Robust Visual Tracking Via Incremental Low-Rank Learning

Visual tracking is one of the fundamental tasks in computer vision society with many ap-plications such as surveillance, robotics, human computer interaction, and medical imaging, etc. Many efforts have been devoted to this topic, but visual tracking remains challenging due to significant change of target’s appearance caused by heavy occlusion, drastic pose, scale and illumination variation, background clutter, and varying viewpoints, etc.Subspace representation is easy to compute and robust for in-plane rotation, scale and pose changes and illumination variation. So subspace representation is possibly the most common choice for appearance models in visual tracking. By modeling the low-dimensional subspace which target appearance lies in, the appearance model can online adapt to the change of target. However, there is no global constraint on the target object in the video sequences. Therefore, these methods still can not exactly capture the global subspace structure for the appearance of the target. Recently, some new robust subspace learning methods which is based on the properties of low-rankness are proposed, such as RPCA and LRR. They can successfully recover the intrinsic low-rank subspace structure from corrupted observations. However, these low-rank models are difficult to be directly utilized for incremental visual processing. Inspired by above discussions, this paper proposed two visual tracking algorithms via incremental low-rank learning.First, this paper proposed a novel robust visual tracking method via incremental low-rank features learning. We show that the proposed strategy is actually an online extension of Robust PCA(RPCA). Compared with previous methods, which directly learn subspace from corrupted observations, our model can incrementally pursuit the low-rank features for the target and detect the occlusions by the sparse errors. Then another novel subspace learning method for robust visual tracking is proposed. Under particle filter tracking framework, an online scheme is devel-oped to incrementally pursue the optimum projection, and the columns of the projection defined in the latent feature space are a set of redundant basis, treating an observation as its coefficient. As a result, the low-rank property of the pursued optimum projection can exactly reveal the in-trinsic low-dimensional structure of the global feature space, contributing to the high precision of capturing appearance changes. Experimental results on various challenging videos validate the superiority over other state-of-the-art methods.