A multi-target tracking method based on target confidence hierarchical association strategy

Abstract

The present invention discloses a multi-target tracking method based on a target confidence layered association strategy, comprising: inputting a video V , the video V Split into separate frame sequences f k ; Get frame f k The set of all pedestrian targets in D ; Using the occluded target motion state prediction method based on the detection frame change trend, predict the existing trajectory in the current frame f k According to the confidence layered association strategy of the predicted trajectory and the target, the predicted trajectory is associated with the detected target; and the set of successfully matched trajectories is saved. T matched , for each successfully matched trajectory in the current frame t k , using the observation value of the matched target in the current frame, update to obtain a more accurate motion state X k and speed of change P k ; Video V All frame sequences in the image are processed and multi-target recognition and tracking is output. The present invention has the characteristics of improving the accuracy and robustness of target tracking in crowded, complex and dynamic scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a multi-target tracking method based on a target confidence hierarchical association strategy. Background Art

[0002] The increasing use of cameras in our daily lives has created a ubiquitous network of digital eyes. These cameras are widely used in surveillance systems, autonomous driving systems, and robotic vision systems. However, a key requirement for camera data to support autonomous intelligent decision-making is the rapid and accurate identification and tracking of potential targets in a scene, particularly in systems with high real-time requirements. Furthermore, the intelligence of intelligent robots is inseparable from machine vision technology. Multi-object Tracking (MOT), which accurately identifies and tracks specific targets in complex, dynamic scenes with multiple people, is crucial for ensuring accurate behavioral perception and monitoring, and for providing personalized, human-like services.

[0003] The multi-target tracking task is usually decomposed into two sub-tasks: target detection and data association. Thanks to the rapid development of current target detection technology, good tracking speed has been achieved with the help of powerful detectors, and clues are combined in the data association task to assist in the association of targets and trajectories. In the existing technology, the association ability of the tracking algorithm is improved by utilizing "weak clues" such as detection box confidence and height changes. In the data association stage, cascade matching is a commonly used method in MOT research. DeepSORT prioritizes matching younger trajectories with detection objects, while ByteTrack prioritizes matching high-confidence detection targets with existing trajectories. However, the following problems still exist: the scale changes of the target itself during the tracking process and the frequent overlap between targets lead to tracking loss and identity ID mismatching, resulting in poor accuracy and robustness in tracking targets in complex dynamic scenes. Summary of the Invention

[0004] The purpose of the present invention is to overcome the above shortcomings and propose a multi-target tracking method based on a target confidence hierarchical association strategy that can improve the accuracy and robustness of target tracking in crowded, complex and dynamic scenes.

[0005] The present invention provides a multi-target tracking method based on a target confidence hierarchical association strategy, which comprises the following steps:

[0006] Step 1: Input video V and split the video V into separate frame sequences, each frame is denoted as f k , where k represents the frame number;

[0007] Step 2: For the current frame f k , use the target detector to detect f k Processing is performed to obtain frame f kThe set D of all pedestrian targets in k , set D k Each target d in k The detection confidence including the target detection box is recorded;

[0008] Each target d k , its detection confidence and the preset detection confidence threshold τ det Compare; if the target d k The detection confidence is greater than or equal to the threshold τ det , then the target is divided into the high confidence target set D high ; If the target d k The detection confidence is less than the threshold τ det , then the target is divided into the low confidence target set D low ;

[0009] Step 3: Motion state prediction: Use the motion state prediction method of the occluded target based on the detection frame change trend to predict the existing trajectory in the current frame f k The position in the , the set of trajectories up to the k-1th frame is T k-1 , the specific steps include:

[0010] Step 3.1: For the set T k-1 Each existing trajectory t k-1 , using a ten-tuple vector Model the motion state of the detection box corresponding to the target of the predicted trajectory, where x, y represent the center coordinates of the detection box corresponding to the target of the predicted trajectory in the image plane, w and s represent the width and area of ​​the detection box respectively, and c represents the detection confidence. Represent the speed of the corresponding variable respectively, and input the ten-tuple vector into the Kalman filter to predict the motion state of the existing trajectory in the current frame;

[0011] Step 3.2: For each current frame, predict the trajectory The prediction confidence and the preset trajectory confidence threshold τ tracklet Compare; if The prediction confidence is greater than or equal to the threshold τ tracklet , then divide this predicted trajectory into the high confidence trajectory set T high ;if The prediction confidence is less than the threshold τ tracklet , then the target is divided into a low confidence trajectory set T low ;

