Hierarchical Deep Feature for Visual Tracking via Discriminative Correlation Filter
Jian Wei, Yue Wang, Feng Liu, Ziguan Cui · 2019
Visual tracking is an important research topic in the field of computer vision. It plays a vital role in camera focus tracking, drone target recognition, pedestrian tracking, vehicle tracking, etc. Due to various challenging attributes in the real-world scenario, the accuracy and robustness of most existing tracking algorithms cannot be considered at the same time. The main reason is to use the hand-crafted features to present the target appearance model. Hand-crafted features are typical underlying information used to present the target. The high resolution of the underlying feature information is conducive to accurate target location, but lacks high-level semantic information and cannot handle serious appearance changes, that is, poor robustness. In this work, we use the VGGNet19 model to extract the underlying and high-level features of the target. In order to predict the state of the target while considering the accuracy and robustness, the underlying and high-level feature response maps obtained by discriminative correlation filter (DCF), which can estimate the mapped positions of different layers of the target, and then adaptively fuse the mapped positions of different layers, and finally estimate the new state of the target. Extensive experiments show that the proposed algorithm outperforms the state-of-the-art tracking algorithms in terms of quantitative and qualitative evaluation.