Design of a Spacecraft Pose Estimation System Using Convolutional Neural Networks
V. Fanizza (TU Delft - Aerospace Engineering)
E Mooij – Mentor (TU Delft - Astrodynamics & Space Missions)
Dominic Dirkx – Graduation committee member (TU Delft - Astrodynamics & Space Missions)
E. Van Kampen – Graduation committee member (TU Delft - Control & Simulation)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Performing crucial activities for space exploration, e.g., debris removal and on-orbit servicing, systems for Rendezvous and Proximity Operations (RPO) are required to be autonomous and scalable. Within this context, learning-based relative navigation has gained significant traction due to the latest advancements in AI and the cost-effectiveness of monocular cameras.
This thesis introduces a pose estimation system composed of an Object Detection (OD) and a Keypoint Detection (KD) network for keypoint extraction, coupled with a pose solver, and trained purely on synthetic images. When applying realistic data augmentations, the system achieves a reduction in KD error by 80%, further improved by training on photorealistic images. After an architecture optimization step, the final system consistently meets the inference time requirements on the Myriad X edge processor. This proves the feasibility of developing RPO systems with artificial data, showcasing a scalable approach, while complying with the limitations of onboard hardware.