Unmanned aerial vehicle vision multi-target tracking method based on trend change

Abstract

The invention belongs to the technical field of target tracking, and particularly relates to an unmanned aerial vehicle vision multi-target tracking method based on trend change. The method comprises the following steps: firstly, training a target detection model, and performing target detection on an unmanned aerial vehicle visual image by using the trained target detection model; then, whether the track needs to be compensated or not is judged; if so, calculating a global offset to compensate the trajectory, and matching the trajectory with the target after the compensation is completed; if not, directly matching; then, matching the trajectory with the target, and mapping a matching result to a target detection frame to obtain a multi-target tracking result; and finally, each trajectory is updated by using a trend change adaptive Kalman filter, and the updated trajectory is used for multi-target tracking of a next frame of unmanned aerial vehicle visual image. According to the method, a trend change factor capable of reflecting a target state change trend is provided, a trend change adaptive Kalman filter, a trend change grouping matching strategy and a rapid state compensation mechanism are provided based on the trend change factor, the problems of mutual shielding, view angle change and camera shake are effectively solved, and the tracking effect is further improved.

Description

Technical Field

[0001] The invention belongs to the technical field of multi-target tracking of computer vision, and specifically is a method for multi-target tracking of unmanned aerial vehicle vision based on trend changes. Background Art

[0002] As an important research direction in computer vision, visual target tracking is widely used in the fields of autonomous driving, video surveillance, drone aerial tracking, etc. Most existing multi-target tracking methods follow the detection-based tracking paradigm. Due to the lack of clues such as trend change information, it is often necessary to add additional neural network models and computing modules in the data association step, or improve relatively independent original modules. These methods usually bring additional training costs or have limited performance improvement.

[0003] In addition, the current multi-target tracking methods are mainly studied for conventional multi-target tracking scenarios, while target tracking in UAV scenarios is mainly to directly improve tracking performance by optimizing the general performance of multi-target tracking. Compared with conventional multi-target tracking scenarios, multi-target tracking under UAV vision has its own particularities. For example, in most UAV scenarios, UAVs often remain at high altitudes and move continuously, with the characteristics of mutual occlusion, camera shake, and perspective changes. In addition, the UAV equipment itself also has some limitations, resulting in the existing multi-target tracking methods often being unable to achieve a good balance between detection and tracking, and thus unable to achieve excellent tracking performance.

[0004] For camera motion compensation, traditional methods include ORB, SIFT, etc. There are also deep learning methods that use optical flow detection models for compensation. These methods are direct responses to surface problems, and traditional methods are not fast, while deep learning methods often require additional training costs.

[0005] Based on the above problems, the present invention conducts an in-depth analysis on the characteristics of multi-target tracking under UAV vision and proposes a UAV vision multi-target tracking method based on trend changes. Summary of the invention

[0006] In view of the deficiencies in the prior art, the technical problem that the present invention intends to solve is to provide a UAV visual multi-target tracking method based on trend changes.

[0007] The present invention solves the technical problem by adopting the following technical solution: A method for tracking multiple targets using UAV visual system based on trend change, comprising the following steps: Step 1: Train the target detection model and use the trained target detection model to detect targets in UAV visual images; divide the targets into high-confidence targets and low-confidence targets according to the confidence level, and obtain a high-confidence target set and a low-confidence target set; Step 2: Determine whether the trajectory needs to be compensated; if necessary, calculate the global offset to compensate the trajectory, and then proceed to step 3 after the compensation is completed; if not, proceed directly to step 3; Global offset The calculation formula is as follows: (1) (2) (3) (4) in, is the reliability weight of the i-th successfully matched trajectory, , is the matching cost and confidence of the i-th successfully matched trajectory, and is a hyperparameter, , is the confidence change value and matching cost change value of the i-th successfully matched trajectory, , is the coordinate offset and size offset of the i-th successfully matched trajectory, n is the number of successfully matched trajectories, is the confidence of the i-th trajectory at time t-1 and t-2, , is the matching cost of the i-th trajectory at time t-1 and t-2; Step 3: Match the trajectory with the target, map the matching result to the target detection frame, and obtain the multi-target tracking result; Step 4: Use the trend change adaptive Kalman filter to update each trajectory, and the updated trajectory is used for multi-target tracking in the next frame of UAV visual image; The trend change adaptive Kalman filter updates the trajectory according to the tracking result of the current frame UAV visual image, and its noise covariance update formula is: (6) (7) in, is the noise covariance of the i-th trend change adaptive Kalman filter at time t, is the influence factor of the ith trajectory at time t, is a hyperparameter; , , and are the noise covariance of the abscissa, ordinate, width and height of the ith trajectory at time t. If the trajectory matches the target successfully, the abscissa, ordinate, width and height of the trajectory are the abscissa, ordinate, width and height of the target; if the trajectory does not match successfully, the abscissa, ordinate, width and height of the trajectory remain at the current values.

