Pedestrian tracking method based on iterative extension IOU association strategy

Abstract

The invention discloses a pedestrian tracking method based on an iterative extension IOU association strategy, and particularly relates to the field of computer vision tracking, and the method comprises the steps: detecting a pedestrian target in a first frame of a video sequence through a Cstrake algorithm, carrying out the decoupling of a feature map through a cross neural network, constructing a detection and Re-ID branch, obtaining the information of the pedestrian target through a YOLOv5 algorithm, and carrying out the recognition of the pedestrian target. The appearance features are aggregated through a scale perception attention network; a tracker is initialized, and a Kalman filter is used for prediction; in the data association stage, two times of matching are carried out based on iterative extension IOU, firstly, a high-score detection frame is processed, and similarity calculation is carried out in combination with IOU distance and appearance characteristics; carrying out iterative extension IOU matching on a low-score detection frame and an unmatched track; in the tracking process, iterative association is carried out by gradually increasing the expansion scale bt so as to adapt to pedestrian tracking in different scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision tracking, and more specifically, to a pedestrian tracking method based on an iteratively extended IOU association strategy. Background Art

[0002] Multi-target tracking is an important branch of computer vision, which aims to assign and maintain a unique ID for the target of interest in the video sequence and predict the target's position. Pedestrian tracking requires associating the targets of the previous and next frames through data association technology to achieve the purpose of continuous tracking of pedestrians in the video, thereby further obtaining information such as the pedestrian's movement speed, trajectory and direction, which is convenient for further application to the field of individual or large-scale group behavior research.

[0003] However, it still has some shortcomings in actual use, such as mutual occlusion between targets, similar appearance features of targets, and too fast target movement. In actual tracking scenarios, these complex situations will lead to a decrease in tracking performance, boundary drift, and identity ID switching. Therefore, a multi-target pedestrian tracking algorithm suitable for complex scenarios and with robustness is proposed. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a pedestrian tracking method based on an iteratively extended IOU association strategy to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] Step 1: The Cstrack algorithm detects pedestrians in the first frame of the video sequence;

[0007] Step 2: Initialize the tracker;

[0008] Step 3: Use Kalman filter for tracking prediction;

[0009] Step 4: Perform data association twice based on iteratively expanded IOU.

[0010] Preferably, the step A1 further comprises the following steps:

[0011] Step 1.1: First read the input image;

[0012] Step 1.2: Decouple the feature map obtained by the detector backbone network into two different feature maps through a cross neural network, which are used to construct the detection branch and the Re-ID branch;

[0013] Step 1.3: Use the YOLOv5 detection algorithm to detect the feature map passed through the cross neural network to obtain pedestrian target information;

[0014] Step 1.4: For the Re-ID feature map, the scale-aware attention network is used to aggregate the appearance feature information of different scales to alleviate the problem of target scale change in multi-target tracking;

[0015] Step 1.4.1: The appearance feature updating module is used to enhance the feature association of the target and update the appearance features of the target;

[0016] In step 1.4.1, an appearance feature updating unit is also included, wherein as the target moves, the features of the target will change, which is not conducive to the learning of appearance features and affects subsequent matching; in order to enhance the association between target features, the appearance feature updating unit is used to learn the appearance features between adjacent frames, and the appearance feature updating module uses temporal correlation to associate the features of the previous frame with the current frame, and adaptively updates the target appearance features;

[0017] In the appearance feature updating unit, the appearance feature updating unit first extracts the Re-ID feature Ht-1 in the previous frame It-1. In order to simplify the calculation, it is assumed that the displacement of the corresponding pixel point exists in a fixed range. When calculating the associated information, a search window of a fixed size d is maintained, and the feature Ht-1 is compressed from 128 dimensions to 16 dimensions for feature updating; then, the feature Ht-1 of the It-1 frame is associated with the feature Ht of the It frame through the cross-correlation layer to obtain the cross-correlation attention weight to guide the network's attention in the current frame; finally, the cross-correlation attention weight is multiplied with the compressed It-1 frame feature map to obtain the related attention feature map of the It-1 frame, which is then spliced ​​with Ht and convolved to complete the feature update; the essence of the cross-correlation operation is to guide the network to focus on the most similar features between the two frames by convolving the feature maps of the previous and next frames with each other, thereby ensuring the robustness of the appearance features.

