Fast and Robust Registration of Partially Overlapping Point Clouds

Real-time registration of partially overlapping point clouds has emerging applications in cooperative perception for autonomous vehicles and multi-agent SLAM. The relative translation between these is higher than traditional SLAM odometry applications, which challenges the identification correspondences a successful registration. In this paper, we propose novel method where are learned using an efficient point-wise feature encoder, refined graph-based attention network. This network exploits geometrical relationships key points to improve matching with low overlap. At inference time, pose transformation obtained by robustly fitting through sample consensus. evaluation performed on KITTI dataset synthetic including low-overlapping displacements up 30 m. proposed achieves on-par performance state-of-the-art methods dataset, outperforms existing clouds. Additionally, significantly faster times, as 410 ms, 5 35 times competing methods. Our code available at https://github.com/eduardohenriquearnold/fastreg .