关键词: 低重叠场景, 点云配准, 位置编码, 重叠掩模, 自注意力, 交叉注意力

Abstract: Abstract: To address the issues of insufficient keypoint feature representation and limited overlapping regions in low-overlap point cloud registration scenarios, which often result in high mismatch rates and low registration accuracy, a low-overlap point cloud registration network integrating positional encoding and overlap mask was proposed to reduce mismatches and enhance registration precision. First, point cloud keypoints were extracted using a PointNet++ point-wise feature encoder, and their feature information, coordinates, and positional encoding were fused to generate more discriminative keypoint descriptors. Then, the fused features were fed into self-attention and cross-attention modules to enhance the descriptive ability of features and strengthen contextual information interaction, thereby alleviating the lack of keypoint description. Next, an overlap mask module was introduced after the attention modules. This module was designed to identify overlapping regions and filter out non-overlapping keypoints, reducing the mismatch rate. Finally, the Sinkhorn algorithm was applied for optimal matching, and the Iterative Closest Point (ICP) algorithm was used for refinement to further improve registration accuracy. Comparative experiments were conducted on the CODD and KITTI datasets against several existing low-overlap point cloud registration methods. The results show that the proposed network achieves superior performance. In particular, on the CODD dataset, the Relative Translation Error (RTE) and Relative Rotation Error (RRE) were reduced by 53.29% and 42.72%, respectively, while the Registration Recall (RR) was increased by 0.2 percentage points compared with the state-of-the-art method CoFiI2P. It is demonstrated that the fusion of positional encoding and overlap mask effectively enhances the descriptiveness of keypoint features, leading to significant improvements in registration accuracy under low-overlap conditions.

Key words: low-overlap scenarios, point cloud registration, positional encoding, overlap mask, self-attention