Computer vision applied to underwater robotics

Ocean and seafloors are today probably the less known and unexplored places on earth.Nowadays, the continuous technological improvements on underwater inspection offer new challenges and possibilities. Beside the lassic acoustic sensors, modern cameras are playing an ever increasing role in autonomous underwater navigation. In particular, The capability to perform a context-driven navigation, based on what the vehicle is actually seeing on the seafloor, is of great interest in many research fields, spanning from marine archaeology and biology to environment preservation. Industrial companies on oil and gas or submarine cabling, also have a strong interest in underwater robotics. The peculiarities of the underwater environment offer new opportunities to computer vision and pattern analysis researchers. This thesis analyses, discusses and extends computer vision techniques applied to the underwater environment. The main topic is the semantic classification of the seabed. A framework that may actually be embedded in an underwater vehicle and made to work in real time during the navigation was developed. The first part of this work addresses the problem of semantic image labelling. For this purpose a deep analysis of feature sets and related classification algorithms was carried out. The physical properties of light propagation in water need to be properly considered. Inspired by techniques for terrestrial single image dehazing, a new approach for underwater scenarios was developed. This approach is capable to significantly remove both the marine snow and the haze effects in images, and to effectively handle non-uniform and artificial lighting conditions. By jointly combining the results of underwater classification and the physical modelling of light transmission in water, a new feature set, more robust and with better discriminative performance was defined. Experimental results confirmed the accuracy improvements, Over the state-of-the-art obtained with the new feature set, in most critical environmental conditions. This work is largely based on original images and data, acquired during the European project ARROWS. The novelties introduced by this thesis may represent a basis for future applications, stimulating novel directions for research in computer vision and its applications to the underwater environment.