Symmetric Joint Detection and Re-Identification for UAV-Based Multiple Object Tracking

With the development of multi-tasking learning, multi-object tracking (MOT) has been dominated by the model of Joint Detection and Re-Identification (JDR). FairMOT is a milestone work under such a paradigm. It contends that the anchor-free methods are more suitable than the anchor-based ones for building JDR models. However, the anchor-based detectors are widely used in unmanned aerial vehicle (UAV) applications since they tend to achieve higher performance for detecting small targets which are universal in drone videos. In this paper, we propose an anchor-based JDR model for the MOT task related to UAV vision. Firstly, to balance the detection task and Re-Identification (Re-ID) task, a symmetric two-branch architecture is proposed. Secondly, to deal with the correspondence problem between the two tasks, we learn Re-ID features for every anchor box. Thirdly, to alleviate the conflicts between the detection problem and the Re-ID problem, we train the model in two steps, where the detection branch is trained first, then is the Re- Id branch. Our model based on the popular anchor-based detector YOLOv7 has achieved great performance on the Visdrone- MOT20 19 dataset. The experimental results demonstrate that our model can validly complete the MOT task related to UAV visions.