Handling Inter-object Occlusion for Multi-object Tracking Based on Attraction Force Constraint

This paper presents a novel social interaction relation, attraction (interaction that would lead to occlusion for inter-object) for multi-object tracking to handle occlusion issue. We propose to build attraction by utilizing spatial-temporal information from 2D image plane, such as decomposed distance between objects. Then pairwise attraction force is obtained by the modeled attraction. Lastly, the attraction force is used to improve tracking when hierarchical data association performs. To meet requirements of practical application, we have our method evaluated on widely used PETS 2009 datasets. Experimental results show that our method achieves results on par with, or better than state-of-the-art methods.