| Many underwater robot tasks performed currently are observational in nature, including those for inspection and exploration applications. The information of consequence for these tasks is visual imagery; therefore, visual sensing is an ideal sensing approach since it directly measures visual information. Despite the potential benefits offered by visual sensing, several complications have kept it from being utilized to its full potential for automating the control of underwater robots. General issues that must be faced (independent of the medium) include unstructured scenery, geometric ambiguities (because images are two-dimensional projections of three-dimensional information), limited field of view, and the processing of massive amounts of image data. Additional underwater-specific difficulties include limited viewing range, low contrast, lighting variations, and marine snow.; This disseration describes efforts to develop new visual sensing technologies useful for automatic control of underwater robots. Laplacian-of-Gaussian sign-correlation, a previously-developed computer-vision technique, is established as an effective approach for determining correspondences in underwater imagery. Its robustness to low contrast, brightness and contrast variation, nonuniform lighting, and marine snow are analyzed. In addition, the basic theory of Laplacian-of-Gaussian sign-correlation is extended to predict the effects of image scaling and rotation on correlation degradation.; Image correspondences are used to compute geometric image quantities including stereo disparity, optical flow, and optical displacement; these measurements are in turn used to determine terrain-relative and object-relative robot state. Terrain-relative state is calculated by combining optical-displacement and stereo-disparity measurements with nonvisual measurements of the camera pointing-direction. Object-relative state is determined by locating an object in the scene using motion (optical flow) and range (stereo disparity) segmentation.; To demonstrate the benefit of visual sensing for underwater robot control, this research investigates the automation of three tasks: fish tracking, station keeping, and video mosaicking. In each of the tasks, visual sensing is used as the primary feedback for control. The tasks were demonstrated experimentally using OTTER, a semiautonomous underwater robot designed specifically for automatic control research. In addition, automatic video mosaicking of the ocean floor was performed using the Monterey Bay Aquarium Research Institute's Ventana vehicle. The results of these experimental demonstrations conclusively establish that visual sensing can be used effectively for automatic underwater-robot control. |
