Detecting a signal in the presence of noise with an unknown covariance matrix is the problem considered in this paper. Although it considers the particular application of radar space-time adaptive processing (STAP), the problem occurs in a variety of fields, in both single and multi-dimensional signal processing. The author’s choice of the STAP application allows him to customize two previously known approaches, with a modest improvement in results.
The two previously known approaches that the author builds on are the adaptive matched filter (AMF) and the adaptive cosine/coherence estimator (ACE) methods. Both of these methods are based on estimating the covariance matrix of the noise from a limited number of noise samples, collected in the absence of the signal being detected.
The primary contribution of this paper is the forcing of a specific matrix structure on the noise covariance matrix. The result is that the matrix is non-singular, even when the number of samples available to estimate the covariance matrix is less than the size of the noise vector. Since the matrix structure chosen is itself valid for realistic STAP scenarios, this represents a customization of the approach for STAP. The authors use simulations to demonstrate that, assuming this forced covariance matrix structure for the noise is valid, their approach outperforms AMF and ACE, when the number of noise samples are small.
Quite meaningful, modest improvements are presented in this paper, in view of the physical basis for the author’s choice of matrix structure for the noise covariance matrix. However, the paper lacks an evaluation of the performance of the algorithm when the forced matrix structure is not valid. Furthermore, the authors consider a simplistic signal (uniform) for their simulations, which is not representative of target echoes from real world targets.