Multi-target tracking method based on EKF-ANA and multi-distance trajectory matching

Abstract

The invention provides a multi-target tracking method based on EKF-ANA and multi-distance trajectory matching, relates to the technical field of multi-target tracking, and provides an adaptive extended Kalman filter algorithm for a trajectory prediction stage and a multi-distance trajectory matching method for a trajectory matching stage based on a two-stage multi-target tracking framework. And a multi-target tracking function under a complex background is realized. According to the method, a two-stage detection-based multi-target tracking mode is used, firstly, the position, the category and the confidence coefficient of a detected target are obtained through a detection algorithm, and then a track is predicted through a self-adaptive extended Kalman filtering algorithm. And finally, performing association operation between the target and the track by using a Hungary matching algorithm based on a multi-distance weighted data association cost matrix. According to the method, the accuracy of target tracking can be improved, the problem of partial ID switching caused by shielding is reduced, and certain real-time performance is met while a multi-target tracking task is completed.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-object tracking, and in particular to a multi-object tracking method based on EKF-ANA and multi-distance trajectory matching. Background Art

[0002] Among various research directions in the field of computer vision, multi-object tracking (MOT) is an important and popular research direction. Its main task is to accurately and quickly identify the positions of multiple targets in consecutive video frames, assign a unique identification ID to each identical target, and finally form continuous target trajectories. This technology is usually used to analyze video data to understand the behavior of targets or related activities. Due to the great potential of multi-object tracking technology in academic research and practical applications, it has received increasing attention and become a hot research direction in computer vision. Multi-object tracking has a wide range of applications in many industries, such as environmental monitoring, sports, transportation, animal husbandry, military field, etc., bringing great convenience to production and life.

[0003] In recent years, with the booming development of deep learning technology, this wave has pushed two-stage multi-object tracking methods onto the mainstream stage, and a series of tracking algorithms derived from the SORT algorithm have emerged, such as DeepSORT, ByteTrack, BoT-SORT, and StrongSORT, etc., which have stood out in many application scenarios. The key to the operation of these algorithms lies in accurately discriminating the unique identity of targets by means of a data association algorithm based on the past records of target trajectories and current detection data. In this process, Kalman filtering and Hungarian matching are the two most relied-on "weapons". Usually, the two cooperate closely and work together. Among them, Kalman filtering gives full play to its expertise in state prediction and future trend prediction, providing forward-looking guidance for subsequent tracking; Hungarian matching focuses on the accurate association of data, organically connecting fragmented detection information with target trajectories. Only by such a two-pronged approach can high-precision synchronous tracking of multiple targets be achieved. However, once encountering complex and changeable tracking scenarios, especially when there are frequent mutual occlusions between targets, even with the support of advanced algorithms, it is still difficult to avoid a large number of target ID jumps and missed or false tracking problems, bringing great challenges to accurate tracking. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a multi-object tracking method based on EKF-ANA and multi-distance trajectory matching, which reduces the phenomena of false tracking, missed tracking, and ID switching in tracking based on adaptive extended Kalman filtering (EKF-ANA) and target detection and tracking technologies, and improves the accuracy of the tracking algorithm in complex scenarios.

[0005] A multi-target tracking method based on EKF-ANA and multi-distance trajectory matching, comprising the following steps:

[0006] Step 1: Initialize the input video frame using the AEKTrack tracking algorithm;

[0007] Step 1.1: Use the AEKTrack tracking algorithm to track the input video frame, specifically including the tracking trajectory and the current frame detection box;

[0008] The tracking trajectory starts from the first frame of the input video through the AEKTrack tracking algorithm, and creates all tracking trajectories including continuous tracking, interrupted tracking, and stopped tracking for the detected target. Specifically, it includes six trajectory states: active trajectory, unactivated trajectory, new trajectory, tracked trajectory, missing trajectory, and deleted trajectory; the current frame detection box is all target boxes obtained by the YOLOv8 target detector in the current frame, including target position, category, and confidence information, and does not include any other frame target information;

