in highway tunnels and use DarkLabel software to annotate the data;

[0050] S6.2 Use the target detection model to train the multi-target tracking model for vehicles in the tunnel on the prepared dataset, use the Adam optimizer to optimize the model loss, and obtain and save the model weights;

[0051] S6.3 Loading weights for the multi-target tracking model of vehicles in tunnels.

[0052] Beneficial effects of the present invention:

[0053] Considering that most of the existing predictions for the target trajectory of vehicles in high-speed tunnels use Kalman filtering, and Kalman prediction is only applicable to linear motion. Because the Lost state Lost trajectory is within the retained 60 frames, due to the rapid movement of vehicles in the tunnel, possible curved roads and the driver's lane changing behavior, its trajectory is nonlinear, which ultimately leads to the failure of IoU matching. To solve this problem, the present invention designs a two-branch prediction algorithm, uses the Cubic Spline interpolation method to fit the historical data of the Lost state Lost trajectory, predicts the position of the current frame of the trajectory, so as to re-associate the detection target and realize the continuous tracking of the target, thereby improving the stability and continuity of target tracking and reducing the loss of target trajectory.

[0054] In addition, when matching the low-confidence detection results in the tunnel with the trajectory, the motion blur caused by occlusion and excessive movement speed is the main reason for the low confidence of the detection results, and the traditional IoU matching will have a poor matching effect when the target shape changes greatly. To address this problem, the present invention improves the similarity matching method and replaces IoU with CDIoU. CDIoU is more adaptable to changes in target shape while ensuring real-time performance. It not only considers the overlap of the target frame, but also pays attention to the distance between the center points of the target frame, and is more suitable for highway tunnel scenarios. The vehicle multi-target tracking method of the present invention enhances the robustness of vehicle multi-target tracking in tunnels, reduces target mismatching, and improves the accuracy of target tracking.

[0055] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 The technical roadmap of the vehicle multi-target tracking method applicable to the highway tunnel monitoring scene of the present invention;

[0057] Figure 2 The overall flow chart of the vehicle multi-target tracking method of the present invention is applicable to the highway tunnel monitoring scene;

[0058] Figure 3 Design diagram of the prediction algorithm under different trajectory states. DETAILED DESCRIPTION

[0059] In order to make the technical solutions, advantages and purposes of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of this application.

[0060] like Figure 1 and Figure 2 As shown, the present invention provides a vehicle multi-target tracking method suitable for a highway tunnel monitoring scenario, comprising the following steps:

[0061] S1. Use the target detection model to detect the vehicle target of the current frame, and classify the vehicle target detection results and vehicle target trajectories;

[0062] The vehicle target detection results are divided into: high confidence detection results D high And low confidence detection results D low ;

[0063] The confidence test results are judged based on the low score threshold and the high score threshold;

[0064] Among them, the low score threshold θ L =0.2, high score threshold θ H =0.6;

[0065] The detection score of the vehicle target Score ≥ θ H D is a high confidence detection result high ;

[0066] The detection score of the vehicle target θ L ≤Score<θ H D is a low confidence detection result low ;

[0067] The vehicle target trajectory is divided into: confirmed trajectory T confirmed and unconfirmed trajectory T unconfirmed ;

[0068] Confirmed track T confirmed Divided into: Confirmed tracks T in the activated state confirmed and the confirmed trajectory T of the lost state confirmed ;

[0069] For the detection result of the first frame of the video, directly use the high confidence detection result D high Updated to confirmed track T confirmed , and set it to the active state Tracked;

[0070] For the detection results of non-first frames of the video, the high confidence detection result D of the current frame high The first occurrence will be marked as an unconfirmed track T unconfirmed , it is necessary to successfully match the vehicle target detection result again in the next frame to obtain the unconfirmed trajectory T unconfirmed Convert to confirmed trajectory T confirmed .

