Multi-pedestrian tracking method and system based on short trajectory splicing

Abstract

The invention discloses a multi-pedestrian tracking method and system based on short track splicing. The method comprises the following steps: firstly, detecting a video frame by using a YOLOX detector; secondly, performing detection and short-term tracking on the target in the video frame in each time period to generate a short track; then, the similarity between the tracks in the adjacent time periods is measured by calculating the appearance features and the motion features of the tracks, and the similar short tracks are spliced into a long track based on the similarity measurement; and finally, the spliced tracks are cleaned and smoothed, and the problem of track breakage caused by shielding is solved. According to the method, the robustness of multi-target tracking is effectively improved, and the method is particularly suitable for complex scenes with dense targets and serious shielding. A video sequence is segmented to generate a short-term target trajectory, and trajectory splicing is realized by using trajectory similarity measurement, so that a longer and more robust target trajectory is formed.

Description

Technical Field

[0001] The present invention relates to the technical field of monitoring target tracking, and in particular to a multi-pedestrian tracking method and system based on short trajectory splicing. Background Art

[0002] As a mid-level task in the field of computer vision, multi-target tracking has broad application prospects, such as security monitoring, behavior analysis, unmanned driving, and posture estimation. The task of multi-target tracking is to input a video and output the trajectory of the target appearing in the video. Since pedestrian tracking has a wide range of research value, multi-pedestrian tracking has become the mainstream research in the field of multi-target tracking.

[0003] In recent years, with the continuous improvement of the performance of detection algorithms, multi-pedestrian tracking based on the detection tracking framework has become the mainstream method for multi-pedestrian tracking. The principle of the detection tracking framework is to first detect pedestrians in each frame of the video, then extract the appearance or motion features of the pedestrians for data association and form the final motion trajectory. Existing methods usually learn reliable motion patterns to match the same target between adjacent frames, and learn discriminative appearance features to re-identify targets that have been lost for a long time. However, during the tracking process, dense crowds and extreme occlusions can easily affect the reliability of motion prediction and the discriminability of appearance. Therefore, designing a method to learn robust short-term and long-term motion to associate trajectories from short distances to long distances, and learning reliable long-term motion from the historical trajectory of the target is very important for the accuracy of multi-pedestrian tracking. Summary of the invention

[0004] To solve the above problems, the present invention proposes a multi-pedestrian tracking method based on segment short trajectory splicing. The method effectively improves the tracking robustness and accuracy in dense crowds, occlusion and target loss through segmented tracking and trajectory splicing. The method of the present invention first divides the input video sequence into multiple overlapping time periods, each of which contains a certain number of continuous frames. Short-term trajectories are generated in each time period to avoid interference with the tracking process caused by long-term occlusion and target loss. After the short-term trajectory is generated, the similarity between the trajectories is measured. These short-term trajectories are spliced ​​to form a longer target trajectory. In the process of trajectory management, the trajectory associated with the detection frame will be continuously updated. If a target cannot be matched with an existing trajectory in multiple frames, the trajectory will be terminated or deleted. At the same time, when a new target is detected and cannot be matched with an existing trajectory, a new trajectory will be initialized. For targets that reappear after being occluded, the present invention re-identifies them through a trajectory splicing module, and uses the similarity between their short-term trajectories and previous trajectories to determine their identities, thereby effectively adapting to changes in the appearance model and improving recognition accuracy. After trajectory splicing, the present invention further performs trajectory smoothing and error correction steps. By smoothing the trajectory, the errors and discontinuities that may occur during the stitching process can be reduced, thereby improving the overall continuity and accuracy of the trajectory. This process can effectively correct trajectory instability problems caused by occlusion, detection errors or other interference factors.

[0005] The purpose of the present invention can be achieved by the following technical solution: A multi-pedestrian tracking method based on short trajectory splicing, comprising the following steps:

[0006] Step S1, video frame processing and segmentation;

[0007] Step S2, target detection: in each time period, pedestrians in each frame are detected and multiple detection frames are output;

[0008] Step S3, short-term tracking: assign a unique ID to each detected pedestrian so that it can be tracked in a short period of time, and use the SORT tracker for short-term preliminary tracking to obtain a short trajectory;

[0009] Step S4, feature extraction: after obtaining the detection frame, extract the features of each detection frame for data association, including the appearance features and motion features of the pedestrian;

[0010] Step S5, data association: Calculate the cost matrix between the target detection frames in each pair of adjacent frames based on the appearance features and motion features, use the Hungarian algorithm to perform data association, and match the targets in adjacent frames;

[0011] Step S6, short track splicing: calculating the similarity of short tracks in adjacent time periods to perform short track splicing;

[0012] Step S7, trajectory smoothing and error correction: smoothing the final long trajectory to improve the continuity and accuracy of the trajectory;

[0013] Step S8, repeat steps S3 to S7 until all frames are processed and the target trajectory is output.

