YOLO-based crater detection for cislunar autonomous optical navigation: Characterization and parametric analysis of angular and area measurement errors

Optical navigation (OpNav) will play an important role in the development of cislunar (i.e., Earth–Moon system) autonomous navigation. The introduction of deep learning in recent years has paved a new path for computer vision to be explored. Convolutional neural networks (CNNs) can be trained to interpret images of a desired destination and accurately mark key features on its surface. In this paper, the application of CNNs to detect lunar craters for autonomous optical navigation is explored beyond previous implementations such as terrain-relative navigation (TRN), which are constrained to low lunar orbits ( < 100 km). More specifically, the state-of-the-art, real-time object detection algorithm YOLO (You Only Look Once) is trained with a comprehensive dataset of realistic, synthetically generated Moon images at a range of illumination conditions and distances (up to 71 , 250 km). A comprehensive analysis of line-of-sight and crater-area estimations is also performed as a function of relevant factors (e.g., distance, crater radii, solar-incidence angle, etc.) to inform on the usability of crater detections for autonomous navigation. We found that for a spacecraft orbiting the Moon, a CNN can robustly generalize the detection of surface craters for various illumination conditions and outside of its training range ( 2500 − 100 , 000 km ( ≈ 0 . 1 − 14 km/pixel ground sample distance (GSD))), with performance primarily influenced by crater size, distance to the surface, solar-phase angle, and spacecraft-emission angle. For the following factor ranges (evaluated independently), crater radius 32 − 64 px, distance 3250 − 45 , 000 km, solar-phase angles 30 ° − 135 ° , and spacecraft-emission angles < 60 ° , the model achieved mean crater area error 4 . 9 % − 5 . 8 % , mean normalized angular error 0 . 03 % − 0 . 07 % , and mean confidence 87 % − 91 % . Findings demonstrate the potential of CNNs applied to cislunar OpNav in orbit and provide a comprehensive characterization of measurement accuracies that could be further leveraged for feasibility analyses of autonomous missions. • Demonstrated a crater detection model that could detect craters up to 100 , 000 km. • Quantified how certain factors (radius, distance, etc.) drive detection accuracy. • Defined an OpNav Usability Envelope for crater detection up to 71 , 250 km. • Offered guidance for crater selection and imaging geometries for future missions.