<p>Point cloud registration is a key technology for 3D reconstruction. Its core task is to achieve spatial alignment of multi-view point cloud data to reconstruct a complete 3D structure. However, in practical small scenarios such as industrial component recognition and pose estimation, targets often face challenges such as lack of texture, repeated structures, and severe occlusion, which leads to problems such as poor robustness and high registration failure rate in traditional methods. To address the above difficulties, this paper proposes a robust registration framework based on statistically weighted feature fusion and dual-constraint graph matching: (1) Constructing a similarity feature fusion model of curvature features and Fast Point Feature Histograms (FPFH) features, and enhancing the discrimination ability of local descriptors through collaborative optimization of feature space; (2) Designing a dual-constrained weighted bipartite graph (DWBG) matching mechanism guided by the global optimal distribution to achieve dual-domain alignment of local features and global distribution. Experimental verification shows that on the ROBI benchmark dataset and the actual industrial scene dataset, this method improves the registration accuracy by 46.68%, 42.47%, and 35.29% respectively compared with the 4PCS, RANSAC, and SAC-IA algorithms, and the RMSE standard deviation is stable in the range of 0.122–0.229&#xa0;mm. When cascaded with the ICP algorithm, the Ours + ICP scheme improves the robustness index by 57.1% while maintaining the same accuracy as LSG-CPD, showing significant advantages especially in high-noise and low-overlap scenarios.</p>