A traffic vehicle tracking method based on video frame skipping processing

Abstract

The application provides a traffic vehicle tracking method based on video frame skipping processing, uses a DeepSort algorithm, introduces a cascade matcher, improves the ID-Switch phenomenon caused by only one matching in the Sort algorithm, and simultaneously considers distance measurement and appearance measurement between vehicles, adds a gating matrix to limit possible abnormal values of a certain measurement, and if the cascade matching fails, a certain track or target is not immediately deleted, but IOU-Dis matching is continuously performed; the Euclidean distance is introduced in the IOU-Dis matching, the vehicle displacement is obviously increased in the frame skipping case, the Euclidean distance measurement can prevent error matching of a single IOU distance, the maximum life cycle of the track is set, the target shielding situation can be coped with, the target is matched back to the original track in the maximum life cycle, and the ID-Switch phenomenon is reduced. The application has the advantage that a large displacement target can be tracked.

Description

Technical Field

[0001] This invention relates to target tracking and traffic parameter extraction technologies, specifically to a traffic vehicle tracking method based on video frame skipping processing. Background Technology

[0002] In traffic management, traffic flow parameters are essential analytical data. Extracting traffic flow parameters from intersection surveillance video requires vehicle tracking. Based on the number of tracking tasks, this can be categorized into single-target tracking and multi-target tracking. In the context of traffic flow calculation, multi-target tracking algorithms are necessary. Multi-target processing can be further divided into online tracking and offline tracking from a process perspective. Online tracking processes video frame by frame, applying only the current frame image for tracking. Offline tracking, on the other hand, can store and utilize information from historical frames and even future frames to optimize the entire video, but it is often limited by speed issues in practical applications.

[0003] Due to the uncertainty of road conditions, multi-object tracking modeling is quite complex. Besides associating identical targets between two frames, multi-object tracking also requires handling frequent occlusion, false positives and false negatives by the detector, and high target similarity. In particular, edge computing devices are easily deployed at intersections to directly calculate traffic flow parameters. Although edge devices are small and easy to deploy, their limited computing power makes processing multiple video streams very slow.

[0004] Traditional tracking algorithms often have their limitations, such as frequent ID hopping, slow speed, and simplistic matching strategies. Deep learning-based feature extraction can achieve more accurate matching. Furthermore, in edge devices with extremely limited computing power, performance requirements are very high, which traditional tracking algorithms cannot meet. Summary of the Invention

[0005] The purpose of this invention is to provide a traffic vehicle tracking method based on video frame skipping processing.

[0006] The technical solution to achieve the purpose of this invention is as follows: a traffic vehicle tracking method based on video frame skipping processing, which improves the matching mechanism, utilizes the directional characteristics of traffic flow for matching optimization, and uses the Kalman filter algorithm and the Hungarian matching algorithm for vehicle tracking, including the following steps:

[0007] Step 1: Obtain road video through roadside monitoring, determine the direction of traffic flow, and perform frame skipping processing;

[0008] Step 2: Obtain the target detection box of the vehicle in each frame of the video using the target detection algorithm, including position and size information, and input the target detection box into the cascade matcher;

[0009] Step 3: Use the cascade matcher to perform an initial match between each input target detection box and the predicted box of each trajectory in the cascade matching trajectory library. The cascade matching trajectory library is empty when matching in the first frame. If the match is successful, proceed to step 5; if the match fails, proceed to step 4.

[0010] Step 4: Perform a secondary matching between the input target detection box and the IOU-Dis trajectory library, i.e., IOU-Dis matching. If IOU-Dis matching fails, determine the target status based on the type of matching failure, and determine whether to add it to the trajectory set based on the target status. The matching of this frame ends, and jump to step 2. If IOU-Dis matching is successful, jump to step 5.

[0011] Step 5: Perform Kalman filtering to update the predicted bounding boxes of each trajectory in the successfully matched IOU-Dis trajectory library to obtain the updated target position and velocity information. Then, confirm the target's marking information and store the updated trajectory into the trajectory set.

[0012] Step 6: Perform Kalman filter prediction on each trajectory in the trajectory set to predict the bounding box of each trajectory at the next time step and save it in the trajectory.

