Cascaded-Scoring Tracklet Matching for Multi-object Tracking

Multi-object tracking (MOT) aims at locating the object of interest in a successive video sequence and associating the same moving object frame by frame. Most existing approaches to MOT lack the integration of both motion and appearance information, which limits the effectiveness of tracklet association. The conventional approaches for tracklet association often struggle when dealing with scenarios involving multiple objects with indistinguishable appearances and irregular motions, leading to suboptimal performance. In this paper, we introduce an appearance-assisted feature warper (AFW) module and a motion-guided based target aware (MTA) module to efficiently utilize the appearance and motion information. Additionally, we introduce a cascaded-scoring tracklet matching (CSTM) strategy that seamlessly integrates the two modules, combining appearance features with motion information. Our proposed online MOT tracker is called CSTMTrack. Through extensive quantitative and qualitative results, we demonstrate that our tracker achieves efficient and favorable performance compared to several other state-of-the-art trackers on the MOTChallenge benchmark.