[0008] Furthermore, for the matching of trajectories and targets, the priority weight of the trajectories is calculated according to formula (5). The smaller the priority weight, the higher the priority of the trajectory; (5) in, It is The priority weight of each trajectory; Set the number of matches to C, the priority range to Z, initialize the set of participating matching trajectories to empty, and the set of non-participating matching trajectories to the set formed by all trajectories; Match the trajectories according to the priority. For the kth match, add the trajectories with priority weight less than k×Z to the set of participating matching trajectories, and remove these trajectories from the set of unparticipated matching trajectories. Match each trajectory in the set of participating matching trajectories with each high-confidence target in the set of high-confidence targets. If the intersection and union ratio of the trajectory and the high-confidence target is greater than the set threshold, the match is successful. The successfully matched trajectory is removed from the set of participating matching trajectories, and the unsuccessfully matched trajectory is retained in the set of participating matching trajectories to continue to participate in the next match, and the matching cycle is repeated C-1 times. For the last match, merge the current set of unmatched trajectories with the set of matched trajectories to obtain a merged set of unmatched trajectories; merge the unmatched high-confidence targets in the high-confidence target set with the low-confidence target set to obtain a mixed set of high- and low-confidence targets; combine each target in the mixed set of high- and low-confidence targets with each trajectory in the merged set of unmatched trajectories to obtain multiple candidate matching pairs; calculate the intersection-and-union ratio of each candidate matching pair, use the intersection-and-union ratio as a matrix element, and obtain a matching cost matrix; for low-confidence targets, if the intersection-and-union ratio of the candidate matching pair where the low-confidence target is located is less than the filtering threshold, it is an unreliable candidate matching pair, set the intersection-and-union ratio of the unreliable candidate matching pair to zero, update the matching cost matrix, and traverse all candidate matching pairs where the low-confidence targets are located to obtain an updated matching cost matrix; According to the updated matching cost matrix, the Hungarian algorithm is used to match the trajectory with the target, thereby completing the matching of all trajectories.

[0009] Furthermore, for the setting of the filtering threshold, all low-confidence targets are grouped according to the confidence interval of the low-confidence target and the set confidence interval interval, and a filtering threshold is set for each group according to the value range of the confidence interval; the higher the confidence group, the lower the filtering threshold.

[0010] Furthermore, the process of judging whether the trajectory needs to be compensated is as follows: 1) screening the trajectories to select stable trajectories; 2) matching the screened trajectories with each high-confidence target in the high-confidence target set, if the intersection and union ratio of the trajectory and the high-confidence target is greater than the set threshold, the match is successful; the average matching cost of all successfully matched trajectories is calculated; if the ratio of the average matching cost of the successfully matched trajectories at the previous moment and the current moment is less than or equal to the set threshold, no compensation is required, otherwise compensation is required.

[0011] Furthermore, the screening condition of the trajectory is at least one of confidence, matching cost and change range.

