Multi-target tracking method based on threshold separation and feature enhancement

Abstract

The invention discloses a multi-target tracking method based on threshold separation and feature enhancement, and the method achieves the association between a target and a track in a current frame through a data association algorithm based on a target detection result through the motion features and appearance features of the target. The method comprises the following steps: dividing a high-score detection frame and a low-score detection frame according to a result of a detector and a detection confidence score threshold value; extracting motion features and appearance features of the high-score detection frame, and predicting positions of all tracks in the track set in the current frame; the first association occurs between a high-score detection frame and a track set, motion features and appearance features are fused and used, the second association occurs between a low-score detection frame and a track which is not matched, and only the motion features are used for calculating a cost matrix; trajectory management includes initializing new trajectories, deleting expired trajectories, and removing repeated trajectories. According to the method, a feature extraction mode and a feature fusion mode are improved, and the high-order accuracy of multi-target tracking is improved.

Description

Technical Field

[0001] The invention relates to the field of multi-target tracking, and more specifically to a multi-target tracking method based on threshold separation and feature enhancement. Background Art

[0002] The main task of multi-target tracking is to output the motion trajectories of all targets in a given video sequence and maintain the identity information of each target. Multi-target tracking technology has made significant breakthroughs in the fields of intelligent monitoring systems, autonomous driving, and sports event analysis, further exploring application scenarios and improving the intelligence level of security, transportation, and sports. In addition, multi-target tracking is the basis of advanced computer vision tasks such as posture estimation, behavior recognition, behavior analysis, and video analysis.

[0003] Compared with target detection, multi-target tracking needs to complete more tasks on the basis of obtaining all targets, including obtaining the target's motion information and appearance information and matching the target set and trajectory set based on the above information. Therefore, in the actual tracking process, there are a lot of challenging factors that will cause the tracking performance to degrade, such as target occlusion, high similarity between targets, interaction between targets, target scale change and illumination change. Among the above challenging factors, the difficulty of stable tracking in complex scenes can be attributed to the following three points: First, due to the occlusion of the target or the frequent interaction between targets, the target is difficult to track accurately, which leads to frequent changes in the target ID; second, the high similarity between targets makes the target feature information degraded or invalid; third, the target scale change, ambient light changes and camera movement may cause the target detection frame to drift.

[0004] With the continuous breakthroughs and widespread applications of deep learning in the field of computer vision, multi-target tracking methods based on deep learning have emerged in recent years. According to the tracking model adopted by the model to complete the three subtasks of target detection, feature extraction and data association, the multi-target tracking algorithms in recent years can be divided into separate detection and feature extraction methods (Separate Detection and Embedding, SDE), joint detection and feature extraction methods (Joint Detection and Embedding, JDE) and joint detection and tracking methods (Joint Detection and Tracking, JDT). Figure 1As shown in the figure, the multi-target tracking method based on separate detection and feature extraction completes three subtasks in succession, namely, firstly locating the target through a target detection network, then extracting the target's features, and finally calculating the similarity between targets and associating the targets through a data association algorithm. The multi-target tracking method based on separate detection and feature extraction has become the mainstream of research due to its simple structure and excellent tracking effect. SORT is one of the earliest multi-target tracking algorithms that uses convolutional neural networks to detect targets. It models the position and velocity of each target, and then uses the Hungarian Algorithm (HA) to associate the targets frame by frame based on the intersection over union (IoU) between the predicted box obtained by the Kalman Filter (KF) and the detection box obtained by Faster R-CNN. DeepSORT introduces the appearance information of the target on the basis of the SORT algorithm, uses the cosine distance and Mahalanobis distance to measure the similarity between targets, and uses the cascade matching strategy to greatly reduce the IDs (ID Switches) phenomenon. ByteTrack optimizes the multi-target tracking algorithm in terms of the matching process. Unlike previous tracking methods that discard low-scoring detectors at the target detection stage, ByteTrack retains all detection frames and divides them into high-scoring detection frames and low-scoring detection frames. ByteTrack first associates high-scoring detection frames with existing tracks, then associates low-scoring detection frames with unmatched tracks, and finally creates new tracks for unmatched high-scoring detection frames. ByteTrack makes full use of the detection results, making it robust to target occlusion.