[0013] Step 7: Check the Kalman filter prediction results generated in Step 6. If the result is marked as confirmed, proceed to Step 3 to continue tracking; if the result is marked as unconfirmed, proceed to Step 4 to continue tracking.

[0014] Furthermore, in step one, the road video is a top-down view of the intersection, including all lanes in one direction, with the traffic flow direction perpendicular to the shooting image. Frame skipping refers to discarding some frames in the video stream and using only the remaining frames for tracking processing.

[0015] Furthermore, in step two, the target detection algorithm selected is the YOLOv4 recognition algorithm.

[0016] Furthermore, in step three, the cascaded matcher employs a cyclic matching strategy, and the specific matching process is as follows:

[0017] Step 3-1: For the input data, namely the target detection box position and the target detection box size output by the target detection algorithm, perform distance measurement and appearance measurement calculation on each detection box. The distance measurement for the initial matching uses Mahalanobis distance. The appearance measurement refers to the comparison of appearance features between targets. That is, the Re-ID re-identification algorithm is used to extract appearance features of the region within the detection box and calculate the cosine distance of the appearance features.

[0018] Step 3-2: Calculate the cost matrix. The cost matrix is ​​used for Hungarian matching and contains information on both distance and appearance metrics, weighted accordingly. The cost matrix is ​​calculated as follows:

[0019] c ij =λd m (A i B j )+(1-λ)d c (A i B j )

[0020] Where A i B j Let d represent the i-th target detection box to be calculated and the predicted box of the j-th trajectory in the cascaded matching trajectory library, respectively. m (A i B j ) represents target A i With B j The Mahalanobis distance between them, d c (A,B) represents target A i With B j The cosine distance between them, where λ represents the set weight parameter, which is a value between 0 and 1;

[0021] Step 3-3: Calculate the threshold matrix. The threshold matrix is ​​used to limit the range of metric values. The threshold matrix is ​​a matrix with only 0 and 1 values. If a value in the cost matrix corresponds to 0 in the threshold matrix, then the two objectives are considered to be unmatched. The threshold matrix is ​​calculated as follows:

[0022]

[0023] Indicates the validity of Mahalanobis distance. The validity of the cosine distance is indicated by only two values: 0 and 1.

[0024] Step 3-4 involves performing Hungarian matching, which is done using the cost matrix calculated in step 3-2 and the threshold matrix calculated in step 3-3.

[0025] Furthermore, in step four, the secondary matching, IOU-Dis matching, requires calculating the IOU distance and Euclidean distance between each input target detection box and each trajectory prediction box in the IOU-Dis trajectory library. These two distances are weighted and used as a cost matrix, which is then calculated using the Hungarian matching algorithm to output the pairing status of each target box. The cost matrix is ​​calculated as follows:

[0026] c ij =λd e (A i B j )+(1-λ)d IOU (A i Bj )

[0027] Where A i B j Let d represent the i-th object detection box to be calculated and the predicted box of the j-th trajectory in the IOU-Dis trajectory library, respectively. e (A i B j ) represents the Euclidean distance, d IOU (A i B j ) represents the IOU distance, and λ represents the set weight parameter, which is a value between 0 and 1;

[0028] For the result of IOU-Dis matching failure, that is, the target and trajectory did not match successfully, there are two specific situations:

[0029] 1) If the trajectory is not matched, that is, a trajectory prediction box cannot match any detection box, the trajectory marking status is checked. If it is marked as confirmed, the life cycle judgment is continued. If the current trajectory life is less than the set life cycle, the trajectory is stored in the trajectory set. If the current trajectory life is greater than the life cycle, the trajectory is discarded directly. If it is marked as unconfirmed, the trajectory is discarded directly.

[0030] 2) If the target is not matched, that is, a target cannot be matched with any trajectory prediction box, it is considered that a new trajectory may have been generated. The target is added to the trajectory set as a new trajectory and marked as unconfirmed.

[0031] Furthermore, in step five, the Kalman filter update process takes the successfully matched target as input and requires the following data: the position and size of the target detection box, and the noise value. After the update, a confirmation mark is made, which means that the trajectory is successfully matched for three consecutive frames, and the trajectory is marked as confirmed. Otherwise, it is marked as unconfirmed.

