Automatic morphological pre-alignment and global hybrid registration of LiDAR close range images

The point cloud alignment problem has always attracted the interest of the researchers. Two are the procedures usually applied for the close range surveys: the ICP method with all its variants, and the method based on the use of tie points identified by reflecting targets properly located on the overlapping part of two range images. The two methods are not concurrent. ICP requires an initial sufficiently accurate pre-alignment and does not work with regular surfaces; the tie points method requires the materialization and the recognition of the corresponding points, task that it is not always feasible and realizable in economical terms. This paper proposes a new hybrid technique to automatically execute the alignment of close range point clouds by evidencing the corresponding morphological singularities in the various scanning models. The point recognition is based first on the study of the local surface Gaussian curvature values, second by running a clustering procedure of laser points having extreme curvature values, and third by determining the centroids of each cluster. The computation of the local Gaussian curvature is carried out by applying to each sampled point a Taylor’s expansion local nonparametric algorithm with respect to its surrounding points. This makes it possible to locally estimate the surface function value, and its first and second order partial derivatives. The computation of the Weingarten map matrix elements, from second order Taylor’s expansion differential terms, allows to easily determine the Gaussian curvature for each point. For each point cloud, the above defined centroids, generate a vertex configuration. The punctual correspondence with the analogous vertices of an adjacent point cloud are automatically defined, according to the analysis of the respective adjacency matrices. From these sets of pairs, the pre-alignment rototranslation parameters are computed by a SVD algorithm. The final alignment is completed with an ICP method. The experimental results obtained for the alignment of the various parts of one of the well known Stanford models are shown. The paper also reports a general model, derived from the Generalized Procrustes Analysis, to obtain the simultaneous global registration of a set of point clouds. The method is able to simultaneously consider pairs of correspondent points automatically obtained by the ICP algorithm, pairs of tie points manually defined, and control points.