Visual mapping of underwater structures using autonomous underwater vehicles

Abstract

With continuous advances in robotics and sensing technologies, unmanned underwater vehicles (UUVs) have been attracting increasing research attention in recent years, and they are expected to perform various missions in regions that are inaccessible or unsafe for humans. These vehicles have been of particular use in performing missions to obtain visual information of the surveyed areas for scientific research and engineering applications. In this context, visual inspections of underwater structures, which used to be performed by human divers, are a potential application for UUVs.

This dissertation addresses an automated underwater visual inspection system that includes an autonomous underwater vehicle (AUV) and vision-based global mapping algorithms developed in the framework of simultaneous localization and mapping (SLAM). Specifically, an AUV system as an experimental test-bed platform was newly developed for autonomous surveys around underwater structures and for obtaining optical images from the target surface. A visual mapping algorithm for planar surface structures is then described. Here, a selective image registration (SIR) scheme is proposed to effectively use the underwater image pairs where visual features might not be evenly distributed. Because the proposed SIR approach evaluates the potential utility of given image pairs and discards meaningless image registration attempts, the overall computational efficiency of the visual mapping algorithm can be substantially improved without significant decreases in the resulting accuracy. Finally, a novel visual mapping algorithm is proposed for three-dimensional (3D) reconstruction of curved surface structures using a monocular camera as a primary mapping sensor. The proposed 3D visual mapping method uses the fact that the local scenes obtained from different camera perspectives on the target surface reflect the information of a combination of piecewise-planar panels constituting the moderately curved surface. The nonsequential relative measurements between the cameras are extracted from the 3D panels associated with the local images and are incorporated into a SLAM framework as loop-closure factors to prevent unbounded error growth in the estimated camera trajectory. The validity and practical feasibility of the proposed methods are demonstrated using field experiment datasets obtained with full-scale ships in real sea environments.