Nc5-PCS Based Heterogeneous Point Cloud Matching for Roads

Abstract

Point cloud registration represents a fundamental task in geospatial informatics and 3D computer vision, aiming to align heterogeneous point clouds through rigid transformation estimation. While Super-4PCS serves as an efficient coarse registration method, it exhibits limitations when handling large-scale datasets, planar-distributed point clouds, and scenarios with unknown scale differences. To overcome these challenges, this paper proposes the Nc-5PCS (Neighborhood-constrained 5-Point Congruent Sets) algorithm. Nc-5PCS first performs approximate scale estimation through concavity-convexity similarity analysis within coarse overlap regions, addressing the inherent scale limitation in 4PCS-based approaches. Subsequently, the algorithm employs 3D Harris feature point extraction to significantly reduce data volume while preserving critical geometric characteristics. The core innovation lies in designing a non-coplanar 5-point basis with a corresponding hash-based retrieval mechanism, effectively resolving the feature degradation problem caused by coplanar 4-point bases. Furthermore, normal vector angular constraints are incorporated to enhance consensus evaluation during correspondence selection, substantially improving registration accuracy. Experimental validation demonstrates that Nc-5PCS achieves a point-to-point RMS error of ≤ 0.227 m, outperforming Super-4PCS to provide superior initial alignment for subsequent ICP refinement.