Coarse-to-Fine UAV Image Geo-Localization Using Multi-stage Lucas-Kanade Networks

Unmanned aerial vehicles (UAVs) have become an indispensable part of our intelligent society. Accurate UAV image geo-localization is fundamental in many applications such as search and rescue, precision agriculture, change detection. Previously, the UAV image geo-localization was usually calculated by photogrammetry method or by matching with the well-located reference images. The photogrammetry method relies heavily on the accuracy of the UAV positioning parameters, however, the construction and use of high-precision positioning equipment are expensive and difficult to be widely used. The method based on image matching requires good similarity between matching images, which is hard to be satisfied as the reference image and UAV image are heterologous. The existing methods are difficult to achieve accurate image geo-location on UAVs equipped with economical positioning devices. To address this issue, we propose a coarse-to-fine UAV image geo-localization pipeline. Firstly, we implement the coarse UAV image geo-localization by utilizing the UAV's position parameters. Then, a deep learning algorithm multi-stage Lucas-Kanade (MS-LK) is proposed to refine the UAV image geo-localization results iteratively. MS-LK can make full use of the UAV images' global texture. We validate the effectiveness of the proposed method on two real UAV data sets. The results show that the proposed UAV image geo-localization method can promote the results of the traditional photogrammetry-based approaches.