Leveraging Learned Monocular Depth Prediction for Pose Estimation and Mapping on Unmanned Underwater Vehicles

Marco Job^1*David Botta¹Victor Reijgwart¹Luca Ebner²Andrej Studer²Roland Siegwart¹Eleni Kelasidi^3,4

• ¹ETH Zürich, Zurich, Switzerland
• ²Thetys Robotics, Zurich, Switzerland
• ³NTNU, Trondheim, Sør-Trøndelag, Norway
• ⁴SINTEF Ocean, Trondheim, Sør-Trøndelag, Norway

This paper presents a general framework that integrates visual and acoustic sensor data to enhance localization and mapping in complex, highly dynamic underwater environments, with a particular focus on fish farming. The pipeline enables net-relative pose estimation for Unmanned Underwater Vehicles (UUVs) and depth prediction within net pens solely from visual data by combining deep learning-based monocular depth prediction with sparse depth priors derived from a classical Fast Fourier Transform (FFT)-based method. We further introduce a method to estimate a UUV's global pose by fusing these net-relative estimates with acoustic measurements, and demonstrate how the predicted depth images can be integrated into the wavemap mapping framework to generate detailed 3D maps in real-time. Extensive evaluations on datasets collected in industrial-scale fish farms confirm that the presented framework can be used to accurately estimate a UUV's net-relative and global position in real-time, and provide 3D maps suitable for autonomous navigation and inspection 1 .