[0071] S2. The high confidence detection result D in step S1 high and all confirmed trajectories T confirmed Matching is performed to obtain the successfully matched association pair 1, the unmatched high confidence detection results and the unmatched trajectories, such as Figure 3 As shown;

[0072] S2.1 Design prediction algorithm for confirmed trajectory T confirmed Make predictions and obtain predicted values;

[0073] In step S2.1, the prediction algorithm has two different prediction branches:

[0074] Prediction branch 1: Confirmed trajectory T confirmed When in the active state, Kalman filtering is used for motion prediction;

[0075] Its state prediction equation and covariance prediction equation are as follows:

[0076]

[0077] P k|k-1 =F k .P k-1|k-1 .F k T +Q k

[0078] In the formula, x represents the state vector; and Respectively represent the mean value of the trajectory predicted in the current frame and the mean value of the previous frame; F k is the state transfer matrix, describing how the state changes from the previous frame to the current frame; P represents the covariance matrix; P k|k-1 and P k-1|k-1 They represent the covariance of the trajectory predicted in the current frame and the covariance of the previous frame respectively; Q k is the process noise covariance matrix, which represents the impact of system model uncertainty on the prediction.

[0079] Prediction branch 2: Confirmed trajectory T confirmed When in lost state, check confirmed track T confirmed If the length of the historical trajectory data is greater than 4, the Cubic Spline interpolation method is used for trajectory prediction, otherwise the Kalman filter is used for motion prediction;

[0080] Assume that the trajectory has n+1 historical data: (x0, y0), (x1, y1), ..., (x n ,y n ), where x0<x1<...<x n , x and y represent the horizontal and vertical coordinates of each data point. Then a cubic polynomial S can be generated between each adjacent data point i (x),

[0081] S i (x) = a i +b i (xx i )+c i (xx i )2 +d i (xx i ) 3

[0082] Cubic polynomial S i (x) in the interval [x i ,x i+1 ], where x i ≤x≤x i+1 .a i ,b i ,c i ,d i are the coefficients to be determined. To determine these coefficients, S i (x i ) must satisfy the following interpolation and smoothing conditions:

[0083]

[0084]

[0085] Thus, it is ensured that the data value predicted by the Cubic Spline interpolation method in this interval is on the same curve as the historical data value before this trajectory, and at the same time, it is ensured that the trajectory curve predicted in the adjacent interval is continuous and smooth.

[0086] S2.2 Use IoU to calculate the predicted value and high confidence detection result D high The similarity between them generates the distance matrix Dis;

[0087] In step S2.2, the calculation expression of the distance matrix Dis is as follows:

[0088] Dis=1-IoU

[0089]

[0090] Where: a is the high confidence detection result D high The bounding box of the confirmed trajectory T confirmed The bounding boxes obtained after prediction; Area(a) and Area(b) represent the areas of a and b respectively.

[0091] S2.3 Using the Hungarian algorithm, the distance matrix Dis is used to convert the high confidence detection results D high With confirmed trajectory T confirmed Matching is performed to obtain successfully matched association pairs 1, unmatched high-confidence detection results, and unmatched trajectories.

[0092] S3. The low confidence detection result D in step S1 is lowMatching with the trajectory that is in the activated state and not successfully matched in step S2, so as to obtain the successfully matched association pair 2, the unsuccessfully matched high confidence detection result and the unsuccessfully matched trajectory;

[0093] S3.1 uses CDIoU to calculate the active and unmatched trajectory in step S2 and the low confidence detection result D in step S1 low The similarity between them generates the distance matrix Dis CD ;

[0094] For low confidence detection results D low Occlusion causes changes in the appearance of vehicle targets, and improves the motion similarity matching method in data association, using CDIoU;

[0095] Distance matrix Dis CD The calculation expression is as follows:

[0096] Dis CD =1-CDIoU

[0097] CDIoU=IoU+λ(1-diou)

[0098]

