Learning salient features to prevent model drift for correlation tracking

Abstract Correlation Filter (CF) based algorithms play an important role in the field of Visual Object Tracking (VOT) due to their high accuracy and low computational complexity. While existing CF tracking algorithms suffer performance degradation due to inaccurate object modeling. In this paper, we improve the object modeling accuracy in both CF training stage and target detection procedure to preventing the drift problem. Specifically, we propose a multi-model structure for CF trackers to capture the target appearance changes, where different appearance models are trained with specific samples to catch the salient features of the target and reduce the computational cost. Furthermore, a space filter for detection features is designed to suppress the boundary effect under Gaussian motion prior, which contributes to improving the accuracy of position estimation. We deploy our method to three hand-crafted features based CF trackers to perform real-time visual tracking on popular benchmarks. The experimental results demonstrate the efficacy of our proposed scheme and the efficiency of our trackers. In addition, we provide a comprehensive analysis of the proposed method to facilitate application.