|
Abstract
|
Cylinder point extraction is a crucial prerequisite for applications, such as reverse engineering and clearance analysis, in industrial scenarios based on laserpoint clouds. The current methods for extracting cylinder points still suffer from data missing, noise, and interference from the planar objects. In addition, cylindricalobjects with different radii, such as pipes and tanks, require the capability of generalization. To address these challenges, this paper proposes a region growthalgorithm using multi-scale prospecting for cylinder point extraction, considering a cylindrical object as a continuously extended cylinder. The multi-scale prospectingis used to detect the potential cylinder patches as kernels. The iterative growth is subsequently performed with the kernels as seeds to compute cylinder parametersused for allocating the raw points to diverse cylinders, thereby completing the cylinder point extraction. The point clouds of 4 industrial scenarios are collected to create a real-world dataset for qualitative, quantitative evaluation, and comparative analysis. Simulated point clouds with varying noise levels areused to validate the robustness to noise. Experimental results demonstrate the algorithm's effectiveness in extracting cylinder points from industrial scenario pointclouds while maintaining robustness to noise and cylinders with different radii. The average precision, recall, and F1-Score on the real-world dataset are 0.879, 0.894,and 0.885, respectively
|