[0014] Furthermore, in step S1, the public dataset data is used to obtain each frame of the image, and the video frame is segmented and preprocessed. The number of frames in each time period is set for the video frame of the dataset, and the overlapping area is defined to ensure the continuity of the trajectory. The division method is:

[0015] T={F t ,F t+1 ,…,F t+N-1}

[0016] Where T represents the time period, F t Represents the frames in a time period, and N is the number of frames in each time period.

[0017] Furthermore, in step S2, a YOLOX detector is used to detect pedestrian targets in each frame, and each detection frame includes coordinates, size, and confidence.

[0018] Furthermore, the processing process of the SORT tracker in step S3 is: using Kalman filtering to predict the position state of the target at the next moment, constructing a global cost matrix based on the current detection frame and the Kalman prediction result, and using the Hungarian algorithm to obtain the global optimal solution.

[0019] Furthermore, in step S4, a convolutional neural network (CNN) is used to extract the appearance features of the detection frame, and the motion features are to calculate the speed and direction of the detection frame in consecutive frames; the calculation formulas of the appearance features and the motion features are as follows:

[0020] Appearance i =f CNN (B i )

[0021]

[0022] Among them, Appearance i is the appearance feature of the i-th detection box, f CNN is the feature extraction function, B i is the detection box coordinate, Motion i is the motion feature of the ith target, where the ith target corresponds to the ith detection frame, Δt is the time interval, and x i+1 -x i and i+1 -yi Indicates the motion offset of the target between the previous and next frames.

[0023] Furthermore, each element in the cost matrix of step S5 represents the matching cost between objects in adjacent frames, and the cost formula is as follows:

[0024] Cost i,j =α·Dist appearance (B i ,B j )+β·Dist motion (B i ,B j )

[0025] Cost i,j Dist is the cost between the detection box in the i-th frame and the detection box in the j-th frame. appearance and Dist motion They represent the distance metrics of appearance and motion features respectively, and α and β are weight coefficients.

[0026] Furthermore, in step S6, the short trajectories are matched using a similarity metric, and the calculation formula is as follows:

[0027] Sim(T i ,T j )=α·Sim appearance (T i ,T j )+β·Sim motion (T i ,T j )

[0028] Among them, Sim(T i ,T j ) represents the short trajectory T i and T j Similarity, Sim appearance and Sim motion They are the similarities of motion and appearance features respectively. The appearance similarity is measured by the cosine similarity between the two feature vectors. The motion similarity uses the Euclidean distance to measure the similarity of the motion trajectories of the two detection boxes. α and β are weight coefficients.

[0029] The formula for concatenating short trajectories in adjacent time periods into a longer target trajectory is:

[0030] T long =Concatenate(T i ,T j )

[0031] Among them, T long Concatenate(Ti ,T j ) represents the short trajectory T i and T j Stitched together.

[0032] Furthermore, after the segment trajectory splicing, it also undergoes trajectory management and re-identification, that is, during the trajectory splicing process, trajectory updating, re-identification, trajectory termination and trajectory initialization operations are performed to generate the final long trajectory;

[0033] The trajectory update operation refers to updating the short trajectory associated with the detection box, updating the existing trajectory according to the latest detection information, and updating the new detection box and the existing trajectory;

[0034] The re-identification operation refers to identifying the target that reappears after being blocked, re-identifying the target using the previously generated short trajectory, and comparing the similarity of the current short trajectory with the historical long trajectory:

[0035] ReID(T current ,T history )=Sim(T current ,T history )

[0036] Among them, ReID(T current ,T history ) represents the re-identification metric of the current short trajectory and the historical long trajectory;

[0037] The track termination operation refers to processing unmatched targets. If the target fails to find a matching detection box within multiple time periods, the track is terminated or deleted. Whether to terminate the track is determined based on the set threshold;

[0038] The trajectory initialization operation is to initialize a new target trajectory. When a new target is detected and cannot be matched with an existing trajectory, a new trajectory is initialized, a unique ID is assigned to the new target, and a new trajectory is created.

