Regressing Poses from Monocular Images in an Underwater Environment

Re-localization is crucial to underwater vehicles especially for short-range missions, such as inspection and maintenance. While re-localization has been tackled successfully in the land and aerial environments, it remains to be a challenge in the underwater environment where radio communications and global positional systems are unavailable. Traditionally, underwater localization methods include dead reckoning (DR), inertial navigation systems (INS) and acoustic sensors. However, DR and INS tend to accumulate errors along time and are vulnerable to changes in water speed while acoustic sensors often require complex architectures which are expensive. Inspired by recent progress in land re-localization methods, we implemented a regression learning method, which is simple and cost-effective, for re-localization in short-range missions. The method is able to regress a 6-DOF pose from a single RGB image. We trained and evaluated the method on datasets collected from an underwater simulator. We also investigated its robustness towards changes in lighting and made improvements.