[0005] like Figure 2 As shown in the figure, the multi-target tracking method based on joint detection and feature extraction adds a parallel feature extraction branch to the detection network, thereby outputting the location and appearance features of the target in a single network at the same time. By using the shared features of the two tasks, some repeated calculations are effectively avoided, and the tracking speed of the model is improved. Figure 3 As shown, the multi-target tracking method based on joint detection and tracking takes multiple adjacent frames of images as input, predicts the current position offset or appearance features of the target based on its previous motion or appearance information, and thus associates the target.

[0006] The present invention discloses a multi-target tracking method based on separation detection and feature extraction of threshold separation and feature improvement. On the basis of dividing the detection frame into a high-score detection frame and a low-score detection frame according to the detection confidence score threshold, the high-score detection frame not only has the right to match the trajectory set first, but also makes the following improvements in motion features and appearance features: First, according to the detection confidence score, the high-score detection frame is given a lower detection noise, thereby making its motion features more accurate; second, for the matched trajectory, the appearance feature update weight is calculated according to its detection confidence score, and the trajectory set matched by the high-score detection frame is given a higher update weight, and the appearance feature of the weight corresponding to the low-score detection frame is not updated. By improving the appearance features and motion features and optimizing the matching process, the algorithm is improved in terms of tracking accuracy and so on. Summary of the invention

[0007] In view of this, an embodiment of the present invention provides a multi-target tracking method based on threshold separation and feature enhancement to accurately and stably track targets.

[0008] In order to achieve the above object, the embodiment of the present invention provides the following solution:

[0009] A multi-target tracking method based on threshold separation and feature enhancement, characterized in that it comprises the following steps:

[0010] Step 1. Object Detection

[0011] The present invention uses the target detection network YOLOX as a detector to output all potential targets in each frame, and its output contains the following information: the confidence score of the target, the category ID of the target, and the bounding box coordinates of the target.

[0012] Step 2. Motion feature extraction and enhancement

[0013] In the field of multi-target tracking, Kalman filtering updates and predicts the target based on the observation equation and the state transfer equation. The prediction step obtains the prior estimate of the current moment from the posterior estimate of the previous moment:

[0014]

[0015] Then the covariance matrix of the prior error is obtained:

[0016]

[0017] In the update step, the Kalman gain coefficient is obtained from the covariance matrix:

[0018]

[0019] Then the posterior estimate is obtained from the Kalman gain coefficient:

[0020]

[0021] Finally update the error covariance matrix:

[0022] P k∣k =(IK k H k ) k∣k-1

[0023] In order to reduce the impact of detection noise of low confidence scores, the present invention uses the following formula to enhance the trajectory motion features:

[0024]

[0025] Among them, R k represents the covariance matrix of the measurement noise, c k Represents the detection confidence score of the detection box.

[0026] Step 3. Enhanced appearance feature extraction

[0027] Step 3.1 Feature extraction:

[0028] The existing trajectory set of the current frame is T = {T1, T2, T3, ..., T N}, the target set of the current frame is D = {D1, D2, D3, ..., D M}, the SBS50 model in the FastReID library calculates the appearance features of the current detection target, represents the appearance features of the i-th target in the k-th frame.

[0029] Step 3.2 Feature fusion:

[0030] Calculate the cosine similarity matrix d between all trajectories and targets in the current frame cos And IOU similarity matrix d iou , and perform feature fusion according to the following formula:

[0031]

[0032]

[0033] Step 3.3 Feature Update:

[0034] For the matched tracks, in the process of updating their appearance features, we adopt the same idea as the motion features, that is, we set the weights according to the detection confidence scores of the tracks to update their appearance features. The update method is as follows:

[0035]

[0036]

[0037] in represents the appearance feature description of the i-th track in the k-th frame image, Indicates the appearance features of the target detection box that has been matched by the i-th track in the k-th frame image, α t is the momentum term, α f is a fixed value weight, σ is the detection confidence score threshold, S det It is the detection confidence score of the matched target detection box.