[0039] Furthermore, in step S7, the final long trajectory is smoothed using a Kalman filter, and the Kalman filter parameters are adjusted to adapt to the dynamic characteristics of the trajectory:

[0040]

[0041] in, is the smoothed trajectory position, F is the state transfer matrix, B is the control matrix, u t is the control input, w t is the process noise, Z t is the observation value, H is the observation matrix, v t is the observation noise.

[0042] The present invention also provides a multi-pedestrian tracking system based on short trajectory splicing, characterized in that it includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a multi-pedestrian tracking method based on short trajectory splicing as described in the above scheme.

[0043] Compared with the existing multi-pedestrian tracking technology, the present invention has the following advantages and beneficial effects:

[0044] 1) Compared with the prior art, the present invention divides the video sequence into multiple time periods for short-term tracking. The present invention can more effectively deal with the loss and reappearance of targets in a short period of time, and avoids the false association caused by occlusion or rapid movement of the target during the tracking of a single long trajectory. The present invention adopts a trajectory splicing strategy to splice short-term trajectories into longer continuous trajectories by calculating the motion and appearance similarity of the trajectories. This method can better handle the problem of target tracking in long-term occlusion and complex scenes, and ensure the continuity and accuracy of tracking.

[0045] 2) The present invention designs a flexible trajectory management mechanism, including operations such as trajectory update, re-identification, termination and deletion. This enables the tracking system to adapt to dynamic changes in targets and complex environmental conditions, such as new targets that suddenly appear or targets that have left the scene. By combining the correlation analysis of short-term and long-term trajectories, the present invention can effectively re-identify the target after it reappears, thereby significantly improving the tracking reliability under occlusion. The present invention introduces smoothing techniques such as Kalman filters to process the generated trajectories, eliminate noise and abnormal data, and generate smoother and continuous trajectories. This not only improves the tracking accuracy, but also makes the final output trajectory more consistent with the target's actual motion pattern.

[0046] 3) The method designed by the present invention can not only handle challenges such as dense crowds, occlusions, and complex backgrounds, but is also applicable to various practical applications, such as security monitoring, unmanned driving, and behavior analysis. This wide applicability makes the present invention have strong practical value in multiple fields. By decomposing the tracking task into short-term tracking and splicing, the present invention reduces the demand for computing resources. Compared with direct tracking over a long period of time, this segmented processing method can complete the target tracking task faster and reduce the computational complexity while maintaining high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0048] In order to facilitate ordinary technicians in the field to understand and implement the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0049] The present invention provides a multi-pedestrian tracking method based on short trajectory splicing, which detects and short-term tracks the targets in the video frames in each time period to generate short trajectories. The similarity between trajectories in adjacent time periods is measured by calculating the motion characteristics of the trajectories, and similar short trajectories are spliced ​​into long trajectories based on the similarity measurement. The spliced ​​trajectories are cleaned and smoothed to deal with the trajectory breakage problem caused by occlusion. The method effectively improves the robustness of multi-target tracking, and is particularly suitable for complex scenes with dense targets and severe occlusion. The video sequence is segmented to generate short-term target trajectories, and the trajectory similarity measurement is used to achieve trajectory splicing, thereby forming a longer and more robust target trajectory.

[0050] The specific implementation includes the following steps:

[0051] Step S1: Use the public dataset data to obtain each frame of the image and perform segmentation preprocessing on the video frame. Set the number of frames in each time period for the video frame of the dataset and define the overlapping area to ensure the continuity of the trajectory. Division method:

[0052] T={F t ,F t+1 ,…,F t+N-1}

[0053] Among them, T represents the time period, F t Indicates the frames in a time period, and N is the number of frames in each time period. If the total number of video frames is 5000, and 30 frames are selected for each time period, with an overlap of 10 frames, then the first time period will contain frames 1 to 30, the second time period will contain frames 21 to 50, and so on.

[0054] Step S2: Target detection. Use the YOLOX detector to detect pedestrian targets in each frame. Apply YOLOX to each frame to obtain the coordinates, confidence scores, and category labels of the detection boxes. For each frame, YOLOX outputs multiple detection boxes, each of which includes coordinates, size, and confidence.

