TOWARDS A GENERAL MULTIVIEW REGISTRATION TECHNIQUE

We present an algorithm that reduces significantly the level of the re gistration errors between all pairs in a set of range views. This algo rithm refines initial estimates of the transformation matrices obtaine d from either the calibrated acquisition setup or a crude manual align ment. It is an instance of a category of registration algorithms known as iterated closest-point (ICP) algorithms. The algorithm considers t he network of views as a whole and minimizes the registration errors o f all views simultaneously. This leads to a well-balanced network of v iews in which the registration errors are equally distributed, an obje ctive not met by previously published ICP algorithms which all process the views sequentially. Experimental results show that this refinemen t technique improves the calibrated registrations and the quality of t he integrated model for complex multipart objects. In the case of scen es comprising man-made objects of very simple shapes, the basic algori thm faces problems common to all ICP algorithms and must thus be exten ded.