Autonomous Optical Processing for Space Surveillance: From Detections to Data-to-Track Association and Orbit Determination

Optical sensors play an increasingly central role in space surveillance by enabling wide-area monitoring of Earth orbit and the detection of objects that are difficult to observe with radar systems. However, transforming raw optical observations into reliable orbital information remains challenging, particularly in dense multi-target scenes and in catalogue-incomplete regimes where prior object identification is unavailable or unreliable. This thesis addresses these challenges through the development, integration, and validation of an end-to-end autonomous optical processing framework designed to transform ground-based observations into catalogue-quality orbital products with minimal human intervention. The proposed framework integrates AI-based detection, astrometric correction and calibration, data correlation and autonomous association, and angles-only orbit determination within a coherent processing architecture. Deep learning-based image processing extracts candidate resident space object detections from raw optical imagery acquired under different tracking modes. These detections are transformed into geometrically and temporally consistent angular measurements through deterministic astrometric corrections and system-level time-bias calibration. A dual-mode correlation and association strategy enables both catalogue-based processing and autonomous reconstruction of object-level tracklets in the absence of catalogue information. Robust angles-only orbit determination is then achieved through a combination of classical initialisation methods, particle swarm optimisation-based global search, and high-fidelity batch differential correction, producing dynamically consistent trajectories and associated uncertainty estimates. The framework is validated on real observational data acquired across LEO, GEO, and HEO regimes using the Sapienza observatory network. In LEO, angles-only orbit determination achieves arcsecond-level post-fit residuals and stable multi-day prediction when multi-night data are assimilated. In GEO survey conditions, large-scale catalogue correlation scales to datasets containing tens of thousands of measurements per night, and progressive multi-night refinement reduces long-term prediction errors from degrees to arcseconds. In the Molniya regime, where catalogue coverage is sparse and short arcs dominate, autonomous association and orbit determination driven cross-night linking enable sustained track custody and retrospective object identification. Across all orbital regimes, the experimental results demonstrate that a fully integrated optical processing framework can autonomously generate accurate orbit solutions from angles-only optical observations, enabling sustained multi-night tracking even in the presence of short observational arcs and high-density measurement environments. Together, these results establish the viability of autonomous optical space surveillance frameworks capable of supporting catalogue maintenance for known objects as well as independent tracking in regimes where catalogue information is incomplete.