Robotic 3D Reconstruction Utilising Structure from Motion

Sensing the real-world is a well-established and continual problem in the field of robotics. Investigations into autonomous aerial and underwater vehicles have extended this challenge into sensing, mapping and localising in three dimensions. This thesis seeks to understand and tackle the challenges of recovering 3D information from an environment using vision alone. There is a well-established literature on the principles of doing this, and some impressive demonstrations; but this thesis explores the practicality of doing vision-based 3D reconstruction using multiple, mobile robotic platforms, the emphasis being on producing accurate 3D models. Typically, robotic platforms such as UAVs have a single on-board camera, restricting which method of visual 3D recovery can be employed. This thesis specifically explores Structure from Motion, a monocular 3D reconstruction technique which produces detailed and accurate, although slow to calculate, 3D reconstructions. It examines how well proof-of-concept demonstrations translate onto the kinds of robotic systems that are commonly deployed in the real world, where local processing is limited and network links have restricted capacity. In order to produce accurate 3D models, it is necessary to use high-resolution imagery, and the difficulties of working with this on remote robotic platforms is explored in some detail.