[0012] Step 4: Associating the predicted trajectory with the detected target: According to the confidence layered association strategy of the predicted trajectory and the target, the predicted trajectory is associated with the detected target. The specific steps are as follows:

[0013] Step 4.1: High confidence matching: Construct a set of high confidence prediction trajectories T high and high confidence detection target set D high The cost matrix C between stage1 , the cost is calculated based on the intersection-over-union (IoU) distance between the detection box and the trajectory prediction box; the Hungarian algorithm is used according to C stage1 In T high and D high Match between them to determine the trajectory of the foreground target in the video frame; retain the unmatched trajectory as T remain , the detection target that was not successfully matched is D remain ;

[0014] Step 4.2: Low confidence matching: Construct a low confidence prediction trajectory set T low And the detection target D that was not successfully matched in the previous stage remain The cost matrix C between stage2 , combined with the IoU distance and the confidence difference between the current frame and the previous frame when calculating the cost to improve the matching accuracy; merge the unmatched trajectories into T remain , the unmatched detection targets are merged into D remain ;

[0015] Step 4.3: Remaining target matching: Construct the trajectory T that was not successfully matched in the first two stages remain and low confidence detection target set D low The cost matrix C between stage3 , matching is performed based on the IoU distance calculation cost between the detection box and the trajectory prediction box, ensuring that all possible targets are considered to avoid omissions;

[0016] Step 5: Save the successfully matched trajectory set T matched , for each successfully matched trajectory t in the current frame k , using the observation value of the matched target in the current frame, update the more accurate motion state X k and the change rate P k ; The detected targets that are not successfully matched will be initialized as new tracks and tracked in the next frame;

[0017] Step 6: Loop through steps 2 to 5 until all frame sequences in the video V are processed, and then perform multi-target recognition and tracking.

[0018] The above-mentioned multi-target tracking method based on target confidence hierarchical association strategy, wherein: in step 2, the target detector is the YOLOX algorithm.

[0019] The above multi-target tracking method based on target confidence hierarchical association strategy, wherein: in step 2, each target d k The center coordinates, width and height of the target detection box and its detection confidence are recorded.

[0020] The above multi-target tracking method based on target confidence hierarchical association strategy, wherein: in step 2, the preset detection confidence threshold τ det is 0.6.

[0021] The above multi-target tracking method based on the target confidence hierarchical association strategy, wherein: in step 3.1, the motion state of the detection box is modeled as a ten-tuple vector input Kalman filter to predict the motion state of the existing trajectory in the current frame, let X = [x, y, w, s, c] T is the directly observed motion state; let Indicates the speed of change of the corresponding state; in the kth frame, the motion state of the previous frame is used to predict the motion state of the current frame. The calculation formula is:

[0022]

[0023] in, and X k-1 They represent the predicted state of the trajectory in the current k-th frame and the motion state observed in the k-1th frame respectively; F k It is the state transfer matrix of the kth frame under the assumption that the movement between detection frames is linear motion; and P k-1 Respectively represent the state change speed of the trajectory in the kth frame and the k-1th frame; Q k is the noise matrix of the system, and the calculation formula is:

[0024]

[0025] in, and They represent the predicted width and predicted area of ​​the detection box corresponding to the target in the k-1 frame, σ p Take 0.05, σ v Take 0.00625.

[0026] The above multi-target tracking method based on target confidence hierarchical association strategy, wherein: in step 3.2, the preset trajectory confidence threshold τ tracklet Based on the comparative experimental results under different values, it is set to 0.8.

[0027] The above multi-target tracking method based on the target confidence hierarchical association strategy, wherein: in step 4.2, the IoU distance and the confidence difference between the current frame and the previous frame are combined when calculating the cost, and the formula is:

[0028]

[0029] Among them, c k The target d detected by the target detector k The confidence value at the kth frame, Kalman filter is based on the trajectory sequence T up to the k-1 frame k-1 The confidence value of the history is used to predict its confidence value at the kth frame.

[0030] The above multi-target tracking method based on target confidence hierarchical association strategy, wherein: in step 5, the set of successfully matched trajectories T is saved. matched , for each successfully matched trajectory t in the current frame k , using the observation value of the matched target in the current frame, update the more accurate motion state X k and the change rate P k , the calculation formula is:

[0031]

[0032] Among them, K k is the Kalman gain, which is used to estimate the importance of the error; R k is the noise matrix of the detector; H k As the measurement matrix, the predicted value of the motion state Mapped to the detection space, combined with the observation value z k Calculating the error between the predicted value and the observed value and correcting the predicted value with the current observed value is conducive to gradually improving the accuracy of target position estimation.

