Object Tracking via Multi-layer Convolutional Features with Adaptive Correlation Weighting

Object tracking is an important task in computer vision and artificial intelligence. Correlation filter trackers have been extensively studied due to their high efficiency, but most of them are limited by the hand-crafted features. Although some works have introduced deep learning to solve the problem, the corresponding methods tend to use the fully-connected feature and cannot fully utilize the rich spatial and semantic information in different levels of deep features. Therefore, a novel tracking method is proposed via multi-layer convolutional features with adaptive correlation weighting. In this study, the convolution feature in low level are extracted to describe the spatial structure information, while the high level is exploited to capture the semantic information of the target. In particular, this paper proposes an adaptive correlation weighting scheme based on peak-to-sidelobe ratio to measure the discriminative ability at each level, so as to realize the final tracking. Extensive experiments demonstrate that the proposed method outperforms several existing methods.