The Doctoral Studies Office is pleased to inform you that
Mr Clémentin BOITTIAUX
Doctoral student at the COSMER laboratory
to Doctoral School 548 “Sea & Sciences”
will submit his thesis in order to obtain the degree of Doctor
under the supervision of
Vincent HUGEL, University Professor, University of Toulon (France)
Aurélien ARNAUBEC, PhD, Ifremer (France), Co-supervisor
Claire DUNE, Associate Professor, University of Toulon (France), Co-supervisor
Ricard MARXER, University Professor, University of Toulon (France), Co-supervisor
Discipline: Automatics, Signal, Productics, Robotics
on the theme
“Visual localization for long-term deep-sea monitoring”.
Thursday, December 14, 2023 at 10:00 a.m.
At Université de Toulon – Campus La Garde Bâtiment Y1 – Amphi Y.008
before a jury composed of :
- Ms Marie-Odile BERGER, Director of Research, INRIA Nancy Grand-Est (France), Rapporteur
- Vincent LEPETIT, University Professor, Ecole Nationale des Ponts et Chaussées (France), Rapporteur
- Cédric DEMONCEAUX,Professor, University of Burgundy (France), Examiner
- Torsten SATTLER,Research Fellow, Czech Institute of informatics Robotics and Cybernetics (Czech Republic), Examiner
- Aurélien ARNAUBEC,PhD, Ifremer (France), Co-supervisor
- Ms Claire DUNE, Associate Professor, University of Toulon (France), Co-supervisor
- Ricard MARXER,Professeur des Universités, Université de Toulon (France), Co-supervisor
- Vincent HUGEL,Professor, University of Toulon (France), Thesis supervisor
Visual localization for deep-sea long-term monitoring
This thesis explores the challenge of localizing underwater vehicles within previously explored environments in long-term scenarios, where significant scene appearance changes may have occurred. Typically, underwater vehicle positioning relies on fusing measurements from acoustic and inertial sensors. While these sensors deliver precise relative pose estimations, their absolute position estimates exhibit notable biases, resulting in position offsets spanning tens of meters between different dives. This limitation impedes the practical use of autonomous underwater vehicles for tasks requiring high precision, like mapping a precise area of interest. In response, this thesis investigates the use of visual observations made by underwater vehicles to enhance absolute positioning accuracy. The underwater environment introduces unique sources of variability absent in terrestrial environments. Consequently, the first contribution of this thesis is a novel dataset designed for benchmarking long-term visual localization algorithms in deep-sea conditions. Another obstacle inherent to underwater images is that their suffer from low contrast and loss of colors because of light propagation in the water medium. To address this issue, the second contribution of this work introduces two underwater color restoration methods, specifically designed to mitigate these phenomena and recover clear images. Independent of the underwater environment, the third contribution of this thesis is a novel loss function tailored for camera pose regression within the context of deep learning applications. This is an important aspect to consider when training visual localization networks. Finally, this thesis concludes with a benchmark of several visual localization methods on the proposed dataset. The obtained results show that applying our underwater color restoration method improves visual localization performance. This work also identifies the major problem encountered by visual localization methods on the proposed underwater dataset, and presents an approach to improve the accuracy of visual localization techniques by making the most of a limited size dataset.
Keywords: visual localization, deep-sea mapping.