Spatially guided multi-scale robust ICP for high-precision reconstruction of large weak-feature curved surfaces

High-precision 3D reconstruction of large-size weak feature curved surfaces has long been a challenge in industrial manufacturing. Noise interference, occlusions, low overlap rates, and sparse geometric features often cause traditional point cloud registration methods to converge to local optima or fail to meet accuracy requirements. To this end, we propose a novel registration framework that integrates spatial positioning guidance with a multi-scale robust iterative closest point algorithm (MR-ICP). This method first establishes a world coordinate system using a global positioning system and achieves coarse registration of multi-view point clouds by introducing a transition coordinate system, thereby providing precise initial registration positions. Subsequently, the innovative MR-ICP algorithm incorporates a multi-scale extended convex hull to identify effective overlapping regions, while integrating a multi-scale feature fusion weighting mechanism to achieve synergistic optimization of global constraints and local details. Furthermore, we design an adaptive robust loss function, which dynamically adjusts residual weights to suppress noise during fine registration effectively. In Gaussian noise simulation stitching, this method reduces the root mean square error (RMSE) by up to 85% compared to the point-to-surface benchmark method. In the experiment, the average error achieved in the reconstruction of the engine hood (1.8 × 1.2 m) reached 0.261 mm, significantly outperforming existing advanced algorithms. This provides an efficient and scalable solution for large-scale industrial surface 3D reconstruction.