[0055] Step S3: Short-term tracking. The SORT tracker performs short-term preliminary tracking to obtain a short trajectory. SORT tracking uses Kalman filtering to predict the position state of the target at the next moment, builds a global cost matrix based on the current detection box and the Kalman prediction result, and uses the Hungarian algorithm to obtain the global optimal solution. In each time period, a unique ID is assigned to the detected target, the ID of each target is recorded, and the ID is used for target identification during the tracking process.

[0056] Step S4: Feature extraction. Use a convolutional neural network (CNN) to extract the appearance features of the detection box. You can use a pre-trained network such as ResNet. Motion features calculate the motion features of the detection box in consecutive frames, such as speed and direction. The calculation formula is as follows:

[0057] Appearance i =f CNN (B i )

[0058]

[0059] Among them, Appearance i is the appearance feature of the i-th detection box, f CNN is the feature extraction function, B i The coordinates of the detection box, including the horizontal and vertical coordinates of the upper left corner of the detection box, as well as the width and height. i is the motion feature of the ith target, where the ith target corresponds to the ith detection frame, Δt is the time interval, and x i+1 -x i and i+1 -y i Indicates the motion offset of the target in the previous and next frames. i , use CNN to extract its appearance features Appearance i , calculate motion feature Motion i as speed and direction.

[0060] Step S5: Data association: Match the objects in adjacent frames, calculate the cost matrix between the object detection frames in each pair of adjacent frames based on the appearance features and motion features, and use the Hungarian algorithm to find the optimal match. Each element in the cost matrix represents the matching cost between the objects in adjacent frames, and the cost formula is as follows:

[0061] Cost i,j =α·Dist appearance (B i ,B j )+β·Dist motion (B i ,B j )

[0062] Cost i,j Dist is the cost between the detection box in the i-th frame and the detection box in the j-th frame. appearance and Dist motionThey represent the distance measurement of appearance and motion features respectively. α and β are weight coefficients, which are parameter settings. They are optimized through continuous parameter adjustment. For example, if the detection confidence is high, α can be 0.3 and β can be 0.7, and vice versa.

[0063] Step S6: short track splicing. Calculate the similarity of short tracks in adjacent time periods to perform short track splicing. Calculate the similarity measure of motion and appearance features for each pair of short tracks, and use the similarity measure to match the short tracks:

[0064] Sim(T i ,T j )=α·Sim appearance (T i ,T j )+β·Sim motion (T i ,T j )

[0065] Among them, Sim(T i ,T j ) represents the short trajectory T i and T j Similarity, Sim appearance and Sim motion They are the similarities of motion and appearance features respectively. The appearance similarity is measured by the cosine similarity between the two feature vectors. The motion similarity uses the Euclidean distance to measure the similarity of the motion trajectories of the two detection boxes. α and β are weight coefficients.

[0066] When splicing short trajectories in adjacent time periods into a longer target trajectory, a calculation-based similarity measurement is performed to connect similar short trajectories to generate a continuous long trajectory:

[0067] T long =Concatenate(T i ,T j )

[0068] Among them, T long Concatenate(T i ,T j ) represents the short trajectory T i and T j Splice together. k and T k+1 The short trajectories are spliced ​​together to form a continuous target trajectory to solve problems such as target loss during tracking.

[0069] Step S7: Track management and re-identification. Update the track associated with the detection box, update the existing track according to the latest detection information, and update it in combination with the new detection box and the existing track (the track that has been tracked and the track after the short track is spliced). Re-identify the target that reappears after being occluded, use the previously generated track (the short track obtained by SORT) to re-identify the target, and compare the similarity of the current short track with the historical long track (that is, the "track" that has been spliced ​​in the previous frames):

[0070] ReID(T current ,T history )=Sim(T current ,T history )

[0071] Among them, ReID(T current ,T history ) represents the re-identification metric of the current short trajectory and the historical long trajectory.

[0072] Processing unmatched targets: If a target fails to find a matching detection box within multiple time periods, the track is terminated or deleted. The set threshold is used to determine whether to terminate the track (i.e., if there is still no target in the unassociated track after a certain number of frames, the track is considered to be terminated). Initializing a new target track: When a new target is detected and cannot be matched with an existing track, a new track is initialized, a unique ID is assigned to the new target, and a new track is created.