[0033] Compared with the prior art, the present invention has obvious beneficial effects. As can be seen from the above scheme, in order to solve the problems of tracking loss and identity ID mismatch caused by the scale change of the target itself and the frequent overlap between targets during the tracking process, the present invention models the size of the target detection frame from the width and area, incorporates the detection confidence of the target into the motion state that needs to be modeled, and realizes the matching between the predicted trajectory and the detected target based on the IoU distance and confidence difference in the data association stage. At the same time, based on the construction of the confidence hierarchical association strategy of the predicted trajectory and the target, the tracker of the present invention is constructed in accordance with the tracking-by-detection paradigm, and the entire tracking task is regarded as the association problem of the current frame detection target and the existing trajectory. A multi-target tracking method (ScoreMOT) based on the predicted trajectory and target confidence hierarchical association strategy is proposed.

[0034] Among them, in order to improve the positioning ability of occluded targets by using width, area and confidence to model the motion state of the detection frame, and then provide confidence difference information between the predicted trajectory and the detected target for the trajectory association of the occluded target, the present invention proposes an occluded target motion state prediction method (MPOBV) based on the detection frame change trend. In addition, with the trajectory as the matching subject, the trajectory and the target are grouped according to the level of confidence, so that the overlapping targets are divided into two levels: foreground and background. The present invention further proposes a confidence hierarchical association strategy (SHAS) for predicted trajectories and targets to achieve separate matching of predicted trajectories and detected targets at different levels in frequently overlapping scenes. The main advantages of the present invention are:

[0035] 1) A multi-target tracking method based on a hierarchical association strategy of predicted trajectories and target confidence (ScoreMOT) is proposed. This method only uses motion information to achieve stable tracking of multiple targets in irregularly moving and dense scenes.

[0036] 2) A motion state prediction method for occluded targets based on the detection frame change trend is proposed. This method uses width, area and confidence to model the motion state of the detection frame to improve the positioning ability of the occluded target, and then provides confidence difference information between the predicted trajectory and the detected target for the trajectory association of the occluded target.

[0037] 3) A confidence-based hierarchical association strategy for predicted trajectories and targets is proposed. This method uses trajectories as the matching subject and groups them by confidence level, thereby dividing overlapping targets into two levels: foreground and background. This allows for matching predicted trajectories and detected targets at different levels in frequently overlapping scenes.

[0038] In summary, the present invention has the characteristics of being able to improve the accuracy and robustness of tracking targets in crowded, complex and dynamic scenes.

[0039] The beneficial effects of the present invention are further illustrated below through specific implementation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Schematic diagram of the multi-target tracking process of the present invention. DETAILED DESCRIPTION

[0041] The following is a detailed description of the specific implementation, features and efficacy of a multi-target tracking method based on a target confidence hierarchical association strategy proposed in the present invention in conjunction with the accompanying drawings and preferred embodiments.

[0042] See also Figure 1 The present invention provides a multi-target tracking method based on a target confidence hierarchical association strategy, which includes the following steps:

[0043] Step 1: Input video V and split the video V into separate frame sequences, each frame is denoted as f k , where k represents the frame number;

[0044] Step 2: For the current frame f k , using the target detector YOLOX to detect f k Processing is performed to obtain frame f k The set D of all pedestrian targets in k , set D k Each target d in k The center coordinates, width and height of the target detection box and its detection confidence are recorded;

[0045] Each target d k , its detection confidence and the preset detection confidence threshold τ det Compare; if the target d k The detection confidence is greater than or equal to the threshold τ det , then the target is divided into the high confidence target set D high ; If the target d k The detection confidence is less than the threshold τ det , then the target is divided into the low confidence target set D low ;

[0046] Step 3: Motion state prediction: Use the motion state prediction method of the occluded target based on the detection frame change trend to predict the existing trajectory in the current frame f k The position in the , the set of trajectories up to the k-1th frame is T k-1 , the specific steps include:

