Centroiding technique using machine learning algorithm for space optical navigation

Kaluthantrige, Aurelio and Feng, Jinglang and Gil-Fernández, Jesús and Pellacani, Andrea (2022) Centroiding technique using machine learning algorithm for space optical navigation. In: 3rd IAA Conference in Space Situational Awareness, 2022-04-04 - 2022-04-06, GMV.

[thumbnail of Kaluthantrige-etal-ICSSA-2022-Centroiding-technique-using-machine-learning-algorithm]
Preview
 Text. Filename: Kaluthantrige_etal_ICSSA_2022_Centroiding_technique_using_machine_learning_algorithm.pdf
Accepted Author Manuscript
License: Strathprints license 1.0
Download (3MB)| Preview

Abstract

Near-Earth-Objects pose a major threat to our planet with potential impacts. Together with the National Aeronautics and Space Administration (NASA), the European Space Agency (ESA) launched the Asteroid Impact Deflection Assessment (AIDA) international collaboration to demonstrate the deflection of the trajectory of binary asteroid system (65803) Didymos with kinetic impact. ESA’s HERA mission will arrive at Didymos in 2026 to observe closely NASA’s DART impact effects . This paper focuses on the Early Characterization Phase (ECP) and the Detailed Characterization Phase (DCP) of the proximity operations of HERA. The objective of these phases is to achieve physical and dynamical characterizations of Didymos. The optical navigation is applied to determine the relative state of the spacecraft with respect to the asteroid by detecting its Centre of Mass (COM). This can be achieved by the centroiding technique with the images captured by the spacecraft on-board camera and the Image Processing (IP) algorithm. Nevertheless, the standard IP algorithms depend thoroughly on the visibility of the target in the images, and they lack of robustness in case of adverse illumination conditions or if the target is partially out of the camera frame. To address these insufficiencies, this paper develops a model of IP based on the High-Resolution Network (HRNet) Machine Learning algorithm. With its convolutional layers, the HRNet is capable of extracting specific information from images without dependency on the quality of the image itself. Furthermore, the HRNet is capable of preserving the high-resolution of the image with superior spatial precision, which is desirable for estimating specific features such as the center of an asteroid. The training, validation and testing datasets are generated using the software Planet and Asteroid Natural scene Generation Utility (PANGU). The performances of the HRNet-based IP algorithm are evaluated in terms of Root Mean Squared Error (RMSE) between the pixel coordinates of the estimated and the true centroid of Didymos. The results shows that the HRNet-based IP algorithm is capable of regressing the position of the centroid with high accuracy and without being affected by the illumination conditions or if the asteroid is partially out of the camera frame.

CORE (COnnecting REpositories)