[0038] Step 5. Data association

[0039] (1) The detector obtains all the objects D in the current frame t, and then divides the detection frame into two parts according to the detection threshold σ: the low-score detection frame D low and high score detection box D high ;

[0040] (2) Use NSAKalmanFilter to predict the position of trajectory set T in the current t-th frame;

[0041] (3) Calculate the high-score detection box D high The IOU cost matrix and appearance similarity matrix of the trajectory set T are combined, and the total cost matrix is ​​calculated according to the fusion rule;

[0042] (4) Use the Hungarian algorithm to associate high-score detection boxes D according to the total cost matrix high and trajectory set T;

[0043] (5) For high-score detection boxes D that have no matching high Set it to D remain ;

[0044] (6) For the trajectory T that has no match, set it to T remain ;

[0045] (7) For the matched trajectory T matched Update its appearance characteristics according to Dynamic EMA;

[0046] (8) Calculate the low-score detection box D low and the remaining unmatched trajectory set T remain The IOU cost matrix between them is used, and the Hungarian algorithm is used for secondary matching;

[0047] (9) The unmatched trajectories are classified as T re-remain , the detection boxes without matching are considered as background and deleted;

[0048] (10) T re-remain Classified as T lost, and delete the traces whose life cycle exceeds 30;

[0049] (11) Finally, the unmatched high-score detection frame is initialized as a new trajectory;

[0050] Preferably, the confidence score threshold σ is 0.6.

[0051] Preferably, the appearance similarity threshold θ emb is 0.25.

[0052] Preferably, the appearance similarity threshold θ iou is 0.5.

[0053] Preferably, the fixed weight value α f is 0.95.

[0054] The present invention discloses a multi-target tracking method based on threshold separation and feature enhancement, aiming at accurately and stably tracking targets. On the basis of dividing into high-score detection frames and low-score detection frames according to the detection confidence score, the feature extraction of high-score detection frames and high-score trajectories is enhanced. In terms of motion features, Kalman filtering is used to predict the possible positions of all trajectories in the current frame. In order to reduce the interference of process noise on the prediction, NSAKalmanFilter is used to predict their positions. In terms of appearance features, according to the confidence score of the detection frame, its update weight is dynamically calculated, and then its appearance features are updated. In terms of association strategy, the principle of hierarchical matching is adopted, and high-score detection frames are matched first, and then low-score detection frames and unmatched trajectories are matched. By enhancing feature expression and optimizing the matching process, the present invention improves the accuracy of multi-target tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] 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 ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0056] Figure 1 A schematic diagram of a multi-target tracking method for separation detection and feature extraction provided by an example of the present invention;

[0057] Figure 2 A schematic diagram of a multi-target tracking method for joint detection and feature extraction provided by an embodiment of the present invention;

[0058] Figure 3 A flowchart of a multi-target tracking method for joint detection and tracking provided by an embodiment of the present invention;

[0059] Figure 4 A schematic diagram of a framework of a multi-target tracking method based on threshold separation and feature enhancement provided by an embodiment of the present invention;

[0060] Figure 5 A schematic diagram of an IOU provided in an embodiment of the present invention;

[0061] Figure 6 A flowchart of a multi-target tracking method based on threshold separation and feature enhancement provided by an embodiment of the present invention; DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] The multi-target tracking algorithm based on deep learning aims to detect targets containing specific semantic information in the video sequence and associate targets of the same identity in different frames, thereby forming the trajectory of the target in the video sequence. Its formal definition is as follows:

[0064] T t =M(T t-1 ,F t ),t∈[0,N]#(1)

[0065] Where T t represents the tracking result at the tth frame, usually represented by the bounding box and identity ID of the tracked target at the tth frame, F t represents the input image of the tth frame, M(·) represents the multi-target tracking method, and N represents the total number of frames in the video sequence.

[0066] The embodiment of the present invention discloses a multi-target tracking method based on threshold separation and feature enhancement to achieve target tracking. Figure 4 , the above method includes at least the following 5 steps:

[0067] Step 1. Object Detection

[0068] The present invention uses the target detection network YOLOX as a detector to output all potential targets in each frame, and its output result includes the following information: the confidence score of the target, the category ID of the target, and the bounding box coordinates of the target.

