A Bayesian denoising mehod for complex radar signal with application to classification of human individuals

We propose a novel noise-robust classification scheme to discriminate one walking person and two walking persons. Non-parametric Bayesian techniques with the Beta process prior are applied to the Principal Component Analysis (PCA) model for the automatic choice of the number of principal components. Noise reduction is accomplished via reconstructing the echo within the subspace composed of the selected principal components. Then a 2-dimensional feature vector is extracted from the time frequency (TF) spectrograms of the low resolution radar echoes to depict the different micro-Doppler characteristics of one walking person and two walking persons. Compared with the existing CLEAN based noise reduction method, the new method can work without SNR prior information. In the experiments based on measured data, the proposed method shows the good noise reduction and classification performance even under the relative low Signal-Noise-Ratio (SNR) cases.