Regalignsort: Multi-Object Tracking Based on Location Regression and Local Feature Association

Recently, multi-object tracking (MOT) has attracted more attention, and it has made significant progress with the development of various object detection methods as well as various REID models. However, motion models in most recent tracking algorithms usually assume that the object does linear motion in a tiny time window, so these methods are sensitive to nonlinear motion and irregular random motion, resulting in tracking accuracy being easily disturbed by occlusion. In addition, not powerful enough global appearance feature makes it difficult to recover the trajectory from the lost state, which leads to a large number of ID switches in the tracking process. In view of the above two problems, we propose a new multi-object tracking (MOT) algorithm named RegAlignSORT based on the DeepSORT framework. Aiming at the occlusion or irregular movement that linear motion models are susceptible to, we use the regression head of Faster-RCNN to realize more stable motion estimation, avoiding the tracking uncertainty of the linear motion model. Aiming at the problem of limited matching accuracy of global appearance features, we use our own previously proposed LocalOptReid network to extract local features, and the local optimal alignment strategy is used to calculate the alignment appearance loss between trajectories and detections to achieve more accurate data association. The experimental results show that our proposed RegAlignSORT achieve state-of-the-art HOTA and IDF1 on MOT17 and MOT20 datasets.