Studies on Range Image Segmentation using Curvature Signs

Abstract

This paper studies on range image segmentation using curvature signs. The signs of mean curvature and Gaussian curvature can determine local surface shapes and they have been used for the range image segmentations. Range images are usually distorted by noise which is caused in range data acquisition, quantization and so on. The noise contamination affects curvature computation. This paper first evaluates the accuracy and the noise robustness of several curvature computation methods. The experimental results show that the robustness depends on the number of neighborhood points from which the curvatures are computed and that the method which uses first and second derivatives computed from quadric fitting is more robust than others. This paper then studies the characteristics of surface types and thresholds, which determine the curvature signs. The study shows that the noise robustness of plane, cylinder and sphere shape are different and that the smaller thresholds segment curved regions correctly.