Fig. 6.7 A taxonomy of advanced registration techniques

6.3.1 Registration of More than Two Views

Once registration has been performed pairwise, all the views need to be transformed into a global reference system by applying a multiple-view registration technique. There are two main issues: (i) error accumulation and (ii) automation of the process.

Reducing Error Accumulation When the ordering of the sequence of views N1, . . . , Np is available, registration can be performed pairwise between consecutive views (i.e., between views Ni and Ni+1). In general, even if all the pairs are apparently well registered, some misalignment typically appears when the full model is reconstructed due to the accumulation and propagation of the registration error. The general idea of multiple-view registration techniques is to solve simultaneously for the global registration by exploiting the interdependences between all views at the same time. This introduces additional constraints which reduce the global error. A comparative study of similar multiple-view registration schemes was performed [27]. In [71] a method is presented that first aligns the scans pairwise with each other and then uses the pairwise alignments as constraints in a multi-view step. The aim is to evenly distribute the pairwise registration error, but the method itself is still based on pairwise alignments. In [17] a method that distributes registration errors evenly across all views was proposed. It operates in the space of estimated pairwise registration matrices, however ordering of the views is required. More recently, [91] proposed a new approach based on the well-known Generalized Procrustes Analysis, seamlessly embedding the mathematical theory in an ICP framework. A variant of the method, where the correspondences are non-uniformly weighted using a curvature-based similarity measure was also presented.

Automating Registration Especially when the full model is composed of a large number of scans, the view order might not be available and therefore should be manually specified. Many methods have been proposed to improve the automation