Spatio-Temporal Correlation Graph for Association Enhancement in Multi-object Tracking

摘要

Due to the frequent interaction between targets in real-world scenarios, various data association problems, such as association ambiguities and association failure, are caused by potential correlation between interactive tracklets, especially during crowded and cluttered scenes. To overcome the non-intuitionistic of tracklet interaction, spatio-temporal correlation graph (STCG) is proposed to model the potential correlation between pairwise tracklets. Three primitive interactions (aggregation, abruption, stability) are defined to model the completed period of the tracklet interaction. Furthermore, STCG model is applied into network flow tracking to exploit the potential correlation between tracklets and enhance the association of the interactive tracklets, especially when overlapping or occlusion is happened. Our method is effective on MOT challenge benchmarks and achieves considerable competitive results with current state-of-the-art trackers.