Real-time Topology-Aware Augmented Reality.

Abstract

Augmented Reality (AR) technology fuses virtual information with the real-world en- vironment to enhance the way people interact with digital information in their physical world. This thesis is concerned with topology-aware AR systems designed to be aware of the topology changes in the surroundings and explore the topological features of scenes. Topological structures, such as graphs, can provide information on the relationship between point clouds to improve the quality of point cloud-based real-world 3D map reconstruc- tions for topology-aware AR systems. The reconstructed 3D maps provide information to improve the registration accuracy between virtual objects and the physical environment. Furthermore, 3D maps also help to reduce registration failures caused by complex and dynamic scenes, such as object occlusions, object motion, and object deformation. This thesis explores algorithms, computational methods, and frameworks for dense 3D surface reconstructions based on monocular videos and images for augmented reality applications. The main contributions of this PhD work are: 1) Proposed a graph deep learning-based framework for monocular depth estimation, which learns non-Euclidean features and improves the accuracy of depth estimations. Mathematical background on group equivariance, including translation equivariance and permutation equivariance, is also introduced to provide theoretical support for the proposed network; 2) Conducted two use cases to demonstrate the capabilities of the proposed methods in improving fine details of depth estimation for complex and unstructured environments with free camera motions; 3) A further improved the framework to address low-illumination endoscopy videos; 4) Proposed a statistical method to handle the non-rigid point cloud registration with special topology changes. Within which, a clustering and refinement scheme is proposed to deal with distribution irregularities of point sets; 5) Developed a framework to demonstrate the functionality of the proposed method in AR. Under challenging scenes such as endoscopy and unmanned aerial vehicle videos, the proposed methods outperform the state-of-the-art algorithms with robustness and accuracy. For example, the proposed depth estimation method improves the 3D data acquisition, the Break and Splice framework improves the 3D dynamic reconstruction, and the proposed AR framework provides a solution in dynamic scenes for medical applications.