Fast Correlation Tracking Using Low-Dimensional Scale Filter and Local Search Strategy

Abstract

Recently, correlation filter-based tracking algorithms have attracted much attention for its high efficiency and robustness. However, achieving fast and accurate scale estimation remains a challenging problem. Most existing scale estimation approaches are inefficient and time-consuming. Besides, these existing trackers perform poorly when the object is under fast motion and partial occlusion due to limited searching area. In this paper, an independent scale filter is proposed to estimate the scale of an object, and the dimensionality reduction strategy is used to reduce computational cost. In addition, a local search strategy is proposed to expand the searching area of the tracker, which can effectively solve the problem caused by fast motion and occlusion. Extensive experiments have been conducted on three large-scale benchmarks, and it is shown that our proposed tracker outperforms the most state-of-the-art trackers. It achieves an average precision score of 87.2%, and an average success score of 65.2% on the object tracking benchmark-2013 benchmark. Moreover, our proposed tracker can run at a speed of nearly 80 frames/s on a single CPU, exceeding most competitive trackers by several times.