Multi-target tracking system and method combined with selection mechanism

Abstract

The invention discloses a multi-target tracking system and method combined with a selection mechanism, and belongs to the technical field of computer vision, the multi-target tracking system comprises a detection processing module and a track processing module, the detection processing module and the track processing module are simultaneously connected with a matching association module, and the matching association module is connected with a track management module. The modules are used for obtaining target information from an initial frame video image to obtain an initial track; obtaining a target detection frame in the current video frame, and dividing the detection frame into a high-confidence group and a low-confidence group; obtaining a trajectory prediction frame in the current video frame; respectively matching and associating the two groups of detection frames with the track, further comparing, and determining an optimal matching pair; updating an active track state; recovering the lost track; processing unmatched detection frames and tracks; and displaying the tracking result in the video. According to the method, the accuracy and robustness of matching association can be effectively improved, and the re-association effect after the track is lost is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a multi-target tracking system and method combined with a selection mechanism. Background Art

[0002] Multi-target tracking is an important task in computer vision, which is applied in fields such as video surveillance, autonomous driving, and sports analysis. Its core is to identify and track multiple targets in video frames. The multi-target tracking system using the "detection-tracking" mode divides the task into two steps: target detection and target association. The traditional "two-stage matching" strategy usually prioritizes the use of high-confidence detection boxes for matching, and then uses low-confidence detection boxes to supplement. However, this matching strategy relies on the confidence of the detection box, and prioritizes matching high-confidence detection boxes, which may lead to incorrect associations. Even if the low-confidence detection box is more accurate, the high-confidence detection will still be selected, which affects the association accuracy of subsequent frames.

[0003] In addition, in scenes where the target is occluded, the detection frame often cannot accurately reflect the state of the target. Unscreened low-quality detection frames will have a negative impact on the tracker state update and reduce tracking accuracy.

[0004] Another common problem is re-association after track loss. Occlusion may cause the target to be temporarily lost. Although some studies have tried to recover the track using high-confidence detection boxes, the difficulty of re-association increases because the target may be partially visible or have low confidence. The longer the track is lost, the more difficult it is to recover.

[0005] To address these issues, it is necessary to propose a more robust matching strategy to evaluate the reliability of the detection box in order to improve the accuracy and robustness of multi-target tracking. Summary of the invention

[0006] The technical problem to be solved by the present invention is to overcome the suboptimal matching problem caused by two-stage matching in traditional multi-target tracking methods, the problem that detection fluctuations affect the accuracy of the Kalman filter, and the problem that lost targets are difficult to recover, and to provide a multi-target tracking system and method combined with a selection mechanism to improve the accuracy and robustness of matching associations, enhance the prediction accuracy of the Kalman filter and the re-association effect after trajectory loss.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is: a multi-target tracking system combined with a selection mechanism, including a detection processing module and a trajectory processing module arranged side by side, the detection processing module and the trajectory processing module are simultaneously connected to a matching association module, and the matching association is connected to a trajectory management module.

[0008] The further improvement of the technical solution of the present invention is that each module is specifically composed as follows:

[0009] Detection processing module: including a target detection unit and a detection frame classification unit composed of detectors connected in sequence, wherein the detector is trained using target images in different scenes before use to obtain the detector required for multi-target tracking;

[0010] Trajectory processing module: including a trajectory prediction unit composed of a Kalman filter and a camera motion compensation unit connected in sequence;

[0011] Matching association module: including a high-confidence detection box matching unit and a low-confidence detection box matching unit arranged side by side, and a detection selection unit is connected to the high-confidence detection box matching unit and the low-confidence detection box matching unit, wherein the high-confidence detection box matching unit and the low-confidence detection box matching unit both have built-in association cost calculation functions based on IoU (intersection over union) and linear allocation operations based on the Hungarian matching algorithm, and the detection selection unit has a built-in comparison function to compare and calculate the association cost;

[0012] Trajectory management module: includes a state selection unit, a recovery selection unit, an unmatched trajectory management unit and an unmatched detection frame management unit arranged side by side, wherein the state selection unit has a built-in calculation function, the recovery selection unit has a built-in calculation function, the unmatched trajectory management unit has a built-in logic judgment function for the trajectory loss time, and the unmatched detection frame management unit has a built-in logic judgment function for the detection frame confidence.

