Multi-object Tracking Algorithm Based On Joint Detection And Data Association

Multi-object tracking is a key issue in the field of digital image processing and vision,including multiple tasks such as detection,recognition and data association.And it has broad application prospects in social life,industrial production and military fields.With the development of detection algorithms,detection-based multi-object tracking algorithms have gradually become popular and have resulted in different tracking paradigms.Among them,the One-shot paradigm,that is,the multi-object tracking paradigm of joint detection and embedding,embeds the appearance feature extraction process into detector through network sharing,which better balances the tracking performance and efficiency.However,this paradigm has problems such as detection dependence and poor tracking robustness,and it is difficult to guarantee the real-time performance of online tracking.In addition,there are still challenges in how to reasonably utilize detection results in data association and reduce the impact of missing detection information on tracking performance.Aiming at the above problems,this thesis studies the multi-object tracking algorithm based on joint detection and data association,the main contents are as follows:(1)The one-stage multi-target tracking algorithm based on center point detection has problems such as heat map positioning deviation and small learning range of appearance feature map,which leads to the excessive dependence of appearance feature extraction on detection accuracy and affects the tracking effect.Aiming at this problem,a multi-object tracking algorithm with enhanced re-identification is proposed.The algorithm designed detection deviation loss for the problem of deviation in the predicted target center point.By suppressing the size of the response value at the non-true position in the predicted heatmap,the high response value is made to approach the true position,thereby reducing the influence of the detection bias on the appearance feature extraction.Aiming at the feature range,a dynamic expansion strategy of the learnable feature range is proposed.According to the target scale,the learnable feature range on re-identification feature map is adaptively expanded.Thus,the requirement of feature selection on detection accuracy is reduced.The experimental analysis shows that the tracking accuracy of the algorithm on the MOT16 and 17 datasets reaches 71.5%and 70.6%,and the target recognition accuracy reaches 71.7% and 71.0%.It is verified that the proposed algorithm can enhance the tracking stability without affecting the tracking speed.(2)Aiming at the problems of real-time tracking,easy mixing of appearance features into background noise,and missing detection information during data association,YOLOXs-MOT is proposed.Using the characteristics of its decoupling head,the appearance feature extraction module can be embedded with a very small computational cost.Aiming at the problem that embeddings are easily mixed with background noise,a secondary screening strategy is proposed.During training,the labeled bounding box is used to screen predicted positive samples again,which removes samples with large offset distances.Aiming at the problem of missing detection information,a tracking threshold separation strategy is proposed.Using different confidence thresholds in different matching processes can not only provide highquality appearance features for similarity matching,but also provide more candidate bounding boxes for Io U matching,making detection information allocation more reasonable.Through experimental analysis,the tracking accuracy of the algorithm on the MOT17 and 20 datasets reaches 70.5% and 65.1%,and the online tracking speed on the mobile computer terminal reaches 24.4FPS and 8.5FPS.This verifies the effectiveness of the proposed algorithm,which achieves higher tracking accuracy and faster tracking speed.(3)The multi-target tracking algorithm has shortcomings in dealing with the problem of missing target information,which easily leads to problems such as missed tracking and trajectory fragmentation.Low-quality features during data association will affect the stability of feature matching,thereby affecting the tracking effect.Aiming at the above problems,a multi-object tracking algorithm of fusing trajectory compensation is proposed.The algorithm performs information loss judgment on the trajectory of losing target during association,and uses the trajectory prediction information to expand the trajectory of the missing information to achieve compensation for the missing information.At the same time,a strategy of not updating low-confidence target features is proposed to ensure the stability of feature matching.Experiments show that the tracking accuracy of the algorithm on the MOT17 and 20 datasets reaches 70.7% and 65.0%,the target recognition accuracy is increased by 1.2% and 0.7%compared with YOLOXs-MOT,and the trajectory hit rate is increased by 2.5% and 5%.It can effectively improve the problem of missing heel and track fragmentation.The proposed algorithm is an information supplement to the association process,which has strong generality.Extensibility experiments on several different association algorithms and multi-target tracking algorithms show that the proposed algorithm can effectively improve the problem of missing information without affecting the real-time performance,and has strong scalability.This thesis addresses the problems existing in the multi-object tracking,from improving the detection dependence,tracking robustness and lack of information.Aiming at these problems,the corresponding multi-object tracking algorithms are proposed,and a large number of experiments are carried out to verify the proposed algorithms.The experimental results verify the effectiveness of the algorithms in this thesis,which can stably improve the multi-object tracking performance for different problems.