[0018] Preferably, in step 3, an estimated value is obtained through the state of the target in the previous frame, thereby determining the state of the target in the next frame.

[0019] Preferably, the step 4 further comprises the following steps:

[0020] Step 4.1: First, separate the high-score detection frames from the low-score detection frames. For the first association, the high-score detection frames with a confidence level greater than 0.6 are used to calculate the similarity using the IOU distance and appearance features.

[0021] Step 4.2: The IOU distance between the detection box and the prediction box is used to describe the degree of motion association;

[0022] Step 4.3: Cosine distance is the cosine value of the angle between two vectors in vector space, which can be used as a measure of the difference between individuals. In tracking and matching, the minimum cosine distance is used to characterize the appearance similarity between the detection box and the prediction box.

[0023] Step 4.4: Use the IOU-Re-ID fusion association as the final metric and output the matching result;

[0024] Step 4.5: Iteratively expand the IOU matching of the low-resolution frames with confidence lower than 0.6 and higher than 0.3 and the first unmatched tracking tracks;

[0025] Step 4.6: If the detection frame does not match the existing tracking track but has a higher score, a new tracking track is established to filter the negative sample background and restore the occluded target;

[0026] Step 4.7: For the tracks that are not matched to the detection frame twice, keep 30 frames and delete them if they are not matched to the detection frame during this period.

[0027] Preferably, the step 4.2 includes an IOU-Re-ID fusion association unit, wherein, in the process of data association, two associations are performed according to the confidence of the detection box results. In order to process positive and negative samples, a threshold is usually set, and the detection results with confidence higher than the threshold are retained for the next association, and those below the threshold are directly discarded;

[0028] In the IOU-Re-ID fusion association unit, when performing the first data association, in order to obtain a higher accuracy association, the IOU distance is combined with the appearance feature as the cost matrix; the candidate boxes with low cosine similarity and long IOU distance are deleted. The minimum value of the two matrices is used as the final value of the cost matrix C; the specific calculation method of the IOU-Re-ID fusion mechanism is:

[0029]

[0030] in, It is expressed as the IOU distance between the i-th prediction box and the j-th detection box. It is expressed as the appearance matching degree between the i-th prediction box and the j-th detection box, that is, the cosine distance; Represented as the new appearance similarity measure obtained by calculation; θ IOU and θ emb It is expressed as a threshold, which filters out targets that are far from the trajectory and have low appearance similarity, ignoring irregular target motion and only matching high-quality results;

[0031] The calculation method of the cost matrix is ​​as follows:

[0032]

[0033] Among them, C i , j It is represented as the cost matrix of matching the i-th prediction box with the j-th detection box. Represented as the new appearance similarity measure calculated, Expressed as the IOU distance between the i-th prediction box and the j-th detection box;

[0034] Finally, the IOU-Re-ID fusion mechanism is used as the cost matrix for Hungarian matching, and a linear assignment of high-confidence detections is implemented.

[0035] Preferably, step 4.5 also includes iterative expansion of IOU association units, where the reasons for poor detection results often include occlusion, light and dark changes, scale changes, blurred appearance, etc., which means that the appearance features are greatly affected, the appearance distinguishability is low, and the robustness of appearance matching will be greatly reduced; in addition, during the tracking process, the target moves too fast and irregularly, resulting in no overlap between adjacent frames, which will also cause the target motion feature measurement to fail. If the target motion estimate is inaccurate, then when the unmatched trajectory continues to be matched in subsequent frames, even if there are corresponding detection results, the match will still be missed;