[0073] Step S8: Data association. Smooth the generated long trajectory, reduce noise and anomalies, use the Kalman filter to smooth the trajectory, and adjust the filter parameters to adapt to the dynamic characteristics of the trajectory:

[0074]

[0075] in, is the smoothed trajectory position, F is the state transfer matrix, B is the control matrix, u t is the control input, w t is the process noise, Z t is the observation value, H is the observation matrix, v t is the observation noise.

[0076] Correct the errors that may occur during the trajectory splicing process, backtrack the generated trajectory, identify and correct the errors, and use historical data to adjust the trajectory.

[0077] In order to deal with problems such as dense crowds, occlusion and target loss, this paper proposes a multi-target tracking method based on short trajectory splicing. This method generates target trajectories within a short time range by segmenting the video sequence, and then splices these short trajectories to achieve longer-term target tracking. This method effectively improves the robustness of tracking by extending the traditional frame-by-frame association to trajectory association within a short period of time.

[0078] Specifically, the method of the present invention first divides the video sequence into multiple overlapping time periods, each of which includes a certain number of continuous frames. In each time period, a detection-based tracking framework is used to detect and associate pedestrian targets in the video to generate short-term target trajectories. Through this method, it is possible to avoid interference with the tracking process caused by long-term occlusion and target loss in a shorter time window. Next, the present invention measures the similarity between trajectories by calculating the similarity of motion features (such as speed, direction, etc.) of trajectories in adjacent time periods, and splices short trajectories based on these similarities to form longer and more stable target trajectories. Through this splicing method, the association between short trajectories becomes more robust, especially in complex scenes, and can better cope with target occlusion, motion changes, and background interference.

[0079] In addition, in order to further improve the continuity and accuracy of the trajectory, the present invention also includes trajectory cleaning and smoothing steps to deal with errors and discontinuities that may occur during the trajectory splicing process. This multi-target tracking method based on short trajectory splicing significantly improves the accuracy and robustness of tracking while keeping the computational complexity controllable, and is particularly suitable for multi-pedestrian tracking tasks in dense crowds and complex environments.

[0080] In summary, the present invention effectively solves the key problems faced by traditional multi-target tracking methods in complex scenes through an innovative short trajectory splicing method, and provides a more robust and accurate multi-target tracking solution.

[0081] On the other hand, an embodiment of the present invention also provides a multi-pedestrian tracking system based on short trajectory splicing, characterized in that it includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a multi-pedestrian tracking method based on short trajectory splicing as described in the above scheme.

[0082] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of patent protection of the present invention. Under the enlightenment of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested for the present invention shall be based on the attached claims.

Claims

1. A multi-pedestrian tracking method based on short trajectory splicing, characterized in that: The steps include: Step S1, video frame processing and segmentation; Step S2, target detection: in each time period, pedestrians in each frame are detected and multiple detection frames are output; Step S3, short-term tracking: assign a unique ID to each detected pedestrian so that it can be tracked in a short period of time, and use the SORT tracker for short-term preliminary tracking to obtain a short trajectory; Step S4, feature extraction: after obtaining the detection frame, extract the features of each detection frame for data association, including the appearance features and motion features of the pedestrian; Step S5, data association: Calculate the cost matrix between the target detection frames in each pair of adjacent frames based on the appearance features and motion features, use the Hungarian algorithm to perform data association, and match the targets in adjacent frames; Step S6, short track splicing: calculating the similarity of short tracks in adjacent time periods to perform short track splicing; Step S7, trajectory smoothing and error correction: smoothing the final long trajectory to improve the continuity and accuracy of the trajectory; Step S8, repeat steps S3 to S7 until all frames are processed and the target trajectory is output.

2. The method for tracking multiple pedestrians based on short trajectory splicing according to claim 1, characterized in that: In step S1, the public dataset data is used to obtain each frame of the image, and the video frame is segmented and preprocessed. The number of frames in each time period is set for the video frame of the dataset, and the overlapping area is defined to ensure the continuity of the trajectory. The division method is: T={F t ,F t+1 ,…,F t+N-1 } Where T represents the time period, F t Represents the frames in a time period, and N is the number of frames in each time period.

3. The method for tracking multiple pedestrians based on short trajectory splicing according to claim 1, characterized in that: In step S2, the YOLOX detector is used to detect pedestrian targets in each frame, and each detection frame includes coordinates, size and confidence.

