Architectural 3D Point cloud data registration and fusion combining deep learning and traditional algorithms

This paper combines deep learning (DL) with traditional algorithms to solve the problems of noise interference, sparse data and incomplete local structure in architectural 3D PC (Point Cloud) data registration and fusion. Combining the RPMNet (Robust Point Matching Network)++ model with the ICP (Iterative Closest Point) algorithm improves the accuracy of PC registration and fusion effect, providing accurate 3D data support for building information modeling, bridge construction and smart city construction. RPMNet++ is used to extract high-dimensional features from PC, a bidirectional attention mechanism enhances the global representation of features, and the Copula correlation model analyzes and removes noise points. The ICP algorithm is introduced to fine-tune the initial registration results, achieving an efficient combination of global and local optimization. The experimental design covers noise-free PC, noisy PC and unnormalized PC scenes. The data source is four subsets of Paris-Lille-3D. Comparative experiments are conducted with mainstream methods such as PointNet (Point Network), DCP (Deep Closest Point), DeepICP (Deep Iterative Closest Point), FGR (Fast Global Registration), PRNet (Point Recognition Network) and LCCP-Net (Locally Convex Connected Patches Network). The anisotropic error (rotation error 0.00911, translation error 0.00009) and isotropic error (rotation error 0.01901, translation error 0.00021) of the proposed method in noise-free PC registration are both the lowest. The average inter-cloud chamfer distance of the noise-free points of the proposed method is 0.000002. Ablation experiments show that the average F-norm of the PC data registration method in this paper is 2.58. And it also shows significant advantages in processing time, with an average processing time of 0.47 seconds. This method achieves high robustness and high-efficiency registration for complex PC scenes, providing reliable technical support for the application of three-dimensional PC.