Reconstruction of curved cylindrical point clouds via iterative B-spline approximation and RANSAC-based robust estimation
TL;DR
Accuracy evaluations are conducted on seven simulated datasets and on a real curved cylinder point cloud consisting of 3126 198 points, demonstrating that the method can achieve high-precision and high-efficiency extraction of curved cylinder central axes, providing reliable technical support for related engineering acceptance and quality traceability.
Abstract
In the construction and building of conventional engineering projects, there is inevitably a demand for manufacturing curved cylindrical components in large castings. To address the problem that existing algorithms are prone to fitting divergence due to point cloud missing, non-uniform density, and high‐frequency noise in complex environments, this study proposes a high-precision and robust reverse reconstruction method. The method first employs principal component analysis to determine the principal direction, classify and index the point cloud data, and construct a B-spline as the initial trajectory. An adaptive slicing mechanism based on density field estimation is introduced to control the slice thickness, overcoming the sampling imbalance problem caused by non-uniform density distribution and local data incompleteness. After adjusting slices along this trajectory, RANSAC circle fitting is applied to remove noise, and a new trajectory is then constructed using the centers of least-squares circles, with multiple iterations performed. To verify the algorithm’s applicability under extreme working conditions, this paper conducts accuracy evaluations on seven simulated datasets (covering data missing, multiple types of noise, density gradients, high curvature, and radius variation) and on a real curved cylinder point cloud consisting of 3126 198 points. Experimental results show that on simulated datasets with approximately 2.5 million points each, the mean absolute error of the central axis extracted by the proposed algorithm remains stable within the high-precision range of 1.7 mm to 3.82 mm; for the real point cloud with a diameter of approximately 0.655 m, the maximum spatial deviation is only 5.08 mm, and the reconstructed axis achieves engineering acceptance-level consistency with the design ground truth. The average processing time for the simulated datasets is approximately 21.0 s, demonstrating that the method can achieve high-precision and high-efficiency extraction of curved cylinder central axes, providing reliable technical support for related engineering acceptance and quality traceability.