[0032] Furthermore, in step six, the Kalman filter prediction process takes as input the position, height, width, and velocity of all target detection boxes included in a trajectory, thereby predicting the target detection boxes at the next moment of the current trajectory.

[0033] A traffic vehicle tracking system based on video frame skipping processing, which realizes traffic vehicle tracking based on video frame skipping processing based on the aforementioned traffic vehicle tracking method.

[0034] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements traffic vehicle tracking based on the traffic vehicle tracking method, using video frame skipping processing.

[0035] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements traffic vehicle tracking based on video frame skipping processing according to the traffic vehicle tracking method.

[0036] Compared with existing technologies, the significant advantages of this invention are: 1) It utilizes a cascaded matcher and IOU-Dis matching for trajectory determination and trajectory-target matching. During cascaded matching, object appearance metrics and trajectory metrics between objects are introduced. IOU-Dis matching is used to prevent missed detections in cascaded matching. If the vehicle displacement is large between two tracking attempts, a single IOU matching will not be sufficient for vehicle tracking. Euclidean distance is introduced to predict vehicle displacement. In situations requiring fast video tracking or where computing power is limited, frame skipping can be performed on the video to accelerate video processing. The direct problem caused by frame skipping is the large vehicle displacement between two recognition attempts, which this invention addresses. 2) It assigns a lifecycle attribute to already formed trajectories. If a target cannot be matched within the lifecycle, the trajectory is considered lost and removed from the trajectory set, avoiding unnecessary searching. For trajectories that are not yet formed, a count is assigned. When three consecutive matches are successful, it is marked as a formed trajectory; otherwise, it is considered a false detection and discarded. This approach can handle situations where the target is occluded. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the traffic vehicle tracking method for video frame skipping processing according to the present invention.

[0038] Figure 2 This is a diagram of a cascaded matcher.

[0039] Figure 3 This is a comparison chart for tracking experiments. Detailed Implementation

[0040] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0041] A video frame skipping method for traffic vehicle tracking, the main flowchart is as follows: Figure 1 The specific implementation includes the following steps:

[0042] Step 1: Identify the traffic sections that require traffic flow statistics and collect surveillance video from traffic intersections.

[0043] The road video is a top-down view of the intersection, which must include all lanes in one direction. The direction of traffic flow is perpendicular to the shooting image. Frame skipping means discarding some frames in the video stream and using only the remaining frames for tracking.

[0044] Step two involves determining the frame number and performing frame skipping calculations during video processing. For example, processing only even-numbered frames reduces the amount of data processed by half. Different frame skipping steps can also be set as needed. Video frame skipping can accelerate video processing and improve algorithm efficiency. The position and size information of the vehicle in each frame of the video are obtained through an object detection algorithm, generating object detection boxes. All object detection boxes in the current frame image are used as input to the tracker. The object detection algorithm is a recognizer that can arbitrarily select outputs containing vehicle position and size information; the accuracy of the recognizer has a certain impact on the tracking algorithm. This invention uses YOLOv4 as the recognition algorithm.

[0045] Step 3: Construct the cascaded matcher. The structure of the cascaded matcher is as follows: Figure 2 As shown. Cascaded matching follows the principle of prioritizing the matching of trajectories with the fewest missing frames and rejecting trajectories with more than a threshold missing frames. The cascaded matcher integrates appearance and distance metrics. The appearance metric uses cosine distance between features, while the distance metric uses Mahalanobis distance. Appearance features are extracted from the regions within the detection boxes using a Re-ID re-identification algorithm.

[0046] The Mahalanobis distance and cosine distance are calculated for the input target detection box and the detection boxes predicted from the cascaded matching trajectory library, respectively, and weights are assigned to calculate the cost matrix. The cost matrix is ​​calculated as follows:

[0047] c ij =λd m (A i B j )+(1-λ)d c (A i B j )

[0048] Where A i B j Let λ represent the target detection box of the i-th input and the trajectory prediction box in the j-th cascaded matching trajectory library, respectively, where λ is the fusion weight and d is the fusion weight. m (A i B j ) represents the Mahalanobis distance, d c (A i B j ) represents the cosine distance. λ represents the manually set weight, with a value between 0 and 1.

