3D reconstruction and object recognition from 2D SONAR data

Abstract

Accurate and meaningful representations of the environment are required for autonomy in underwater applications. Thanks to favourable propagation properties in water, acoustic sensors are commonly preferred to video cameras and lasers but do not provide direct 3D information. This thesis addresses the 3D reconstruction of underwater scenes from 2D imaging SONAR data as well as the recognition of objects of interest in the reconstructed scene. We present two 3D reconstruction methods and two model-based object recognition methods. We evaluate our algorithms on multiple scenarios including data gathered by an AUV. We show the ability to reconstruct underwater environments at centimetre-level accuracy using 2D SONARs of any aperture. We demonstrate the recognition of structures of interest on a medium-sized oil-field type environment providing accurate yet low memory footprint semantic world models. We conclude that accurate 3D semantic representations of partially-structured marine environments can be obtained from commonly embedded 2D SONARs, enabling online world modelling, relocalisation and model-based applications.