Flow-Guided Single Object Tracking Framework in UAV Aerial Video

摘要

Tracking target in UAV aerial video is challenging for at least two reasons: (1) the target have small scale relative to the image which leads to have less feature information; and (2) compared to the traditional tracking, drones bring new challenges to the tracking methods, even the combination of several difficult scenarios. Notice that the most existing trackers only consider appearance feature of the current frame, and hardly benefit from information between time and motion sequence. In this paper, we address both issues inspired by time sequence and motion information. In particular, we proposed a network that unifies Siamese subnetwork and optical flow estimation in a framework. Given an aerial video, a CNN subnetwork is used for feature extraction and a flow subnetwork is used for flow estimation. Then according to target historical motion speed, EA module is used to reduce the redundant motion information by scale estimation and FA module is used for feature fusion. Finally each aggregated feature and template feature is fed into region proposal subnetwork. Extensive experiments are performed on popular aerial video datasets, and our proposed method achieves promising performance in comparison with state-of-the-art trackers.