[0047] Step 3.1: For the set T k-1 Each existing trajectory t k-1 , using a ten-tuple vector Model the motion state of the detection box corresponding to the target of the predicted trajectory, where (x, y) represents the center coordinates of the detection box corresponding to the target of the predicted trajectory in the image plane, w and s represent the width and area of ​​the detection box respectively, and c represents the detection confidence. Let X = [x, y, w, s, c] T , is the directly observed motion state; let Indicates the speed of change of the corresponding state. In the kth frame, the motion state of the previous frame is used to predict the motion state of the current frame. The calculation formula is:

[0048]

[0049] in, and X k-1 They represent the predicted state of the trajectory in the current k-th frame and the motion state observed in the k-1th frame, respectively.k It is the state transition matrix of the kth frame under the assumption that the movement of the detection frame between frames is linear motion. Similarly, and P k-1 They represent the state change speed of the trajectory in the kth frame and the k-1th frame, Q k is the noise matrix of the system, and the calculation formula is:

[0050]

[0051] in, and They represent the predicted width and predicted area of ​​the detection box corresponding to the target in the k-1 frame, σ p Take 0.05, σ v Take 0.00625.

[0052] This implementation differs from the conventional motion prediction method of modeling motion states based on center position, width, and aspect ratio. Considering that changes in the detection box area during target movement are more stable than changes in aspect ratio, this implementation utilizes center position, width, and area for modeling. Furthermore, based on the variation in detection confidence when the target is occluded, a "weak clue" such as detection confidence is incorporated to assist in associating the trajectory with the detected target in the subsequent matching stage.

[0053] Step 3.2: For each current frame, predict the trajectory The prediction confidence and the preset trajectory confidence threshold τ tracklet Compare; if The prediction confidence is greater than or equal to the threshold τ tracklet , then divide this predicted trajectory into the high confidence trajectory set T high ;if The prediction confidence is less than the threshold τ tracklet , then the target is divided into a low confidence trajectory set T low ;

[0054] Step 4: Associating the predicted trajectory with the detected target: According to the confidence layered association strategy of the predicted trajectory and the target, the predicted trajectory is associated with the detected target. The specific steps are as follows:

[0055] Step 4.1: High confidence matching: Construct a set of high confidence prediction trajectories T high and high confidence detection target set D high The cost matrix C between stage1 , the cost is calculated based on the intersection-over-union (IoU) distance between the detection box and the trajectory prediction box; the Hungarian algorithm is used according to C stage1 In T high and D highMatch between them to determine the trajectory of the foreground target in the video frame; retain the unmatched trajectory as T remain , the detection target that was not successfully matched is D remain ;

[0056] Step 4.2: Low confidence matching: Construct a low confidence prediction trajectory set T low And the detection target D that was not successfully matched in the previous stage remain The cost matrix C between stage2 , combined with the IoU distance and the confidence difference between the current frame and the previous frame when calculating the cost to improve the matching accuracy; merge the unmatched trajectories into T remain , the unmatched detection targets are merged into D remain ; The cost is calculated by combining the IoU distance and the confidence difference between the current frame and the previous frame. The formula is:

[0057]

[0058] Among them, c k The target d detected by the target detector k The confidence value at the kth frame, Kalman filter is based on the trajectory sequence T up to the k-1 frame k-1 The historical confidence value predicts its confidence value at the kth frame;

[0059] Step 4.3: Remaining target matching: Construct the trajectory T that was not successfully matched in the first two stages remain and low confidence detection target set D low The cost matrix C between stage3 , matching is performed based on the IoU distance calculation cost between the detection box and the trajectory prediction box, ensuring that all possible targets are considered to avoid omissions;

[0060] In the absence of depth information, the front-to-back relationship of overlapping people can be distinguished by detecting differences in confidence. Targets with higher confidence are more likely to be targets in the foreground of the current frame, while targets with lower confidence are more likely to be background targets. Targets with higher confidence are more likely to be tracked correctly, while targets with lower confidence are more likely to be lost and have association errors. Therefore, this embodiment first divides both trajectories and targets into two groups, high and low, based on their confidence, to divide overlapping targets in the real world into two levels, foreground and background. Then, the predicted trajectories and detected targets at different levels are matched separately to avoid mutual interference between overlapping targets.

[0061] Step 5: Save the successfully matched trajectory set T matched , for each successfully matched trajectory t in the current frame k, using the observation value of the matched target in the current frame, update the more accurate motion state X k and the change rate P k , the calculation formula is:

[0062]

[0063] Among them, K k is the Kalman gain, which is used to estimate the importance of the error; R k is the noise matrix of the detector; H k As the measurement matrix, the predicted value of the motion state Mapped to the detection space, combined with the observation value z k Calculating the error between the predicted value and the observed value and correcting the predicted value with the current observed value will help to gradually improve the accuracy of target position estimation;

[0064] The detected targets that are not successfully matched will be initialized as new tracks and tracked in the next frame.