[0012] Compared with the prior art, the advantages and beneficial effects of the present invention are: 1. The present invention conducts an in-depth analysis of the drone scene, observes the characteristics of the target motion in the scene, and adds the matching cost as a new clue to the attributes of the trajectory. The matching cost is the motion similarity between the trajectory and the target when it was successfully matched for the most recent time. The trend change factor is constructed using the confidence of the trajectory and the confidence of the trajectory. The trend change factor includes the confidence change value and the matching cost change value. Because the change in confidence can intuitively reflect the change in the clarity of the target for the target detection model, and the change in matching cost can reflect the stability of the target motion, the trend change factor can reflect the trend of target state changes. The trend change factor can not only simply and intuitively reflect the change in the target state, but also adapt to different situations after expansion, more effectively estimate the target state, and further determine whether the target is about to be obscured by a large area or about to emerge from the occlusion. This method can extract and utilize the potential information of the target motion without complex deep learning.

[0013] 2. The present invention improves and constructs a Kalman filter and matching strategy by introducing a trend change factor, and constructs a fast and effective fast compensation mechanism, thereby combining the entire data association step into an organic whole, thereby achieving multi-target tracking with excellent speed and superior performance. The trend change adaptive Kalman filter of the present invention combines the trend change factor and the analysis of the characteristics of the drone scene to enhance its motion perception ability, thereby enhancing the motion prediction performance. The trend change grouping matching strategy uses the trend change factor to enhance the distinction of the trajectory, and groups the trajectories precisely and effectively, thereby enhancing the matching accuracy and reducing the tracking error caused by the mutual occlusion of dense targets. The fast compensation mechanism adopts the idea of ​​the optical flow method, but does not use numerous pixels with low average information validity as the analysis target, but a small number of trajectories with rich information as the analysis target, and adjusts the compensation intensity in conjunction with the trend change factor to achieve fast and effective camera motion compensation. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Schematic diagram of the overall multi-target tracking framework of the present invention; Figure 2 is a schematic diagram of the rapid compensation mechanism of the present invention; DETAILED DESCRIPTION

[0015] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods, but the protection scope of the present application is not limited thereto.

[0016] The present invention is a method for tracking multiple targets of unmanned aerial vehicle vision based on trend change (hereinafter referred to as method, see Figure 1~2 ), including the following steps: Step 1: Train the target detection model, use the trained target detection model to detect targets in the UAV visual image, and obtain multiple targets; divide the targets into high-confidence targets and low-confidence targets according to the confidence level, and obtain a high-confidence target set. and low confidence target set .

[0017] Step 2: Based on the Fast State Compensation (FSC) mechanism, determine whether the trajectory needs to be compensated; if necessary, calculate the global offset, compensate the trajectory according to the global offset, and then enter the third step after the compensation is completed; if not, directly enter the third step; The fast state compensation mechanism of the present invention is used to compensate for the error caused by camera shake due to the characteristics of the drone itself, and has the characteristics of fast operation and no need for additional hardware equipment. FSC has a good effect in suppressing camera shake caused by the special motion trajectory of the drone. The main theoretical basis of FSC is similar to the optical flow method, which uses the changes in a large number of pixels in adjacent frames to reflect the movement of the object. FSC only quantifies camera shake by calculating the offset between the trajectory and detection, thereby avoiding a large number of pixel calculations, and compared with pixels, the trajectory in the multi-target tracking scene has richer and easier to understand states, such as confidence, size, position and survival time. Therefore, FSC can achieve the effect of camera motion compensation in a concise and efficient way. The FSC mechanism is roughly divided into four steps: screening, matching, calculation and compensation.

