Range Image Segmentation and Classification via Split-and-Merge Based on Surface Curvature

Surface curvatures have been widely used for the purpose of segmenting and classifying range images. However, since surface curvatures include the second order derivatives, they become unreliable in the presence of noise, yielding false segmentation and/or classification of the range images. In this paper we investigate the sensitivity of the surface curvatures to the noise, and provide some observations on the characteristics of surface curvatures in the presence of noise. Following these observations, we then propose a scheme for reliable range image segmentation and classification. This scheme first differentiate planar region from curved region using planarity test. Surface curvatures are then computed only from the points belonging to the curved region, and mean curvature sign image (MCSI) and Gaussian curvature sign image (GCSI) are generated. Segmentation is done using split-and-merge operations on a quadtree representation of the MCSI. The sign of the Gaussian curvature is incorporated with the segmented MCSI to give a classification of the region. In the preliminary classification stage, the distribution of the Gaussian curvature signs are considered for each segmented region of the segmented MCSI. In the secondary classification stage, unclassified regions from the preliminary classification are classified using split-and-merge operations based on predominant Gaussian curvature signs.