Exploring AI-Based Satellite Pose Estimation: from Novel Synthetic Dataset to In-Depth Performance Evaluation

Vision-based pose estimation using deep learning offers a promising cost effective and versatile solution for relative satellite navigation purposes. Using such a solution in closed loop to control spacecraft position is challenging from validation and performance verification viewpoint, because of the complex specification and development process. The validation task entails bridging the gap between the dataset and real-world data. In particular, modelling of Sun power and spectrum, Earth albedo, and atmospheric absence effects, is costly to replicate on ground. This article suggests a novel approach to produce synthetic space scene images. Fine statistical balancing is ensured to train and assess pose estimation solutions. A physically based camera model is used. Synthetic images incorporate realistic light flux, radiometric properties, and texture scatterings. The dataset comprises 120000 images supplemented with masks, distance maps, celestial body positions, and precise camera parameters (dataset publicly available https://www.irt-saintexupery.com/space_rendezvous/ created in the frame of a project called RAPTOR: Robotic and Artificial intelligence Processing Test On Representative target). An analysis method using a dedicated metric library has been developed to help the assessment of the solution performance and robustness. A deeper comprehension of algorithm behavior through distribution law fitting and outlier identification is then facilitated. Finally, it is shown that implementing Region-of-Interest (RoI) training can drastically increase the performance of the Convolutional Neural Networks (CNNs) for long-range satellite pose estimation tasks.