Visual tracking with representative templates based on low-rank matrix
Deqian Fu, Shunbo Hu, Seong Tae Jhang · 2013
Robust visual tracking, as a critical problem in community of computer vision, is still knotty, especially in challenging scenarios. In this paper, using the nature of low-rank matrix recovery, we propose a tracker with structured appearance model consisting of multiple representative models. By exploring the signal recovery power of Low-Rank matrix, we get effective representation of target and background for tracking; at the same time maintain a robust appearance model with multiple representative templates. Benefitting from low-rank recovery power, the representation matrix of candidates w.r.t the low-rank dictionary shows low-rank and sparse. Meanwhile, by our update strategy, a novel dictionary is maintained with low-rank models derived from multiple representative templates, which further encourages the sparse representation of particles. The proposed algorithm is demonstrated by extensive experiments on several challenging databases.