[0013] A multi-target tracking method combined with a selection mechanism, the specific steps are as follows:

[0014] Step 1: Obtain target information from the initial frame video image and obtain the initial trajectory;

[0015] Step 2: Get the target detection frame in the current video frame and divide the detection frame into a high confidence group and a low confidence group;

[0016] Step 3: Get the initial prediction frame of the trajectory in the current video frame, and obtain the trajectory prediction frame after further processing;

[0017] Step 4: Match and associate the two sets of detection boxes with the trajectories respectively, and further compare them to determine the best matching pair;

[0018] Step 5: Update the active track status;

[0019] Step 6: Recover lost tracks;

[0020] Step 7: Process unmatched detection frames and trajectories;

[0021] Step 8: Display the tracking results in the video.

[0022] The further improvement of the technical solution of the present invention is that step 1 is specifically as follows:

[0023] Step 1.1: Obtain the detection box information of each target from the initial frame video image, including the location information and confidence of the detection box;

[0024] Step 1.2: Confidence screening is performed on all detection boxes. According to the given threshold, high-confidence detection boxes are selected and initialized as trajectories.

[0025] The further improvement of the technical solution of the present invention is that step 2 is specifically as follows:

[0026] Step 2.1: Obtain the target detection frame in the current video frame through the detector;

[0027] Step 2.2: Divide all detection boxes into high-confidence detection boxes and low-confidence detection boxes according to a certain confidence threshold.

[0028] The further improvement of the technical solution of the present invention is that step 3 is specifically as follows:

[0029] Step 3.1: Use the Kalman filter to predict the next position of all activated trajectories and construct the position information of the trajectories;

[0030] Step 3.2: Perform camera motion compensation on the position of the prediction box to ensure the accuracy of the detection box and trajectory matching under camera motion conditions, and output the final trajectory prediction box.

[0031] The further improvement of the technical solution of the present invention is: Step 4.1: inputting the high confidence detection frame and the trajectory prediction frame into the high confidence detection frame matching unit, and calculating the association cost between the high confidence detection frame and the trajectory through the association cost calculation function;

[0032] Step 4.2: Construct a cost matrix between the high-confidence detection box and the trajectory through the associated cost between the high-confidence detection box and the trajectory, and match the high-confidence detection box and the trajectory according to the cost matrix between the high-confidence detection box and the trajectory based on the linear assignment operation of the Hungarian matching algorithm to obtain a preliminary high-confidence detection box-trajectory matching pair;

[0033] Step 4.3: Input the low-confidence detection box and the trajectory prediction box into the low-confidence detection box matching unit, and calculate the association cost between the low-confidence detection box and the trajectory through the association cost calculation function;

[0034] Step 4.4: Construct a cost matrix between the low-confidence detection box and the trajectory through the association cost between the low-confidence detection box and the trajectory, and match the low-confidence detection box and the trajectory according to the cost matrix between the low-confidence detection box and the trajectory based on the linear assignment operation of the Hungarian matching algorithm to obtain a preliminary low-confidence detection box-trajectory matching pair;

[0035] Step 4.5: Traverse the high-confidence detection box-track matching pairs and the low-confidence detection box-track matching pairs to find two matching pairs of the same track;

[0036] Step 4.6: The association costs of the two matching pairs of the same trajectory are input into the detection selection unit. In the detection selection unit, the association cost of the low-confidence detection box-trajectory matching pair is multiplied by a given coefficient and then compared with the association cost of the high-confidence detection box-trajectory matching pair. If the result is less than the association cost of the high-confidence detection box-trajectory matching pair, the low-confidence detection box-trajectory matching pair is taken as the optimal matching pair, and the low-confidence detection box in the low-confidence detection box-trajectory matching pair is taken as the optimal matching detection box. Otherwise, the high-confidence detection box-trajectory matching pair is taken as the optimal matching pair, and the high-confidence detection box in the high-confidence detection box-trajectory matching pair is taken as the optimal matching detection box.

