| Multiple Object Tracking(MOT)consists of object detection and data association,which are two modules used to evaluate the location and the respective identity of targets in a video sequence respectively.The output result of object detection will be used as the input of data association for tracking targets,so the performance of object detection will directly affect the overall effect of MOT,and it is important to complete the detection task and improve the performance of the detector to better accomplish the MOT task.In addition,with the improvement of the performance of re-identification models used for data association,MOT is becoming more and more effective on pedestrians and vehicles and other targets with easily distinguishable appearance,and this trend in algorithmic development makes the tracking performance of existing algorithms for the targets have highly similar appearance degraded due to the inability to effectively distinguish targets by re-identification model.To address the above background and problems,this paper selects the Dance Track public dataset and the Soccer Net dataset which meet the characterstic of high similarity of target appearance,the main research contents of this paper are as follows:For the problem of missed detection and inaccurate bounding box regression in scenes of target with diverse motion,in order to improve the detection accuracy of targets,a feature decoupling-based object detection algorithm is proposed.In the improved network,the detection head uses Anchor-free mechanism,while the detection head is decoupled to solve the problem of inaccurate bounding box regression arising from conflicting regression and classification tasks.In addition,in order to retain more real targets in the network prediction results,this paper adds bounding boxes center information in the process of non-maximal suppression of redundant bounding boxes,helps alleviate the target omission problem and strengthen the detection performance of the algorithm by fusing the center distance with the intersection of union for calculation.Comparative experiments were carried out on the Dance Track dataset and the Soccer Net dataset,and the experimental results show that the detection accuracy of the two targets had a better improvement and was more conducive to the subsequent data correlation,and ablation experiments were designed to verify the effectiveness of each improved moudle.For the problem of performance degradation of tracking algorithm for targets with similar appearance,in order to improve the tracking accuracy of the targets,a data association algorithm based on cascade matching is proposed.For the targets with highly similar appearance,it is proved through comparative experiments that adding appearance information to the association metric will negatively affect the tracking performance,so only motion information is considered in the subsequent study.In this paper,the buffer is expanded for the intersection over union matching,and cascade matching is performed with different buffer ratios to make full use of motion information to achieve correct tracking of the target.and the low quality detection bounding boxes is additionally matched with the trajectory again after this cascade matching to find the real target from the targets that get lower detection scores and reduce the trajectory fragmentation.In addition,this paper mitigates the noise problem caused by the introduction of low quality detection bounding boxes to the algorithm by adding an adaptive noise scale to the Kalman filtering process.Finally,the MOT algorithm in this paper consists of an improved object detection algorithm and a data association algorithm based on cascade matching.The hyperparameters of the algorithm are confirmed by grid experiment,and comparative experiments are carried out on the Dance Track dataset and the Soccer Net dataset,and the experiment results show that the model in this paper has advantages in tracking performance compared with the baseline model and MOT models. |
PDF Full Text Request