Robustness analysis of data-driven image-processing algorithms applied to the Hera mission

Space proximity operations around asteroids demand precise relative navigation and high dynamic response achieved with the onboard implementation of autonomous visual-based navigation systems, which comprise image-processing algorithms that extract information from images taken by onboard cameras. This work presents a series of functional tests of two data-driven image-processing algorithms based on two different convolutional neural network architectures and designed for the application to the European Space Agency’s Hera mission with the target of binary asteroid system (65803) Didymos. The two data-driven methods estimate the position of the centroid of Didymos and its range from the spacecraft. Through different image datasets and comparative analyses, this work evaluates the two algorithms’ performance under conditions of adverse illumination, different shapes of the target asteroid, and different noise levels of the images, addressing questions on performance deviations, architectural discrepancies, and fine-tuning requirements upon encountering real-world scenarios. The analyses indicate that algorithms with more sophisticated and complex architectures exhibit greater robustness across various contingencies, despite being less accurate in their estimations. Furthermore, the results show that fine-tuning datasets improves the performances of the algorithms in the specific mission scenario they are generated for, while reducing the performances in other circumstances.