Set to Set Visual Tracking

Wencheng Zhu, Pengfei Zhu, Qinghua Hu, Changqing Zhang

Abstract: Sparse representation has been widely used in visual tracking and achieves superior tracking results. However, most sparse representation models represent the target candidate as a linear combination of target templates and need to solve a sparse optimization problem. In this paper, we propose a novel set to set visual tracking (SSVT) method. Under the particle filter framework, we consider both the target candidates and target templates as image sets, and model them as convex hulls. Then the distance between two image sets is minimized and the tracking result is the target candidate with the maximum coefficient. As the target candidates are modeled as one convex hull, SSVT utilizes the underlying relationship of the target candidates. Moreover, SSVT is very efficient in that it only needs to solve one quadratic optimization problem rather than sparse optimization problems. Both qualitative and quantitative analyses on several challenging image sequences show that the proposed SSVT algorithm outperforms the state-of-the-art trackers.