Underwater robot visual place recognition in the presence of dramatic appearance change

Abstract

This paper reports on an algorithm for underwater visual place recognition in the presence of dramatic appearance change. Long-term visual place recognition is challenging underwater due to biofouling, corrosion, and other effects that lead to dramatic visual appearance change, which often causes traditional point-based feature methods to perform poorly. Building upon the authors' earlier work, this paper presents an algorithm for underwater vehicle place recognition and relocalization that enables an autonomous underwater vehicle (AUV) to relocalize itself to a previously-built simultaneous localization and mapping (SLAM) graph. High-level structural features are learned using a supervised learning framework that retains features that have a high potential to persist in the underwater environment. Combined with a particle filtering framework, these features are used to provide a probabilistic representation of localization confidence. The algorithm is evaluated on real data, from multiple years, collected by a Hovering Autonomous Underwater Vehicle (HAUV) for ship hull inspection.