[0049] To eliminate outliers in the Mahalanobis and cosine distances, a threshold matrix is ​​introduced. The threshold matrix restricts the cost matrix by setting a maximum value for the metric. The threshold matrix is ​​a matrix containing only 0s and 1s; if a value in the cost matrix corresponds to a 0 in the threshold matrix, the two objectives are considered unmatched. The threshold matrix is ​​calculated as follows:

[0050]

[0051] Indicates the validity of Mahalanobis distance. The validity of the cosine distance is indicated by only two values: 0 and 1.

[0052] The cost matrix, obtained through the threshold matrix, is fed into the Hungarian matcher for cascaded matching, ultimately outputting the matching results between each trajectory and the input detection box. Depending on the matching results from the cascaded matcher, different cases are handled: if matching fails, proceed to step four; if matching succeeds, proceed to step five.

[0053] Step four involves performing IOU-Dis matching between the input target detection boxes and the trajectory database. The matching algorithm uses the Hungarian algorithm, where the cost matrix is ​​calculated as follows:

[0054]

[0055] d e (A,B) represents the Euclidean distance, d IOU (A,B) represents the IOU distance. λ represents a manually set weight, with a value between 0 and 1. The cost matrix is ​​fed into the Hungarian matching algorithm for calculation, and the matching result is output.

[0056] Matching results can be categorized as successful or unsuccessful. Unsuccessful matching can result in two scenarios: trajectory not matched or target not matched. If the trajectory is unmatched and already confirmed, it is added to the trajectory set. If the trajectory is unmatched and unconfirmed, it is deleted from the entire tracking algorithm. If the target is unmatched, it is considered a potentially new trajectory, treated as such, and marked as unconfirmed before being added to the trajectory set. A failed match terminates the tracking process for the current frame. A successful match proceeds to step five.

[0057] Step 5 involves updating the successfully matched target detection boxes and trajectories using Kalman filtering. The update process involves taking a trajectory as input and the target detection boxes that successfully match it, requiring the following data: the position and size of the target detection boxes, and noise values. The noise values ​​can be initialized with random values. The positions and sizes of the predicted trajectory boxes from the Kalman filter are then updated to match the positions and sizes of the successfully matched target detection boxes. The Kalman filter equation can be modified after the update. A confirmation marking operation is then performed; target trajectories that match successfully for three consecutive frames (including cascaded matching and IOU-Dis matching) are marked as confirmed. Finally, the trajectories with status markings are merged into the trajectory set.

[0058] Step six involves performing Kalman filtering prediction on each trajectory within the trajectory set. This prediction operation yields the position of the target detection box at the next time step for each trajectory. The inputs to the Kalman filtering prediction are the position, height, width, and velocity of the target detection box.

[0059] Step 7: Mark and check the trajectory Kalman filter prediction results generated in Step 6. If a trajectory is marked as confirmed, it is transferred to the cascaded matching trajectory library in Step 3 and waits for the next frame to track. If it is marked as unconfirmed, it is transferred to the IOU-DIS matching trajectory library in Step 4 and waits for the next frame to track.

[0060] This invention also proposes a traffic vehicle tracking system based on video frame skipping processing, which realizes traffic vehicle tracking based on video frame skipping processing based on the aforementioned traffic vehicle tracking method.

[0061] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements traffic vehicle tracking based on the traffic vehicle tracking method, using video frame skipping processing.

[0062] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements traffic vehicle tracking based on video frame skipping processing according to the traffic vehicle tracking method.

[0063] In summary, this invention improves performance during the data input stage and reduces the number of tracking algorithm calls by processing video frame skipping. This design can greatly improve the overall efficiency of the algorithm, but it increases the accuracy requirements of the tracking algorithm. This invention improves the matching mechanism of the DeepSort algorithm. The first matching adopts a cascaded matching strategy, prioritizing highly stable trajectories and handling occlusion. Since the displacement between two vehicle tracking operations increases in frame skipping cases, Euclidean distance is introduced in the second matching to reduce the IOU weight, and vehicle tracking is performed based on distance prediction. Simultaneously, a maximum lifetime for the trajectory is designed, which not only provides tolerance for target occlusion but also avoids invalid tracking after the target leaves.