[0069] Step 2. Motion feature extraction and enhancement

[0070] In the present invention, the target detection box is represented by an octet, namely Where x, y, a, and h represent the center coordinates, aspect ratio, and height of the detection frame, respectively. Represents the speed of its corresponding value. Its state matrix and transfer matrix are shown as follows:

[0071]

[0072] In the field of multi-target tracking, Kalman Filter updates and predicts the target based on the observation equation and state transfer equation. The prediction step obtains the prior estimate at the current moment from the posterior estimate at the previous moment:

[0073]

[0074] Then the covariance matrix of the prior error is obtained:

[0075]

[0076] In the update step, the Kalman gain coefficient is obtained from the covariance matrix:

[0077]

[0078] Then the posterior estimate is obtained from the Kalman gain coefficient:

[0079]

[0080] Finally update the error covariance matrix:

[0081] P k∣k =(IK k H k ) k∣k-1 #(6)

[0082] In order to reduce the impact of detection noise with low confidence scores, the present invention uses NSA Kalman Filter to enhance the trajectory motion features:

[0083]

[0084] Among them, R k represents the covariance matrix of the measurement noise, c kRepresents the detection confidence score of the detection frame. Therefore, for a detection frame with a high score, the covariance matrix value of its measurement noise is smaller, which means that in the update step, it has a larger Kalman gain coefficient, that is, in the posterior estimate, the current measurement value accounts for a larger proportion and has the greatest impact, and vice versa, so the accuracy of the update state can be improved. The above method is also in line with the actual situation. For a detection frame with a high score, it usually means that the target picture is clear and without any occlusion. In this case, its measurement value is more accurate and does not require correction of past motion state data.

[0085] The motion characteristics of the target are reflected in the motion information of the target. After obtaining the predicted position of the trajectory and the detected position of the target in the current frame, the motion characteristics between the target and the trajectory in the current frame can be measured by the intersection and union ratio. Figure 5 As shown, the black area is the intersection of the two rectangular boxes, and the white area plus the black area in the two rectangular boxes is the union. Then the detection box B d And trajectory prediction box B t The IOU calculation is as follows:

[0086]

[0087] The final IOU similarity is shown in Formula 9:

[0088]

[0089] Step 3. Enhanced appearance feature extraction

[0090] Step 3.1 Feature extraction:

[0091] Appearance features, also known as re-identification (ReID) features, can be used to identify specific targets in different images. The similarity measure between appearance features is usually cosine similarity. Assume that the existing trajectory set of the current frame is T = {T1, T2, T3, ..., T N}, the target set of the current frame is D = {D1, D2, D3, ..., D M}, the appearance features of the existing N trajectories are as follows:

[0092]

[0093] Then the appearance features of the M detected targets in the current frame are shown in Formula 11:

[0094]

[0095] Then the appearance feature of the i-th trajectory is For the appearance features of the jth detection box The cosine similarity calculation formula is as follows:

[0096]

[0097] The SBS50 model in the FastReID library calculates the appearance features of the current detection target. represents the appearance features of the i-th target in the k-th frame.

[0098] Step 3.2 Feature fusion:

[0099] Calculate the cosine similarity matrix d between all trajectories and targets in the current frame cos And IOU similarity matrix d iou , and perform feature fusion according to the following formula:

[0100]

[0101] When the cosine cost between the trajectory and the detection box is less than the appearance feature cost threshold θ emb And the IOU cost is less than the motion feature cost threshold θ iou When Substitute the value for the cosine Half of , otherwise it is 1. Finally, the total cost is the cosine cost and The minimum value of .

[0102] Step 3.3 Feature Update:

[0103] For the matched tracks, in the process of updating their appearance features, we adopt the same idea as the motion features, that is, we set the weights according to the detection confidence scores of the tracks to update their appearance features. The update method is as follows:

[0104]

[0105]

[0106] in represents the appearance feature description of the i-th track in the k-th frame image, Indicates the appearance features of the target detection box that has been matched by the i-th track in the k-th frame image, α t is the momentum term, α f is a fixed value weight, σ is the detection confidence score threshold, S det is the detection confidence score of the matched target detection box. Consider the following two extreme cases:

[0107] When S det =1, Update weights;

[0108] When S det =σ, Do not update weights;

[0109] In the above way, the weights are calculated according to the detection confidence scores, so as to dynamically and adaptively update the weights. For detection boxes with high detection confidence scores, it is also considered that they are not blocked or do not interact with other targets, so their appearance features are updated; for weights with low detection confidence scores, it is considered that they are blocked or have changed shape, and the target appearance is incomplete, so their appearance features are not updated.