[0099] Where: IoU represents the low confidence detection result D in step S1 low and the intersection-over-union ratio of the trajectories that are active and not successfully matched in step S2; λ is the weight coefficient; diou represents the distance metric; DL represents the low confidence detection result D low ; TH represents the bounding box of the trajectory that is active and not successfully matched in step S2; ||DL-TH||2 represents the sum of the Euclidean distances of the four vertices of the two bounding boxes; MBR represents the minimum matrix surrounding the two bounding boxes DL and TH; Diagonal(MBR) represents the diagonal length of the MBR matrix. The best effect can be obtained when the weight coefficient λ = 0.001.

[0100] S3.2 Using the Hungarian algorithm, the distance matrix Dis CD The low confidence detection result D in step S1 low Matching is performed with the trajectory that is in the activated state and has not been successfully matched in step S2, so as to obtain the successfully matched association pair 2, the unsuccessfully matched low-confidence detection result and the unsuccessfully matched trajectory.

[0101] S4. Combine the unmatched high confidence detection results in step S2 and the unconfirmed trajectory T in step S1 unconfirmed Matching is performed to obtain successfully matched association pairs 3, unmatched high-confidence detection results and unmatched trajectories;

[0102] S4.1 Use IoU to calculate the unconfirmed track T in step S1 unconfirmed The distance matrix between the high confidence detection results that were not successfully matched in step S2;

[0103] S4.2 Using the Hungarian algorithm, the distance matrix calculated in step S4.1 is used to match the unmatched high-confidence detection results in step S2 and the unconfirmed trajectories T in step S1. unconfirmed To allocate, take the threshold θ of the allocation algorithm thresh = 0.7, the best effect can be obtained, thus obtaining the successfully matched association pair 3, the unmatched high-confidence detection results and the unmatched trajectories.

[0104] S5. Update the matching results of steps S2 to S4, and repeat steps S1 to S4 and the update of step S5;

[0105] S5.1 uses Kalman filtering to update the trajectories of association pairs 1, association pairs 2, and association pairs 3 that are successfully matched in steps S2 to S4; specifically, the gain calculation, state update, and covariance update equations of the Kalman filter are as follows:

[0106]

[0107]

[0108] P k|k =(IK k H k ) k|k-1

[0109] In the formula, H k is the observation matrix, which represents the mapping of the mean vector of the associated center trajectory to the detection space; R k is the noise matrix, which is used to describe the uncertainty in the measurement process; K k is the Kalman gain; z k Represents the mean vector of detection results in the associated pairs.

[0110] S5.2 updates the unmatched trajectories in step S3 to a lost state. The lost trajectories can be retained for 60 frames. If the number of frames exceeds 60, the trajectories are updated to a removed state. The unmatched low-confidence detection results in step S3 are deleted.

[0111] S5.3 In step S4, if no trajectory T is confirmed in the association pair Match3 unconfirmed If the matching is successful twice in a row, it will be updated to the confirmed trajectory T in the activated state. confirmed ; The high confidence detection results that are not matched successfully are updated to new trajectories, that is, unconfirmed trajectories Tunconfirmed ; Unmatched trajectories are updated to removed status;

[0112] S5.4 Repeat steps S1 to S4 and steps S5.1 to S5.3 to implement the design of a vehicle multi-target tracking method suitable for highway tunnel monitoring scenarios, that is, construct a vehicle multi-target tracking model in the tunnel to implement the design of a vehicle multi-target tracking method suitable for highway tunnel monitoring scenarios.

[0113] S6. Create a highway tunnel vehicle target tracking dataset, and use the dataset together with the target detection model to train a multi-target tracking model for vehicles in the tunnel.

[0114] S6.1 Create a highway tunnel vehicle target tracking dataset;

[0115] Collect video data in highway tunnels and use DarkLabel software to annotate the data;

[0116] The annotation work included about 15,000 frames of video, covering three types of vehicles: cars, buses and trucks, a total of about 160 vehicles. In order to ensure the diversity of training data, clips were selected for annotation in different scenes of the tunnel, including tunnel curves, straight sections, low-light environments, and exit strong-light environments.