[0064] Example

[0065] To verify the effectiveness of the present invention, the following experiment was conducted.

[0066] Simulation object: A 1200-frame video segment, used for vehicle tracking. The original DeepSort algorithm is compared with the tracking algorithm of this invention. The frame skipping step size is set to 2, meaning only odd-numbered frames are processed, and even-numbered frames are skipped.

[0067] Simulation environment: GTX2080Ti graphics card, 12GB VRAM; 128GB RAM; Intel(R) Xeon(R) Gold 5218 sixteen-core processor. The YOLOv4 recognizer performs vehicle identification on each frame of the input video, and the detection bounding box of the current frame is passed to the tracker of this invention. Once the tracking of the current frame is complete, the trajectory set of the current frame is saved, and the detection and tracking of the next incoming frame continues.

[0068] Experimental results: as attached Figure 3 As can be seen, the tracking algorithm of this invention outperforms the original DeepSort in all three metrics: IDF, IDR, and MOTA.

[0069] Experimental conclusions: IDF represents the F-value of trajectory recognition within the bounding box, summing the precision and recall of trajectory recognition; a higher value is better. IDP represents the precision of trajectory recognition within the bounding box; a higher value is better. IDR represents the recall of trajectory recognition within the bounding box. MOTA represents multi-target tracking accuracy. MOTP represents multi-target tracking precision. (See appendix...) Figure 3 As can be seen, the present invention can track vehicles more stably in frame-skipping acceleration scenarios, with higher tracking accuracy and recall rate, ensuring accurate tracking of vehicles in frame-skipping situations and minimizing false tracking.

[0070] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0071] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A traffic vehicle tracking method based on video frame skipping processing, characterized in that, The matching mechanism is improved by utilizing the directional nature of traffic flow for matching optimization. Vehicle tracking is performed using the Kalman filter algorithm and the Hungarian matching algorithm, including the following steps: Step 1: Obtain road video through roadside monitoring, determine the direction of traffic flow, and perform frame skipping processing; Step 2: Obtain the target detection box of the vehicle in each frame of the video using the target detection algorithm, including position and size information, and input the target detection box into the cascade matcher; Step 3: Use the cascade matcher to perform an initial match between each input target detection box and the predicted box of each trajectory in the cascade matching trajectory library. The cascade matching trajectory library is empty when matching in the first frame. If the match is successful, proceed to step 5; if the match fails, proceed to step 4. Step four involves performing a secondary matching process between the input target detection boxes and the IOU-Dis trajectory library, i.e., IOU-Dis matching. If IOU-Dis matching fails, the target state is determined based on the type of failure, and the target state is used to determine whether to add it to the trajectory set. The matching for this frame ends, and the process jumps to step two. If IOU-Dis matching succeeds, the process jumps to step five. The secondary matching, IOU-Dis matching, requires calculating the IOU distance and Euclidean distance between each input target detection box and each predicted trajectory box in the IOU-Dis trajectory library. These two distances are weighted and used as a cost matrix, which is then calculated using the Hungarian matching algorithm to output the pairing status of each target box. The cost matrix is ​​calculated as follows: ； in , They represent the numbers that need to be calculated. The target detection box and the IOU-Dis trajectory library in the first target detection box and the first object detection box. Predicted bounding boxes for each trajectory, Represents Euclidean distance. Indicates the IOU distance. This represents the set weight parameter, which is a value between 0 and 1; Step 5: Perform Kalman filtering to update the predicted bounding boxes of each trajectory in the successfully matched IOU-Dis trajectory library to obtain the updated target position and velocity information. Then, confirm the target's marking information and store the updated trajectory into the trajectory set. Step 6: Perform Kalman filter prediction on each trajectory in the trajectory set to predict the bounding box of each trajectory at the next time step and save it in the trajectory. Step 7: Check the Kalman filter prediction results generated in Step 6. If the result is marked as confirmed, proceed to Step 3 to continue tracking; if the result is marked as unconfirmed, proceed to Step 4 to continue tracking.

