An automated multi-target tracking algorithm for urban bus based on motion model
By employing Kalman filtering techniques with adaptive state transition matrix and camera motion compensation, the problem of multi-target tracking accuracy caused by jitter and complex scenes in automated urban buses was solved, achieving efficient target position prediction and improved tracking performance.
Patent Information
- Application Number
- CN202311012250.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-10
- Publication Date
- 2025-11-11
- Estimated Expiration
- 2043-08-10
AI Technical Summary
The accuracy of Kalman filter prediction of target position in automated urban buses is affected by vibration and complex tracking scenarios, resulting in poor multi-target tracking performance, especially at stops and densely populated areas where accurate tracking is difficult.
By employing Kalman filtering with an adaptive state transition matrix and camera motion compensation techniques, the detection box is separated through the initialization module, and image registration and matrix splitting correction are performed using the prediction module. Combined with the trajectory association module, data association is performed to improve the accuracy of target position prediction.
In automated urban buses, the accuracy and robustness of multi-target tracking algorithms have been improved, the computational load has been reduced, and the tracking performance has been enhanced by adapting to complex scenarios and the effects of jitter.
Smart Images
Figure CN116843720B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent vehicle technology, and in particular relates to an automated urban bus multi-target tracking algorithm based on a motion model. Background Technology
[0002] Intelligent vehicles have become an important direction for current and future transportation development. Their core can be divided into three layers: environmental perception, decision-making and planning, and control execution. Multi-target tracking is a hot research topic in the environmental perception layer. Its main implementation process involves outputting the motion trajectories of all targets in real time from onboard cameras and maintaining the identity information of each target for subsequent trajectory prediction and accurate location.
[0003] Motion modeling refers to modeling and estimating the trajectory of a target in a video. Since the displacement of a target between adjacent frames is often small, its position can be predicted by estimating its motion trend. In recent years, the Kalman filter, based on the assumption of a constant velocity model, has gradually become a popular choice for modeling object motion. However, some complex scenes, including camera motion, can affect the accuracy of the Kalman filter in predicting the target bounding box. Therefore, many researchers employ camera motion compensation, adjusting frames through image registration.
[0004] Research on autonomous driving is developing rapidly, and most related research has focused on autonomous vehicles, neglecting a more environmentally friendly version of autonomous vehicles, namely automated city buses. As a result, few people have paid attention to the additional challenges that multi-object tracking algorithms face in automated city buses.
[0005] From a structural perspective, automated city buses bear heavy loads and have high requirements for braking and power. Therefore, the chassis of automated city buses are all medium and heavy-duty truck chassis. The poor shock absorption of truck chassis causes significant vibration during operation, which causes the onboard camera to vibrate along with the vehicle. From an operational perspective, automated city buses need to stop at multiple stations, resulting in more frequent speed changes compared to ordinary cars. This causes the onboard camera to vibrate more frequently, leading to more noise and nonlinear motion for the tracked target, affecting the accuracy of the Kalman filter in predicting the target's position, and ultimately deteriorating the tracking performance. Furthermore, to facilitate public transportation, automated city buses often place their stops in areas with high pedestrian traffic, making the tracking scene from the onboard camera's perspective more complex. In such cases, the fixed state transition matrix of the Kalman filter is insufficient to accurately handle the state transitions of targets with varying motion patterns. Summary of the Invention
[0006] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.
[0007] To address the aforementioned technical problems, this invention provides the following technical solution: An initialization module reads the detection file to obtain the detection box information for each target in the current frame. Based on a threshold, the detection boxes are divided into high-resolution and low-resolution detection boxes and sent to the trajectory association module. A prediction module sets corresponding adaptive state transition matrices for the trajectories in the previous frame. Kalman filtering is used to obtain the predicted bounding boxes and covariance matrices of the trajectories in the previous frame in the current frame. Camera motion compensation is used to register images between adjacent frames. Affine transformation matrices between adjacent frames are obtained through image registration. These matrices are split into scaling, rotation, and transformation components. These two components are combined with the predicted bounding boxes and covariance matrices obtained through Kalman filtering to correct and obtain predicted bounding boxes and covariance matrices that better match the true position of the target. The trajectory association module performs intersection-union (IoU) calculations on the corrected predicted bounding boxes of the current frame with the high-resolution and low-resolution detection boxes. Predicted bounding boxes and detection boxes with high overlap exceeding a threshold are then fed into the Hungarian algorithm for data association. Detection boxes successfully associated with the corrected predicted bounding boxes are considered trajectories in the current frame.
