Star Field Feature Characterization for Initial Acquisition by Neural Networks

This paper presents an analysis of star field image features for star field recognition using neural networks during initial acquisition. This is a critical mode in star tracker operation. A learning vector quantization network is investigated. This is an alternative to routines that browse pre-compiled star feature databases for identification because the network structure itself contains the information about star feature vectors. A set of 200 circular sectors, partially overlapping and uniformly distributed over the celestial sphere, was selected as prototypes for recognition during network operation. Then, a number of candidate features was evaluated in each sector. The data have been analyzed to assess feature capability of addressing star field recognition when used as network input. Three basic statistical analyses have been performed: uniformity of distribution, feature cross-correlation, and stability. The most representative and suitable features for the considered application were selected: total number of stars, total radiant flux, four statistics of nearest neighbor angular separations in the field of view (min, max, mean and standard deviation), three second order moments of the star field image. A preliminary validation of feature selection has been performed by running a neural network