[0018] 1) Screen the trajectories and select stable ones; FSC uses high confidence, low matching cost, low change amplitude and other screening conditions. The conditions can be in an OR relationship or an AND relationship. For example, if the confidence of the trajectory is > 0.7 or the matching cost of the trajectory is < 0.37 or the number of consecutive successful matches of the trajectory (also called trajectory length) is > 7, then the trajectory is considered to be stable; 2) Combine the selected trajectories with the high confidence target set Match each high-confidence target in the trajectory. If the intersection-over-combination ratio of the trajectory and the high-confidence target is greater than the set threshold, the match is successful. Calculate the average matching cost of all successfully matched trajectories The smaller the value, the smaller the camera shake and the more stable the drone state. If the average matching cost of the trajectories that are successfully matched at adjacent moments is / If the average matching cost ratio is less than or equal to the set threshold, it means that the camera shake is small and the trajectory does not need to be compensated; if the average matching cost ratio is greater than the set threshold, it means that the camera shake is greater and the trajectory needs to be compensated, then go to step 3); 3) Calculate the global offset , compensate the trajectory according to the global offset; calculate the reliability weight of the successfully matched trajectory based on the trend change factor The larger the value, the more obvious the impact of the successfully matched trajectory on the global offset. The global offset is calculated according to the reliability weight, and the calculation formula is as follows: (1) (2) in, is the reliability weight of the i-th successfully matched trajectory, , is the matching cost and confidence of the i-th successfully matched trajectory, and is a hyperparameter, , is the confidence change value and matching cost change value of the i-th successfully matched trajectory, , is the coordinate offset and size offset of the i-th successfully matched trajectory, and n is the number of successfully matched trajectories; The confidence change value of the i-th trajectory and matching cost change Satisfy the following formulas respectively: (3) (4) in, is the confidence of the i-th trajectory at time t-1 and t-2, , is the matching cost of the i-th trajectory at time t-1 and t-2; in the initial stage, both values ​​are initialized to 0.

[0019] Add a global offset to the coordinates and dimensions of each track , complete trajectory compensation, thereby suppressing the impact of camera shake.

[0020] Step 3: Match the trajectory and the target according to the trend change group matching strategy to achieve multi-target tracking; The present invention proposes a trend grouping matching strategy (TGM) based on trend change to improve matching accuracy and alleviate the errors caused by mutual occlusion in target-dense scenes. The strategy evaluates the priority of trajectories based on trend change factors, and then groups and matches trajectories according to the priority. Compared with other conventional group matching methods, the grouping basis constructed based on trend change factors is more effective, not limited to the difference between targets on a single frame image, but more focused on the potential motion change trend between frames, so that targets have better distinction. High-priority trajectories will be matched earlier, which can effectively ensure that the trajectories match the correct targets, and unmatched trajectories in the high-priority trajectory group will participate in the subsequent matching process of the low-priority trajectory group. In short, TGM relies on trend change factors to accurately assign priorities according to the stability of trajectory states, ensure the accuracy of relatively stable object matching, and alleviate the problem of mutual occlusion in drone scenes.

[0021] Calculate priority weights based on trend change factors , which is used to evaluate the priority of the trajectory. The smaller the priority weight, the higher the priority of the trajectory. The priority weight calculation formula is: (5) in, It is The priority weight of each trajectory; From formula (5), we can see that the matching cost of the trajectory is is the main parameter, matching cost change value As the main correction parameter, the confidence change value As The volatility factor; and the single matching cost Compared with the priority weight Combining the trend change factor between frames can more effectively measure the stability of the drone state and improve the matching accuracy of trajectory grouping. For the priority of the trajectory, each priority is assigned a value range. Trajectories are assigned the same priority when they are within the priority range.

[0022] Set the number of matches to C and the priority range to Z, and the set of trajectories that will participate in the matching Initialized to empty, not involved in matching trajectory set Initialize to the set formed by all trajectories; match the trajectories according to priority. For the kth match, add the trajectories with priority weight less than k×Z to the set of trajectories participating in the matching. and never participated in the matching trajectory set Eliminate these trajectories from the matching trajectory set Each trajectory in the high confidence target set is matched with each high confidence target in the high confidence target set. If the intersection and union ratio of the trajectory and the high confidence target is greater than the set threshold, the trajectory is considered to be successfully matched with the high confidence target and is removed from the participating matching trajectory set. The successfully matched trajectories are removed from the set, while the unmatched trajectories are retained in the set of participating matching trajectories. Continue to participate in the next match, and cycle matching C-1 times. For the last match, merge the current set of unmatched trajectories with the set of matched trajectories to obtain the merged set of unmatched trajectories; at the same time, merge the unmatched high-confidence targets in the high-confidence target set with the low-confidence target set. Merge to obtain a mixed set of high and low confidence targets; perform a final match on the trajectories in the merged set of unmatched trajectories and the targets in the mixed set of high and low confidence targets; first, combine each target in the mixed set of high and low confidence targets with each trajectory in the merged set of unmatched trajectories to obtain multiple candidate matching pairs; calculate the intersection-and-union ratio of each candidate matching pair, use the intersection-and-union ratio as a matrix element, and obtain a matching cost matrix; for low confidence targets, filter unreliable candidate matching pairs according to the low confidence matching screening strategy, set the intersection-and-union ratio of unreliable candidate matching pairs to zero, update the matching cost matrix, and obtain an updated matching cost matrix; The low-confidence matching screening strategy refers to: traversing the intersection-and-union ratios of all candidate matching pairs where low-confidence targets are located. If the intersection-and-union ratio is less than the filtering threshold, it is considered to be an unreliable candidate matching pair, and the matching of the corresponding low-confidence target and the trajectory is not considered.