[0008] The purpose of this invention is to improve the accuracy of the algorithm in predicting the target position by combining Kalman filtering with adaptive state transition matrix and camera motion compensation technology. To this end, an automated multi-target tracking algorithm for urban buses based on motion model is provided.
[0009] The present invention achieves the above-mentioned technical objectives through the following technical means.
[0010] As a preferred embodiment of the motion model-based automated urban bus multi-target tracking algorithm described in this invention, the initialization module specifically includes:
[0011] The video is converted into image frames, and the detector obtains the detection box information of each target in all image frames. The information includes the coordinates and confidence score of the detection box. The detection box information is stored in the detection file. The detection boxes are divided into high-scoring detection boxes and low-scoring detection boxes according to the confidence score threshold and sent to the trajectory association module.
[0012] As a preferred embodiment of the motion model-based automated urban bus multi-target tracking algorithm described in this invention, the prediction module specifically includes:
[0013] Retrieve the trajectories stored in the tracking list of the previous frame, set a corresponding adaptive state transition matrix for each trajectory, obtain the predicted bounding box and covariance matrix of the trajectory in the previous frame in the current frame through Kalman filtering, perform image registration between adjacent frames using camera motion compensation, obtain the affine transformation matrix, split the affine transformation matrix into two parts: scaling, rotation and transformation, and combine these two parts with the predicted bounding box and covariance matrix obtained by Kalman filtering respectively to correct and obtain a predicted bounding box and covariance matrix that more closely matches the true position of the target.
[0014] As a preferred embodiment of the motion model-based automated urban bus multi-target tracking algorithm described in this invention, the step of setting a corresponding adaptive state transition matrix F for each trajectory specifically includes:
[0015]
[0016] Where a represents the aspect ratio of the predicted bounding box, h represents the height of the predicted bounding box, and d... t This represents the difference between the current frame and the previous frame. and This represents the velocity change value of the corresponding variable.
[0017] As a preferred embodiment of the motion model-based automated urban bus multi-target tracking algorithm described in this invention, the step of obtaining the predicted bounding box and covariance matrix of the trajectory in the previous frame in the current frame through Kalman filtering specifically includes:
[0018] x′=Fx
[0019] p′=FpF T +Q
[0020] Where x represents the coordinates of the trajectory bounding box in the previous frame in the image frame, p represents the covariance matrix of the trajectory in the previous frame, F represents the adaptive state transition matrix, Q represents the noise matrix of the system, representing the reliability of the entire system, x′ represents the coordinates of the predicted bounding box in the current frame in the image frame, and p′ represents the covariance matrix of the predicted bounding box in the current frame.
[0021] As a preferred embodiment of the motion model-based automated urban bus multi-target tracking algorithm described in this invention, the step of using camera motion compensation to register images of adjacent frames and obtain the affine transformation matrix specifically includes:
[0022] The color space of the previous and current image frames is converted from BGR to grayscale. Based on the maximum number of corner points and the quality factor, corner points in the previous and current image frames are selected. The optimal affine transformation matrix is obtained from multiple pairs of corner points in the previous and current image frames using the RANSAC random sampling algorithm.
[0023] As a preferred embodiment of the motion model-based automated urban bus multi-target tracking algorithm described in this invention, the splitting of the affine transformation matrix into two parts—scaling / rotation and transformation—and the combination of these two parts with the predicted bounding box and covariance matrix obtained by Kalman filtering specifically includes:
[0024] A = [Z, T]
[0025] Where A represents the affine transformation matrix, Z represents the scaling and rotation matrix, and T represents the transformation matrix.
