Multi-person tracking by discriminative affinity model and hierarchical association

Occlusion is probably the most challenging part of multi-person tracking. In this paper, we exploit a new discriminative affinity model which combines deep convolution appearance features, random walk motion model, and object shape to robustly handle occlusion during a long period of time. Moreover, We propose a hierarchical association framework based on the observation that association increase the uncertainty when the object is occluded over time, and decompose a data association problem into several sub-problems. Our approach runs online from a single camera and does not require extra calibration information. Experiments on public Multiple Object Tracking 2016 benchmark datasets show distinct performance improvement compared to state-of-the-art online and batch multi-person tracking methods.