Segment-based Cross-domain Localization between Aerial and Ground Robots

Abstract

Cross-domain localization is the foundation and key link for collaborations of unmanned aerial vehicles (UAVs) and unmanned ground vehicles (UGVs). A major challenge of this problem is the extremely large difference in view-points of aerial and ground robots. In this work, we present a segment-based cross-domain localization solution that accurately calculates the pose of a UGV in the reference map generated by a UAV. Especially, in scenarios with few small-size structures, our method exploits the common information of large structures in aerial and ground view-points to match, which leads to a satisfactory recall of place recognition. Additionally, we build a large scale outdoor simulation urban environment at high fidelity and present three accurate synthetic benchmark datasets with easy, medium and difficult levels, to quantitatively evaluate the performance of our solution. Experimental results demonstrate that our method achieves notable better performance than the baseline method SegMap. Our implementation is publicly available at https://www.github.com/5john/AirGroundSegMap.