Direction of Motion-Based Optical Navigation Using Deep Learning: Methods and Results from Dawn at Vesta

Direction of motion observables, extracted from landmarks tracked across successive image pairs, provide valuable constraints for spacecraft orbit determination. Unlike state-of-the-art optical navigation techniques, this method is independent from a priori knowledge of the target’s shape model or georeferenced landmark catalogs, making it suitable for the exploration of uncharted bodies. In this study, deep learning techniques are integrated into the navigation framework to detect and match keypoints across successive acquisitions, enabling direction of motion estimation. The framework is designed to process direction of motion data, either alone or in combination with other measurement types. By using real data from NASA’s Dawn mission at Vesta, we demonstrate that deep learning-derived direction of motion observables, when combined with radio tracking data, can significantly enhance trajectory reconstruction in scenarios with limited tracking coverage.