Range image registration for free-form surfaces using curvature information as invariance

A curvature based technique for fast range image registration is described in this work. With this technique the rotation and the displacement of a surface with respect to another one is calculated. The proposed algorithm uses Gaussian curvature information from each image point to extract a number of feature points. Two bounds are applied to the curvature values. Points with these bound curvature value are classified into two classes of feature points. In order to calculate the rotation angle between the two range images, the feature points are paired. That means the angle and the distance of the point pairs have to be transformed into a general view coordinate system. In the orientation matching procedure, we compare each pairing in the first image with each possible in the second one. This process is independent of the translation calculation which is calculated in a sequencing step. This pipelining principle speeds up the registration process if correspondences are calculated constantly.