[0009] The active trajectory refers to the target box that has been actively tracked for more than two frames; the unactivated trajectory refers to the trajectory that appears in the middle of the video frame and the second point of the trajectory has not been matched yet; the new trajectory is the trajectory newly generated during the tracking process; the tracked trajectory refers to the trajectory that was successfully tracked in the previous frame; the missing trajectory refers to the trajectory that has not been successfully tracked but is still retained within the first 30 frames; the deleted trajectory refers to the trajectory that has not been successfully tracked for more than 30 frames;

[0010] Step 1.2: In the first frame, the AEKTrack tracking algorithm creates trajectory objects for all detection results of the detector and labels them as tracked trajectories;

[0011] Step 1.3: Starting from the second frame, divide all tracking trajectories into two categories: active trajectories and unactivated trajectories, and divide all current frame detection boxes into two categories according to the score confidence threshold of the bounding box. Detection boxes with a score greater than or equal to 0.8 are classified as high-confidence detection boxes, and detection boxes with a score lower than 0.8 are classified as low-confidence detection boxes;

[0012] Step 2: Perform the first tracking match on the active trajectories and the current frame high-confidence detection boxes;

[0013] Step 2.1: Combine all tracked trajectories and missing trajectories into preliminary tracking trajectories, obtain the target detection confidence information from the detection results of the YOLOv8 target detector, and use the extended Kalman filter EKF-ANA with adaptive noise adjustment to predict the possible size and position of the preliminary tracking trajectories in the next frame;

[0014] The extended Kalman filter with adaptive noise adjustment uses an adaptively adjustable measurement noise, introduces the target detection confidence as a noise adjustment factor, and calculates the adaptive measurement noise covariance. Its calculation formula is shown in Equation (1):

[0015]

[0016] where R k is the measurement noise covariance of the original fixed value, and c k is the target detection confidence, which is the measurement noise covariance after adaptive adjustment;

[0017] Step 2.2: Calculate the multi-distance weighted data association cost matrix between the high-confidence detection boxes in the current frame and the predicted trajectory bounding boxes in the next frame; where the multi-distance includes the IoU distance, the relative movement distance, and the shape change distance;

[0018] The calculation of the relative movement distance is shown in Equation (2):

[0019]

[0020] In the formula and (x c , y c ) represent the center point coordinates of the detection box and the predicted trajectory bounding box, and c is the diagonal distance of the minimum bounding box of the detection box and the predicted trajectory bounding box;

[0021] The calculation of the shape change distance is shown in Equation (3);

[0022]

[0023] In the formula, θ is the scale factor, e is the base of the natural logarithm, and the calculation of ω w and ω h is shown in Equation (4). In the formula, w and h are the width and height of the predicted trajectory bounding box, and w gt and h gt are the width and height of the detection box;

[0024]

[0025] Then the calculation of the multi-distance weighted data association cost matrix is shown in Equation (5):

[0026] C i,j = 0.8×IoU i,j + 0.15×distance i,j + 0.05×Ω i,j (5)

[0027] where \(i\) represents the \(i\)-th detection box, \(j\) represents the \(j\)-th predicted trajectory bounding box, IoU i,j represents the IoU distance between the \(i\)-th detection box and the \(j\)-th predicted trajectory bounding box, distance i,j represents the relative movement distance between the \(i\)-th detection box and the \(j\)-th predicted trajectory bounding box, \(\Omega\) i,j represents the shape change distance between the \(i\)-th detection box and the \(j\)-th predicted trajectory bounding box, \(C\) i,j is the desired multi-distance weighted data association cost matrix.

[0028] Step 2.3: According to the multi-distance weighted data association cost matrix, use the Hungarian matching algorithm to match the high-confidence detection boxes in the current frame and the preliminary tracking trajectories. There are three possible matching results: the successfully matched trajectories and detection boxes, the trajectories that are not successfully matched, and the high-confidence detection boxes in the current frame that are not successfully matched; change the bounding box of the preliminary tracking trajectory to the high-confidence detection box in the current frame that has been successfully matched, and assign the same ID to complete the update of the preliminary tracking trajectory;

[0029] Step 3: Conduct a second tracking match for the active trajectories and the low-confidence detection boxes in the current frame.

