CycleGAN Enhanced Computer Vision Pipeline for Spacecraft Pose Estimation

Rai, Nitesh (2025) CycleGAN Enhanced Computer Vision Pipeline for Spacecraft Pose Estimation. [Laurea magistrale], Università di Bologna, Corso di Studio in Astrophysics and cosmology [LM-DM270], Documento ad accesso riservato.
Documenti full-text disponibili:
[thumbnail of Thesis] Documento PDF (Thesis)
Full-text accessibile solo agli utenti istituzionali dell'Ateneo
Disponibile con Licenza: Salvo eventuali più ampie autorizzazioni dell'autore, la tesi può essere liberamente consultata e può essere effettuato il salvataggio e la stampa di una copia per fini strettamente personali di studio, di ricerca e di insegnamento, con espresso divieto di qualunque utilizzo direttamente o indirettamente commerciale. Ogni altro diritto sul materiale è riservato

Download (4MB) | Contatta l'autore

Abstract

The rapid expansion of space activities has created unprecedented orbital congestion, with over 11,700 active spacecraft demanding sophisticated autonomous operations and enhanced Space Situational Awareness capabilities. This thesis addresses the challenge of accurate six-dimensional spacecraft pose estimation —essential for autonomous docking and proximity operations— through a computer vision pipeline enhanced by domain adaptation techniques. The methodology integrates YOLO11n for object detection, EfficientNet-B0 for keypoint regression, and Perspective-n-Point algorithms for pose estimation, creating a computationally efficient framework suitable for onboard deployment. The synthetic-to-real domain gap is addressed using Cycle-Consistent Generative Adversarial Networks (CycleGAN) for unsupervised domain adaptation through adversarial image-to-image translation. Experimental validation using the SPEED+ dataset revealed severe performance degradation when transferring from synthetic to real domains, with keypoint localization accuracy deteriorating by 13-14 fold and pose estimation orientation error increasing by 19-22 fold. The proposed CycleGAN-based framework achieved substantial improvements: 67% reduction in pose estimation error for lightbox conditions, improving orientation accuracy from 34.13° to 11.37°. Sunlamp domain results showed mixed performance, indicating operational limits under extreme illumination conditions. Ablation studies identified keypoint regression as the primary performance bottleneck, contributing 52% additional improvement beyond object detection adaptation alone. The research establishes that cross-domain generalization is achievable through unsupervised domain adaptation using only unpaired imagery, providing a viable pathway for deploying synthetic-trained models in real space scenarios without extensive annotation campaigns, advancing autonomous spacecraft operations in increasingly congested orbital environments.

Abstract
Tipologia del documento
Tesi di laurea (Laurea magistrale)
Autore della tesi
Rai, Nitesh
Relatore della tesi
Correlatore della tesi
Scuola
Corso di studio
Ordinamento Cds
DM270
Parole chiave
satellite pose estimation space situational awareness deep learning computer vision generative AI CycleGAN
Data di discussione della Tesi
18 Luglio 2025
URI

Altri metadati

Statistica sui download

Gestione del documento: Visualizza il documento

^