Improved Camera-Sonar Combination With Scale-Invariant Camera Elevation Estimation for Underwater SLAM in Close Visual Inspection

Abstract

Accurate localization is crucial for autonomous underwater vehicles (AUVs) performing close visual inspection (CVI) in underwater environments. Visual simultaneous localization and mapping (visual SLAM) based on a single monocular camera is commonly used for this purpose but suffers from scale ambiguity. Thus, additional underwater range sensors such as sonars are needed for an effective solution. This study aims to improve the recovery of the scale in a camera and multi-beam sonar combined system, under the CVI constraint. The proposed method leverages both sonar range measurements and scale-invariant camera elevation estimations, despite elevation ambiguity in a multi-beam sonar. Through a series of simulations, and under the CVI stand-off limit of 30 cm, the proposed system demonstrated a notable and consistent improvement of at least 31.13% on average (across different stand-off ranges and motion scenarios) in navigation error, over the existing state-of-the-art visual-sonar SLAM and other baselines such as visual-inertial or traditional monocular approaches. The proposed system effectively corrects visual distance approximation errors and ensures reliable depth ratio estimation in a simulation-validated manner, pending real-world verification, making it potential for navigation in real CVI missions.