[0030] Find all the tracked trajectories from the trajectories that were not successfully matched in Step 2, and calculate the multi-distance weighted data association cost matrix between the adaptive noise-adjusted extended Kalman filter EKF-ANA predicted trajectories of these trajectories and the low-confidence detection boxes in the current frame. Use the Hungarian algorithm to match the low-confidence detection boxes in the current frame and the predicted trajectories to obtain three possible results, and use the successfully matched low-confidence detection boxes in the current frame to update the matched trajectories. If there are still trajectories that are not successfully matched at this time, mark them as missing trajectories and leave them for matching in the next frame;

[0031] Step 4: Conduct a tracking match for the unactivated trajectories;

[0032] Calculate the multi-distance weighted data association cost matrix between the low-confidence detection boxes in the current frame that were not successfully matched in Step 3 and the unactivated trajectories. Based on this cost matrix, use the Hungarian algorithm for matching, and use the successfully matched low-confidence detection boxes in the current frame to update the matched trajectories. If there are still unactivated trajectories that are not successfully matched at this time, mark them as deleted trajectories;

[0033] Step 5: Create new tracking trajectories.

[0034] If there are high-confidence detection boxes in the current frame that have not been successfully matched so far, consider them as new tracking targets, mark them as new trajectories and assign new IDs, while directly discarding the low-confidence detection boxes;

[0035] Step 6: Return the tracking results. At this time, the AEKTrack tracking algorithm returns all successfully tracked trajectories, and each trajectory is assigned a unique ID as the result of continuous tracking to achieve the multi-object tracking process.

[0036] The beneficial effects of adopting the above technical solutions are as follows:

[0037] The present invention provides a multi-object tracking method based on EKF-ANA and multi-distance trajectory matching. In view of the disadvantages of the conventional two-stage tracking algorithm, such as many problems of mis-tracking, missing tracking, and ID switching during the tracking process, the present invention proposes an extended Kalman filter algorithm with adaptive noise adjustment and a multi-distance weighted trajectory matching method. The extended Kalman filter algorithm with adaptive noise adjustment can adapt to the multi-object tracking task in the actual non-linear scenario through the non-linear improvement of the linear Kalman filter, and introduces the detection target confidence into the noise model through noise adaptation, so that it can adaptively adjust the noise during the tracking process, enhance the anti-interference effect, and overcome the tracking instability; the multi-distance weighted data association matching proposed in the trajectory matching stage overcomes the problems of missing tracking, mis-tracking, and frequent ID switching caused by single-distance matching, and improves the tracking persistence. Experimental results show that the present invention can effectively improve the tracking stability and persistence and reduce the ID switching problem. Description of the Drawings

[0038] Figure 1 is a flowchart of the multi-object tracking method provided for the specific embodiment of the present invention;

[0039] Figure 2 is the tracking effect diagram of the ByteTrack multi-object tracking method in the specific embodiment of the present invention;

[0040] where (a) - the 136th frame, (b) - the 143rd frame, (c) - the 144th frame, (d) - the 166th frame, (e) - the 169th frame;

[0041] Figure 3 is the diagram of the AEKTrack multi-object tracking method based on EKF-ANA and multi-distance trajectory matching in the specific embodiment of the present invention;

[0042] where (a) - the 136th frame, (b) - the 143rd frame, (c) - the 144th frame, (d) - the 166th frame, (e) - the 169th frame. Specific Embodiment

[0043] The following combines the drawings and embodiments to further describe the specific embodiments of the present invention in detail. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0044] A multi-object tracking method based on EKF-ANA and multi-distance trajectory matching, asFigure 1 As shown, it includes the following steps:

[0045] Step 1: Initialize the input video frames using the AEKTrack tracking algorithm;

[0046] Step 1.1: Use the AEKTrack tracking algorithm to track the input video frames, specifically including the tracking trajectory and the current frame detection box;

