Robust object tracking via sparse representation based on compressive collaborative Haar-like feature space
Ming Zhao, Hanming Qian, Rong Yingjiao, Chen Guo · 2016
Robust object tracking is a challenging problem as data streams change over time. In this paper we proposed a robust yet fast sparse tracking method using compressive collaborative Haar-like feature space of the model collection including positive (foreground), negative (background) and square templates. As traditional methods, the target is sparsely represented in the models with positive templates only and directly. But, as the complexity computation of sparse representation, we use the compressive Haar-like feature of positive and square templates to deal with occlusion. In order to improve the reliability during the tracking process, we develop a sparsity-based discriminative expression using foreground and background information. The expression with foreground reconstruction error and background reconstruction error is formulated to find the most correct sample. At last, a new update scheme related to the coefficients of templates of our appearance model is presented. Numerous challenging image sequences demonstrate that the proposed algorithm is able to adapt to scale, pose variation, rotation, occlusion and shows excellent real-time performance.