Automatic segmentation of unorganized noisy point clouds based on the Gaussian map

A nonparametric clustering algorithm, called cell mean shift (CMS), is developed to extract clusters of a set of points on the Gaussian sphere S^2. It is computationally more efficient than the traditional mean shift (MS). Based on the singular value decomposition, the dimensional analysis is introduced to classify these clusters into point-, curve-, and area-form clusters. Each cluster is the Gaussian image of a set of points which will be examined by a connected search in R^3. An orientation analysis of the Gaussian map to area-form clusters is applied to identify hyperbolic and elliptical regions. A signed point-to-plane distance function is used to identify points of convex and concave regions. Segmentation results of several real as well as synthetic point clouds, together with complexity analyses, are presented.