Line feature detection from 3D point clouds via adaptive CS-RBFs shape reconstruction and multistep vertex normal manipulation

This paper presents a multistep based on vertex normal manipulation method for ridge line detection on 3D surfaces, which is reconstructed from 3D point clouds via adaptive CS-RBF. Based on point density, centers of RBF are randomly chosen from the points. The support size of each RBF is obtained adaptively depending on surface geometry, which surrounds the RBF center. After the reconstructed shape is polygonized, a classification-based method for vertex normal approximation is executed. Searching the entire neighborhood in geodesic distance, we calculate the standard deviation of angles between a vertex and its geodesic neighbors, decide which kind of neighbor can be employed for final normal estimation, with the help of a user-specified parameter. With continuing curvature tensor estimation and ridge vertex evaluation, ridge-valley lines of salient features on polygon meshes are detected. Comparing with common vertex normal approximation, experiment results show that this method works satisfactorily, especially for 3D human faces.