[0037] The further improvement of the technical solution of the present invention is: Step 5.1: Calculate the ratio of the intersection area of ​​the detection frame in the best matching pair and other high-confidence detection frames to the area of ​​the detection frame in the best matching pair, take the maximum area ratio as the occlusion degree value, if the occlusion degree value exceeds a certain threshold and the confidence of the detection frame in the best matching pair is lower than a certain threshold, use the trajectory prediction frame to update the trajectory;

[0038] Step 5.2: Calculate the height ratio between the detection box and the trajectory prediction box in the best matching pair as the shape change value. If the shape change value is greater than a certain threshold, use the trajectory prediction box to update the trajectory;

[0039] Step 5.3: If the confidence of the detection box in the best matching pair is lower than a certain threshold, the trajectory prediction box is used to update the trajectory. In other cases, the detection box in the best matching pair is used to update the trajectory.

[0040] The further improvement of the technical solution of the present invention is that the specific steps of step 6 are as follows:

[0041] Step 6.1: If the best matching pair is a high confidence detection box-track matching pair, the detection box in the best matching pair is used to restore the lost track;

[0042] Step 6.2: If the best matching pair is a low-confidence detection frame-trajectory matching pair, the occlusion degree value of the detection frame in the best matching pair is calculated. When the occlusion degree value is lower than the determined threshold and the confidence of the detection frame in the best matching pair is higher than the determined threshold, the detection frame is used to restore the lost trajectory, otherwise the lost trajectory is not restored.

[0043] The further improvement of the technical solution of the present invention is that the specific steps of step 7 are as follows:

[0044] Step 7.1: For unmatched trajectories whose loss time exceeds a certain threshold, delete them, and unmatched trajectories that do not exceed a certain threshold are marked as lost;

[0045] Step 7.2: For unmatched detection boxes, if the confidence is higher than the initialization threshold, initialize it as a track;

[0046] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is: through innovative modules and a two-stage selection matching strategy, the optimal detection frame and trajectory matching can be selected more accurately, reducing the suboptimal matching situation in the traditional method. Before updating the trajectory based on the matching result, the state selection unit screens the detection frame based on multi-dimensional conditions such as occlusion degree, shape difference and confidence, ensuring that only qualified detection frames are used to update the trajectory. The introduction of the state selection module effectively avoids erroneous updates caused by target occlusion or degradation of the detection frame quality, thereby maintaining the continuity and accuracy of the trajectory and further enhancing the stability of the system in dynamic scenes. The recovery selection unit is specifically used to handle the re-association of lost trajectories. By screening the occlusion degree and confidence of the low-confidence detection frame, the qualified detection frame is used to restore the lost trajectory, effectively improving the success rate of re-association. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work.

[0048] Figure 1 It is a schematic diagram of the structure of the multi-target tracking system of the present invention;

[0049] Figure 2 It is a flow chart of the multi-target tracking method of the present invention;

[0050] Figure 3 It is a schematic diagram of using the frame multi-target tracking method of the present invention to track multiple targets. DETAILED DESCRIPTION

[0051] The present invention is further described in detail below in conjunction with embodiments:

[0052] like Figure 1 The figure is a schematic diagram of the structure of a multi-target tracking system combined with a selection mechanism, including a detection processing module and a trajectory processing module arranged side by side, a matching association module is connected to the detection processing module and the trajectory processing module, and a trajectory management module is connected to the matching association. The specific structure of each module is as follows:

[0053] Detection processing module: includes a target detection unit and a detection frame classification unit composed of detectors connected in sequence, wherein the detector is trained using target images in different scenes before use to obtain the detector required for multi-target tracking.

[0054] Trajectory processing module: includes a trajectory prediction unit composed of a Kalman filter and a camera motion compensation unit connected in sequence.

