Our ability to image extended underwater scenes is severely limited by attenuation and backscatter. Generating a composite view from multiple overlapping images is usually the most practical and flexible way around this limitation. In this thesis we look at the general constraints associated with imaging from underwater vehicles for scientific applications - low overlap, non-uniform lighting and unstructured motion - and present a methodology for dealing with these constraints toward a solution of the problem of large area 3D reconstruction. Our approach assumes navigation data is available to constrain the structure from motion problem. We take a hierarchical approach where the temporal image sequence is broken into subsequences that are processed into 3D reconstructions independently. These submaps are then registered to infer their overall layout in a global frame. From this point a bundle adjustment refines camera and structure estimates. We demonstrate the utility of our techniques using real data obtained during a SeaBED AUV coral reef survey. Test tank results with ground truth are also presented to validate the methodology.
展开▼
