Generalised Discriminative Optimisation Algorithms for Augmented Reality: Theory and Practice.

Abstract

Augmented Reality (AR) technology achieves the seamless registration between virtual scenes and the real world. Three-dimensional registration is a core method for integrat- ing virtual information and the real world. Registration has always been treated as an optimisation process, involving the design of objective functions and the estimation of gradient directions. Learning-based optimisation methods learn the gradient directions via the least-square methods acting on a parameter vector space and extracted features from data. Learning-based optimisation methods do not require the design of objective functions or the calculation of derivations. This advantage reduces the high complexity and the large storage requirement for inverse Hessian approximation, which is common when using traditional optimisation methods. This thesis explores learning-based optimisation methods for point cloud registration with the aim for augmented reality applications. Three methods and a computational framework have been proposed: (1) A General Discriminative Optimisation method (GDO) has been proposed to reduce the effect of perturbations on updating gradient directions. The existing learning-based optimisation methods have several drawbacks, one of which is using a single feature to learn gradient paths, which makes the learning vulnerable to perturbations; (2) A Reweighted Discriminative Optimisation method (RDO) has been put forward to explore the asymmetrical contributions of each component of parameter vectors on registration to capture the influence of the component to improve the registration accuracy; (3) A Graph-based Discriminative Optimisation method (GRDO) has been proposed to reduce the storage requirement and computational cost; (4) Finally, a computational framework, SGRTmreg, has been devised to achieve multiple point clouds registration, which is a step forward in the state-of-the-art, since previous learning-based optimisation algorithms have mainly focused on single point cloud registration. Each of the new algorithms has been comprehensively compared with several state-of-the-art traditional registration algorithms as well as recently developed deep learning-based algorithms using public point cloud data sets in order to demonstrate key features (robustness, accuracy, efficiency, stability) of algorithms and its registration performance. In this thesis, theoretical convergence proofs for the proposed algorithms are also provided in appendixA. The potential of GDO, RDO, and GRDO for 3D point cloud registration is demonstrated through applications of 3D registration in real scenes and object tracking.