[0036] In the iteratively extended IOU association unit, the extended IOU association increases the matching range of detection and trajectory by increasing the area of ​​the IOU, which makes up for the deficiency of the IOU metric, so as to directly match the non-overlapping detection and trajectory of the same target in adjacent frames, compensate for the motion estimation deviation in the matching space, alleviate the influence of irregular motion on tracking, have robustness to fast motion and motion estimation deviation, and reduce the dependence on the appearance characteristics of the target; the extended IOU retains the position coordinates and proportions of the original detection trajectory, assuming that the target frame coordinates are o=(x, y, w, h), where (x, y) are the coordinates of the upper left corner of the target frame, (w, h) are the width and height of the target frame, and let the extended scale be b, then the extended target frame is (xb w ,yb h ,w+2b w ,h+2b h ). However, in the tracking process, the choice of hyperparameter expansion scale b is very important, and the choice of hyperparameter will greatly affect the performance of the tracker. In real scenarios, it is impossible to modulate the expansion scale according to the tracking effect when tracking pedestrians online. Therefore, it is used to iteratively expand the IOU association. In the tracking process, based on a gradually increasing b t The association is performed iteratively; in each zoom-in iteration, the expansion scale of the current iteration is calculated as follows:

[0037] Among them, binitial represents the initial expansion scale, λ represents the step size of the iterative expansion process, and t is the iteration count, starting from 0.

[0038] Technical effects and advantages of the present invention:

[0039] 1. The present invention belongs to general pedestrian tracking and can be applied to various practical environments such as traffic and monitoring to achieve long-term multi-target tracking with high accuracy and high robustness, effectively reducing missed detection and improving matching accuracy.

[0040] 2. The present invention improves the tracker from the aspects of re-identification features and data association, and proposes a pedestrian tracking method based on iteratively extended IOU association strategy, thereby improving the tracking accuracy and being able to adapt to complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the process method of the present invention.

[0042] Figure 2 Schematic diagram of the framework of the pedestrian tracking method of the present invention.

[0043] Figure 3 It is a schematic diagram of the appearance update module framework of the present invention.

[0044] Figure 4 It is a schematic diagram of iteratively extended IOU of the present invention. DETAILED DESCRIPTION

[0045] 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.

[0046] See also Figure 1 As shown, the present invention provides a pedestrian tracking method based on iteratively extended IOU association strategy, and the method is:

[0047] Step 1: The Cstrack algorithm detects pedestrians in the first frame of the video sequence;

[0048] The step A1 further includes the following steps:

[0049] Step 1.1: First read the input image;

[0050] Step 1.2: Decouple the feature map obtained by the detector backbone network into two different feature maps through a cross neural network, which are used to construct the detection branch and the Re-ID branch;

[0051] Step 1.3: Use the YOLOv5 detection algorithm to detect the feature map passed through the cross neural network to obtain pedestrian target information;

[0052] Step 1.4: For the Re-ID feature map, the scale-aware attention network is used to aggregate the appearance feature information of different scales to alleviate the problem of target scale change in multi-target tracking;

[0053] Step 1.4.1: The appearance feature updating module is used to enhance the feature association of the target and update the appearance features of the target;

[0054] In step 1.4.1, an appearance feature updating unit is also included, wherein as the target moves, the features of the target will change, which is not conducive to the learning of appearance features and affects subsequent matching; in order to enhance the association between target features, the appearance feature updating unit is used to learn the appearance features between adjacent frames, and the appearance feature updating module uses temporal correlation to associate the features of the previous frame with the current frame, and adaptively updates the target appearance features;

[0055] In the appearance feature updating unit, the appearance feature updating unit first extracts the Re-ID feature Ht-1 in the previous frame It-1. In order to simplify the calculation, it is assumed that the displacement of the corresponding pixel point exists in a fixed range. When calculating the associated information, a search window of a fixed size d is maintained, and the feature Ht-1 is compressed from 128 dimensions to 16 dimensions for feature updating; then, the feature Ht-1 of the It-1 frame is associated with the feature Ht of the It frame through the cross-correlation layer to obtain the cross-correlation attention weight to guide the network's attention in the current frame; finally, the cross-correlation attention weight is multiplied with the compressed It-1 frame feature map to obtain the related attention feature map of the It-1 frame, which is then spliced ​​with Ht and convolved to complete the feature update; the essence of the cross-correlation operation is to guide the network to focus on the most similar features between the two frames by convolving the feature maps of the previous and next frames with each other, thereby ensuring the robustness of the appearance features.

[0056] Step 2: Initialize the tracker;

[0057] Step 3: Use Kalman filter for tracking prediction;

[0058] In step 3, an estimated value is obtained through the state of the target in the previous frame, thereby determining the state of the target in the next frame.

