Autonomous Navigation for Underwater Robots with Camera-Based Mapping and Depth Estimation

This paper presents an autonomous navigation framework for underwater robots that are equipped with a monocular camera and pressure sensors. This sensor configuration is low-cost and has low computational overhead, capable of providing reliable scale information and achieving effective navigation. Visual-based Simultaneous Localization and Mapping (SLAM) and depth estimation methods are used for obtaining the relative position and pose of the robot and perceiving the underwater obstacles respectively. To address the scale problem inherent in monocular cameras, we propose an initialization method of this sensor configuration for scale recovery in real-world environments, along with a fitting procedure for more accurate depth estimation results. Implementation on a real-world underwater robot platform validates the effectiveness pf the proposed framework.