A Two-Stage Correspondence-Free Algorithm for Partially Overlapping Point Cloud Registration

Abstract

Point cloud registration is a key task in the fields of 3D reconstruction and automatic driving. In recent years, many learning-based registration methods have been proposed and have higher precision and robustness compared to traditional methods. Correspondence-based learning methods often require that the source point cloud and the target point cloud have homogeneous density, the aim of which is to extract reliable key points. However, the sparsity, low overlap rate and random distribution of real data make it more difficult to establish accurate and stable correspondences. Global feature-based methods do not rely on the selection of key points and are highly robust to noise. However, these methods are often easily disturbed by non-overlapping regions. To solve this problem, we propose a two-stage partially overlapping point cloud registration method. Specifically, we first utilize the structural information and feature information interaction of point clouds to predict the overlapping regions, which can weaken the impact of non-overlapping regions in global features. Then, we combine PointNet and the self-attention mechanism and connect features at different levels to efficiently capture global information. The experimental results show that the proposed method has higher accuracy and robustness than similar existing methods.