Confidence Based Hierarchical Online Ensemble Visual Tracking
Wen Hua He, Zhi Jing Liu, Jian Ming Qu · Advanced materials research · 2014
In order to adapt to the target appearance changes during visual tracking, feature model needs to be updated by online learning. However, online adaptive methods suffer from the drifting problem: error data are used for updating and will finally lead to tracking failure. In this paper, we propose a novel hierarchical online ensemble tracking method. Optical flow tracker is employed to predict the rough location. Online learning classifier is employed to detect the target. Template match is used to filter the data for updating. All these parts are combined together hierarchically by their confidence to ensure the stability of online learning and tracking. In contrast to the individual online learning and semi-supervised online learning, our method show more adaptability and stability. We demonstrate the performance on several public video sequences, which shows the improvement of our method over other online tracking approaches.