This paper analyzes and summarizes the issues in various Transformer-based multi-object tracking architectures,considering Transformer-based multi-object tracking as "query-based" methods. It focuses on the structure and functionality of these systems,with the "query" mechanism as their core. The different Transformer-based multi-object tracking systems are categorized into four main types:DETR-based trackers,memory-based trackers,temporal-spatial motion trackers,and other miscellaneous trackers. The performance of each algorithm is assessed using the MOTA metric on the MOT17 benchmark dataset. Based on these evaluations,the paper also suggests future trends in Transformer-based multi-object tracking.
维普
万方数据