Numerical analysis of differential operators on raw point clouds (Q2249597)

Smooth surfaces can be provided as raw point clouds. The objective of the paper is to give methods to obtain (the equivalents of) mean and principal curvature from these data. It starts with a survey of the existing literature dominated by polynomial regression and local moments. Elements for asymptotic analysis (as the neighborhood shrinks to a point) are introduced. A main contribution is a section in which 4 different methods are proposed to obtain discrete analogues of a second fundamental form matrix. They are based on covariance matrices for the point cloud or normals and they asymptotically compute local differential operators giving curvatures. Another way to obtain curvature information is by moving least squares (MLS) regression. Smoothing can be obtained by projecting each point on a polynomial regression of degree 1 (MLS1) or 2 (MLS2). MLS2 is accurate to order 4, while MLS1 yields mean curvature. When iterated MLS1 results in a scale space for the surface, but MLS2 does not. Most reliable curvature results are obtained by combining scale space and MLS2.