Robust visual tracker integrating adaptively foreground segmentation into multi-feature fusion framework

Abstract

Existing Discriminative Correlation Filter (DCF) based methods suffer from the limitations of rectangular shape assumptions. Aiming at this issue, in this paper, we propose an effective tracking approach which integrates a pixel-wise foreground segmentation mask into the correlation filter within a multi-feature fusion framework. Specifically, we first propose a novel segmentation algorithm which combines the color histogram with the spatial prior. On this basis, we implement a target-masked correlation filter (TMCF) tracker by introducing the foreground mask into a ridge regression, which successfully suppresses unexpected background information inside the bounding box. Secondly, we apply the alternating direction method of multipliers (ADMM) to solve our TMCF model efficiently to obtain the closed-form solution. Finally, a complementary fusion tracker by the combining of TMCF and color histogram scores (fTMCFCH) is formulated, which is robust to deformations and illumination changes simultaneously. The fusion factor is determined adaptively by the reliability derived from the target resolution of the trackers separately in each frame. We perform extensive experiments on three benchmarks: OTB-2013, OTB-2015 and Temple-Color-128. The concrete experimental results demonstrate that our tracker outperforms several state-of-the-art trackers.