AI-Based Optical Navigation for Rendezvous and Proximity Operations (RPOs) Missions of Small Satellites

On-orbit proximity operations are becoming increasingly important for both present and future missions, particularly OOS (On-Orbiting-Services) and ADR (Active-Debris Removal). In this framework, a high-accuracy estimation of the relative pose (position and attitude) between spacecraft is mandatory to successfully and safely achieve inspection/observation, rendezvous and docking phases. Spacecraft pose estimation is defined as the capability of an active spacecraft (chaser) to estimate the relative position and attitude of a known non-cooperative one (target). It involves both theoretical and technological challenges related to the search for the most suitable sensor architecture and algorithm solutions. On this behalf, visual navigation consists of estimating the pose of the target from a single image. It has recently become the most popular technique for this purpose, thanks to the employment of increasingly compact, precise, reliable, and low systems requirements monocular cameras, which represent the ideal sensor for such application. Traditional approaches focused on hand-engineered feature matching, but these methods have been known, among others, to suffer from lack of robustness, harsh lighting conditions and poor generalization capabilities. In recent years, improvements in terrestrial computer vision applications for object pose estimation, have pushed on employing Convolutional Neural Networks for space applications as well, with impressive results, such as higher robustness and resilience to noise and unseen scenarios. This work aims at proposing a CNN-based architecture for non-cooperative spacecraft pose estimation using a multiple-heads architecture, merging existing literature solutions with new features, consistent with the case study. A custom-made Blender-based synthetic dataset of about 9,500 synthetic images of a specific target has been generated to train the CNN. The obtained results achieve centimeter-level position accuracy and near-degree-level attitude accuracy, with robustness to changes in illumination conditions and background textures. In addition, a comparison with previously developed algorithms is conducted to assess the computational time and pose estimation performances. Moreover, in order to limit the computing cost and enable the proposed solution for real-time applications, a preliminary study towards Network Optimization is conducted. Side studies are also performed on the identification of the most significant hyperparameters to be tuned among all, and their correlations with each other. This study can serve as a starting point for future analysis.