Research on target tracking method based on deep reinforcement learning

Abstract

Recently, the target tracking technology has become a popular scheme, which can assist skaters to achieve better results in training and competition. However, However, this is still a challenging task for smaller, fast-moving targets, such as skaters. To solve above problems, a target tracking method based on the combination of action-decision networks for visual tracking with deep reinforcement learning and the algorithm for precising bounding box estimation, and this method is used for skaters. The proposed method consists of two phases. In the first phases, the proposed method uses a tracker to obtain rough target tracking results, which search area is the position and size of the previous frame tracking result, and the tracker is controlled by sequentially pursuing actions. In the second phases, the alpha-refine, an accurate bounding box estimation algorithm, is applied to precise tracking of the target via the result of the first phases. That is, expand the prediction of the current tracking result to a concentric search area twice the size, and predict a more accurate bounding box in this area as the final tracking result. In addition, to train and test the proposed method, we produced a dataset for skaters. Compared with the traditional target tracking methods, the results of presented dataset show that the proposed method has higher tracking accuracy.