Online human tracking via superpixel-based collaborative appearance model

In this paper, we propose a novel collaborative appearance model for robust human tracking by exploiting both object and motion information in the bayesian framework. In contrast to most existing methods which use low or high-level visual cues, we use mid-level visual cues via superpixel with sufficient structure information to represent the object. In our work, the collaborative appearance is modeled by the static confidence map and the motion map. We present a spatial clustering method (SCM) to group superpixels with local features, and evaluate the static confidence map by the clustering results. The motion map is computed by predicting the direction and velocity of moving object. We can handle the appearance change adaptively to alleviate the drift problem, and reduce the influence of the occlusion by the occlusion detection. We employ both quantitative and qualitative evaluations on various challenging video sequences, and demonstrate that the proposed tracking method performs favorably against several state-of-the-art methods.