Paper_id810.pdf

ROBUST OBJECT TRACKING VIA INCREMENTAL SUBSPACE DYNAMIC SPARSE MODEL Zhangjian Ji, Weiqiang Wang, Ning Xu University of Chinese Academy of Sciences, School of Computer and Control Engineering, Beijing, China Email: jizhangjian08@ wqwang@ xuning12@ ABSTRACT Sparse representation has been widely applied to some gener- ative tracking methods. However, few methods consider the inter-frame correlation between sparse representations of tar- gets. Based on the inter-frame correlation, this paper presents a novel incremental subspace dynamic sparse tracking (IS- DST) model with the error term of Gaussian-Laplacian dis- tribution. The proposed method can effectively deal with the occlusion in scenes by integrating two aspects of techniques. First, we adopt the l1,2 mixed norm to characterize the spatial structured sparsity, i.e., the the group effect of the outlier en- tries caused by occlusion. Second, we maintain an occlusion mask and eliminate the influence of occlusion pixels in the process of calculating the likelihood probability. The exten- sive experiments are carried out to compare our model with 5 state-of-the-art methods on 12 challenging sequences. The experimental results demonstrate that the proposed method consistently outperforms the existing state-of-the-art method- s. Index Terms— incremental subspace learning, dynamic sparse model, spatial structured sparsity, occlusion detection 1. INTRODUCTION Object tracking is one of the most important research prob- lems in computer vision, and is tightly related with numer- ous practical applications, e.g., intelligent surveillance, au- tonomous navigation of vehicles, human computer interac- tion, action recognition. Although the related research of object tracking has made significant progress during the last decades, it still remains a rather challenging problem. In order to design robust visual tracking algorithms, a lot of complex factors (e.g., partial occlusion, illumination variation, pose changes, background