Abstract
A system of registering multiple range images obtained from viewpoints whose relative position is unknown is proposed. We deal with a complex object such as a human hand in this system. In this system, we determine transformation from each respective viewpoints to the common world coordinate by correspondences of their planes. To deal with non-polyhedral objects, we construct three dimensional convex hull from each range images. Then, we find corresponding planes between two successive views. For that purpose, we build adjacency graphs of large faces contained in the each convex hull, and match graphs of two successive views based on their areas. We find transformation between all views simultaneously. We represent the 3D transformation by rotation matrix and translation vector. First, we determine the rotation matrices using normal vectors of the planes by constrained least squares fitting. The rotation matrix must constrained to be orthonormal with determinant 1. Then, translation vectors are obtained by simple least squares fitting.