[0047] The tracking trajectory is created from the first frame of the input video by the AEKTrack tracking algorithm for the detection target, including all tracking trajectories of continuous tracking, interrupted tracking, and stopped tracking, specifically including six trajectory states: active trajectory, unactivated trajectory, new trajectory, tracked trajectory, missing trajectory, and deleted trajectory; the current frame detection box is all target boxes obtained by the YOLOv8 object detector in the current frame, containing target position, category, and confidence information, and does not contain any other frame target information;

[0048] The active trajectory refers to the target box that has been actively tracked for more than two frames; the unactivated trajectory refers to the trajectory that appears in the middle of the video frame and the second point of the trajectory has not been matched yet; the new trajectory is the newly generated trajectory during the tracking process; the tracked trajectory refers to the trajectory that was successfully tracked in the previous frame; the missing trajectory refers to the trajectory that has not been successfully tracked but is still retained within the first 30 frames; the deleted trajectory refers to the trajectory that has not been successfully tracked for more than 30 frames;

[0049] Step 1.2: In the first frame, the AEKTrack tracking algorithm creates trajectory objects for all detection results of the detector and marks them as tracked trajectories;

[0050] Step 1.3: Starting from the second frame, all tracking trajectories are divided into two categories: active trajectories and unactivated trajectories, and all current frame detection boxes are divided into two categories according to the score confidence threshold of the bounding box. Detection boxes with a score greater than or equal to 0.8 are classified as high-confidence detection boxes, and detection boxes with a score lower than 0.8 are classified as low-confidence detection boxes;

[0051] Step 2: Perform the first tracking match on the active trajectories and the current frame high-confidence detection boxes;

[0052] Step 2.1: Combine all the tracked trajectories and missing trajectories into preliminary tracking trajectories, obtain the target detection confidence information from the detection results of the YOLOv8 object detector, and use the Extended Kalman Filter with Adaptive Noise Adjustment (EKF-ANA) to predict the possible size and position of the preliminary tracking trajectories in the next frame;

[0053] The extended Kalman filter with adaptive noise adjustment abandons the fixed measurement noise method used in the traditional Kalman filter and instead adopts an adaptively adjustable measurement noise. The target detection confidence is introduced as a noise adjustment factor to calculate the adaptive measurement noise covariance, and its calculation formula is shown in Equation (1):

[0054]

[0055] where R k is the measurement noise covariance of the original fixed value, c k is the target detection confidence, is the measurement noise covariance after adaptive adjustment;

[0056] The extended Kalman filter with adaptive noise adjustment can adapt to the real-time measurement noise that changes with the environment or time in the actual tracking scenario, and has more accurate motion modeling and stronger anti-interference performance.

[0057] Step 2.2: Calculate the multi-distance weighted data association cost matrix between the high-confidence detection box in the current frame and the predicted trajectory bounding box in the next frame; where the multi-distance includes the IoU distance, relative movement distance, and shape change distance;

[0058] The calculation of the relative movement distance is shown in Equation (2):

[0059]

[0060] In the formula and (x c , y c ) represent the center point coordinates of the detection box and the predicted trajectory bounding box, and c is the diagonal distance of the minimum bounding box of the detection box and the predicted trajectory bounding box;

[0061] The calculation of the shape change distance is shown in Equation (3);

[0062]

[0063] In the formula, θ is a scale factor used to control the degree of attention to shape changes, e is the base of the natural logarithm, and the calculation of ω w and ω h is shown in Equation (4). In the formula, w and h are the width and height of the predicted trajectory bounding box, and w gt and h gt are the width and height of the detection box;

[0064]

[0065] Then the calculation of the multi-distance weighted data association cost matrix is shown in Equation (5):

[0066] C i,j = 0.8 × IoU i,j + 0.15 × distance i,j + 0.05 × Ω i,j (5)

[0067] Where i represents the i-th detection box, j represents the j-th predicted trajectory bounding box, and IoU i,j represents the IoU distance between the i-th detection box and the j-th predicted trajectory bounding box, and distance i,j represents the relative movement distance between the i-th detection box and the j-th predicted trajectory bounding box, and Ω i,j represents the shape change distance between the i-th detection box and the j-th predicted trajectory bounding box, and C i,j is the desired multi-distance weighted data association cost matrix.