[0055] Matching association module: includes a high-confidence detection box matching unit and a low-confidence detection box matching unit arranged side by side, and a detection selection unit is commonly connected to the high-confidence detection box matching unit and the low-confidence detection box matching unit, wherein the high-confidence detection box matching unit and the low-confidence detection box matching unit both have built-in association cost calculation functions based on IoU (intersection over union) and linear allocation operations based on the Hungarian matching algorithm. The association cost calculation function calculates the association cost between the detection box and the trajectory based on the intersection over union between the detection box and the trajectory, and the linear allocation operation is based on the Hungarian matching algorithm according to the association cost between the detection box and the trajectory to achieve the optimal linear matching between the detection box and the trajectory. The detection selection unit has a built-in comparison function that can compare and calculate the association cost.

[0056] Track management module: including a state selection unit, a recovery selection unit, an unmatched track management unit and an unmatched detection frame management unit arranged side by side, wherein the built-in calculation function of the state selection unit can further make an association selection between the best matching detection frame and the prediction frame of the matching track according to the degree of occlusion of the best matching detection frame determined in the detection selection unit by other high-confidence detection frames, the degree of shape change between the track matched by the best matching detection frame and the best matching detection frame, and the detection confidence of the best matching detection frame, and the built-in calculation function of the recovery selection unit can further determine whether to use the best matching detection frame to restore the track lost according to the degree of occlusion of the best matching detection frame by other high-confidence detection frames and the confidence of the best matching detection frame. The unmatched track management unit has a built-in logic judgment function for the track loss time, which is used to determine whether to allow the unmatched track to continue to participate in the next detection frame-track matching association process, and the unmatched detection frame management unit has a built-in logic judgment function for the detection frame confidence, which is used to determine whether to initialize the unmatched detection frame as a new track.

[0057] like Figure 2 As shown, with the help of the above multi-target tracking system, a multi-target tracking method combined with a selection mechanism can be implemented. The specific steps are as follows:

[0058] Step 1: Obtain target information from the initial frame video image and obtain the initial trajectory;

[0059] Step 1.1: Obtain the detection box information of each target from the initial frame video image, including the location information and confidence of the detection box;

[0060] Step 1.2: Confidence screening is performed on all detection boxes, and high-confidence detection boxes are selected according to a given threshold, such as 0.7, and the selected high-confidence detection boxes are initialized as trajectories.

[0061] Step 2: Get the target detection frame in the current video frame and divide the detection frame into a high confidence group and a low confidence group;

[0062] Step 2.1: Input the current video frame into the detector, and obtain the target detection frame in the current video frame through the detector;

[0063] Step 2.2: All target detection frames are input into the detection frame classification unit and divided into high-confidence detection frames and low-confidence detection frames according to a certain confidence threshold, such as 0.7.

[0064] Step 3: Get the initial prediction frame of the trajectory in the current video frame, and obtain the trajectory prediction frame after further processing;

[0065] Step 3.1: Use step 1 to obtain the multi-target tracking trajectory, input the multi-target tracking trajectory into the trajectory prediction unit, use the Kalman filter to predict the next position of all activated trajectories, and construct the initial trajectory prediction frame;

[0066] Step 3.2: Input the initial trajectory prediction frame into the camera motion compensation unit, perform camera motion compensation on the position of the initial trajectory prediction frame, and output the final trajectory prediction frame.

[0067] Step 4: Match and associate the high-confidence detection box and the low-confidence detection box with the trajectory respectively, and further compare them to determine the best matching pair;

[0068] Step 4.1: Input the high confidence detection box and the trajectory prediction box into the high confidence detection box matching unit, and calculate the association cost between the high confidence detection box and the trajectory through the association cost calculation function;

[0069] Step 4.2: Construct a cost matrix between the high-confidence detection box and the trajectory through the associated cost between the high-confidence detection box and the trajectory, and match the high-confidence detection box and the trajectory according to the cost matrix between the high-confidence detection box and the trajectory based on the linear assignment operation of the Hungarian matching algorithm to obtain a preliminary high-confidence detection box-trajectory matching pair;

[0070] Step 4.3: Input the low-confidence detection box and the trajectory prediction box into the low-confidence detection box matching unit, and calculate the association cost between the low-confidence detection box and the trajectory through the association cost calculation function;

[0071] Step 4.4: Construct a cost matrix between the low-confidence detection box and the trajectory through the association cost between the low-confidence detection box and the trajectory, and match the low-confidence detection box and the trajectory according to the cost matrix between the low-confidence detection box and the trajectory based on the linear assignment operation of the Hungarian matching algorithm to obtain a preliminary low-confidence detection box-trajectory matching pair;

