Inverse Sparse Object Tracking via Adaptive Representation

Jian-Xun Mi, Yun Gao, Renjie Li

Abstract: Sparse representation has been widely applied in object tracking, which is challenging due to various factors, such as occlusion, illumination variation, and background clutter. In this paper, we propose an inverse sparse tracking algorithm based on adaptive representation. Unlike the conventional tracking methods, we explore the low-rank property from the group structure of the candidate particles and calculate their similarity between foreground and background, which makes our representations more discriminative and robust. Specifically, we employed the candidate particles as the dictionary atoms to sparse encode the templates, which makes less cost on optimization. The adaptive regularization based on the rank of each group of candidate sub-matrix is employed in our method to adapt to appearance changes. Moreover, foreground templates and background templates are collected as positive and negative dictionaries, respectively, to make the tracker more discriminate from the complex environment. The experiment is conducted on a public dataset OTB100, and the quantitative results are reported, which demonstrate the favorable performance of the proposed method over some state-of-the-art trackers.