A Background-Aware Tracker with Multi-Features Correlation Filters

Abstract

Discriminative Correlation Filter (DCF) based trackers have shown great potential in visual tracking because of their excellent accuracy and high speed. However, traditional DCF framework applies cyclic samples and ignores the importance of background samples, which can lead to serious boundary effect and unsatisfactory tracking results against complex tracking challenges such as background clutter, deformation, occlusion, etc. Besides, the blind update strategy sometimes results in model drift. In order to handle such challenges, a background-aware tracker constructed via multiple features is proposed. The correlation filter based on color histogram feature is designed to be trained with abundant background samples to improve target appearance precision while reducing serious boundary effect. Based on an odd merge function, the correlation filter is also combined with a Histogram of Oriented Gradient (HOG) correlation filter to get precise tracking results. Finally, the two correlation filters are updated adaptively by the tracking result of each frame. Experiments show that the proposed tracker can act ~50FPS in OTB2015 benchmark and achieves good tracking accuracy.