Neighborhood Structure Consistency for Point Cloud Registration With Severe Outliers

Point cloud registration is a fundamental task in 3D sensor technology and computer vision, aiming to estimate the rigid pose between two point clouds and align them. Correspondence-based methods generate and verify hypothetical poses by sampling consistent correspondences. However, the presence of severe outliers in feature correspondences poses substantial challenges during the sampling process. It may result in misjudgment of inlier consistency relationships, leading to incorrect sampling and causing point cloud registration failure. In this paper, we introduce the Neighborhood Structure Consistency (NSC) metric to guide consensus excavation. It incorporates tighter global consistency constraints rather than local consistency, yielding a more distinct inlier cluster and significantly mitigating the impact of outliers. After that, we propose a Neighborhood Structure Consistency based-Graph guided Sampling Consensus (NSC-GSAC) algorithm, which performs the sampling process on a consistency graph based on the NSC metric. This graph provides a dense representation of connections among inlier correspondences, narrowing the search space and facilitating cleaner consensus subsets sampling, thus improving the registration robustness and efficiency. Additionally, to further improve efficiency, we introduce a Pruning Redundancy Hypothesis Selection (PRHS) strategy to prioritize highly promising representative hypothetical poses, thus accelerating the registration process. Extensive experiments on real-world indoor and outdoor datasets demonstrate that our method achieves superior registration accuracy and efficiency compared to state-of-the-art methods, including both traditional and learning-based methods. Furthermore, experimental results also indicate that our method exhibits more robust performance in the presence of severe outliers, outperforming all other methods.