Learning a global underwater geolocalization model with sectoral transformer.

Area of Science:

• Robotics and Autonomous Systems
• Navigation and Positioning
• Optical Oceanography

Background:

• Global Positioning System (GPS) is ineffective for underwater navigation.
• Current deep neural network (DNN) methods for polarization-based geolocalization lack generalizability across different sites.
• A domain gap exists between training and testing data in cross-site geolocalization tasks.

Purpose of the Study:

• To develop an advanced DNN methodology for robust underwater geolocalization using polarization patterns.
• To enhance the generalizability of polarization-based geolocalization to unseen locations.
• To improve positioning accuracy and inference speed for unmanned autonomous devices.

Main Methods:

• A novel DNN architecture based on the Transformer model, similar to large language models.
• Integration of an unscented Kalman filter (UKF) for precise geolocation estimation.
• Utilizing polarization-based images to simulate the sun's daily trajectory.

Main Results:

• The proposed DNN-UKF method significantly enhances cross-site geolocalization accuracy by approximately 28% compared to traditional DNN methods.
• The technique demonstrates improved performance across diverse global locations.
• Faster inference speeds were observed compared to existing benchmarks.

Conclusions:

• The advanced DNN methodology effectively addresses the domain gap in cross-site underwater geolocalization.
• This approach offers a more reliable and efficient solution for autonomous underwater navigation.
• The integration of Transformer architecture and UKF shows great promise for future geolocalization research.