[0059] Step 4: Perform data association twice based on iteratively expanded IOU;

[0060] The step 4 also includes the following steps:

[0061] Step 4.1: First, separate the high-score detection frames from the low-score detection frames. For the first association, the high-score detection frames with a confidence level greater than 0.6 are used to calculate the similarity using the IOU distance and appearance features.

[0062] Step 4.2: The IOU distance between the detection box and the prediction box is used to describe the degree of motion association;

[0063] Step 4.3: Cosine distance is the cosine value of the angle between two vectors in vector space, which can be used as a measure of the difference between individuals. In tracking and matching, the minimum cosine distance is used to characterize the appearance similarity between the detection box and the prediction box.

[0064] Step 4.4: Use the IOU-Re-ID fusion association as the final metric and output the matching result;

[0065] Step 4.5: Iteratively expand the IOU matching of the low-resolution frames with confidence lower than 0.6 and higher than 0.3 and the first unmatched tracking tracks;

[0066] Step 4.6: If the detection frame does not match the existing tracking track but has a higher score, a new tracking track is established to filter the negative sample background and restore the occluded target;

[0067] Step 4.7: For the tracks that are not matched to the detection frame twice, keep 30 frames and delete them if they are not matched to the detection frame during this period.

[0068] The step 4.2 includes an IOU-Re-ID fusion association unit, wherein, in the process of data association, two associations are performed according to the confidence of the detection box results. In order to process positive and negative samples, a threshold is usually set, and the detection results with confidence higher than the threshold are retained for the next association, and those below the threshold are directly discarded;

[0069] In the IOU-Re-ID fusion association unit, when performing the first data association, in order to obtain a higher accuracy association, the IOU distance is combined with the appearance feature as the cost matrix; the candidate boxes with low cosine similarity and long IOU distance are deleted. The minimum value of the two matrices is used as the final value of the cost matrix C; the specific calculation method of the IOU-Re-ID fusion mechanism is:

[0070]

[0071] in, It is expressed as the IOU distance between the i-th prediction box and the j-th detection box. It is expressed as the appearance matching degree between the i-th prediction box and the j-th detection box, that is, the cosine distance; Represented as the new appearance similarity measure obtained by calculation; θIOU and θ emb It is expressed as a threshold, which filters out targets that are far from the trajectory and have low appearance similarity, ignoring irregular target motion and only matching high-quality results;

[0072] The calculation method of the cost matrix is ​​as follows:

[0073]

[0074] Among them, C i , j It is represented as the cost matrix of matching the i-th prediction box with the j-th detection box. Represented as the new appearance similarity measure calculated, Expressed as the IOU distance between the i-th prediction box and the j-th detection box;

[0075] Finally, the IOU-Re-ID fusion mechanism is used as the cost matrix for Hungarian matching, and a linear assignment of high-confidence detections is implemented.

[0076] The step 4.5 also includes iterative expansion of IOU association units, where the reasons for poor detection results often include occlusion, light and dark changes, scale changes, blurred appearance, etc., which means that the appearance features are greatly affected, the appearance distinguishability is low, and the robustness of appearance matching will be greatly reduced; in addition, during the tracking process, the target moves too fast and irregularly, resulting in no overlap between adjacent frames, which will also cause the target motion feature measurement to fail. If the target motion is not accurately estimated, then when the unmatched trajectory continues to be matched in subsequent frames, even if there are corresponding detection results, the match will still be missed;

[0077] In the iteratively extended IOU association unit, the extended IOU association increases the matching range of detection and trajectory by increasing the area of ​​the IOU, which makes up for the deficiency of the IOU metric, so as to directly match the non-overlapping detection and trajectory of the same target in adjacent frames, compensate for the motion estimation deviation in the matching space, alleviate the influence of irregular motion on tracking, have robustness to fast motion and motion estimation deviation, and reduce the dependence on the appearance characteristics of the target; the extended IOU retains the position coordinates and proportions of the original detection trajectory, assuming that the target frame coordinates are o=(x, y, w, h), where (x, y) are the coordinates of the upper left corner of the target frame, (w, h) are the width and height of the target frame, and let the extended scale be b, then the extended target frame is (xb w ,yb h ,w+2b w ,h+2b h). However, in the tracking process, the choice of hyperparameter expansion scale b is very important, and the choice of hyperparameter will greatly affect the performance of the tracker. In real scenarios, it is impossible to modulate the expansion scale according to the tracking effect when tracking pedestrians online. Therefore, it is used to iteratively expand the IOU association. In the tracking process, based on a gradually increasing b t The association is performed iteratively; in each zoom-in iteration, the expansion scale of the current iteration is calculated as follows:

[0078] Among them, b initial represents the initial expansion scale, λ represents the step size of the iterative expansion process, and t is the iteration count, starting from 0.

