UAV Tracking with Proposals Based on Optical Flow

Min Jia, Zheng Gao, Zhisong Hao, Qing Guo

Abstract: UAV tracking is aimed to infer the location of the object from the videos captured by an aerial viewpoint. The challenges mainly focus on fast motion, scale variation and aspect ratio variation. The region proposal in image detection can detect the object candidates in the image, which can be leveraged to find the optimal location of the object. In this paper, a tracking algorithm using Farneback optical flow is proposed to provide object proposals for correlation filter for robust tracking under aerial scenarios. The Farneback flow estimates the motion of the object between adjacent frames and an improved FAST detector is adopted to detect the keypoints that contain the local patterns of the object from the last frame. The object proposal is obtained by computing translations of the keypoints. The final proposal is determined by computing the bounding box that encloses the keypoints. A correlation filter from KCF is used to detect the object on the proposal. The quantitative evaluation results on OTB100 show the advantage of the proposed tracker to state-of-the-art trackers in accuracy, especially under fast motion.