Learning correlation filters in independent feature channels for robust visual tracking

摘要

Although existing correlation filter-based tracking algorithms have shown competitive performance, most of them suffer from defects of insufficient learning, inauthentic combination and inflexible update when it comes to multiple feature channels. To tackle these problems, a specific correlation filter is learned for each feature channel and then the final response map is generated according to the confidence of channels. The confidence is decided by both spatial and temporal distribution of response maps. Under such circumstances, correlation filters are allowed to be updated independently with their own learning rates which can vary with the diversification of feature space in corresponding channels. We evaluate our work on OTB-2013, OTB-2015 and VOT-2017. Our approach outperforms the baseline fDSST by 6.1% in Mean Overlap Precision (OP) especially on OTB-2015 and shows competitive performance compared to state-of-the-art trackers with hand-crafted features while running at 70 FPS. (C) 2018 Elsevier B.V. All rights reserved.