Generalized Likelihood Ratio Test for Detection of Gaussian Rank-One Signals in Gaussian Noise With Unknown Statistics

Abstract

We consider the classical radar problem of detecting a target in Gaussian noise with unknown covariance matrix. In contrast to the usual assumption of deterministic target amplitudes, we assume here that the latter are drawn from a Gaussian distribution. The generalized likelihood ratio test (GLRT) is derived based on multiple primary data and a set of secondary data containing noise only. The new GLRT is shown to be the product of Kelly's GLRT and a corrective, data dependent term. We also investigate two-step approaches where the GLRT for a known disturbance covariance matrix is first derived. In order to come up with detectors that provide a good tradeoff between detection of matched signals and rejection of mismatched signals, we also investigate the two-step GLRT when a fictitious signal is included in the null hypothesis. The constant false alarm rate properties of the detectors are analyzed. Numerical simulations are presented, which show that for small sample sizes the newly-proposed GLRT can outperform Kelly's GLRT and, in addition, that detectors including a fictitious signal are very powerful, at least for low-to-intermediate clutter to noise ratio values.