Landmark characterization based on lunar image using deep artificial neural network

Abstract

This thesis addresses a formulation and develops a novel approach to solve a landmark characterization problem for spacecraft flying over the moon with the ability to autonomously compute their absolute position based on a measured two-dimensional image of the lunar terrain. The goal of landmark characterization is to enhance the availability of image data in a harsh natural environment before flight. In a surface image, unique features can be localized to unknown landmarks that represent absolute positions, and an analyzed level of discrimination between regions of the landmarks can explore reliable landmarks. In particular, the proposed algorithms and procedures are structured to improve the performance of existing techniques by taking into account changes in the location of the sensor and the natural environment. The problem is formulated as a regional characterization method based on deep neural network (DNN), and four incremental approaches are proposed: a) the DNNs are trained and validated to quantify regional uniqueness (i.e., differences between regions)

b) the DNNs are applied to create characterized performance maps for evaluating the integrity of the landmarks

c) flight simulation demonstrates the usefulness of the landmark decision process using the performance map even in dark highlands

d) through a camera-based indoor experiment considering a test case of a flat plain on the near side of the moon, it is demonstrated that landmarks selected using the proposed method can compensate for the shortcomings of a classical algorithm. In addition, it is demonstrated that performance maps can be utilized to confine a marginal window of usable landmarks due to changes in the mission time of the spacecraft.