Nonstationary Hidden Markov Models for Multiaspect Discriminative Feature Extraction From Radar Targets

This paper presents a new scheme for radar target recognition, in which we fuse sequential radar echoes from multiple target-radar aspect angles. The nonstationary hidden Markov model (NSHMM) is employed to characterize the sequential information contained in multiaspect radar echoes. Features from echoes are extracted via the multirelax algorithm, and moments are used to reduce the extracted-feature dimensionality. The proposed NSHMM has many parameters and states to be estimated, so the Markov chain Monte Carlo sampling algorithm is adopted. Finally, this new scheme is demonstrated with experiments on inverse synthetic aperture radar data

I. Introduction

There are many applications for radar target recognition based on the radar echo such as the high-range resolution profile (HRRP) [1]–[7]. The HRRP is a strong function of the target-radar aspect angle [8], [9]. Nevertheless, there are generally contiguous sets of orientations (aspect angles) for which the scattering physics varies slowly and the associated HRRPs can be approximated as statistically stationary signal. As shown in Fig. 1, each such set of angles is termed as a hidden “state” of the hidden Markov model (HMM) [10], because the target is distant and possibly concealed so that both the target identity and pose are unknown. The visible observations in an HMM are the radar echoes received from multiple aspects, specifically HRRPs previously mentioned.