[0023] For low-confidence targets, conventional methods usually match them directly with limited active trajectories. This matching method is neither sufficient nor rigorous, and it fails to deeply explore the correlation performance between low-confidence targets and trajectories. However, this invention analyzes the matching rules based on the characteristics of low-confidence targets, and proposes a low-confidence matching screening strategy based on the special performance of multi-target tracking in drone scenarios. The filtering threshold is based on the confidence interval of the low-confidence target. All low-confidence targets are grouped according to the set confidence interval interval. A filtering threshold is set for each group according to the value range of the confidence interval; the higher the confidence group, the lower the filtering threshold; for example, if the confidence of the low-confidence target is less than 0.6 and greater than 0.4, it is divided into two groups according to the confidence interval interval of 0.1, and the value range of the confidence interval is [0.4,0.5) and [0.5,0.6] respectively. The filtering thresholds are set to 0.6 and 0.5 respectively, so that truly reliable low-confidence targets can be matched successfully, and unreliable ones are filtered out, which increases the matching accuracy, reduces the error, and effectively improves the association performance between low-confidence targets and trajectories.

[0024] According to the updated matching cost matrix, the Hungarian algorithm is used to match the trajectory with the target, and then the trajectory matching is completed to obtain the matching result; the matching result is mapped to the target detection frame to obtain the multi-target tracking result; at this point, the multi-target tracking of the current frame UAV visual image is completed.

[0025] Step 4: Use the Trend-NSA Kalman Filter (T-NSA) to update each trajectory. The updated trajectory is used for multi-target tracking in the next frame of UAV visual image. The trend change adaptive Kalman filter is an improvement on the adaptive Kalman filter. The adaptive Kalman filter uses the confidence of the detected object to adaptively adjust the noise covariance of the Kalman filter. However, this form is relatively simple and does not fully consider other potential influencing factors. It is difficult to cope with complex and changeable situations by relying solely on the confidence of the trajectory as an influencing parameter. In order to improve the prediction accuracy and robustness, the present invention proposes a trajectory influencing factor based on the trend change factor, and uses the trajectory influencing factor to influence the noise covariance of the Kalman filter, thereby obtaining a trend change adaptive Kalman filter to better adapt to more complex situations in drone scenes. The trend change adaptive Kalman filter updates the trajectory according to the tracking results of the drone visual image of the current frame, and its noise covariance update formula is: (6) (7) in, is the noise covariance of the i-th trend change adaptive Kalman filter at time t, is the influence factor of the ith trajectory at time t, is a hyperparameter; , , and are the noise covariance of the abscissa, ordinate, width and height of the ith trajectory at time t. If the trajectory matches the target successfully, the abscissa, ordinate, width and height of the trajectory are the abscissa, ordinate, width and height of the target; if the trajectory does not match successfully, the abscissa, ordinate, width and height of the trajectory remain at the current values.

[0026] It can be seen from formula (6-7) that the matching cost change value is used as the cost change coefficient of the trajectory, and the confidence change value is used as the confidence change coefficient of the trajectory, so that the Kalman filter can perceive the trajectory trend change while controlling the noise covariance, and further optimize the motion prediction accuracy; in addition, the trajectory influencing factor only affects the noise covariance of the horizontal axis and the noise covariance of the vertical axis. This is because from the perspective of the drone, objects usually do not deform significantly in a short period of time, so excessive influence on the noise covariance of the width and the noise covariance of the height will lead to performance degradation.

