FPGA Hardware Acceleration for Deep Learning-based Satellite Pose Estimation

Accurate and robust pose determination is a key capability for spacecraft proximity operations in various mission classes such as On Orbit Servicing, Active Debris Removal, and Formation Flying. Depending on the characteristics of the main and target spacecraft, the required accuracy, and the operational environment, several methods and approaches exist for spacecraft pose estimation. When the main spacecraft is a very small satellite, such as a nanosatellite, with strict mass and power constraints, monocular vision is typically the most common solution. In this context, monocular approaches for pose estimation based on Deep Learning models represent a promising option to overcome the limitations of traditional approaches, especially in terms of robustness against harsh and variable lighting conditions. This paper presents the design, implementation, and validation of the FPGA-based acceleration of a Convolutional Neural Network (CNN)-based algorithm for pose estimation of a non-cooperative spacecraft. The FPGA hardware design is developed within the ANHEO research project, proposed by TSD-Space (Techno System Developments) company, and co-financed by the Italian Space Agency (ASI). The ANHEO system is a highly integrated unit for autonomous absolute and relative navigation of nano- and microsatellites, from low Earth orbit and high Earth orbits, up to lunar altitude. Designed for deployment in 6U-12U CubeSats, the ANHEO system optimizes form factor, interface compatibility, and mass versus power consumption. It features a novel deep learning-based monocular vision algorithm developed by the research team from the University of Naples “Federico II” to ensure robust pose estimation capabilities. This paper details the innovative architecture of the ANHEO unit and describes the FPGA hardware design developed for efficient real-time inference—a key capability for advanced navigation tasks in space.