[0072] Step 4.5: Traverse the high-confidence detection box-track matching pairs and the low-confidence detection box-track matching pairs to find two matching pairs of the same track;

[0073] Step 4.6: The association costs of the two matching pairs of the same trajectory are input into the detection selection unit. In the detection selection unit, the association cost of the low-confidence detection box-trajectory matching pair is multiplied by a given coefficient, such as 1.3, and then compared with the association cost of the high-confidence detection box-trajectory matching pair. If the result is less than the association cost of the high-confidence detection box-trajectory matching pair, the low-confidence detection box-trajectory matching pair is taken as the optimal matching pair, and the low-confidence detection box in the low-confidence detection box-trajectory matching pair is taken as the optimal matching detection box. Otherwise, the high-confidence detection box-trajectory matching pair is taken as the optimal matching pair, and the high-confidence detection box in the high-confidence detection box-trajectory matching pair is taken as the optimal matching detection box.

[0074] Step 5: If the trajectory in the best matching pair is an active trajectory (a trajectory that is continuously tracked and not lost), the best matching pair is input to the state selection unit to update the active trajectory state.

[0075] Step 5.1: Calculate the ratio of the intersection area of ​​the detection box in the best matching pair and other high-confidence detection boxes to the area of ​​the detection box in the best matching pair, and take the maximum area ratio as the occlusion degree value. If the occlusion degree value exceeds a certain threshold, such as 0.9, and the confidence of the detection box in the best matching pair is lower than a certain threshold, such as 0.3, then use the trajectory prediction box to update the trajectory;

[0076] Step 5.2: Calculate the height ratio of the detection box and the trajectory prediction box in the best matching pair as the shape change value. If the shape change value is greater than a certain threshold, such as 1.2, the trajectory prediction box is used to update the trajectory;

[0077] Step 5.3: If the confidence of the detection box in the best matching pair is lower than a certain threshold, such as 0.2, the trajectory prediction box is used to update the trajectory. In other cases, the detection box in the best matching pair is used to update the trajectory.

[0078] Step 6: If the trajectory in the best matching pair is the lost trajectory, the best matching pair is input into the recovery selection unit to recover the lost trajectory;

[0079] Step 6.1: If the best matching pair is a high confidence detection box-track matching pair, the detection box in the best matching pair is used to restore the lost track;

[0080] Step 6.2: If the best matching pair is a low confidence detection box-track matching pair, calculate the occlusion degree value of the detection box in the best matching pair. When the occlusion degree value is lower than a certain threshold, such as 0.9, and the confidence of the detection box in the best matching pair is higher than a certain threshold, such as 0.3, use the detection box to restore the lost track. Otherwise, do not restore the lost track.

[0081] Step 7: Process unmatched detection frames and trajectories;

[0082] Step 7.1: For unmatched trajectories, the unmatched trajectory management unit determines the loss time of the unmatched trajectory. If the loss time exceeds a certain threshold, such as 30 frames, the unmatched trajectory is deleted to prevent affecting the tracking of active trajectories and the performance of the tracking system. Otherwise, the unmatched trajectory is marked as lost and continues to participate in the next detection box-trajectory matching association process;

[0083] Step 7.2: For the unmatched detection box, if the confidence of the unmatched detection box is higher than a given initialization threshold, such as 0.7, it is initialized as a trajectory;

[0084] Step 8: Display the tracking results in the video.

[0085] like Figure 3 FIG. 1 is a schematic diagram of using the multi-target tracking method of the present invention to track multiple targets. Figure 3 When the multi-target tracking method of the present invention is used for tracking, the target video is input into the multi-target tracking system combined with the selection mechanism, the system completes the tracking of multiple targets in the video, and visualizes the tracking results.