[0026] As a preferred embodiment of the motion model-based automated urban bus multi-target tracking algorithm described in this invention, the step of correcting the predicted bounding box and covariance matrix obtained by the Kalman filter after splitting the affine transformation matrix specifically includes:
[0027] x′ 校 =Zx′+T
[0028] p′ 校 =Zp′Z
[0029] Where, x′ 校 p′ represents the coordinates of the corrected predicted bounding box in the image frame. 校 Let represent the covariance matrix of the corrected predicted bounding box, Z represent the scaling and rotation matrix, T represent the transformation matrix, x′ represent the coordinates of the original predicted bounding box in the image frame, and p′ represent the covariance matrix of the original predicted bounding box.
[0030] As a preferred embodiment of the motion model-based automated urban bus multi-target tracking algorithm described in this invention, the trajectory association module specifically includes:
[0031] The corrected predicted bounding boxes of the current frame are then subjected to intersection-union (IoU) calculations with the high-scoring and low-scoring detection boxes. Predicted bounding boxes and detection boxes with high overlap exceeding a threshold are then fed into the Hungarian algorithm for data association. Detection boxes that are successfully associated with the corrected predicted bounding boxes are considered to be trajectories in the current frame.
[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0033] (1) Due to the heavy load and high requirements for braking and power of automated urban buses, medium and heavy truck chassis with poor shock absorption are used, which will generate large vibrations during operation. At the same time, the buses need to stop at multiple stations, and the speed will vary more than that of ordinary cars. Under the dual influence of structure and operation mechanism, the on-board camera faces more noise and nonlinear motion along with the bus vibration, which affects the accuracy of Kalman filter prediction of target position and ultimately makes the tracking effect worse. Therefore, camera motion compensation is adopted to correct the prediction results of Kalman filter under the frequent vibration of on-board camera and make up for the poor tracking effect caused by the bus itself.
[0034] (2) To facilitate people's travel, automated urban buses often place their stops in areas with high pedestrian traffic, making the tracking scene from the perspective of the onboard camera more complex. In this case, the fixed state transition matrix of the Kalman filter is difficult to accurately perform state transitions for targets with changing motion patterns. Therefore, an adaptive state transition matrix is adopted to accurately perform state transitions for targets with changing motion models, thereby improving tracking accuracy.
[0035] (3) Currently, mainstream tracking algorithms often combine appearance and motion models. However, given the limited hardware resources of automated urban buses, such as small embedded hardware platforms, their computing power is poor, and complex tracking models are often difficult to apply to the vehicle. This requires reducing the computational load of the algorithm while ensuring accuracy. Therefore, this invention achieves the best tracking effect with the least amount of computation by using only motion models without considering appearance models, thus enabling industrial implementation. Attached Figure Description
[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:
[0037] Figure 1 This is a flowchart illustrating the multi-target tracking algorithm for automated urban buses based on motion models, as described in an embodiment of the present invention.
[0038] Figure 2 The flowchart below illustrates the process of obtaining the affine transformation matrix using a motion model-based automated urban bus multi-target tracking algorithm as described in an embodiment of the present invention.
[0039] Figure 3 This is a schematic diagram illustrating the effect of the motion model-based automated urban bus multi-target tracking algorithm described in an embodiment of the present invention. Detailed Implementation
[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0041] The purpose of this invention is to overcome the shortcomings of the prior art and propose an automated multi-target tracking algorithm for urban buses based on a motion model. It adopts Kalman filtering with an adaptive state transition matrix and camera motion compensation to improve the problem that the prediction module of urban buses is unable to accurately predict the target position due to the platform.
[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0043] Example
[0044] To facilitate public transportation, automated city buses often have bus stops located in densely populated areas. This makes the tracking scene from the onboard camera's perspective more complex. In such cases, the fixed state transition matrix of the Kalman filter struggles to accurately handle the state transitions of targets with varying motion patterns. Furthermore, urban roads often have multiple bus stops, resulting in more speed changes for automated city buses compared to ordinary cars. This leads to increased motion for the onboard camera, affecting the accuracy of the Kalman filter's prediction of the target bounding box. Therefore, this invention combines an adaptive state transition matrix and camera motion compensation to improve the accuracy of the algorithm's target position prediction, thereby enhancing the overall multi-target tracking performance on automated city buses.
