A novel multiplex rotational attention-based network for point cloud registration and place recognition

SHI Cheng-hao; CHEN Xie-yuanli; GUO Rui-bin; XIAO Jun-hao; DAI Bin; LU Hui-min

doi:10.7641/CTA.2023.30286

Abstract：Point cloud registration and place recognition are critical tasks for localization in robotics and autonomous driving.There are few methods that can achieve efficient place recognition while providing accurate 6-degree-of-freedom pose.In this paper,we propose a novel multi-head network that simultaneously addresses both of these tasks.The network first extracts discriminative sparse points from the point cloud using a backbone network,and then solves the point cloud registration task in a dense point matching head and the place recognition task in a global descriptor head.In the backbone,we apply a novel 3D-RoFormer mechanism that explicitly encodes the relative pose information of points efficiently,result-ing in more discriminative and robust point features and significantly improving network performance.In the dense point matching head,the network establishes reliable correspondences between sparse points and progressively finds coarse-to-fine dense point correspondences to improve final pose estimation.In the global descriptor head,the network compresses the sparse point features into a global descriptor to describe the features of the current point cloud and achieves place recognition.We extensively evaluate our method on multiple datasets collected by different sensors in various environ-ments.Experimental results show that our method depicts strong generalization ability on all the datasets,outperforming or performing comparably to the state-of-the-art methods,among which the continuous point cloud registration error is reduced by about 27％,and the closed-loop point cloud registration error is reduced by about 37％.