[0086] The present invention aims at the suboptimal matching problem caused by two-stage matching in the existing multi-target tracking method, the problem that the fluctuation of detection affects the accuracy of the Kalman filter, and the problem that lost targets are difficult to recover. A multi-target tracking method combined with a selection mechanism is proposed to effectively solve the above problems. By matching all trajectories with a high-confidence detection frame and a low-confidence detection frame, and then comparing the associated cost values ​​of successfully matching the two detection frames at the same time, the optimal matching pair is determined; by measuring the degree of occlusion, shape change and confidence of the detection frame, it is determined whether to update the trajectory with the detection frame or the prediction frame of the Kalman filter, so as to maintain the accuracy of the trajectory and the stability of the Kalman filter; by measuring the degree of occlusion and confidence of the detection frame, it is determined whether it can be used to recover the trajectory, so as to improve the recovery rate of the lost trajectory while ensuring the correct association of the trajectory. The collaborative work of different modules improves the continuity and accuracy of the trajectory.

[0087] The embodiments described above are merely descriptions of preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A multi-target tracking system combined with a selection mechanism, characterized in that: It comprises a detection processing module and a trajectory processing module which are arranged side by side. The detection processing module and the trajectory processing module are connected with a matching association module at the same time, and the matching association module is connected with a trajectory management module.

2. A multi-target tracking system combined with a selection mechanism according to claim 1, characterized in that: The specific components of each module are as follows: Detection processing module: including a target detection unit and a detection frame classification unit composed of detectors connected in sequence, wherein the detector is trained using target images in different scenes before use to obtain the detector required for multi-target tracking; Trajectory processing module: including a trajectory prediction unit composed of a Kalman filter and a camera motion compensation unit connected in sequence; Matching association module: including a high-confidence detection frame matching unit and a low-confidence detection frame matching unit arranged side by side, and a detection selection unit is connected to the high-confidence detection frame matching unit and the low-confidence detection frame matching unit, wherein the high-confidence detection frame matching unit and the low-confidence detection frame matching unit both have built-in association cost calculation functions based on intersection-over-union ratio and linear allocation operations based on the Hungarian matching algorithm, and the detection selection unit has a built-in comparison function to compare and calculate the association cost; Trajectory management module: includes a state selection unit, a recovery selection unit, an unmatched trajectory management unit and an unmatched detection frame management unit arranged side by side, wherein the state selection unit has a built-in calculation function, the recovery selection unit has a built-in calculation function, the unmatched trajectory management unit has a built-in logic judgment function for the trajectory loss time, and the unmatched detection frame management unit has a built-in logic judgment function for the detection frame confidence.

3. A multi-target tracking method combined with a selection mechanism, implemented based on the tracking system of claim 1 or 2, characterized in that: The specific steps are as follows: Step 1: Obtain target information from the initial frame video image and obtain the initial trajectory; Step 2: Get the target detection frame in the current video frame and divide the detection frame into a high confidence group and a low confidence group; Step 3: Get the initial prediction frame of the trajectory in the current video frame, and obtain the trajectory prediction frame after further processing; Step 4: Match and associate the two sets of detection boxes with the trajectories respectively, and further compare them to determine the best matching pair; Step 5: Update the active trajectory status; Step 6: Recover lost tracks; Step 7: Process unmatched detection frames and trajectories; Step 8: Display the tracking results in the video.

4. The multi-target tracking method combined with a selection mechanism according to claim 3, characterized in that: Step 1 is as follows: Step 1.1: Obtain the detection box information of each target from the initial frame video image, including the location information and confidence of the detection box; Step 1.2: Confidence screening is performed on all detection boxes. According to the given threshold, high-confidence detection boxes are selected and initialized as trajectories.

5. The multi-target tracking method combined with a selection mechanism according to claim 3, characterized in that: Step 2 is as follows: Step 2.1: Obtain the target detection frame in the current video frame through the detector; Step 2.2: Divide all detection boxes into high-confidence detection boxes and low-confidence detection boxes according to a certain confidence threshold.

6. The multi-target tracking method combined with a selection mechanism according to claim 3, characterized in that: Step 3 is as follows: Step 3.1: Use the Kalman filter to predict the next position of all activated trajectories and construct the position information of the trajectories; Step 3.2: Perform camera motion compensation on the position of the prediction box to ensure the accuracy of the detection box and trajectory matching under camera motion conditions, and output the final trajectory prediction box.