[0045] Reference Figure 1 and Figure 2 , Figure 3 As an embodiment of the present invention, an automated urban bus multi-target tracking algorithm based on a motion model is provided, comprising the following steps:
[0046] Step 1: Convert the video into image frames, use the detector to obtain the detection box information of each target in all image frames, including the coordinates and confidence score of the detection box, store the detection box information in the detection file, divide the detection boxes into high-scoring detection boxes and low-scoring detection boxes according to the confidence score threshold, and send them to the trajectory association module.
[0047] Step 2: Retrieve the trajectories stored in the tracking list of the previous frame, set the corresponding adaptive state transition matrix for each trajectory, obtain the predicted bounding box and covariance matrix of the trajectory in the previous frame in the current frame through Kalman filtering, perform image registration between adjacent frames using camera motion compensation, obtain the affine transformation matrix, split the affine transformation matrix into two parts: scaling, rotation and transformation, and combine these two parts with the predicted bounding box and covariance matrix obtained by Kalman filtering respectively to correct and obtain a predicted bounding box and covariance matrix that better fits the true position of the target.
[0048] The adaptive state transition matrix in step 2 is:
[0049]
[0050] Where a represents the aspect ratio of the predicted bounding box, h represents the height of the predicted bounding box, and d... t This represents the difference between the current frame and the previous frame. and This represents the velocity change value of the corresponding variable.
[0051] In step 2, the predicted bounding box and covariance matrix of the trajectory in the previous frame obtained through Kalman filtering in the current frame are as follows:
[0052] x′=Fx
[0053] p′=FpF T +Q
[0054] Where x represents the coordinates of the trajectory bounding box in the previous frame in the image frame, p represents the covariance matrix of the trajectory in the previous frame, F represents the adaptive state transition matrix, Q represents the noise matrix of the system, representing the reliability of the entire system, x′ represents the coordinates of the predicted bounding box in the current frame in the image frame, and p′ represents the covariance matrix of the predicted bounding box in the current frame.
[0055] In step 2, image registration between two adjacent frames is performed using camera motion compensation, resulting in the affine transformation matrix:
[0056] A = [Z, T]
[0057] Where A represents the affine transformation matrix, Z represents the scaling and rotation matrix, and T represents the transformation matrix.
[0058] After splitting the affine transformation matrix in step 2, the predicted bounding box and covariance matrix obtained by the corrected Kalman filter are:
[0059] x′ 校 =Zx′+T
[0060] p′ 校 =Zp′Z
[0061] Where, x′ 校 p′ represents the coordinates of the corrected predicted bounding box in the image frame. 校 Let represent the covariance matrix of the corrected predicted bounding box, Z represent the scaling and rotation matrix, T represent the transformation matrix, x′ represent the coordinates of the original predicted bounding box in the image frame, and p′ represent the covariance matrix of the original predicted bounding box.
[0062] Step 3: Perform intersection and union calculations on the corrected predicted bounding boxes of the current frame with the high-resolution and low-resolution detection boxes. Predicted bounding boxes and detection boxes with high overlap and exceeding the threshold are then fed into the Hungarian algorithm for data association. Detection boxes that are successfully associated with the corrected predicted bounding boxes are considered to be trajectories in the current frame.
[0063] Preliminary experimental results
[0064] To verify the feasibility and effectiveness of this invention, it was validated on the training sets of MOT17 and MOT20, and the preliminary results are as follows:
[0065]
[0066] The table above compares the performance of this invention on the MOT17 and MOT20 datasets with two other benchmark algorithms. Regarding the datasets: the MOT17 dataset captures the movement of urban buses at busy intersections, thus representing the camera motion problem faced by automated urban buses; the MOT20 dataset captures the behavior of crowded pedestrians in a train station, thus representing the complex tracking scenario of automated urban buses near platforms. Regarding evaluation metrics: MOTA represents the accuracy of multi-target tracking; higher accuracy indicates better tracking performance. IDS represents the total number of target ID switching caused by occlusion in multi-target tracking; fewer switching times indicate better tracking performance.