[0068] Step 2.3: According to the multi-distance weighted data association cost matrix, use the Hungarian matching algorithm to match the high-confidence detection boxes in the current frame and the preliminary tracking trajectories. There are three possible matching results: the successfully matched trajectories and detection boxes, the trajectories that are not successfully matched, and the high-confidence detection boxes in the current frame that are not successfully matched. Change the bounding boxes of the preliminary tracking trajectories to the high-confidence detection boxes in the current frame that have been successfully matched and assign the same ID to complete the update of the preliminary tracking trajectories;

[0069] Step 3: Perform a second tracking match for the active trajectories and the low-confidence detection boxes in the current frame.

[0070] Find all the tracked trajectories from the trajectories that were not successfully matched in Step 2, and calculate the multi-distance weighted data association cost matrix between the EKF-ANA predicted trajectories with adaptive noise adjustment of these trajectories and the low-confidence detection boxes in the current frame. Use the Hungarian algorithm to match the low-confidence detection boxes in the current frame and the predicted trajectories to obtain three possible results, and use the successfully matched low-confidence detection boxes in the current frame to update the matched trajectories. If there are still trajectories that are not successfully matched at this time, mark them as missing trajectories and leave them for matching in the next frame;

[0071] Step 4: Perform a tracking match for the unactivated trajectories;

[0072] Calculate the multi-distance weighted data association cost matrix between the low-confidence detection boxes in the current frame that were not successfully matched in Step 3 and the unactivated trajectories. Based on this cost matrix, use the Hungarian algorithm for matching, and use the successfully matched low-confidence detection boxes in the current frame to update the matched trajectories. If there are still unactivated trajectories that are not successfully matched at this time, mark them as deleted trajectories;

[0073] Step 5: Create new tracking trajectories.

[0074] If there are current frame high-confidence detection boxes that have not been successfully matched so far, they are considered new tracking targets, marked as new trajectories and assigned new IDs, while low-confidence detection boxes are directly discarded;

[0075] Step 6: Return the tracking results. At this time, the AEKTrack tracking algorithm returns all successfully tracked trajectories, and each trajectory is assigned a unique ID. As the result of continuous tracking, the AEKTrack tracking realizes the tracking process of multiple targets in this way of frame-by-frame prediction and matching.

[0076] In this embodiment, a program is written in PyCharm, and the AEKTrack multi-object tracking method based on adaptive extended Kalman filter and multi-distance trajectory matching is implemented according to the above process. Taking 7 validation set video sequences in the VisDrone2019-MOT dataset as an example, the MOTA index and the number of ID switches of the common advanced algorithm ByteTrack are compared. The comparison results of the MOTA values are shown in Table 1, and the comparison results of the ID switches are shown in Table 2.

[0077] Table 1 Comparison of MOTA index results

[0078]

[0079] Table 2 Comparison of ID switch results

[0080]

[0081]

[0082] Table 1 gives the comparison of the MOTA indexes of the two methods. The MOTA index can reflect the comprehensive tracking performance of the algorithm, and the higher the index value, the better. In the present invention, the AEKTrack multi-object tracking method based on EKF-ANA and multi-distance trajectory matching can effectively improve the MOTA value under each video sequence, and the average index increase reaches 4.49%, which proves that the AEKTrack multi-object tracking method based on adaptive extended Kalman filter and multi-distance trajectory matching can improve the tracking performance of the algorithm. From the results, the uav0000182_00000_v and uav0000268_05773_v video sequences are for tracking large targets with simple backgrounds, and the effect improvement is weak. Other video sequences are for tracking small targets with complex backgrounds, and the tracking ability is effectively improved. Therefore, the method of the present invention will play a better role in the small target tracking task under complex backgrounds.

[0083] Table 2 presents the comparison of the number of ID switches between the two methods. In the present invention, the AEKTrack multi-object tracking method based on EKF-ANA and multi-distance trajectory matching can effectively reduce the number of ID switches under each video sequence, with a total reduction of 69 ID switches and an average reduction of 9.86 ID switches, proving that the AEKTrack multi-object tracking method based on adaptive extended Kalman filtering and multi-distance trajectory matching can effectively reduce the number of target ID switches during tracking, extend the tracking time, and enhance the continuous tracking ability.