[0065] Step 6: Loop through steps 2 to 5 until all frame sequences in the video V are processed, and then perform multi-target recognition and tracking.

[0066] Performance Analysis:

[0067] ScoreMOT was implemented using the PyTorch framework on Ubuntu 20.04.6. To ensure fair comparison with other methods, in all subsequent tests, the publicly available YOLOX was used as the detector for the model, and the same detector parameters as ByteTrack were used. The confidence threshold τ for the detection box was det Set to 0.6, the trajectory confidence threshold τ tracklet The lag is set to 0.8, and all experiments are run on a single NVIDIA A10 GPU with 24GB RAM.

[0068] The DanceTrack dataset primarily contains group dance videos. Its main characteristics are the high similarity of the appearance of the people in the videos, their varied movement paths, and their complex joint movements.

[0069] In the comparison on the DanceTrack dataset, ScoreMOT is mainly compared with current SOTA online tracking methods such as ByteTrack, FairMOT, StrongSORT, GHOST, OC-SORT, etc.

[0070] Table 1 shows that on the DanceTrack test set, ScoreMOT achieved the best results in its three key metrics, MOTA, IDF1, and HOTA, at 92.7, 55.2, and 55.5, respectively, among all the CNN-based tracking methods compared. Among all the methods compared, ScoreMOT surpassed the MOTA score on the DanceTrack dataset with a score of 92.7, while achieving the second-best results in IDF1 and HOTA, at 55.2 and 55.5, respectively.

[0071] Table 1 Comparative performance results of the proposed method and other methods on the DanceTrack dataset

[0072]

[0073] In summary, in order to solve the problems of tracking loss and identity ID mismatch in scenes with dense crowds, similar appearances, and complex motion, the present invention proposes a simple and efficient online multi-target tracking method ScoreMOT. In the motion prediction module, ScoreMOT uses MPOBV to achieve target positioning when multiple similar targets frequently intersect, thereby providing useful clues for subsequent tracking and matching of occluded and overlapping targets. In the association stage, ScoreMOT uses SHAS to divide the predicted trajectory and the detected target into different levels for staged matching, which improves the accuracy of trajectory and target association in scenes where targets are frequently occluded and overlapped. In the overall structure, only a single deep network is used, which has a low computational load and is easy to deploy on edge devices.

[0074] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A multi-target tracking method based on a target confidence hierarchical association strategy, characterized by: The steps include: Step 1: Input video V and split the video V into separate frame sequences, each frame is denoted as f k , where k represents the frame number; Step 2: For the current frame f k , use the target detector to detect f k Processing is performed to obtain frame f k The set D of all pedestrian targets in k , set D k Each target d in k The detection confidence including the target detection box is recorded; Each target d k , its detection confidence and the preset detection confidence threshold τ det Compare; if the target d k The detection confidence is greater than or equal to the threshold τ det , then the target is divided into the high confidence target set D high ; If the target d k The detection confidence is less than the threshold τ det , then the target is divided into the low confidence target set D low ; Step 3: Motion state prediction: Use the motion state prediction method of the occluded target based on the detection frame change trend to predict the existing trajectory in the current frame f k The position in the , the set of trajectories up to the k-1th frame is T k-1 , the specific steps include: Step 3.1: For the set T k-1 Each existing trajectory t k-1 , using a ten-tuple vector Model the motion state of the detection box corresponding to the target of the predicted trajectory, where x and y represent the center coordinates of the detection box corresponding to the target of the predicted trajectory in the image plane, w and s represent the width and area of ​​the detection box respectively, and c represents the detection confidence. Represent the speed of the corresponding variable respectively, and input the ten-tuple vector into the Kalman filter to predict the motion state of the existing trajectory in the current frame; Step 3.2: For each current frame, the predicted trajectory The prediction confidence and the preset trajectory confidence threshold τ tracklet Compare; if The prediction confidence is greater than or equal to the threshold τ tracklet , then divide this predicted trajectory into the high confidence trajectory set T high ;if The prediction confidence is less than the threshold τ tracklet , then the target is divided into a low confidence trajectory set T low ; Step 4: Associate the predicted trajectory with the detected target: According to the confidence layered association strategy of the predicted trajectory and the target, associate the predicted trajectory with the detected target. The specific steps are as follows: Step 4.1: High confidence matching: Construct a set of high confidence prediction trajectories T high and high confidence detection target set D high The cost matrix C between stage1 , the cost is calculated based on the intersection-over-union (IoU) distance between the detection box and the trajectory prediction box; the Hungarian algorithm is used according to C stage1 In T high and D high Match between them to determine the trajectory of the foreground target in the video frame; retain the unmatched trajectory as T remain , the detection target that was not successfully matched is D remain ; Step 4.2: Low confidence matching: Construct a low confidence prediction trajectory set T low And the detection target D that was not successfully matched in the previous stage remain The cost matrix C between stage2 , combined with the IoU distance and the confidence difference between the current frame and the previous frame when calculating the cost to improve the matching accuracy; merge the unmatched trajectories into T remain , the unmatched detection targets are merged into D remain ; Step 4.3: Remaining target matching: Construct the trajectory T that was not successfully matched in the first two stages remain and low confidence detection target set D low The cost matrix C between stage3 , matching is performed based on the IoU distance calculation cost between the detection box and the trajectory prediction box, ensuring that all possible targets are considered to avoid omissions; Step 5: Save the successfully matched trajectory set T matched , for each successfully matched trajectory t in the current frame k , using the observation value of the matched target in the current frame, update the more accurate motion state X k and the change rate P k ; The detected targets that are not successfully matched will be initialized as new tracks and tracked in the next frame; Step 6: Loop through steps 2 to 5 until all frame sequences in the video V are processed, and then perform multi-target recognition and tracking.

