Deep-Learning-Based Optical Navigation Framework Using Direction of Motion Measurements

Optical navigation plays a critical role in enabling challenging operations near small bodies. Existing techniques for processing optical data often depend on surface-related information, such as geological feature catalogs or shape models, which are limited by scale variations and low resolution of onboard datasets. To address these limitations, this study proposes a visual odometry-inspired method that uses motion direction observables extracted from the analysis of image pairs. These observables are used to constrain the spacecraft’s trajectory through a least-squares batch filtering approach. Leveraging deep learning models to reliably detect and match common features in partially overlapping images acquired at successive epochs, the proposed framework enables accurate spacecraft localization in the inertial reference frame. A comprehensive validation pipeline was developed to evaluate the network’s performance on datasets generated under varying illumination conditions and orbital configurations, demonstrating robustness in identifying correspondences between keypoints. This process supports the precise computation of the spacecraft’s estimated direction of motion. Numerical simulations in a synthetic environment, modeled on mission scenarios involving the Moon and the asteroid Vesta, provide an initial assessment of the achievable accuracy of the navigation solution and demonstrate its robustness to dynamical errors and camera pointing uncertainties representative of real mission conditions.