Relative Pose Estimator based on Semi-Dense Homography for UAV Landing on Moving Target using Monocular Camera

Abstract

Autonomous landing has been studied on various types of landing targets. In particular, relative pose estimation between a moving landing target and a UAV is important for stable landing in various environments. Most of the studies on UAV landing for a moving landing target used templates given in advance and feature-based methods to estimate the relative pose between the UAV and the landing target. However, the feature-based methods are vulnerable to ambiguity due to the symmetric shape of the helipad or camera noise. In this paper, we propose the algorithm that accurately estimates pose by applying the direct method. The algorithm does not need to use a given template image, which enables landing in various environments. Since the algorithm does not use the feature-based method, there is no concern about the ambiguity. To evaluate the proposed algorithm, we collect our dataset and confirm that the proposed algorithm is more accurate than the feature-based method.