Automatic registration of point-based surfaces

Shape modelling of 3D (three-dimensional) structures is an important issue in several fields, among other, manufacturing and quality control, reverse engineering, medicine and cultural heritage. In the first step of the shape modelling pipeline, the data acquisition, the increasing use of range measurement devices as laser scanners, structured light and stereo systems makes available large amounts of data sets as clouds of 3D points, or scans. An important issue in the modelling pipeline is the alignment, or registration, of partially overlapping scans in a common coordinate system. A new approach to automatic coarse pair-wise registration of partially overlapping scans is presented in this paper. The approach is based on the characterization of the pair of scans by two sets of sparse feature points invariant to Euclidean transformations, which are robustly extracted from the data by means of a multiscale procedure. The feature points are grouped in sets of triplets and the triplets are then characterized by signatures, which are vectors of parameters with the same invariance properties of the feature points. Pairs of triplets are matched in the signature space according to their signature distance. The alignment transformation able to register the two scan pairs is then estimated from the matched pairs of triplets. A verification process asses the quality of the match, which is defined as the number of feature points put in correct correspondence by the estimated transformation associated to a matched pair. This number of features must be greater than a prescribed threshold, defined by the expected percentage of overlapping of the two data sets. The proposed approach to registration has been evaluated on original scan sets and on standard ones available on the web. Preliminary results, presented and discussed in the paper, confirm the validity of the approach. Other work has to be done to analyze the robustness of the approach to noisy data and to optimize the computational time.