Image-Based Lunar Hazard Detection in Low Illumination Simulated Conditions via Vision Transformers

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Space Exploration

Background:

• Autonomous lunar landing requires robust hazard detection systems.
• Current methods using 2D images and 3D Lidar face challenges in lunar south pole's low-light conditions.
• Shadows obscure critical terrain features, hindering vision-based navigation.

Purpose of the Study:

• To develop and evaluate a vision transformer (ViT) model for extracting terrain features from low-light lunar images.
• To improve autonomous hazard detection capabilities for lunar missions.
• To address the limitations of existing vision-based systems in challenging lighting.

Main Methods:

• Utilized a deep learning architecture, the vision transformer (ViT), inspired by natural language processing.
• Trained the ViT model to analyze low-light RGB images for terrain feature extraction.
• Compared ViT performance against the UNet convolutional neural network.

Main Results:

• The ViT model demonstrated strong performance in extracting terrain information from low-light images.
• Performance was particularly effective at higher altitudes.
• The ViT model outperformed the UNet model across all tested scenarios.

Conclusions:

• The vision transformer (ViT) model offers a promising solution for autonomous lunar hazard detection in challenging low-light environments.
• ViT-based systems can enhance the safety and reliability of lunar landings.
• This deep learning approach surpasses conventional methods like UNet for this specific application.