Efficient Point Cloud Registration via Stratified Sampling and Bidirectional Correspondence Validation

Abstract

Traditional point cloud registration pipelines based on the Random Sample Consensus (RANSAC) and the Iterative Closest Point (ICP) often suffer from unstable correspondences and excessive computational costs. This paper presents an optimized coarse-to-fine registration framework that explicitly leverages feature saliency and bidirectional correspondence consistency to boost robustness and efficiency. In the coarse registration stage, instead of random keypoint sampling, a saliency-aware stratified sampling strategy based on normalized feature statistics is proposed. This strategy accounts for regions with multiple outlier features simultaneously, enhancing sampling globality and evaluation efficiency while preserving prominent geometric structures. For fine registration, a bidirectional correspondence verification mechanism is integrated to ensure mutual nearest neighbors between the source and target point clouds, effectively resolving cross-matching ambiguity. Extensive experiments on standard benchmark point cloud datasets show that the proposed method markedly outperforms the classical RANSAC+ICP approach and recent fast registration variants in robustness and convergence efficiency.