Scale Adaptive Dense Structural Learning for Visual Object Tracking

Abstract

Object tracking has long been a hot topic in computer vision. However, existing trackers are still too far away from solving the visual tracking problem because of their limited robustness, inadequate precision and low efficiency. The correlation filters are able to acquire high speed as well as moderate tracking performance. However, they fail to deal with fast motion. Recent advantages on dense structural learning based tracking algorithms suggest new solutions to the tracking problem. In this paper, we combine an online structural learner with correlation filters for robust and accurate tracking. The proposed method is composed of two components. First, we search the target translation parameter in a large range by a structural classifier. Second, we estimate the target scale with a discriminatively trained correlation filter. The proposed tracker is then exhaustively experimented on a latest UAV (Unmanned Aerial Vehicle) dataset with 123 low framerate and highly challenging videos. We show superior tracking performance against 13 other trackers on the dataset.