Learing spatial associate temporal correlation filter for visual tracking

Abstract

Discriminative Correlation Filters (DCF) have proved efficient and useful in the tracking community. The success of DCF trackers is attributed to their ability to efficiently generate large numbers of negative examples by shifting a single sample. However, when training the filter, large amounts of impractical samples are also acquired, which results in boundary effects. Traditional methods trained the filter with larger sets to suppress boundary effects, but sacrificed computational efficiency, which affected the real-time performance of the tracker. Recently, a Background-Aware Correlation Filter (BACF) was suggested to mitigate the boundary effect. It proved robust and efficient, but the BACF tracker suffers from drifting when the target rotates in the scene. This occurs because BACF ignores the temporal connection between the frames in the video. In this paper, we propose a Spatial Associate Temporal Correlation Filter (SATCF) by formulating spatial and temporal regularization terms in the BACF model. We trained the filter to recognize patches that belong to the target as positive, foreground samples and other patches from the background as the negative examples. This setup, promotes the discriminative ability of the filter. In addition, we also exploited a spatial term to measure the spatial difference between two consecutive frames - a technique that is capable of alleviating model over-fitting. Our tracker enables computational efficient and achieves superior performance in complex video sequences. We demonstrate the efficiency of the SATCF tracker by comparing with state-of-the-art trackers. Experimental results on the OTB-2013 dataset demonstrate the favorable performance of our tracker against other state-of-the-art methods.