An Enhanced 'Optimization-on-a-Manifold' Framework for Global Registration of 3D Range Data

In this paper we present a robust global registration technique which is suitable to accurately align sets of high-resolution range images. Our approach is based on the `Optimization-on-a-Manifold', OOM framework proposed by Krishnan et al. to which we contribute with both systemic and computational improvements. The original OOM algorithm performs an error minimization over the manifold of rotations through an iterative scheme based on Gauss-Newton optimization, provided that a set of exact correspondences is known beforehand. As a main contribution, we relax this requirement, allowing to accept sets of inexact correspondences that are dynamically updated after each iteration. Other improvements are directed toward the reduction of the computational burden of the method while maintaining its robustness. The modifications we have introduced allow to significantly improve both the convergence rate and the accuracy of the original technique, while boosting its computational speed. Meaningful comparisons with a classic global registration approach are also provided.