4. The method for tracking multiple pedestrians based on short trajectory splicing according to claim 1, characterized in that: The processing process of the SORT tracker in step S3 is: use Kalman filtering to predict the position state of the target at the next moment, build a global cost matrix based on the current detection box and the Kalman prediction result, and use the Hungarian algorithm to obtain the global optimal solution.

5. The method for tracking multiple pedestrians based on short trajectory splicing according to claim 1, characterized in that: In step S4, a convolutional neural network (CNN) is used to extract the appearance features of the detection frame. The motion features are Calculate the speed and direction of the detection box in consecutive frames; the calculation formulas for appearance features and motion features are as follows: Appearance i =f CNN (B i ) Among them, Appearance i is the appearance feature of the i-th detection box, f CNN is the feature extraction function, B i is the detection box coordinate, Motion i is the motion feature of the ith target, where the ith target corresponds to the ith detection frame, Δt is the time interval, and x i+1 -x i and i+1 -y i Indicates the motion offset of the target between the previous and next frames.

6. The method for tracking multiple pedestrians based on short trajectory splicing according to claim 1, characterized in that: Each element in the cost matrix of step S5 represents the matching cost between targets in adjacent frames, and the cost formula is as follows: Cost i,j =α·Dist appearance (B i ,B j )+β·Dist motion (B i ,B j ) Cost i,j Dist is the cost between the detection box in the i-th frame and the detection box in the j-th frame. appearance and Dist motion They represent the distance metrics of appearance and motion features respectively, and α and β are weight coefficients.

7. The method for tracking multiple pedestrians based on short trajectory splicing according to claim 1, characterized in that: In step S6, the short trajectories are matched using a similarity metric, and the calculation formula is as follows: Sim(T i ,T j )=α·Sim appearance (T i ,T j )+β·Sim motion (T i ,T j ) Among them, Sim(T i ,T j ) represents the short trajectory T i and T j Similarity, Sim appearance and Sim motion They are the similarities of motion and appearance features respectively. The appearance similarity is measured by the cosine similarity between the two feature vectors. The motion similarity uses the Euclidean distance to measure the similarity of the motion trajectories of the two detection boxes. α and β are weight coefficients. The formula for concatenating short trajectories in adjacent time periods into a longer target trajectory is: T long =Concatenate(T i ,T j ) Among them, T long Concatenate(T i ,T j ) represents the short trajectory T i and T j Stitched together.

8. The method for tracking multiple pedestrians based on short trajectory splicing according to claim 1, characterized in that: After the segment trajectory is spliced, it also undergoes trajectory management and re-identification, that is, during the trajectory splicing process, trajectory update, re-identification, trajectory termination and trajectory initialization operations are performed to generate the final long trajectory; The trajectory update operation refers to updating the short trajectory associated with the detection box, updating the existing trajectory according to the latest detection information, and updating the new detection box and the existing trajectory; The re-identification operation refers to identifying the target that reappears after being blocked, re-identifying the target using the previously generated short trajectory, and comparing the similarity of the current short trajectory with the historical long trajectory: ReID(T current ,T history )=Sim(T current ,T history ) Among them, ReID(T current ,T history ) represents the re-identification metric of the current short trajectory and the historical long trajectory; The track termination operation refers to processing unmatched targets. If the target fails to find a matching detection box within multiple time periods, the track is terminated or deleted. Whether to terminate the track is determined based on the set threshold; The trajectory initialization operation is to initialize a new target trajectory. When a new target is detected and cannot be matched with an existing trajectory, a new trajectory is initialized, a unique ID is assigned to the new target, and a new trajectory is created.

9. The method for tracking multiple pedestrians based on short trajectory splicing according to claim 1, characterized in that: In step S7, the final long trajectory is smoothed using a Kalman filter, and the Kalman filter parameters are adjusted to adapt to the dynamic characteristics of the trajectory: in, is the smoothed trajectory position, F is the state transfer matrix, B is the control matrix, u t is the control input, w t is the process noise, Z t is the observation value, H is the observation matrix, v t is the observation noise.

10. A multi-pedestrian tracking system based on short trajectory splicing, characterized in that: The method comprises a processor and a memory, wherein the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a multi-pedestrian tracking method based on short trajectory splicing as described in any one of claims 1 to 9.

Images