Sparse Bayesian Network for Fast Micro-Doppler Analysis

Abstract

Micro-Doppler Analysis (MDA) of rigid-body targets is crucial for various practical downstream tasks such as target imaging and recognition. Radar echoes from micro-moving targets typically represent non-stationary signals and are often described using the parameterized Time-Varying Auto Regressive (TVAR) model. Sparse Bayesian Learning (SBL) is commonly employed to estimate time-invariant coefficients, thereby achieving high-resolution micro-Doppler time-frequency distributions. Despite its effectiveness, SBL-based optimization methods often face inefficiency due to the computational burden of inverse operation. To address this challenge and enhance the efficiency of MDA based on TVAR model, we propose a Sparse Bayesian Network (SBN) that unfolds a fast Mean Field SBL (MF) using a deep variational autoencoding framework. This method retains the optimization effectiveness of SBL while incorporating the inference efficiency of Deep Neural Networks (DNNs). Our proposed method demonstrates strong generalization capabilities, performing well on both simulated and measured radar echoes.