Cross-Source Place Recognition for Unmanned Aerial-Ground Vehicles With Low-Overlap and Varying-Density Point Cloud

Abstract

Cross-source place recognition is the foundation of collaborative mapping tasks between Unmanned Aerial Vehicles (UAVs) and Unmanned Ground Vehicles (UGVs) to achieve loop closure and global localization in outdoor environments. Due to the dynamic changes in communication bandwidth, the point density of point clouds transmitted from the robots to the server will also change simultaneously, which significantly affects the accuracy of place recognition. This article proposes a hierarchical BEV fusion network for unmanned aerial-ground vehicles with low-overlap and varying-density point clouds, called HBFusion. To improve the accuracy of place recognition for low-overlap point clouds, a multi-resolution voxel feature aggregation module based on sparse convolution is proposed to aggregate voxel features from multi-scale receptive fields to point features, which improves the similarity of global descriptors between low-overlap point clouds using rich point features. To adapt to point density variations caused by dynamic changes in communication bandwidth, we propose a hierarchical BEV fusion module to extract multi-layer BEV features at different heights, which improves the adaptability of global descriptors to point density variations using BEV representations. Extensive experiments performed on GrAco, a public aerial-ground dataset, and DUT-GA, our self-recorded aerial-ground dataset with low-overlap point clouds, indicate that our network achieves state-of-the-art performance in accuracy, even under the condition of varying point densities of point clouds.