Planetary defence against asteroid is a booming field in the space sector. High-level
of autonomy is required on board spacecraft during their proximity operations around
these small bodies, both to cope with the highly uncertain non-linear dynamical environment which surrounds them, and to reduce ground operation complexity related to
the significant delays in communication covering the distances involved in these missions. This thesis presents an Artificial Intelligence (AI)-assisted Image Processing (IP)
algorithm to support the optical navigation of asteroid rendezvous missions during their
close proximity operations. By focusing on the case scenario of the ESA’s Hera mission
to binary asteroid system (65803) Didymos, this work aims to tackle challenges of the
current paradigm of methodologies involved in standard and intelligent IP algorithms.
Firstly, by exploiting Convolutional Neural Networks, the algorithm is designed and
developed to cope with scenarios involving adverse illumination conditions, irregular
shape of the target body and the presence of external bodies. Secondly, the algorithm is
refined and implemented in an Open Loop navigation system to assess its performances
in the context of proximity operations. Finally an incremental validation test campaign
is performed to assess the applicability of the developed algorithm on board asteroid
rendezvous missions spacecraft. The test campaign objective is twofold: on one hand
it aims to solve standard AI-related issues, i.e. bridging domain gaps to account for
contingencies; on the other it aims to validate the algorithm on board spaceborne computers within the Guidance, Navigation and Control system of the spacecraft. This
thesis primarily contributes by designing and implementing a structured pipeline for
deploying AI-based IP algorithm in asteroid optical navigation, enabling a systematic
evaluation of its suitability and performance.