Leveraging learned monocular depth prediction for pose estimation and mapping on unmanned underwater vehicles

Area of Science:

• Robotics and Automation
• Underwater Systems Engineering
• Computer Vision

Background:

• Complex underwater environments, especially fish farms, pose significant challenges for Unmanned Underwater Vehicle (UUV) navigation and mapping.
• Accurate localization and 3D mapping are crucial for autonomous operations like inspection and navigation in these dynamic settings.

Purpose of the Study:

• To develop a general framework integrating visual and acoustic sensor data for enhanced UUV localization and mapping.
• To enable real-time net-relative pose estimation and depth prediction using solely visual data.
• To achieve global pose estimation by fusing visual and acoustic data and generate detailed 3D maps.

Main Methods:

• Deep learning-based monocular depth prediction combined with Fast Fourier Transform (FFT)-based sparse depth priors for net-relative pose estimation.
• Fusion of net-relative estimates with acoustic measurements for global UUV pose estimation.
• Integration of predicted depth images into the wavemap framework for real-time 3D map generation.

Main Results:

• Accurate real-time estimation of UUV net-relative and global positions demonstrated in industrial-scale fish farm datasets.
• Generation of detailed 3D maps suitable for autonomous navigation and inspection tasks.
• Successful integration of visual and acoustic data for robust underwater perception.

Conclusions:

• The proposed framework effectively enhances UUV localization and mapping capabilities in challenging underwater environments.
• The system provides accurate real-time positioning and 3D mapping essential for autonomous underwater operations in aquaculture.
• This approach offers a significant advancement for UUV applications in dynamic and complex marine settings.