Robust object tracking using semi-supervised appearance dictionary learning

It is a challenging task to develop robust object tracking methods to overcome dynamic object appearance and background changes. Online learning-based methods have been widely applied to cope with the challenges. However, online methods suffer from the problem of drifting. Sparse appearance representation has recently shown promising object tracking results. However, it lacks of information update to accurately track objects in long sequences or when object appearance drastically changes. In this paper, we propose a novel framework for tracking objects using a semi-supervised appearance dictionary learning method. Firstly, an object appearance dictionary is learned on the initial frame. Secondly, a graph model is employed in the proposed method for learning new bases when detecting object appearance change. The selected bases automatically replace the current rarely used bases. The proposed method is quantitatively compared with state-of-the-art methods on several challenging data sets. Results have shown that our proposed framework outperforms other methods even when drastic appearance variations happen.