[0067] This invention discloses an automated urban bus multi-target tracking algorithm based on a motion model, belonging to the field of intelligent vehicle technology, and mainly includes an initialization module, a prediction module, and a trajectory association module. The initialization module reads the detection file to obtain the detection box information of each target in the current frame. Based on a threshold, the detection boxes are divided into high-resolution and low-resolution detection boxes and sent to the trajectory association module. The prediction module includes setting corresponding adaptive state transition matrices for the trajectories in the previous frame, obtaining the predicted bounding boxes and covariance matrices of the trajectories in the current frame through Kalman filtering, performing image registration between adjacent frames using camera motion compensation, obtaining the affine transformation matrices between adjacent frames through image registration, splitting the affine transformation matrix into scaling, rotation, and transformation parts, and combining these two parts with the predicted bounding boxes and covariance matrices obtained from Kalman filtering to correct and obtain predicted bounding boxes and covariance matrices that better fit the true position of the target. The trajectory association module performs intersection-union (IUU) calculations on the corrected predicted bounding boxes of the current frame with the high-resolution and low-resolution detection boxes. Predicted bounding boxes and detection boxes with high overlap exceeding a threshold are then fed into the Hungarian algorithm for data association. Detection boxes successfully associated with the corrected predicted bounding boxes are considered trajectories in the current frame. Because automated city buses often stop in high-traffic areas to facilitate people's travel, the tracking scene from the perspective of the onboard camera becomes more complex. In this case, the fixed state transition matrix of the Kalman filter is difficult to accurately perform state transitions for targets with variable motion patterns. Similarly, in urban roads, automated city buses stop at multiple stations, and their speed varies more than that of ordinary cars, causing the onboard camera to face more motion, thus affecting the accuracy of the Kalman filter in predicting the target bounding box. This invention improves the tracking performance of multi-target tracking algorithms on automated city buses by combining Kalman filtering with an adaptive state transition matrix and camera motion compensation.
[0068] Examples of the present invention have been described above with reference to the accompanying drawings. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be defined by the claims.
Claims
1. A multi-target tracking algorithm for automated urban public buses based on a motion model, characterized in that: The algorithm employs Kalman filtering with an adaptive state transition matrix and camera motion compensation, and specifically includes the following steps: Step 1: Initialization module, used to read the detection file generated by the detector, obtain the detection box information of each target in the current frame, and divide the detection boxes into high-scoring detection boxes and low-scoring detection boxes according to the threshold, and send them into the trajectory association module; Step 2: Prediction module, including setting corresponding adaptive state transition matrices for the trajectories in the previous frame, obtaining the predicted bounding box and covariance matrix of the trajectories in the previous frame in the current frame through Kalman filtering, performing image registration between two adjacent frames using camera motion compensation, obtaining the affine transformation matrix between the two adjacent frames through image registration, splitting the affine transformation matrix into scaling, rotation and transformation parts, and combining these two parts with the predicted bounding box and covariance matrix obtained by Kalman filtering respectively, and correcting to obtain a predicted bounding box and covariance matrix that better fits the true position of the target; Step 22 sets the corresponding adaptive state transition matrix F for each trajectory, which includes the following operations: Step 221: The adaptive state transition matrix F is as follows: Where a represents the aspect ratio of the predicted bounding box, h represents the height of the predicted bounding box, and d... t This represents the difference between the current frame and the previous frame. and This represents the velocity change value of each corresponding variable; Step 3: The trajectory association module performs intersection-union calculations on the corrected predicted bounding boxes of the current frame with the high-scoring and low-scoring detection boxes. Predicted bounding boxes and detection boxes with high overlap and exceeding the threshold are then fed into the Hungarian algorithm for data association. Detection boxes that are successfully associated with the corrected predicted bounding boxes are considered to be trajectories in the current frame.
