RGB-D Perception-Enhanced 3D Gaussian Splatting SLAM: A Robust Framework for Mapping Underground Spaces

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Geospatial Information Science

Background:

• Efficient underground space utilization is critical for mitigating land scarcity.
• 3D Gaussian Splatting (3DGS) offers advantages for robotic perception and digitalization in underground environments.
• Existing Simultaneous Localization and Mapping (SLAM) systems struggle with accuracy due to poor illumination, sensor noise, and geometric degradation in underground spaces.

Purpose of the Study:

• To develop an RGB-D perception-enhanced 3DGS SLAM method for robust underground localization and mapping.
• To address challenges posed by low-light conditions, noise, and geometric degradation in underground environments.
• To improve the accuracy and efficiency of SLAM systems for creating digital twins of underground spaces.

Main Methods:

• Implemented a data enhancement pipeline using Multi-Scale Retinex with Color Restoration (MSRCR), Side Window Filtering (SWF), and adaptive gamma correction in the HIS color space.
• Developed a depth completion network for hole-filling and a hybrid metric for keyframe selection balancing efficiency and completeness.
• Introduced a dual-constraint Gaussian management strategy with opacity threshold and observation frequency, coupled with loop closure detection for global consistency.

Main Results:

• Achieved a 13.8% improvement in Peak Signal-to-Noise Ratio (PSNR) compared to state-of-the-art benchmarks.
• Demonstrated competitive trajectory accuracy and computational efficiency in experiments within coal mine tunnels and underground parking garages.
• Validated the method's effectiveness using a custom-designed underground mobile robot platform.

Conclusions:

• The proposed RGB-D perception-enhanced 3DGS SLAM method significantly improves localization and mapping robustness in challenging underground environments.
• The data enhancement, keyframe selection, and Gaussian management strategies effectively address illumination variations and sensor degradation.
• The findings strongly support the advancement of digital twin systems for underground spaces, enhancing their practical applicability.