2. The traffic vehicle tracking method based on video frame skipping processing according to claim 1, characterized in that, In step one, the road video is a top-down view of the intersection, including all lanes in one direction. The direction of traffic flow is perpendicular to the shooting image. Frame skipping means discarding some frames in the video stream and using only the remaining frames for tracking processing.

3. The traffic vehicle tracking method based on video frame skipping processing according to claim 1, characterized in that, In step two, the target detection algorithm selected is the YOLOv4 recognition algorithm.

4. The traffic vehicle tracking method based on video frame skipping processing according to claim 1, characterized in that, In step three, the cascaded matcher employs a cyclic matching strategy, and the specific matching process is as follows: Step 3-1: For the input data, namely the target detection box position and the target detection box size output by the target detection algorithm, perform distance measurement and appearance measurement calculation on each detection box. The distance measurement for the initial matching uses Mahalanobis distance. The appearance measurement refers to the comparison of appearance features between targets. That is, the Re-ID re-identification algorithm is used to extract appearance features of the region within the detection box and calculate the cosine distance of the appearance features. Step 3-2: Calculate the cost matrix. The cost matrix is ​​used for Hungarian matching and contains information on both distance and appearance metrics, weighted accordingly. The cost matrix is ​​calculated as follows: ； in , They represent the numbers that need to be calculated. The target detection bounding box and the cascaded matching trajectory library are in the first target detection bounding box and the cascaded matching trajectory library. Predicted bounding boxes for each trajectory, Indicate target and Mahalanobis distance between them Indicate target and The cosine distance between them This represents the set weight parameter, which is a value between 0 and 1; Step 3-3: Calculate the threshold matrix. The threshold matrix is ​​used to limit the range of metric values. The threshold matrix is ​​a matrix with only 0 and 1 values. If a value in the cost matrix corresponds to 0 in the threshold matrix, then the two objectives are considered to be unmatched. The threshold matrix is ​​calculated as follows: ； Indicates the validity of Mahalanobis distance. The validity of the cosine distance is indicated by only two values: 0 and 1. Step 3-4 involves performing Hungarian matching, which is done using the cost matrix calculated in step 3-2 and the threshold matrix calculated in step 3-3.

5. The traffic vehicle tracking method based on video frame skipping processing according to claim 1, characterized in that, In step four, for the result of IOU-Dis matching failure, that is, the target and trajectory did not match successfully, there are two specific situations: 1) If the trajectory is not matched, that is, a trajectory prediction box cannot match any detection box, the trajectory marking status is checked. If it is marked as confirmed, the life cycle judgment is continued. If the current trajectory life is less than the set life cycle, the trajectory is stored in the trajectory set. If the current trajectory life is greater than the life cycle, the trajectory is discarded directly. If it is marked as unconfirmed, the trajectory is discarded directly. 2) If the target is not matched, that is, a target cannot be matched with any trajectory prediction box, it is considered that a new trajectory may have been generated. The target is added to the trajectory set as a new trajectory and marked as unconfirmed.

6. The traffic vehicle tracking method based on video frame skipping processing according to claim 1, characterized in that, In step five, the Kalman filter update process takes the successfully matched target as input and requires the following data: the position and size of the target detection box, and the noise value. After the update, a confirmation mark is made, which means that the trajectory is marked as confirmed if the trajectory matches successfully for three consecutive frames, otherwise it is marked as unconfirmed.

7. The traffic vehicle tracking method based on video frame skipping processing according to claim 1, characterized in that, In step six, the Kalman filter prediction process takes as input the position, height, width, and velocity of all target detection boxes included in a trajectory, and predicts the target detection boxes at the next moment of the current trajectory.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements traffic vehicle tracking based on video frame skipping processing according to the traffic vehicle tracking method of any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, wherein when executed by a processor, the computer program implements traffic vehicle tracking based on video frame skipping processing according to the traffic vehicle tracking method of any one of claims 1-7.

Images