Learning Spatial-Corrected Regularized Correlation Filters for Visual Tracking

Abstract

Discriminative Correlation Filters (DCF) have demonstrated excellent performance for visual tracking. DCFs utilizing a periodic assumption of the samples to train a correlation filter performs efficiently in tracking. However, the DCF framework is still an open-loop system prone to tracking failures caused by model drift. In this paper, we present a Spatial-Corrected Regularized Correlation Filters (SCRCF) which is a DCF-based tracker with a correction mechanism. SCRCF exploit the advantages of the spatial regularization method to design a corrective feedback mechanism. The mechanism calculates the reasonable offset to influence the spatial regularization coefficients by observing the output responses of multiple trackers, so as to correct the unreasonable tracking results. Compared to conventional DCF-based trackers, SCRCF is more robust to handle some complicated tracking scenes, such as occlusion and motion blur. Extensive experiments on OTB-2015 benchmarks demonstrate our tracker outperforms most state-of-the-art trackers.