AdaBoost Tracker Embedded in Adaptive Particle Filtering

Due to computational simplicity, low-level visual cues (such as color, contour and corner) have been widely integrated into various visual trackers. However, the robustness of these trackers will be challenged by cluttered backgrounds, partial occlusion and varying illuminations. In many applications, the classes of the interested objects to be tracked are usually known in advance. Hence, high-level information from trained classifiers can be fused into the visual tracker to overcome the above limitations. In this paper, a novel approach is proposed to integrate all weak and strong classifiers of the trained AdaBoost cascade into the observation model of sampled particles in particle filtering. Furthermore, in order to track objects undergoing non-stationary movement, decisions of the whole boosted cascade on all sampled particles are incorporated to adapt the proposal distribution of sampling for better approximation of the desired posterior distribution. Experimental results on tracking different objects over indoor and outdoor video sequences have shown the proposed tracker is able to effectively handle rapid movements, partial occlusion, varying illuminations, and large changes of scale and viewpoint