Autonomous underwater vehicle navigation and mapping in dynamic, unstructured environments

Abstract

This thesis presents a system for automatically building 3-D opticaland bathymetric maps of underwater terrain using autonomous robots.The maps that are built improve the state of the art in resolutionby an order of magnitude, while fusing bathymetric information fromacoustic ranging sensors with visual texture captured by cameras.As part of the mapping process, several internal relationships betweensensors are automatically calibrated, including the roll and pitchoffsets of the velocity sensor, the attitude offset of the multibeamacoustic ranging sensor, and the full six-degree of freedom offsetof the camera. The system uses pose graph optimization to simultaneouslysolve for the robot's trajectory, the map, and the camera locationin the robot's frame, and takes into account the case where the terrainbeing mapped is drifting and rotating by estimating the orientationof the terrain at each time step in the robot's trajectory. Relativepose constraints are introduced into the pose graph based on multibeamsubmap matching using depth image correlation, while landmark-basedconstraints are used in the graph where visual features are available.The two types of constraints work in concert in a single optimization,fusing information from both types of mapping sensors and yieldinga texture-mapped 3-D mesh for visualization. The optimization frameworkalso allows for the straightforward introduction of constraints providedby the particular suite of sensors available, so that the navigationand mapping system presented works under a variety of deploymentscenarios, including the potential incorporation of external localizationsystems such as long-baseline acoustic networks. Results of usingthe system to map the draft of rotating Antarctic ice floes are presented,as are results fusing optical and range data of a coral reef.