SAFT: Real-Time Tracking and Mapping With Self-Supervised Robust Stereo Matching for Underwater Vehicles

Area of Science:

• Robotics
• Computer Vision
• Underwater Navigation

Background:

• Underwater vehicle tracking and mapping face challenges like poor visual quality and limited resources.
• Existing deep learning stereo matching methods often lack speed or generalization.

Purpose of the Study:

• To develop a robust, efficient, and real-time tracking and mapping framework for underwater vehicles.
• To improve stereo matching performance in feature-sparse and textureless underwater environments.

Main Methods:

• SAFT-Stereo: A novel stereo matching network integrating cost aggregation and iterative optimization.
• Spatiotemporal self-supervised loss: Leverages spatial and temporal constraints for stable training.
• SAFT-DSOL: A real-time tracking and mapping algorithm integrating self-supervised models.

Main Results:

• SAFT-Stereo demonstrates superior generalization among real-time methods, with significantly reduced inference time.
• SAFT-DSOL achieves stable, efficient tracking and real-time dense reconstruction in challenging underwater scenarios.
• The framework shows robust performance on diverse underwater datasets.

Conclusions:

• SAFT provides a significant advancement in real-time, self-supervised underwater tracking and mapping.
• The proposed methods effectively address limitations of existing stereo matching and SLAM techniques in underwater environments.