Star identification using regularization and nearest neighbor classification

Erdem Onur Ozyurt,Alim Rustem Aslan

DOI: https://doi.org/10.1016/j.actaastro.2024.01.037

IF: 2.954

2024-02-02

Acta Astronautica

Abstract:Astronomical attitude determination has helped mankind take "giant leaps" throughout history. Many devices have been invented for such use throughout ages, and their accuracy has been improved continuously. Scientific payloads requiring high accuracy have triggered research for even more accurate and precise attitude determination capabilities. Star trackers basically aim to make attitude determination by matching some images captured by a digital device to the reference stars. They provide the most accurate attitude determination results at any point independent from any celestial body. Furthermore, their capability of autonomous operation makes them the most convenient attitude determination sensor for far space missions, which allows trading off for higher complexity and cost. This study presents an algorithm for lost-in-space star identification with high accuracy in real-time implementation by decreasing computational complexity in comparison with state-of-the-art methods. This task is achieved by means of a novel dictionary-based star matching method through an approach that commissions the 1-nearest neighbor classifier through a binary search of Euclidean distances and the regularization method consecutively, accompanied by use of a novel approach of feature extraction in the beginning. A compact database comprising features representing each celestial frame not only ensures need for a smaller storage but also helps reduce computational complexity. The proposed method yields an estimated direction of the normal vector orthogonal to the camera plane and an estimated rotation of the camera plane about the normal vector. A series of experiments are conducted in a parametric-structured simulation medium, in which the Hipparcos catalog is taken as reference for database generation, and the characteristics of the star sensor used in the project SharjahSat-1 is used for generation of test observations. The performance is evaluated using error plots and precision rate curves for different values of the parameters in the algorithm. Complexity analysis is carried out in terms of database size and average run time and compared with the selected state-of-the-art methods. In addition, the simulation allows noise injection to evaluate improvement in performance in the presence of noise in comparison with other methods in the literature. It is shown that the proposed method achieves better accuracy than the given competing methods without noise injection and maintains very high robustness to positional noise and magnitude noise ensuring better accuracy. The high level of accuracy is achieved with an amount of database size and average run time that are competitive in the recent literature. It is additionally demonstrated that the levels of pointing accuracy offered by the state-of-the-art attitude determination and control subsystem models are achieved by the proposed method with the given probabilities.

engineering, aerospace