[0117] S6.2 Use the target detection model to train the multi-target tracking model for vehicles in the tunnel on the prepared dataset, use the Adam optimizer to optimize the model loss, and obtain and save the model weights;

[0118] S6.3 Load weights into the vehicle multi-target tracking model in the tunnel to implement frame-by-frame vehicle multi-target tracking of the video stream data input into the tunnel.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the protection scope of the present invention.

Claims

1. A vehicle multi-target tracking method suitable for highway tunnel monitoring scenarios, characterized in that: The following steps are involved: S1. Use the target detection model to detect the vehicle target of the current frame, and classify the vehicle target detection results and vehicle target trajectories; The vehicle target detection results are divided into: high confidence detection results D high And low confidence detection results D low ; The vehicle target trajectory is divided into: confirmed trajectory T confirmed and unconfirmed trajectory T unconfirmed ; Confirmed track T confirmed Divided into: Confirmed tracks T in the activated state confirmed and the confirmed trajectory T of the lost state confirmed ; S2. The high confidence detection result D in step S1 high and all confirmed trajectories T confirmed Matching is performed to obtain the successfully matched association pair 1, the unmatched high confidence detection result and the unmatched trajectory; Two different prediction branches are designed for the confirmed trajectory T confirmed Make a prediction and get the predicted value: Prediction branch 1: Confirmed trajectory T confirmed When in the active state, Kalman filtering is used for motion prediction; Prediction branch 2: Confirmed trajectory T confirmed When in lost state, check confirmed track T confirmed If the length of the historical trajectory data is greater than 4, the Cubic Spline interpolation method is used for trajectory prediction, otherwise the Kalman filter is used for motion prediction; S3. The low confidence detection result D in step S1 is low Matching with the trajectory that is in the activated state and not successfully matched in step S2, thereby obtaining the successfully matched association pair 2, the unmatched high confidence detection result and the unmatched trajectory, wherein the IoU is used to calculate the similarity during matching; S4. Combine the unmatched high confidence detection results in step S2 and the unconfirmed trajectory T in step S1 unconfirmed Matching is performed to obtain successfully matched association pairs 3, unmatched high-confidence detection results and unmatched trajectories; S5. Update the matching results of steps S2 to S4, repeat steps S1 to S4 and the update of step S5, that is, construct a multi-target vehicle tracking model in the tunnel; S6. Create a highway tunnel vehicle target tracking dataset, and use the dataset together with the target detection model to train a multi-target tracking model for vehicles in the tunnel.

2. The vehicle multi-target tracking method applicable to highway tunnel monitoring scenarios according to claim 1 is characterized in that: In step S1, the confidence test result is judged by a low score threshold and a high score threshold; Among them, the low score threshold θ L =0.2, high score threshold θ H =0.6; The detection score of the vehicle target Score ≥ θ H D is a high confidence detection result high ; The detection score of the vehicle target θ L ≤Score<θ H D is a low confidence detection result low .

3. The vehicle multi-target tracking method applicable to highway tunnel monitoring scenarios according to claim 1 is characterized in that: The step S2 comprises the following sub-steps: S2.1 Design prediction algorithm for confirmed trajectory T confirmed Make predictions and obtain predicted values; S2.2 Use IoU to calculate the predicted value and high confidence detection result D high The similarity between them generates the distance matrix Dis; S2.3 Using the Hungarian algorithm, the distance matrix Dis is used to convert the high confidence detection results D high With confirmed trajectory T confirmed Matching is performed to obtain successfully matched association pairs 1, unmatched high-confidence detection results, and unmatched trajectories.

