Multi-Feature Tracking Via Adaptive Weights

Object tracking is one of the most important components in numerous applications of com-puter vision. It is generally applied to intelligent video surveillance, human-computer interaction and medical image sequence analysis, to name a few. It is a challenging problem to develop a robust object tracking algorithm. With some factors such as partly occlusion, pose change, il-lumination variation, background clutter and appearance change, to develop a fast and robust object tracking algorithm is the key issue.In this work, we present a novel online object tracking algorithm by using multi-feature channels with adaptive weights. Firstly, we exploit intensity, histogram of gradient (HOG) and color naming features to generate a set of confidence maps where the confidence value of each pixel indicates the probability that this pixel belongs to the tracked object. For the robustness of these individual confidence maps, we propose a method to update these confidence maps via a multi-layer cellular automata based system.The intensity feature covers the energy information and HOG feature depicts the contour information of the tracked object and its surrounding background respectively. Color naming features aforementioned not only provide high-level features to build a more stable appearance model, but also handle tracking with cluttered coloring background effectively. Secondly, we learn an online model that denotes the close relationship between center of the target and back-ground context, which represents some statistical correlation in consecutive frames. Finally, we exploit the appearance model, online model and a cellular automata based updating mechanism to generate a confidence map for each feature channel, and then obtain a final confidence map by fusing confidence maps from different channels in an adaptive manner. The optimal location of the tracked object can be determined based on the maximum value in the final fused confidence map.Both qualitative and quantitative evaluations on the recent benchmark dataset demonstrate that the proposed tracking algorithm achieves more favorable performance than several state-of-the-art methods, especially for the color sequences.