Multi-Object Tracking with Cross-Spatial Feature Association Based on Transformer
Multi target tracking algorithms often suffer from inaccurate target recognition and poor tracking performance in complex scenes.In environments with a large number of tracked targets and severe mutual occlusion,the phenomenon of target loss is more pronounced.To this end,a multi-target tracking algorithm based on Transformer cross spatial feature association is proposed,which utilizes the advantages of Transformer structure to extract global features and multi head attention mechanism to improve the extraction ability oftarget features.In addi-tion,to solve the problem of tracking target loss caused by mutual occlusion between tracking targets,a mutual attention mechanism is used to map the features between tracking targets and interfering targets for enhancement and suppression,in order to improve the accuracy and reli-ability of the tracking algorithm;At the same time,the degree of feature enhancement and suppression is determined based on the feature simi-larity between the tracking target and the interfering target.Experiments were conducted on the MOT16 and MOT17 datasets,and the pro-posed algorithm achieved multi-target tracking accuracy of 58.81％ and 60.05％,respectively,with better performance compared to other mainstream algorithms.
万方数据
维普