2. The algorithm according to claim 1, characterized in that: The initialization module in step 1 specifically includes the following sub-steps: Step 11: Convert the video into image frames; Step 12: Use the detector to obtain the detection box information for each target in all image frames, where the information includes the coordinates and confidence score of the detection box; Step 13: Store the detection frame information in the detection file; Step 14: Divide the detection boxes into high-scoring and low-scoring detection boxes according to the confidence threshold and send them to the trajectory association module.
3. The algorithm according to claim 1, characterized in that: The prediction module in step 2 specifically includes the following sub-steps: Step 21: Retrieve the trajectory stored in the tracking list of the previous frame; Step 22: Set the corresponding adaptive state transition matrix for each trajectory; Step 23: Obtain the predicted bounding box and covariance matrix of the trajectory in the previous frame in the current frame using Kalman filtering; Step 24: Use camera motion compensation to register images of two adjacent frames to obtain the affine transformation matrix; Step 25: Split the affine transformation matrix into two parts: scaling, rotation, and transformation. Combine these two parts with the predicted bounding box and covariance matrix obtained by Kalman filtering, respectively. Step 26: Correction yields a predicted bounding box and covariance matrix that better match the true location of the target.
4. The algorithm according to claim 3, characterized in that: Step 23 obtains the predicted bounding box and covariance matrix of the trajectory in the previous frame in the current frame through Kalman filtering, specifically including the following operations: Step 231: The process of obtaining the predicted bounding box and covariance matrix of the trajectory in the previous frame in the current frame through Kalman filtering is as follows: x′=Fx p′=FpF T +Q Where x represents the coordinates of the trajectory bounding box in the previous frame in the image frame, p represents the covariance matrix of the trajectory in the previous frame, F represents the adaptive state transition matrix, Q represents the noise matrix of the system, representing the reliability of the entire system, x′ represents the coordinates of the predicted bounding box in the current frame in the image frame, and p′ represents the covariance matrix of the predicted bounding box in the current frame.
5. The algorithm according to claim 3, characterized in that: Step 24 uses camera motion compensation to register images of two adjacent frames to obtain an affine transformation matrix, which specifically includes the following operations: Step 241: Convert the color space of the previous image frame and the current image frame from BGR to grayscale space; Step 242: Based on the maximum number of corner points and the quality factor, filter out the corner points in the previous image frame and the corner points in the current image frame; Step 243: Obtain the optimal affine transformation matrix from multiple corner points of the previous and current image frames using the RANSAC random sampling algorithm.
6. The algorithm according to claim 3, characterized in that: Step 25 involves splitting the affine transformation matrix into two parts: scaling, rotation, and transformation. These two parts are then combined with the predicted bounding box and covariance matrix obtained by Kalman filtering, respectively. Specifically, this includes the following operations: Step 251: The process of splitting the affine transformation matrix into two parts, scaling, rotation, and transformation, is as follows: A = [Z, T] Where A represents the affine transformation matrix, Z represents the scaling and rotation matrix, and T represents the transformation matrix; Step 252: After splitting the affine transformation matrix, the process of correcting the predicted bounding box and covariance matrix obtained by the Kalman filter is as follows: x′ 校 =Zx′+T p′ 校 =Zp′Z Where, x′ 校 p′ represents the coordinates of the corrected predicted bounding box in the image frame. 校 Let represent the covariance matrix of the corrected predicted bounding box, Z represent the scaling and rotation matrix, T represent the transformation matrix, x′ represent the coordinates of the original predicted bounding box in the image frame, and p′ represent the covariance matrix of the original predicted bounding box.
7. The algorithm according to claim 1, characterized in that: The trajectory association module in step 3 specifically includes the following sub-steps: Step 31: Perform an intersection-union calculation on the corrected predicted bounding boxes of the current frame, the high-resolution detection boxes, and the low-resolution detection boxes; Step 32: Predicted bounding boxes and detection boxes with high overlap and exceeding the threshold are then fed into the Hungarian algorithm for data association; Step 33: The detection boxes associated with the successful and corrected predicted bounding boxes are considered to be the trajectories in the current frame.