[0110] Step 5. Data association

[0111] The present invention adopts a data association method based on threshold separation. This method makes full use of low-score detection frames and overcomes the false positive targets caused by too many low-score detection frames, thereby improving the tracking accuracy of the multi-target tracking algorithm. The specific association process is shown below, and the flow chart is shown below. Figure 6 As shown:

[0112] (12) The detector obtains all the targets D in the current frame t, and then divides the detection frame into two parts according to the detection threshold σ: the low-score detection frame D low and high score detection box D high ;

[0113] (13) Use NSAKalmanFilter to predict the position of trajectory set T in the current t-th frame;

[0114] (14) Calculate the high-score detection box D high The IOU cost matrix and appearance similarity matrix of the trajectory set T are combined, and the total cost matrix is ​​calculated according to the fusion rule;

[0115] (15) Use the Hungarian algorithm to associate high-score detection boxes D according to the total cost matrix high and trajectory set T;

[0116] (16) For the high-score detection box D that has no match high Set it to D remain ;

[0117] (17) For the trajectory T that has no match, set it to T remain ;

[0118] (18) For the matched trajectory T matched Update its appearance characteristics according to Dynamic EMA;

[0119] (19) Calculate the low-score detection box D low and the remaining unmatched trajectory set T remain The IOU cost matrix between them is used, and the Hungarian algorithm is used for secondary matching;

[0120] (20) The unmatched trajectories are classified as T re-remain , the detection boxes without matching are considered as background and deleted;

[0121] (21) T re-remain Classified as T lost , and delete the traces whose life cycle exceeds 30;

[0122] (22) Finally, the high-score detection frame that has not yet been matched is initialized as a new trajectory;

[0123] Preferably, the confidence score threshold σ is 0.6, and the appearance similarity threshold θ is emb is 0.25, and the appearance similarity threshold θ iou is 0.5, and the weight value α is fixed f is 0.95.

[0124] The present invention uses the MOT17 and MOT20 data sets released by MOTChallenge to train and test the model. MOT17 contains a total of 7 groups of test video sequences and training video sequences, totaling 14 video sequences. The video sequences in the data set are taken from scenes such as squares, city streets, roads and shopping malls by surveillance cameras, handheld cameras or vehicle-mounted cameras, and the detection results of SDP, DPM and Fast-RCNN are provided at the same time. The MOT20 data set contains 4 training video sequences and 4 test video sequences, and the videos are taken from crowded scenes such as railway stations, squares and city streets. Compared with the MOT17 data set, the average number of targets contained in each frame of the MOT20 data set reaches 170.9, and the target density is greatly improved. Frequent interactions and occlusions between targets, as well as problems such as ambient light and camera jitter, bring great challenges to the accuracy and reasoning speed of the tracking algorithm. The present invention uses the classic metric method, the "CLEAR metric method", the ID metric method and the HOTA metric method as quantitative evaluation indicators. It mainly includes the following specific evaluation indicators, IDs (ID switch) represents the number of ID exchanges of all targets during the entire tracking process. The multi-target tracking accuracy MOTA (Multiple Object Tracking Accuracy) comprehensively considers the situations of missed tracking and wrong tracking by the tracker, thereby evaluating the degree of match between the tracking trajectory and the true trajectory. The HOTA metric represents the high-order tracking accuracy, comprehensively considers the balance between target detection and association accuracy, and can measure the performance of the multi-target tracker more comprehensively and more balanced. Table 1 and Table 2 are the comparison results of the method disclosed in the present invention and other disclosed methods on the above evaluation indicators. On the MOT17 dataset, the IDF1 index and HOTA index of the algorithm are both improved by 0.8, which shows that NSA KalmanFilter and ReID have certain effects in improving detection accuracy and association accuracy. On the MOT20 dataset, the same is true, with the IDF1 index and HOTA index improved by 0.6 and 1.1 respectively.

[0125] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined in the present invention may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown in the present invention, but will conform to the widest scope consistent with the principles and novel features disclosed in the present invention.

