Sonar SLAM in Structured Underwater Environments

Abstract

Underwater environments pose a challenge to Si-multaneous Localization and Mapping (SLAM) performed by autonomous underwater vehicles (AUV), as underwater struc-tures are typically featureless with few references of motion. In turbid waters, sonar sensors must be used instead of cameras and LiDARs due to the properties of the water medium. A complete framework for underwater sonar SLAM in two dimen-sions is presented, built as a graph-based SLAM with submap registration. The keyframes of the SLAM are submaps of the surroundings, generated from a subset of the most recent frames. Keyframes are incrementally registered and two loop closure strategies are employed to correct the accumulated drift: one based on visual feature matching and bag-of-words, and the other based on refining the point cloud alignment of nearby keyframes. A robust pose graph optimization globally optimizes the path of the robot and the map of the environment, discarding loop closure outliers. The proposed pipeline is evaluated through real-world data collected from a marina underwater environment.