[0079] Example 1

[0080] based on Figure 2 The system framework diagram, input video sequence, detected by YOLOv5 detector, the obtained feature map is sent to CCN (Cross-Correlation Network, CCN), and features more suitable for detection tasks and Re-ID tasks are extracted through cross-correlation and autocorrelation learning mechanisms. The feature map to be detected is sent to the YOLOv5 detection head for target detection and to obtain the initial position information of the target; the Re-ID feature map is sent to SAAN (Scale-aware Attention Network, SAAN) for feature fusion, and the appearance feature update module (Appearance Feature Update, AFU) is used to enhance the feature association of the target and update the appearance features of the target. It is proposed to perform data association based on the iterative extended IOU association strategy to obtain the target trajectory. In the first association, the IOU distance is combined with the Re-ID feature as the cost matrix of Hungarian matching to complete the linear allocation of high-confidence detection. In the second association, iterative extended IOU matching is used to achieve tracking in view of the metric failure caused by irregular target motion and low appearance distinguishability;

[0081] based on Figure 3The framework diagram of the appearance update unit. AFU first extracts the Re-ID feature Ht-1 in the previous frame It-1. In order to simplify the calculation, it is assumed that the displacement of the corresponding pixel point exists in a fixed range. When calculating the associated information, a search window of fixed size d is maintained, and the feature Ht-1 is compressed from 128 dimensions to 16 dimensions for feature update; then the feature Ht-1 of the It-1 frame is correlated with the feature Ht of the It frame through the cross-correlation layer to obtain the cross-correlation attention weight to guide the network's attention in the current frame. Finally, the relevant attention weight is multiplied with the compressed It-1 frame feature map to obtain the relevant attention feature map of the It-1 frame, which is then spliced ​​with Ht and convolved to complete the feature update. The essence of the cross-correlation operation is to guide the network to focus on the most similar features between the two frames by convolving the feature maps of the previous and next frames with each other, thereby ensuring the robustness of the appearance features.

[0082] based on Figure 4 Schematic diagram of iterative expansion IOU. The expanded IOU retains the position coordinates and proportions of the original detection trajectory. Assume that the target box coordinates are o = (x, y, w, h), where (x, y) is the coordinate of the upper left corner of the target box, (w, h) is the width and height of the target box, and let the expansion scale be b, then the expanded target box is (xb w ,yb h ,w+2b w ,h+2b h ).

[0083] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A pedestrian tracking method based on iteratively extended IOU association strategy, characterized in that: include: Step 1: The Cstrack algorithm detects pedestrians in the first frame of the video sequence; Step 2: Initialize the tracker; Step 3: Use Kalman filter for tracking prediction; Step 4: Perform data association twice based on iteratively expanded IOU.

2. The pedestrian tracking method based on iteratively extended IOU association strategy according to claim 1 is characterized in that: The step 1 also includes the following steps: Step 1.1: First read the input image; Step 1.2: Decouple the feature map obtained by the detector backbone network into two different feature maps through a cross neural network, which are used to construct the detection branch and the Re-ID branch; Step 1.3: Use the YOLOv5 detection algorithm to detect the feature map passed through the cross neural network to obtain pedestrian target information; Step 1.4: For the Re-ID feature map, the scale-aware attention network is used to aggregate the appearance feature information of different scales to alleviate the problem of target scale change in multi-target tracking; Step 1.4.1: The appearance feature updating unit is used to enhance the feature association of the target and update the appearance features of the target.

