EagleAI: Estimation of Attitude Geo-localizing Landmarks on Earth

This Master's Thesis presents a visual method based on machine learning to estimate, from a picture of the Earth, the attitude of a satellite at the moment it captured that image. Attitude estimation is a crucial task in satellite operations as it determines the orientation and position of the satellite with respect to its surroundings. Conventional methods for estimating satellite attitude require multiple sensors and complex algorithms, making them prone to errors and limitations, especially in the case of small and low-cost satellites as CubeSats. In this work, a machine learning-based approach is presented, to be deployed on-ground, which leverages image data collected by cameras onboard the satellite to geographically localize the landmarks captured, and provide an estimation of the spacecraft attitude. In the proposed method we can identify three main steps: (1) first, the retrieval of a dataset of reference geolocalized pictures; (2) then, the selection of the best candidate pictures for the matching by means of a convolutional Siamese neural network, trained on a large dataset of Sentinel images synthetically modified; (3) finally, a pixel-level keypoint matching that enables the overlap of the input images and the geo-localization of the query. Results from the experiments demonstrate the feasibility of the proposed method and an in-depth study of the literature allows to point-out possible further developments to enhance its accuracy and robustness.