Underwater simulation and mapping using imaging sonar through ray theory and Hilbert maps

Abstract

Mapping, sometimes as part of a SLAM system, is an active topic of research and has remarkable solutions using laser scanners, but most of the underwater mapping is focused on 2D maps, treating the environment as a floor plant, or on 2.5D maps of the seafloor. The reason for the problematic of underwater mapping originates in its sensor, i.e. sonars. In contrast to lasers (LIDARs), sonars are unprecise high-noise sensors. Besides its noise, imaging sonars have a wide sound beam effectuating a volumetric measurement. The first part of this dissertation develops an underwater simulator for highfrequency single-beam imaging sonars capable of replicating multipath, directional gain and typical noise effects on arbitrary environments. The simulation relies on a ray theory based method and explanations of how this theory follows from first principles under short-wavelegnth assumption are provided. In the second part of this dissertation, the simulator is combined to a continous map algorithm based on Hilbert Maps. Hilbert maps arise as a machine learning technique over Hilbert spaces, using features maps, applied to the mapping context. The embedding of a sonar response in such a map is a contribution. A qualitative comparison between the simulator ground truth and the reconstucted map reveal Hilbert maps as a promising technique to noisy sensor mapping and, also, indicates some hard to distinguish characteristics of the surroundings, e.g. corners and non smooth features.