Overlap confidence-guided fine registration for point cloud reconstruction of large feature-sparse curved components

High-precision point cloud reconstruction of large feature-sparse curved components is essential for industrial metrology, dimensional inspection, and digital manufacturing. However, such components usually contain large smooth surfaces, sparse salient features, low local overlap, boundary truncation, and scanning noise, which make conventional hard-correspondence registration methods prone to false matching, tangential drift, and error accumulation. To address these problems, this paper proposes an overlap confidence-guided local soft-correspondence anisotropic fine registration method. First, bidirectional expanded envelopes are constructed to extract candidate overlapping regions, and cross-level consistency is used to estimate point-wise overlap persistence confidence. Then, candidate points are organized into a multi-level surfel representation that integrates position, normal, covariance, curvature, level weight, and overlap confidence. Based on this representation, a composite matching cost and softmax-based local soft correspondence are introduced to reduce nearest-neighbor ambiguity. A normal-tangential anisotropic residual and adaptive robust weighting are further incorporated to suppress tangential drift, abnormal correspondences, and large-residual points. Real scanning experiments, controlled simulation experiments, ablation studies, and supplementary curved-panel validation were conducted to evaluate the proposed method. On the real engine hood dataset, the proposed method achieves a mean error of 0.1411 mm and an RMSE of 0.1973 mm, reducing the RMSE by 80.3% and 66.9% compared with GICP and Plane-to-Plane ICP, respectively. Under 15% - 25% overlap and Gaussian noise levels of 0.05 - 0.15 mm, it consistently obtains the lowest RMSE among the compared methods. The results demonstrate that the proposed method improves registration accuracy and robustness for the reconstruction of large feature-sparse curved components.