7. The multi-target tracking method combined with a selection mechanism according to claim 3, characterized in that: Step 4 The specific steps are as follows: Step 4.1: Input the high confidence detection box and the trajectory prediction box into the high confidence detection box matching unit, and calculate the association cost between the high confidence detection box and the trajectory through the association cost calculation function; Step 4.2: Construct a cost matrix between the high-confidence detection box and the trajectory through the associated cost between the high-confidence detection box and the trajectory, and match the high-confidence detection box and the trajectory according to the cost matrix between the high-confidence detection box and the trajectory based on the linear assignment operation of the Hungarian matching algorithm to obtain a preliminary high-confidence detection box-trajectory matching pair; Step 4.3: Input the low-confidence detection box and the trajectory prediction box into the low-confidence detection box matching unit, and calculate the association cost between the low-confidence detection box and the trajectory through the association cost calculation function; Step 4.4: Construct a cost matrix between the low-confidence detection box and the trajectory through the association cost between the low-confidence detection box and the trajectory, and match the low-confidence detection box and the trajectory according to the cost matrix between the low-confidence detection box and the trajectory based on the linear assignment operation of the Hungarian matching algorithm to obtain a preliminary low-confidence detection box-trajectory matching pair; Step 4.5: Traverse the high-confidence detection box-track matching pairs and the low-confidence detection box-track matching pairs to find two matching pairs of the same track; Step 4.6: The association costs of the two matching pairs of the same trajectory are input into the detection selection unit. In the detection selection unit, the association cost of the low-confidence detection box-trajectory matching pair is multiplied by a given coefficient and then compared with the association cost of the high-confidence detection box-trajectory matching pair. If the result is less than the association cost of the high-confidence detection box-trajectory matching pair, the low-confidence detection box-trajectory matching pair is taken as the optimal matching pair, and the low-confidence detection box in the low-confidence detection box-trajectory matching pair is taken as the optimal matching detection box. Otherwise, the high-confidence detection box-trajectory matching pair is taken as the optimal matching pair, and the high-confidence detection box in the high-confidence detection box-trajectory matching pair is taken as the optimal matching detection box.

8. The multi-target tracking method combined with a selection mechanism according to claim 3, characterized in that: Step 5 The specific steps are as follows: Step 5.1: Calculate the ratio of the intersection area of ​​the detection frame in the best matching pair and other high-confidence detection frames to the area of ​​the detection frame in the best matching pair, and take the maximum area ratio as the occlusion degree value. If the occlusion degree value exceeds a certain threshold and the confidence of the detection frame in the best matching pair is lower than a certain threshold, use the trajectory prediction frame to update the trajectory; Step 5.2: Calculate the height ratio between the detection box and the trajectory prediction box in the best matching pair as the shape change value. If the shape change value is greater than a certain threshold, use the trajectory prediction box to update the trajectory; Step 5.3: If the confidence of the detection box in the best matching pair is lower than a certain threshold, the trajectory prediction box is used to update the trajectory. In other cases, the detection box in the best matching pair is used to update the trajectory.

9. The multi-target tracking method combined with a selection mechanism according to claim 3, characterized in that: Step 6 The specific steps are as follows: Step 6.1: If the best matching pair is a high confidence detection box-track matching pair, the detection box in the best matching pair is used to restore the lost track; Step 6.2: If the best matching pair is a low-confidence detection frame-trajectory matching pair, the occlusion degree value of the detection frame in the best matching pair is calculated. When the occlusion degree value is lower than the determined threshold and the confidence of the detection frame in the best matching pair is higher than the determined threshold, the detection frame is used to restore the lost trajectory, otherwise the lost trajectory is not restored.

10. The multi-target tracking method combined with a selection mechanism according to claim 3, characterized in that: Step 7 The specific steps are as follows: Step 7.1: For unmatched trajectories whose loss time exceeds a certain threshold, delete them, and unmatched trajectories that do not exceed a certain threshold are marked as lost; Step 7.2: For unmatched detection boxes, if the confidence is higher than the initialization threshold, initialize it as a track.

Images