Deep Learning based Moon Rock Obstacle Detection for Rover Navigation

Safe and autonomous navigation for the lunar rovers will play the most important role in successful lunar exploration missions. Precise detection and avoidance of obstacles, especially rocks, are main challenges for obstacle avoidance systems. This paper proposes a deep learning-based real-time moon rock obstacle detection approach, allowing the rovers to make clear decisions about where to move next over complex lunar terrains and which to avoid. Our approach will employ advanced state-of-the-art semantic segmentation techniques for accurately picking out and segmenting regions of rocks in images taken by cameras mounted on the rover. We train a deep neural network on a diverse collection of lunar images to discern, between rocks and the surrounding lunar surface. The segmented images can be used to generate obstacle maps, which are important for information about path planning and obstacles for the rover. We test our proposed method with a challenging dataset of lunar images, showing that our proposed method efficiently detects rocks under different sizes and shapes under dissimilar lighting conditions. Our experiments show that our approach produces a substantial performance gain over the conventional computer vision techniques and may be employed for safe and efficient navigation of lunar rovers.