The generation of high-quality orthoimages is essential for precision agriculture using UAVs. For quick surveys with little image overlap, georeferenced aerial image-mosaicing methods are effective, since they have shorter processing time than 3D reconstruction-based methods. However, it is difficult in principle to apply them in areas where the correspondence of feature points cannot be obtained in existing stitching methods. We propose a high-quality image stitching method with pose optimization using image similarity and feature point correspondence. The main idea is to correct misalignment using image similarity in overlapped regions unmatched by feature detection. Experiments using a drone with a gimbal-less camera on real fields demonstrate that the proposed method enables image stitching in unmatched areas that have previously been difficult to correct for misalignment and provides a feasible and promising solution for the generation of high-quality orthoimages for quick surveys.
