Residual grounding transformer network for terrain recognition on the lunar surface

Area of Science:

• Robotics and Autonomous Systems
• Planetary Science
• Computer Vision

Background:

• Autonomous navigation is critical for extraterrestrial rovers due to communication delays.
• A scarcity of annotated terrain recognition datasets for extraterrestrial environments hinders algorithm development.
• Existing datasets lack the real-world conditions necessary for robust training and evaluation.

Purpose of the Study:

• To address the lack of annotated datasets for extraterrestrial terrain recognition.
• To develop and evaluate a robust algorithm for identifying hazardous zones on extraterrestrial surfaces.
• To enable autonomous and safe navigation for lunar rovers.

Main Methods:

• Construction of the Chang'e 3 Terrain Recognition (CE3TR) Dataset using Yutu moon rover imagery.
• Proposal of a Residual Grounding Transformer Network (RGTNet) for semantic segmentation and hazard identification.
• Integration of cross-scale feature interactions and a local binary pattern feature fusion module within RGTNet.
• Introduction of a smooth intersection over union loss function to prevent overfitting.

Main Results:

• The CE3TR Dataset captures authentic lunar illumination and terrain conditions.
• RGTNet effectively identifies unsafe areas such as rocks and craters.
• The proposed model demonstrates superior performance in recognizing risky terrain compared to state-of-the-art methods.
• Extensive experiments validate the efficacy of RGTNet on the CE3TR Dataset.

Conclusions:

• The developed CE3TR Dataset and RGTNet significantly advance autonomous terrain recognition capabilities for extraterrestrial rovers.
• The RGTNet model offers a promising solution for enhancing rover safety and operational efficiency on the Moon and beyond.
• This work provides a foundation for future research in extraterrestrial robotic navigation and hazard avoidance.