4. The vehicle multi-target tracking method applicable to highway tunnel monitoring scenarios according to claim 2 is characterized in that: In step S2.2, the calculation expression of the distance matrix Dis is as follows: Dis=1-IoU Where: a is the high confidence detection result D high The bounding box of the confirmed trajectory T confirmed The bounding boxes obtained after prediction; Area(a) and Area(b) represent the areas of a and b respectively.

5. The vehicle multi-target tracking method applicable to highway tunnel monitoring scenarios according to claim 2 is characterized in that: The step S3 comprises the following sub-steps: S3.1 uses CDIoU to calculate the active and unmatched trajectory in step S2 and the low confidence detection result D in step S1 low The similarity between them generates the distance matrix Dis CD ; S3.2 Using the Hungarian algorithm, the distance matrix Dis CD The low confidence detection result D in step S1 low Matching is performed with the trajectory that is in the activated state and has not been successfully matched in step S2, so as to obtain the successfully matched association pair 2, the unsuccessfully matched low-confidence detection result and the unsuccessfully matched trajectory.

6. The vehicle multi-target tracking method applicable to highway tunnel monitoring scenarios according to claim 5 is characterized by: In step S3.1, the distance matrix Dis CD The calculation expression is as follows: Dis CD =1-CDIOU CDIoU=IoU+λ(1-diou) Where: IoU represents the low confidence detection result D in step S1 low and the intersection-over-union ratio of the trajectories that are active and not successfully matched in step S2; λ is the weight coefficient; diou represents the distance metric; DL represents the low confidence detection result D low ; TH represents the bounding box of the trajectory that is activated and not successfully matched in step S2; ||DL-TH||2 represents the sum of the Euclidean distances of the four vertices of the two bounding boxes; MBR represents the minimum matrix surrounding the two bounding boxes DL and TH; Diagonal(MBR) represents the diagonal length of the MBR matrix.

7. The vehicle multi-target tracking method applicable to highway tunnel monitoring scenarios according to claim 5 is characterized by: The step S4 comprises the following sub-steps: S4.1 Use IoU to calculate the unconfirmed track T in step S1 unconfirmed The distance matrix between the high confidence detection results that were not successfully matched in step S2; S4.2 Using the Hungarian algorithm, the distance matrix calculated in step S4.1 is used to match the unmatched high-confidence detection results in step S2 and the unconfirmed trajectories T in step S1. unconfirmed Matching is performed to obtain successfully matched association pairs 3, unmatched high-confidence detection results, and unmatched trajectories.

8. The vehicle multi-target tracking method applicable to highway tunnel monitoring scenarios according to claim 7 is characterized by: The step S5 comprises the following sub-steps: S5.1 uses Kalman filtering to update the trajectories of association pair 1, association pair 2 and association pair 3 that are successfully matched in steps S2 to S4; S5.2 updates the unmatched trajectories in step S3 to a lost state. The lost trajectories can be retained for 60 frames. If the number of frames exceeds 60, the trajectories are updated to a removed state. The unmatched low-confidence detection results in step S3 are deleted. S5.3 Update the successfully matched association pair 3 in step S4 to the confirmed trajectory T in the activated state confirmed ; The high confidence detection results that are not matched successfully are updated to new trajectories, that is, unconfirmed trajectories T unconfirmed ; Unmatched trajectories are updated to removed status; S5.4 repeats the contents of steps S1 to S4 and steps S5.1 to S5.

3.

9. The vehicle multi-target tracking method applicable to highway tunnel monitoring scenarios according to claim 8 is characterized by: The step S6 comprises the following sub-steps: S6.1 Create a highway tunnel vehicle target tracking dataset; Collect video data in highway tunnels and use DarkLabel software to annotate the data; S6.2 Use the target detection model to train the multi-target tracking model for vehicles in the tunnel on the prepared dataset, use the Adam optimizer to optimize the model loss, and obtain and save the model weights; S6.3 Loading weights for the multi-target tracking model of vehicles in tunnels.

Images