Abstract
Descriptor sensitivity to neighborhood scale, unreliable correspondences, and fixed correspondence thresholds pose challenges for partially overlapping point cloud registration under uncertain initial poses. This paper presents a coarse-to-fine registration framework integrating cross-scale consistency, geometric compatibility guidance, and residual-adaptive refinement. After statistical outlier removal and voxel downsampling, Fast point feature histogram (FPFH) descriptors are extracted at three neighborhood scales determined by the estimated point-cloud resolution. Mutual nearest-neighbor matching and cross-scale spatial voting are used to select candidate correspondences. Cross-scale support and pairwise geometric compatibility then guide M-estimator sample consensus (MSAC) sampling for initial pose estimation. The estimated transformation is subsequently refined using point-to-plane iterative closest point (ICP) with an Otsu-adaptive correspondence threshold. On the evaluated 3DMatch subset, the proposed method achieved a registration recall of 77.5%, compared with 57.5% for Fast Global Registration (FGR) + ICP, although it required longer registration time. In the ablation study, registration recall increased from 70.0% for the baseline to 95.0% for the full configuration, with a corresponding increase in runtime. These results support the effectiveness of the integrated framework under the evaluated conditions.