A multi-stage registration method based on dynamic OBBs for industrial 3D point cloud alignment under noise and incompleteness

Traditional point cloud registration methods often struggle with noise sensitivity, poor generalization, and limited stability. To address these issues, we propose a multi-stage rigid registration framework combining dynamic oriented bounding boxes (OBB) with an improved Iterative Closest Point (ICP) algorithm. The initial alignment employs curvature-adaptive OBB segmentation and a dual-modal noise filtering approach based on point density and normal vectors. A curvature-weighted coordinate alignment strategy further enhances initial pose estimation. For fine registration, a dynamic weighted ICP algorithm assigns adaptive weights to point pairs using curvature stability and normal alignment error, improving registration precision. Experiments on public datasets (Stanford Bunny315) and industrial oil pan point clouds demonstrate strong performance under varying overlap ratios and noise conditions. Compared with AGICP, the proposed method reduces error by 72% and time by 80% in missing data scenarios, and by 78% and 83%, respectively, in noisy data. For industrial data, it achieves 69% error reduction compared to NDT + ICP and 81% compared to 3D-SIFT + RANSAC, with faster execution. These results show that the proposed method significantly improves accuracy, robustness, and efficiency, making it well suited for practical industrial applications.