[0084] Taking the uav0000137_00458_v video sequence in the VisDrone2019-MOT dataset as an example, the tracking effect diagram is as Figure 2 and Figure 3 shown, where Figure 2 and Figure 3 (a), (b), (c), (d), and (e) of which are the 136th frame, 143rd frame, 144th frame, 166th frame, and 169th frame respectively. Specifically, taking the pedestrian riding a bicycle above the scene in the figure as an example, during the entire tracking process from the 136th frame to the 168th frame of the video, the ByteTrack algorithm lost the tracking target at the 144th frame; the AEKTrack multi-object tracking method based on adaptive extended Kalman filtering and multi-distance trajectory matching in the present invention did not experience target loss and ID switching phenomena during the entire tracking process from the 136th frame to the 169th frame, and the tracker achieved a complete tracking process, reaching the optimal tracking effect.

[0085] The above description is only the preferred embodiment of the present disclosure and the explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with the (but not limited to) technical features with similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A multi-target tracking method based on EKF-ANA and multi-distance trajectory matching, characterized in that, It includes the following steps: Step 1: Initialize the input video frames using the AEKTrack tracking algorithm; Step 2: Perform the first tracking match on the active trajectories and the high-confidence detection boxes in the current frame; Step 3: Perform the second tracking match on the active trajectories and the low-confidence detection boxes in the current frame; Step 4: Perform tracking match on the unactivated trajectories; Step 5: Create new tracking trajectories; Step 6: Return the tracking results; At this time, the AEKTrack tracking algorithm returns all successfully tracked trajectories, and each trajectory is assigned a unique ID as the result of continuous tracking, realizing the tracking process of multiple targets.

2. The multi-target tracking method based on EKF-ANA and multi-distance trajectory matching according to claim 1, wherein, The said Step 1 includes the following steps: Step 1.1: Use the AEKTrack tracking algorithm to track the input video frames, specifically including tracking trajectories and detection boxes in the current frame; The tracking trajectories are all tracking trajectories created by the AEKTrack tracking algorithm starting from the first frame of the input video for the detection targets, including all tracking states such as continuous tracking, interrupted tracking, and stopped tracking, specifically including six trajectory states: active trajectories, unactivated trajectories, new trajectories, tracked trajectories, missing trajectories, and deleted trajectories; The current frame detection boxes are all target boxes obtained by the YOLOv8 object detector in the current frame, containing target position, category, and confidence information, and do not contain any other frame target information; The active trajectories refer to the target boxes that have been actively tracked for more than two frames; The unactivated trajectories refer to the trajectories that appear in the middle of the video frames and the second point of the trajectory has not been matched yet; The new trajectories are the trajectories newly generated during the tracking process; The tracked trajectories refer to the trajectories that were successfully tracked in the previous frame; The missing trajectories refer to the trajectories that have not been successfully tracked but are still retained within the first 30 frames; The deleted trajectories refer to the trajectories that have not been successfully tracked for more than 30 frames; Step 1.2: In the first frame, the AEKTrack tracking algorithm creates trajectory objects for all detection results of the detector and marks them as tracked trajectories; Step 1.3: Starting from the second frame, divide all tracking trajectories into two categories: active trajectories and unactivated trajectories, and divide all current frame detection boxes into two categories according to the score confidence threshold of the bounding box. Detection boxes with a score greater than or equal to 0.8 are classified as high-confidence detection boxes, and detection boxes with a score lower than 0.8 are classified as low-confidence detection boxes.

