A Cascade Registration Method Based on Transformation Transfer for Multiscale Point Clouds

Point cloud registration is a core technology in fields such as 3D reconstruction and robot navigation. Current methods, however, struggle to balance accuracy with efficiency: the traditional Iterative Closest Point (ICP) algorithm is prone to local minima, while the Coherent Point Drift (CPD) algorithm suffers from high computational complexity. This paper presents a Cascade Registration method based on Transformation Transfer for multi-scale point clouds (CR-TT). Its key innovation lies in constructing a tightly coupled cascade framework where the transformation solved by coarse registration (guided by keypoints for rapid optimization) is converted into a strong prior for fine registration (driven by probability density modeling). This approach reduces the complex global optimization problem to efficient local refinement, achieving a paradigm shift from merely ‘improving the initial guess’ to ‘simplifying the optimization problem itself’. Experiments on multi-scale datasets demonstrate that CR-TT achieves significant advantages across small-, medium-, and large-scale scenes, with its Root Mean Square Error (RMSE) improved by 4.4–25.4 times over traditional ICP and CPD. Compared to state-of-the-art deep learning methods (e.g., DCP, GeoTransformer), CR-TT exhibits superior generalization capability and stability in out-of-distribution, large-scale complex scenes. In the engineering registration of point clouds for arch rib segments of a large concrete-filled steel tubular bridge, the coefficient of determination (R²) reaches 99.64%. The proposed method provides a reliable solution for efficient and robust alignment of cross-scale point clouds.