Effective and Robust Point Cloud Registration in Weak Texture Regions

Point cloud registration, a crucial component of 3D visual tasks such as AR, SLAM, and 3D reconstruction, aligns multiple point clouds by finding transformations. The success of point cloud registration largely hinges on finding high-quality point correspondences, which involves detecting overlapping regions between point clouds. While existing methods have demonstrated excellent performance in both indoor and outdoor scenes, they often falter in weak texture regions with similar structural features, producing numerous erroneous correspondences and leading to registration failure. In this paper, we propose a novel point cloud registration method aimed at reducing matching errors in weak texture regions. We introduce a weak texture score to identify regions with less distinctive structural features, and perform random downsampling based on this score to mitigate interference from weak texture regions. Additionally, we develop a new transformer-based matcher for the matching stage, which incorporates rotary position embedding and bidirectional attention to simplify calculations. This approach maintains high-speed computation while enhancing robustness and registration performance. Experimental results demonstrate that our method not only achieves superior matching accuracy compared to existing methods but also offers faster inference speeds, underscoring the effectiveness and efficiency of the proposed approach.