Research On Multi-Object Tracking Method Based On Anchor-free

Multi-object tracking algorithm as an important research direction in the field of computer vision has important application prospects in the fields of video intelligent surveillance,intelligent robots,unmanned vehicles and other fields.The one-stage multi-object tracking framework of joint detection and tracking has attracted much attention because it can effectively improve the tracking efficiency of the algorithm,but it is limited by the lack of mutual information interaction between the detection and association processes and the inconsistent focus of different tasks,so it is difficult to optimize the effect of both detection and association tasks at the same time.In the face of scenarios such as background interference and dense pedestrian traffic,the problem of track loss due to object occlusion and other conditions is frequent,and the dependence of the tracking process on the detection results can significantly affect the tracking effect.In view of the above problems existing in the joint detection and tracking method,this paper is based on the anchor-free detection method,and the two aspects of multiple feature information fusion and trajectory association correction in the multi-object tracking algorithm are studied in depth,and the main research contents are as follows:The multi-object tracking method with multiple feature information fusion is proposed for the problems that the one-stage multi-object tracking algorithm based on anchor-free centroid detection is prone to heat map localization errors,the difficulty of balancing different tasks in the one-stage method,and the lack of information interaction between tasks.The center point attention features are enhanced by setting the center point deviation loss,and the design neutralize and match associated module use the appearance embedding features to complete the similarity matching to weaken the inter-task conflict problem and further obtain the object motion offset information,and use the deformation adaptive alignment method to fuse the center point attention features and each level of features containing the appearance and motion position information.The method integrates multi-level features to enhance the robustness of tracking features.The proposed method achieves tracking accuracy of 67.4%、66.1% and 53.2% on MOT16,MOT17 and MOT20,and IDF1 of 64.9%,67.5% and 50.1%,with significantly better performance of key metrics than other one-stage tracking methods.The multi-object tracking algorithm is prone to interruption of trajectory due to object occlusion,which leads to the problem of increased object miss-following and trajectory fragmentation.The data association method of tracking relies on detection results,and the objects with poor detection quality due to problems such as occlusion are easily ignored by the matching process.To solve these problems,a multi-object tracking method with trajectory correlation correction is proposed.By adjusting the tracking threshold,the low-quality detection objects were stored in the matching buffer,and the motion position prediction information was used to match the objects and track fragments in the buffer to maintain track integrity.Set the missing trajectory expansion threshold to judge the debris trajectory state,and use the prediction result to expand the missing trajectory to achieve trajectory correction.The proposed method achieves 69.1% and57.9% MOTA and 69.9% and 53.1% IDF1 on MOT17 and MOT20 datasets,which effectively improves the problems of missing tracking and trajectory fragmentation.The method is extended to other algorithms and also optimizes the tracking performance effectively.Compared with the advanced multi-object tracking algorithm,The proposed method achieves better tracking performance.