Evaluation and Integration of YOLO Models for Autonomous Crater Detection

Advancements in deep learning and computer vision are enabling the development of expanded spacecraft capabilities. One field of interest is automated crater detection which has applications in terrain relative navigation, pose estimation, and planetary science. While continued development of learning-based crater detection algorithms (CDAs) has led to more accurate and performant models, there is currently a limited discussion on how these models might be integrated into future mission infrastructure. Specifically, we identify the deployment of CDA models onto resource-constrained, flight-like hardware and the interaction of CDAs with existing flight software as key areas of investigation. To this end, we first introduce a novel Lunar crater dataset based on digital elevation map (DEM) data and 1.2 million known crater positions, leveraging the Blender 3D software to render surface imagery with ground truth bounding box labels. We evaluate the You Only Look Once (YOLO) family of models on this dataset for crater recognition performance while providing runtime and memory analysis on representative flight hardware, consisting of a Teledyne radiation-tolerant LS1046-Space CPU and a Google Coral Edge TPU accelerator. We comment on the choice of activation function in the YOLO architecture as it relates to detection performance and inference time. Carefully considering model operations is essential because Edge TPU compatibility is paramount for near-real-time, onboard deep learning execution. Finally, we put forth an example implementation of YOLO CDA within the On-Board Artificial Intelligence Research (OnAIR) platform, a cognitive architecture for autonomous applications that can interact with flight software frameworks such as NASA's core Flight System (cFS).