Oversaturated part-based visual tracking via spatio-temporal context learning.

Partial occlusion is one of the key challenging factors in a robust visual tracking method. To solve this issue, part-based trackers are widely explored; most of them are computationally expensive and therefore infeasible for real-time applications. Context information around the target has been used in tracking, which was recently renewed by a spatio-temporal context (STC) tracker. The fast Fourier transform adopted in STC equips it with high efficiency. However, the global context used in STC alleviates the performance when dealing with occlusion. In this paper, we propose an oversaturated part-based tracker based on spatio-temporal context learning, which tracks objects based on selected parts with spatio-temporal context learning. Furthermore, a structural layout constraint and a novel model update strategy are utilized to enhance the tracker's anti-occlusion ability and to deal with other appearance changes effectively. Extensive experimental results demonstrate our tracker's superior robustness against the original STC and other state-of-art methods.