[0027] Example In this embodiment, edgeyolo, a variant of the yolox model, is used as the target detection model, and experiments are conducted on two datasets, VisDrone2019 and UAVDT, respectively. The method of the present invention is compared with the current advanced multi-target tracking method. The experimental results are shown in Table 1. Among them, the recognition average ratio takes into account the number of correctly matched targets and the number of unmatched targets, and is used to evaluate the tracking performance at the target level. The larger the value, the longer a target can be accurately tracked. The multi-target tracking accuracy refers to the distance between the predicted trajectory and the true trajectory, which reflects the accuracy of the tracking result. The larger the value, the higher the tracking accuracy. The multi-target tracking accuracy rate comprehensively considers factors such as missed detection, false detection, and trajectory errors. It is one of the main indicators for evaluating the performance of multi-target tracking. The larger the value, the better the tracking performance.

[0028]

[0029] The experimental results show that compared with the current advanced multi-target tracking methods, the method of the present invention shows excellent performance on both data sets and maintains a good running speed. This is because the present invention conducts an in-depth analysis of the target tracking characteristics in the drone scene, proposes a trend change factor that can reflect the trend of target state change, and further proposes a trend change adaptive Kalman filter, a trend change group matching strategy and a fast state compensation mechanism based on the trend change factor, which effectively solves the problems of mutual occlusion, perspective change and camera shake. The present invention not only improves the accuracy of target motion state prediction, but also achieves an overall improvement in speed and performance. Specifically, the trend change adaptive Kalman filter enhances the trajectory stability by introducing a trend change factor, thereby improving the accuracy of motion prediction; the trend change group matching strategy constructs priority weights through trend changes, thereby expanding new dimensions to further distinguish trajectories, effectively reducing the matching error in the case of mutual occlusion of targets; and the fast state compensation mechanism solves the error expansion problem caused by the offset caused by camera shake through simple and efficient camera motion compensation. In addition, the method of the present invention only needs to train the target detection model, without additional training cost, which further highlights its superiority.

[0030] The sources of the above methods are as follows: [1] Liu S, Li X, Lu H, et al. Multi-object tracking meets moving UAV[C] / / Proceedings of the IEEE / CVF Conference on Computer Vision and PatternRecognition. 2022: 8876-8885. [2] Cao J, Pang J, Weng X, et al. Observation-centric sort:Rethinking sort for robust multi-object tracking[C] / / Proceedings of the IEEE / CVF conference on computer vision and pattern recognition. 2023: 9686-9696. [3] Yao M, Wang J, Peng J, et al. FOLT: Fast Multiple Object Tracking from UAV-captured Videos Based on Optical Flow[C] / / Proceedings of the 31stACM International Conference on Multimedia. 2023: 3375-3383. Any matters not described in the present invention are applicable to the prior art.

Claims

1. A method for tracking multiple targets using UAV vision based on trend changes, characterized in that: The steps include: Step 1: Train the target detection model and use the trained target detection model to detect targets in UAV visual images; divide the targets into high-confidence targets and low-confidence targets according to the confidence level, and obtain a high-confidence target set and a low-confidence target set; Step 2: Determine whether the trajectory needs to be compensated; if necessary, calculate the global offset to compensate the trajectory, and then proceed to step 3 after the compensation is completed; If not required, proceed directly to step 3; Global offset The calculation formula is as follows: (1) (2) (3) (4) in, is the reliability weight of the i-th successfully matched trajectory, , is the matching cost and confidence of the i-th successfully matched trajectory, and is a hyperparameter, , is the confidence change value and matching cost change value of the i-th successfully matched trajectory, , is the coordinate offset and size offset of the i-th successfully matched trajectory, n is the number of successfully matched trajectories, is the confidence of the i-th trajectory at time t-1 and t-2, , is the matching cost of the i-th trajectory at time t-1 and t-2; Step 3: Match the trajectory with the target, map the matching result to the target detection frame, and obtain the multi-target tracking result; Step 4: Use the trend change adaptive Kalman filter to update each trajectory, and the updated trajectory is used for multi-target tracking in the next frame of UAV visual image; The trend change adaptive Kalman filter updates the trajectory according to the tracking result of the current frame UAV visual image, and its noise covariance update formula is: (6) (7) in, is the noise covariance of the i-th trend change adaptive Kalman filter at time t, is the influence factor of the ith trajectory at time t, is a hyperparameter; , , and are the noise covariance of the abscissa, ordinate, width and height of the ith trajectory at time t. If the trajectory matches the target successfully, the abscissa, ordinate, width and height of the trajectory are the abscissa, ordinate, width and height of the target; if the trajectory does not match successfully, the abscissa, ordinate, width and height of the trajectory remain at the current values.