3. A multi-target tracking method based on EKF-ANA and multi-distance trajectory matching according to claim 1, characterized in that, The said Step 2 includes the following steps: Step 2.1: Combine all tracked trajectories and missing trajectories into preliminary tracking trajectories, obtain the target detection confidence information from the detection results of the YOLOv8 object detector, and use the extended Kalman filter with adaptive noise adjustment EKF-ANA to predict the possible size and position of the preliminary tracking trajectories in the next frame; The extended Kalman filter with adaptive noise adjustment uses an adaptively adjustable measurement noise, introduces the target detection confidence as a noise adjustment factor, and calculates the adaptive measurement noise covariance. Its calculation formula is shown in Equation (1): where R k is the measurement noise covariance of the original fixed value, c k is the target detection confidence, which is the measurement noise covariance after adaptive adjustment; Step 2.2: Calculate the multi-distance weighted data association cost matrix between the high-confidence detection boxes in the current frame and the predicted trajectory bounding boxes in the next frame; where the multi-distances include the IoU distance, the relative movement distance, and the shape change distance; The relative movement distance is calculated as shown in Equation (2): where and (x c , y c ) represent the center point coordinates of the detection box and the predicted trajectory bounding box, and c is the diagonal distance of the minimum bounding box of the detection box and the predicted trajectory bounding box; The shape change distance is calculated as shown in Equation (3); where θ is the scale factor, e is the base of the natural logarithm, ω w and ω h are calculated according to Equation (4), where w and h are the width and height of the predicted trajectory bounding box, w gt and h gt are the width and height of the detection box; Then the calculation of the multi-distance weighted data association cost matrix is shown in Equation (5): C i,j = 0.8 × IoU i,j + 0.15 × distance i,j + 0.05 × Ω i,j (5) where \(i\) represents the \(i\)-th detection box, \(j\) represents the \(j\)-th predicted trajectory bounding box, and IoU i,j represents the IoU distance between the \(i\)-th detection box and the \(j\)-th predicted trajectory bounding box, and distance i,j represents the relative movement distance between the \(i\)-th detection box and the \(j\)-th predicted trajectory bounding box, and \(\Omega\) i,j represents the shape change distance between the \(i\)-th detection box and the \(j\)-th predicted trajectory bounding box, and \(C\) i,j is the required multi-distance weighted data association cost matrix; Step 2.3: According to the multi-distance weighted data association cost matrix, use the Hungarian matching algorithm to match the high-confidence detection boxes in the current frame and the preliminary tracking trajectories. There are three possible matching results: the successfully matched trajectories and detection boxes, the trajectories that are not successfully matched, and the high-confidence detection boxes in the current frame that are not successfully matched; change the bounding boxes of the preliminary tracking trajectories to the successfully matched high-confidence detection boxes in the current frame and assign the same ID to complete the update of the preliminary tracking trajectories.

4. A multi-target tracking method based on EKF-ANA and multi-distance trajectory matching according to claim 1, characterized in that, The specific content of Step 3 is as follows: Find all the tracked trajectories from the trajectories that were not successfully matched for the first time in Step 2, and calculate the multi-distance weighted data association cost matrix between the EKF-ANA predicted trajectories with adaptive noise adjustment of these trajectories and the low-confidence detection boxes in the current frame. Use the Hungarian algorithm to match the low-confidence detection boxes in the current frame and the predicted trajectories to obtain three possible results, and use the successfully matched low-confidence detection boxes in the current frame to update the matched trajectories. If there are still trajectories that are not successfully matched at this time, mark them as missing trajectories and leave them for matching in the next frame.

5. A multi-target tracking method based on EKF-ANA and multi-distance trajectory matching according to claim 1, characterized in that, The specific content of Step 4 is as follows: Calculate the multi-distance weighted data association cost matrix between the low-confidence detection boxes in the current frame that were not successfully matched in Step 3 and the unactivated trajectories. Based on this cost matrix, use the Hungarian algorithm for matching. According to the matching results, use the successfully matched low-confidence detection boxes in the current frame to update the matched trajectories. If there are still unactivated trajectories that are not successfully matched at this time, mark them as deleted trajectories.

6. The multi-target tracking method based on EKF-ANA and multi-distance trajectory matching according to claim 1, characterized in that The specific content of Step 5 is as follows: If there are high-confidence detection boxes in the current frame that have not been successfully matched so far, consider them as new tracking targets, mark them as new trajectories and assign new IDs, while directly discard the low-confidence detection boxes.

Images