Learning spatial-channel regularization jointly with correlation filter for visual tracking

The boundary effect of correlation filters is one key issue to limit the performance of visual tracking. Most existing methods focus on using regularization to constrain filters in the spatial domain, but less attention is paid to the channel information that is also important to enhance the discriminative ability of the filter. In this paper, we propose to learn the spatial-channel regularization jointly with the filter for visual tracking. Specifically, the channel regularization is integrated into the spatial regularization, which is exploited to make the filter more compact in spatial and channel domains. It benefits to improve the discriminative ability to identify the target from the background, even distractors. Additionally, the temporal coherence of the target is developed to enable the filter to be robust. The temporal regularization can suppress the sudden change of the spatial-channel regularization on the time axis, which can indirectly control the filter to pay attention to the target's robust features. The regularization and filter can be jointly optimized by the alternating direction method of multipliers (ADMM) algorithm. We evaluate our method on a standard database. The results show that compared with traditional methods, our tracker has improved significantly in performance and can achieve real-time tracking.