Strategies for Helping Anchor-Based Trackers Learn re-ID Features for Smart City Surveillance

Abstract

Re-identification has become a crucial issue in computer vision today as it allows for tracking objects in both continuous and discontinuous scenarios. Despite achieving perfect detection results, anchor-based trackers encountered difficulties in effectively learning re-identification features, due to various issues. This research proposes strategies aimed at improving the capability of anchor-based trackers to learn high-quality re-identification (re-ID) features. The model developed through our strategies can extract more distinct features and achieve almost 0.57 Multiple Object Tracking Accuracy (MOTA) on MOT20, even under a limited training dataset. This result indicates that our proposed strategies hold potential for improving the performance of anchor-based trackers.