[0126] Table 1 Experimental results of the method disclosed in the present invention on the MOT17 dataset

[0127]

[0128] Table 2 Experimental results of the method disclosed in the present invention on the MOT20 dataset

[0129]

Claims

1. A multi-target tracking method based on threshold separation and feature enhancement, characterized in that: The following steps are involved: Step 1. Object Detection Use the target detection network YOLOX as a detector to output all potential targets in each frame. Its output contains the following information: the confidence score of the target, the category ID of the target, and the bounding box coordinates of the target. Step 2. Motion feature extraction and enhancement In the field of multi-target tracking, Kalman filtering updates and predicts the target based on the observation equation and the state transfer equation. The prediction step obtains the prior estimate of the current moment from the posterior estimate of the previous moment: Then the covariance matrix of the prior error is obtained: In the update step, the Kalman gain coefficient is obtained from the covariance matrix: Then the posterior estimate is obtained from the Kalman gain coefficient: Finally update the error covariance matrix: P k∣k =(I-K k H k )P k∣k-1 In order to reduce the impact of detection noise on low-confidence target boxes, the NSA Kalman Filter is used to enhance the trajectory motion features: Among them, R k represents the covariance matrix of the measurement noise, c k Represents the detection confidence score of the detection box. Step 3. Enhanced appearance feature extraction Step 3.1 Feature extraction: The existing trajectory set of the current frame is T = {T1, T2, T3, ..., T N }, the target set of the current frame is D = {D1, D2, D3, ..., D M }, the SBS50 model in the FastReID library calculates the appearance features of the current detection target, Represents the appearance features of the i-th target in the k-th frame image. Step 3.2 Feature fusion: Calculate the cosine similarity matrix d between all trajectories and targets in the current frame cos And IOU similarity matrix d iou , and perform feature fusion according to the following formula: Step 3.3 Feature Update: For the matched tracks, in the process of updating their appearance features, we adopt the same idea as the motion features, that is, we set the weights according to the detection confidence scores of the tracks to update their appearance features. The update method is as follows: in Represents the appearance feature description of the i-th track in the k-th frame image; Represents the appearance features of the i-th target in the k-th frame image, which is also the appearance features of the matched target detection frame; α t is the momentum term, α f is a fixed value weight, σ is the detection confidence score threshold, S det It is the detection confidence score of the matched target detection box. Step 5. Data association (1) The detector obtains all the objects D in the current frame t, and then divides the detection frame into two parts according to the detection threshold σ: the low-score detection frame D low and high score detection box D high ; (2) Use NSA Kalman Filter to predict the position of trajectory set T in the current t-th frame; (3) Calculate the high-score detection box D high The IOU cost matrix and appearance similarity matrix of the trajectory set T are combined, and the total cost matrix is ​​calculated according to the fusion rule; (4) Use the Hungarian algorithm to associate high-score detection boxes D according to the total cost matrix high and trajectory set T; (5) For high-score detection boxes D that have no matching high Set it to D remain ; (6) For the trajectory T that has no match, set it to T remain ; (7) For the matched trajectory T matched Update its appearance characteristics according to Dynamic EMA; (8) Calculate the low-score detection box D low and the remaining unmatched trajectory set T remain The IOU cost matrix between them is used, and the Hungarian algorithm is used for secondary matching; (9) The unmatched trajectories are classified as T re-remain , the detection boxes without matching are considered as background and deleted; (10) T re-remain Classified as T lost , and delete the traces whose life cycle exceeds 30; (11) Finally, the unmatched high-score detection boxes are initialized as new tracks.

2. A multi-target tracking method based on threshold separation and feature enhancement as claimed in claim 1, characterized in that: The confidence score threshold σ is 0.

6.

3. A multi-target tracking method based on threshold separation and feature enhancement as claimed in claim 1, characterized in that: The appearance similarity threshold θ emb is 0.

25.

4. A multi-target tracking method based on threshold separation and feature enhancement as claimed in claim 1, characterized in that: The appearance similarity threshold θ iou is 0.

5.

5. A multi-target tracking method based on threshold separation and feature enhancement as claimed in claim 1, characterized in that: The fixed weight value α f is 0.95.

Images