End-to-end Visual Object Tracking with Motion Saliency Guidance
Yucheng Zhang, Kexin Liu, Tian Wang · 2020
In recent years, deep learning-based object tracking methods have achieved excellent performance. Most of existing object tracking algorithms only focus on target appearance features and ignore motion features. However, most of tracked targets are moving, motion features are important in object tracking tasks. In this work, we are committed to extracting deep features with motion saliency to highlight targets from the background and improve tracking accuracy. To this end, a new object tracking network combining correlation filters, siamese network and optical flow network is proposed. In the proposed network, we apply optical flow to extract motion information of targets, and employ an attention network to integrate motion features and appearance features. The evaluations on OTB2013, OTB2015 and UAV123 demonstrate that the proposed method can track targets accurately while ensuring speed.