3. According to claim 2, a pedestrian tracking method based on iteratively extended IOU association strategy is characterized in that: In step 1.4.1, an appearance feature updating unit is also included, in which, as the target moves, the features of the target will change, which is not conducive to the learning of appearance features and affects subsequent matching. In order to enhance the correlation between target features, the appearance feature updating unit is used to learn the appearance features between adjacent frames. The appearance feature updating module uses temporal correlation to associate the features of the previous frame with the current frame, and adaptively updates the target appearance features.

4. According to claim 1, a pedestrian tracking method based on iteratively extended IOU association strategy is characterized in that: In step 3, an estimated value is obtained through the state of the target in the previous frame, thereby determining the state of the target in the next frame.

5. The pedestrian tracking method based on iteratively extended IOU association strategy according to claim 1, characterized in that: The step 4 also includes the following steps: Step 4.1: First, separate the high-score detection frames from the low-score detection frames. For the first association, the high-score detection frames with a confidence level greater than 0.6 are used to calculate the similarity using the IOU distance and appearance features. Step 4.2: The IOU distance between the detection box and the prediction box is used to describe the degree of motion association; Step 4.3: Cosine distance is the cosine value of the angle between two vectors in vector space, which can be used as a measure of the difference between individuals. In tracking and matching, the minimum cosine distance is used to characterize the appearance similarity between the detection box and the prediction box. Step 4.4: Use the IOU-Re-ID fusion association as the final metric and output the matching result; Step 4.5: Iteratively expand the IOU matching of the low-resolution frames with confidence lower than 0.6 and higher than 0.3 and the first unmatched tracking tracks; Step 4.6: If the detection frame does not match the existing tracking track but has a higher score, a new tracking track is established to filter the negative sample background and restore the occluded target; Step 4.7: For the tracks that are not matched to the detection frame twice, keep 30 frames and delete them if they are not matched to the detection frame during this period.

6. A pedestrian tracking method based on iteratively extended IOU association strategy according to claim 5, characterized in that: In step 4.2, including the IOU-Re-ID fusion association unit, in the IOU-Re-ID fusion association unit, when the first data association is performed, in order to obtain a higher accuracy association; the IOU distance is combined with the appearance feature as the cost matrix; the candidate boxes with low cosine similarity and long IOU distance are deleted. The minimum value of the two matrices is used as the final value of the cost matrix C; the specific calculation method of the IOU-Re-ID fusion mechanism is: in, It is expressed as the IOU distance between the i-th prediction box and the j-th detection box. It is expressed as the appearance matching degree between the i-th prediction box and the j-th detection box, that is, the cosine distance; Represented as the new appearance similarity measure obtained by calculation; θ IOU and θ emb It is expressed as a threshold, which filters out targets that are far from the trajectory and have low appearance similarity, ignoring irregular target motion and only matching high-quality results; The calculation method of the cost matrix is ​​as follows: Among them, C i , j It is represented as the cost matrix of matching the i-th prediction box with the j-th detection box. Represented as the new appearance similarity measure calculated, It is expressed as the IOU distance between the i-th predicted box and the j-th detected box.

7. The pedestrian tracking method based on iteratively extended IOU association strategy according to claim 5, characterized in that: The step 4.5 further includes iteratively extending the IOU association unit. In the iteratively extending the IOU association unit, the extended IOU association increases the matching range of the detection and trajectory by increasing the area of ​​the IOU, which makes up for the deficiency of the IOU metric, so as to directly match the non-overlapping detections and trajectories of the same target in adjacent frames, compensate for the motion estimation bias in the matching space, alleviate the influence of irregular motion on tracking, have robustness to fast motion and motion estimation bias, and reduce the dependence on the appearance features of the target; The extended IOU retains the position coordinates and proportions of the original detection trajectory. Assume that the target box coordinates are o = (x, y, w, h), where (x, y) is the coordinate of the upper left corner of the target box, (w, h) is the width and height of the target box, and let the expansion scale be b, then the expanded target box is (xb w ,yb h ,w+2b w ,h+2b h ).

8. The pedestrian tracking method based on iteratively extended IOU association strategy according to claim 7, characterized in that: In each zoom-in iteration, the calculation method of the expansion scale of the current iteration is as follows: Among them, b initial represents the initial expansion scale, λ represents the step size of the iterative expansion process, and t is the iteration count, starting from 0.

Images