Robust Registration of Noisy Point Sets

An algorithm for the pair-wise coarse registration of three-dimensional (3D) point sets sampled from piecewise smooth surfaces is presented. The algorithm is unsupervised and robust and can handle both clean and noisy data sets with outliers. The approach is based on a representation of the two data sets by sparse local features, invariant to rigid transformations, corresponding to the absolute maxima of the mean curvature field. The feature parameters are differential invariants extracted by means of a robust statistic approach. First the normal vector field of the data sets is evaluated by a two-step process to remove outliers and to refine the normal estimation near sharp features. Then the normal field is used to robustly estimate the local curvature tensor and the points of maxima of the mean curvature field. These maxima represent the feature points of the data sets, and are characterized by signature vectors of invariant parameters. The signature vectors are build on relative distance, principal curvatures and projections of normals and principal directions. The estimation of the best rigid transformation which aligns the two 3D data sets is obtained by the matching of pairs of triplets of feature points, one triplet per each data set. Segment matching is performed in a multidimensional signature space. Triplet matches allow the estimation of the roto-translation which align the two data sets. The effectiveness of the proposed registration algorithm is evaluated on synthetic and real 3D data sets and the preliminary results are given.