Machine vision for robust and resource-efficient spacecraft navigation

The proliferation of space activities has substantially intensified the criticality of autonomous spacecraft navigation systems for both on-orbit servicing (OOS) and active debris removal (ADR) missions. This dissertation focuses on developing and validating model-based monocular vision algorithms employing deep learning (DL) architectures for spacecraft pose estimation in close-proximity operations, specifically designed for deployment on computationally constrained embedded processors. The research addresses three main challenges inherent to vision-based navigation systems: achieving real-time performance on resource-limited hardware, bridging the synthetic-to-real domain gap, and ensuring operational reliability through uncertainty-aware decision-making frameworks. The first contribution proposes a DL-based Close-Range Pose Estimation Pipeline (CRPEP) tailored to the European Robotic Orbital Support Services (EROSS) mission. The pipeline combines object detection, keypoint regression, and Perspective-n-Point (PnP) stages with an active switching module that transitions between algorithmic configurations to maintain continuous pose estimation even when the target moves partially outside the camera’s field of view. Validation with photorealistic synthetic datasets and embedded hardware demonstrates real-time performance and accuracy compatible with close proximity mission requirements. The second contribution presents inference acceleration strategies for DL-based models deployment on embedded systems. Through model conversion, quantization, and batch processing, representative lightweight architectures are optimised, achieving significant speedups with minimal accuracy loss — confirming their suitability for satellite avionics. The third contribution tackles the domain gap challenge by employing a CycleGAN based unsupervised image-to-image translation network. Experiments demonstrate that retraining models on CycleGAN-translated synthetic imagery, derived from the existing SPEED+ dataset, results in substantial improvements on HIL test sets, with marked reductions in both keypoint localization and pose estimation errors. These outcomes establish adversarial domain adaptation via CycleGAN as a practical solution for synthetic-to-real transfer when annotated real data are limited and supervised fine-tuning is not feasible.