Online Pedestrian Tracking Using Ensemble Color Feature

In this paper, we propose a novel online visual tracking algorithm using ensemble color feature. Visual tracking has long been one of the most important topics in computer vision. In recent years, significant progress has been achieved by the state-of-the-art tracking algorithms. However, some challenging problems still remain unsolved. One general problem is that it is difficult for a tracking algorithm to achieve both high accuracy and efficiency. More specifically, most tracking algorithms base on the hypothesis that the tracking object is a rigid body. Thus, some visual features are disabled when the hypothesis is invalid. Meanwhile, the existing trackers tend to fail when the target object changes poses, because such changes lead to the dramatic mismatch between the momentaneous appearance model and the learned one. The last problem is especially serious when the tracking targets are pedestrians, which suffer irregular pose and shape changes while they walk or run. The proposed algorithm applies ensemble color feature model. It takes pedestrians' physiological structure into consideration, and takes advantage of color histogram information under several color spaces, including the RGB, normRGB, HSV, and Lab. Since histogram is a statistic-based feature, and the garments of pedestrians provide abundant color information, our appearance model using ensemble color feature is more robust to pose and shape changes. Experiments on Caltech Pedestrian Database with evaluation of the state-of-the-art algorithms show better performance of the proposed algorithm.