2. The multi-target tracking method based on target confidence hierarchical association strategy according to claim 1, characterized in that: In step 2, the object detector is YOLOX.

3. The multi-target tracking method based on target confidence hierarchical association strategy according to claim 1, characterized in that: In step 2, each target d k The center coordinates, width, height, and detection confidence of the target detection box are recorded.

4. The multi-target tracking method based on target confidence hierarchical association strategy according to claim 1, characterized in that: In step 2, the preset detection confidence threshold τ det is 0.

6.

5. The multi-target tracking method based on target confidence hierarchical association strategy according to claim 1, characterized in that: In step 3.1, the motion state of the detection box is modeled as a ten-tuple vector input Kalman filter to predict the motion state of the existing trajectory in the current frame, let X = [x, y, w, s, c] T is the directly observed motion state; let Indicates the speed of change of the corresponding state; in the kth frame, the motion state of the previous frame is used to predict the motion state of the current frame. The calculation formula is: in, and X k-1 They represent the predicted state of the trajectory in the current k-th frame and the motion state observed in the k-1th frame respectively; F k It is the state transition matrix of the kth frame under the assumption that the detection frame δ moves in a linear motion; and P k-1 Respectively represent the state change speed of the trajectory in the kth frame and the k-1th frame; Q k is the noise matrix of the system, and the calculation formula is: in, and They represent the predicted width and predicted area of ​​the detection box corresponding to the target in the k-1 frame, σ p Take 0.05, σ v Take 0.00625.

6. The multi-target tracking method based on target confidence hierarchical association strategy according to claim 1, characterized in that: In step 3.2, the preset trajectory confidence threshold τ tracklet Based on the comparative experimental results under different values, it is set to 0.

8.

7. The multi-target tracking method based on target confidence hierarchical association strategy according to claim 1, characterized in that: In step 4.2, the cost is calculated by combining the IoU distance and the confidence difference between the current frame and the previous frame. The formula is: Among them, c k The target d detected by the target detector k The confidence value at the kth frame, Kalman filter is based on the trajectory sequence T up to the k-1 frame k-1 The confidence value of the history is used to predict its confidence value at the kth frame.

8. A multi-target tracking method based on a target confidence hierarchical association strategy according to any one of claims 1 to 7, characterized in that: In step 5, the set of successfully matched trajectories T is saved. matched , for each successfully matched trajectory t in the current frame k , using the observation value of the matched target in the current frame, update the more accurate motion state X k and the change rate P k , the calculation formula is: Among them, K k is the Kalman gain, which is used to estimate the importance of the error; R k is the noise matrix of the detector; H k As the measurement matrix, the predicted value of the motion state Mapped to the detection space, combined with the observation value z k Calculating the error between the predicted value and the observed value and correcting the predicted value with the current observed value is conducive to gradually improving the accuracy of target position estimation.

Images