Object Tracking by Branched Correlation Filters and Particle Filter

Correlation filter based tracking which is one of the tracking-by-detection methods, has demonstrated competitive performance in recent years. However, once the detected position is incorrect, the parameter of correlation filter is updated based on the appearance of the incorrect position. This causes the target loss, which is a crucial problem in a single object tracking task. In this paper, we propose a novel object tracking method that prevents target loss. In the proposed method, the object is detected using branched correlation filters, and each correlation filter corresponds to a candidate detection position. The optimal correlation filter is estimated by particle filter. The experiments confirmed that the AUC score of proposed method is 2.24 higher than that of the state-of-the-art method.