2. The method for tracking multiple targets using UAV vision based on trend changes according to claim 1 is characterized in that: For the matching of trajectory and target, the priority weight of the trajectory is calculated according to formula (5). The smaller the priority weight, the higher the priority of the trajectory. (5) in, It is The priority weight of each trajectory; Set the number of matches to C, the priority range to Z, initialize the set of participating matching trajectories to empty, and the set of non-participating matching trajectories to the set formed by all trajectories; Match the trajectories according to the priority. For the kth match, add the trajectories with priority weight less than k×Z to the set of participating matching trajectories, and remove these trajectories from the set of unparticipated matching trajectories. Match each trajectory in the set of participating matching trajectories with each high-confidence target in the set of high-confidence targets. If the intersection and union ratio of the trajectory and the high-confidence target is greater than the set threshold, the match is successful. The successfully matched trajectory is removed from the set of participating matching trajectories, and the unsuccessfully matched trajectory is retained in the set of participating matching trajectories to continue to participate in the next match, and the matching cycle is repeated C-1 times. For the last match, merge the current set of unmatched trajectories with the set of matched trajectories to obtain a merged set of unmatched trajectories; merge the unmatched high-confidence targets in the high-confidence target set with the low-confidence target set to obtain a mixed set of high- and low-confidence targets; combine each target in the mixed set of high- and low-confidence targets with each trajectory in the merged set of unmatched trajectories to obtain multiple candidate matching pairs; calculate the intersection-and-union ratio of each candidate matching pair, use the intersection-and-union ratio as a matrix element, and obtain a matching cost matrix; for low-confidence targets, if the intersection-and-union ratio of the candidate matching pair where the low-confidence target is located is less than the filtering threshold, it is an unreliable candidate matching pair, set the intersection-and-union ratio of the unreliable candidate matching pair to zero, update the matching cost matrix, and traverse all candidate matching pairs where the low-confidence targets are located to obtain an updated matching cost matrix; According to the updated matching cost matrix, the Hungarian algorithm is used to match the trajectory with the target, thereby completing the matching of all trajectories.

3. The method for tracking multiple targets using UAV vision based on trend changes according to claim 2 is characterized in that: For the setting of filtering threshold, all low-confidence targets are grouped according to the confidence interval of the low-confidence target and the set confidence interval interval, and a filtering threshold is set for each group according to the value range of the confidence interval; the higher the confidence group, the lower the filtering threshold.

4. The method for tracking multiple targets using UAV vision based on trend changes according to any one of claims 1 to 3, characterized in that: The process of judging whether the trajectory needs to be compensated is as follows: 1) Screen the trajectories to select stable trajectories; 2) Match the selected trajectories with each high-confidence target in the high-confidence target set. If the intersection and union ratio of the trajectory and the high-confidence target is greater than the set threshold, the match is successful; calculate the average matching cost of all successfully matched trajectories; if the ratio of the average matching cost of the successfully matched trajectories at the previous moment and the current moment is less than or equal to the set threshold, no compensation is required, otherwise compensation is required.

5. The method for tracking multiple targets using UAV vision based on trend changes according to claim 4 is characterized in that: The screening condition of the trajectory is at least one of confidence, matching cost and change range.

Images