Fast visual object tracking via distortion-suppressed correlation filtering

Abstract

Visual object tracking is a basic research unit in the construction of smart cities, it focuses on establishing a dynamic appearance model to represent the target in complex scenarios. In this paper, a distortion-suppressed correlation filtering based tracking method (DSCFT) is proposed. Our approach tackles distortions caused by spatial similarity comparison and temporal appearance updating. We establish our method under a Bayesian framework, where spatial and temporal appearance are embedded in likelihood and prior respectively. Firstly, The spatial distortion is handled by modifying weight windows and utilizing a proposal selection strategy to better track targets under fast motion and background clutters. Secondly, temporal information is retained in updating stage as a prior to represent dynamic variations of the target. Moreover, a multi-scale filtering scheme is integrated when updating the temporal appearance to boost the scale sensitivity. Experimental results dedicate the effectiveness and robustness of our DSCFT on benchmark videos.