Speaker
Description
Spacecraft in-orbit servicing is a vital technology for extending satellite lifespans by enabling repairs, upgrades, and refueling. Accurate pose estimation, which determines the relative position and orientation of spacecraft, is crucial for such missions but poses challenges in space environments due to varying lighting, shadows, and high precision requirements. Traditional computer vision and sensor fusion techniques often fall short under these conditions. Recent advancements in Deep Learning (DL), particularly with Convolutional Neural Networks (CNNs), have demonstrated great accuracy and robustness, and emerge as a possible alternative to classic vision algorithms. However, real-time deployment of deep learning models in space is hindered by the limited computational resources of space-qualified hardware.
Field-programmable gate arrays (FPGAs) offer an effective solution, combining flexibility, power efficiency, and high computational performance, making them well-suited for AI acceleration in space applications. This thesis investigates the integration of AI-driven pose estimation algorithms with FPGA acceleration, utilizing Xilinx's Vitis AI and Deep Learning Processing Unit (DPU) to deploy CNNs for real-time space operations. The study focuses on the CAT-MICE system developed by GMV Aerospace and Defence S.A.U., with experiments conducted using Platform-ART® facilities as part of the CAT Breadboard project. The ultimate aim is to achieve high performance and precision in spacecraft pose estimation through FPGA-based deep learning solutions.
Affiliation of author(s)
Carlos III of Madrid University, GMV
| Track | Artificial Intelligence/Machine Learning |
|---|
