Abstract

A 3D geometric model representing the surface of a 3D object is most often used for synthesizing highly realistic 3D images. The advent of 3D scanners has made the acquisition of 3D surface geometry of existing objects relatively simple. In most 3D scanners, it is difficult, to acquire the complete surface geometry in one operation. For example, with a single fixed scan head, the object has to be rotated many times to present different views of the object's surface to the scan-head, so that the entire surface geometry is covered. Then, all the scan data sets from different views have to be combined from their own independent coordinate systems into a single consistent geometric coordinate space. This operation is called scan data registration. Most existing techniques mainly focus on registering two point data sets at a time. For scan data in multiple views, say n views, one has to resort to pair-wise registration, requiring n -1 registration operations. In this research, we study the problem of multiple scan data registration. In particular, we are interested in efficiency, robustness and accuracy of registration techniques. Our main contribution is a new registration technique that can simultaneously register any number of scan data sets as long as there is some overlap amongst them, thus requiring considerably less than n -1 registration operations. To demonstrate the performance of our approach, we have developed a registration system and we have experimented with a number of scan data sets and shown the improvements resulting from the use of our approach.