Pixel-Based Color Model for Robust Tracking

In this paper, we propose a Pixel-based Color Model (PCM) that can deal with the problems of deformation and occlusion in visual object tracking. We use a pixel-based classifier and a feature mapping technique that can distinguish the target pixels from the background. The PCM method is generic as it can be incorporated into any tracker that fails to deal with the problems of deformation and occlusion. The base tracker in this paper is a correlation filter which has achieved great performance in recent years. Compared with the existing trackers, the Pixel-based Color Model combined with Correlation Filter (PCMCF) has several advantages: (1) By learning the color model of the target, it can deal with the problem of deformation. (2) As the model is pixel-based, it is less sensitive to occlusion. (3) The adaptive learning strategy can avoid model drift when the target is partially or fully occluded. (4) When the target recovers from fully occlusion, the color model will find it immediately. Both quantitative and qualitative evaluations have been conducted, and it shows that the proposed PCM method can improve the performance of the base tracker, and the PCMCF algorithm is superior to most state-of-the-art trackers especially in deformation and occlusion.