RU-SLAM: A Robust Deep-Learning Visual Simultaneous Localization and Mapping (SLAM) System for Weakly Textured...

Area of Science:

• Robotics
• Computer Vision
• Marine Technology

Background:

• Simultaneous Localization and Mapping (SLAM) is essential for Autonomous Underwater Vehicles (AUVs) in unknown environments.
• Deep learning-based SLAM faces challenges underwater, including weak textures, image degradation, and inaccurate keypoint annotation.
• Existing SLAM methods struggle with the unique visual characteristics of underwater settings.

Purpose of the Study:

• To develop a robust deep-learning visual SLAM system tailored for underwater AUVs.
• To address the limitations of current SLAM techniques in challenging underwater conditions.
• To improve the accuracy and reliability of AUV navigation in unmapped marine environments.

Main Methods:

• A novel feature generator, UWNet, was designed to extract accurate keypoint features and descriptors, mitigating weak textures and image degradation.
• Knowledge distillation, based on an improved underwater imaging physical model, was employed for self-supervised network training.
• UWNet was integrated into ORB-SLAM3, replacing the traditional feature extractor for enhanced local and global feature utilization.

Main Results:

• The proposed UWNet system demonstrated superior performance in extracting features under challenging underwater conditions.
• Integration with ORB-SLAM3 led to improved feature tracking and closed-loop detection.
• Experimental validation on public and self-collected datasets confirmed the system's high accuracy and robustness.

Conclusions:

• The developed deep-learning visual SLAM system effectively overcomes key challenges in underwater navigation.
• UWNet provides a significant advancement for robust and accurate localization and mapping for AUVs.
• The proposed method offers a promising solution for enabling complex autonomous missions in underwater environments.