On-the-fly dense 3D surface reconstruction for geometry-aware augmented reality.

Abstract

Augmented Reality (AR) is an emerging technology that makes seamless connections between virtual space and the real world by superimposing computer-generated information onto the real-world environment. AR can provide additional information in a more intuitive and natural way than any other information-delivery method that a human has ever in- vented. Camera tracking is the enabling technology for AR and has been well studied for the last few decades. Apart from the tracking problems, sensing and perception of the surrounding environment are also very important and challenging problems. Although there are existing hardware solutions such as Microsoft Kinect and HoloLens that can sense and build the environmental structure, they are either too bulky or too expensive for AR. In this thesis, the challenging real-time dense 3D surface reconstruction technologies are studied and reformulated for the reinvention of basic position-aware AR towards geometry-aware and the outlook of context- aware AR. We initially propose to reconstruct the dense environmental surface using the sparse point from Simultaneous Localisation and Map- ping (SLAM), but this approach is prone to fail in challenging Minimally Invasive Surgery (MIS) scenes such as the presence of deformation and surgical smoke. We subsequently adopt stereo vision with SLAM for more accurate and robust results. With the success of deep learning technology in recent years, we present learning based single image re- construction and achieve the state-of-the-art results. Moreover, we pro- posed context-aware AR, one step further from purely geometry-aware AR towards the high-level conceptual interaction modelling in complex AR environment for enhanced user experience. Finally, a learning-based smoke removal method is proposed to ensure an accurate and robust reconstruction under extreme conditions such as the presence of surgical smoke.