Deformable cylinder extraction from LiDAR point cloud using candidate selection

Cylinder extraction is a fundamental task in point cloud-based environmental mapping such as tree modeling and reverse engineering. However, current methods are hindered by data missing, and their performance on deformable cylinders remains to be improved. To address these challenges, this paper proposes an unsupervised approach for robustly extracting deformable cylinders using a candidate selection strategy. A deformable cylinder is represented by a skeleton and cross-sectional circles. First, preprocessing is performed to downsample the raw point cloud and estimate normals. Second, cylinder extraction is conducted to obtain the initial cylinder segments. Then, candidate selection is applied to evolve the initial cylinder segments, extending the skeletons and generating cross-sectional circles to form the initial deformable cylinders. Finally, a refinement step is performed to optimize these cylinders. Experimental results validate the overall effectiveness and robustness of the algorithm against data missing. The average precision, recall, and F1 score are 0.95, 0.90, and 0.92, respectively, in the point cloud collected from four real-world pipeline scenarios. This algorithm supports practical applications such as clearance space measurement and Building Information Modeling (BIM) generation for pipeline systems.