HRSpecNET: A Deep Learning-Based High-Resolution Radar Micro-Doppler Signature Reconstruction for Improved HAR Classification

Abstract

Micro-Doppler signatures ( $\mu $ -DSs) are widely used for human activity recognition (HAR) using radar. However, traditional methods for generating $\mu $ -DS, such as the short-time Fourier transform (STFT), suffer from limitations, such as the tradeoff between time and frequency resolution, noise sensitivity, and parameter calibration. To address these limitations, we propose a novel deep learning (DL)-based approach to reconstruct high-resolution $\mu $ -DS directly from a 1-D complex time-domain signal. Our DL architecture consists of an autoencoder (AE) block to improve signal-to-noise ratio (SNR), an STFT block to learn frequency transformations to generate pseudo spectrograms, and, finally, a U-Net block to reconstruct high-resolution spectrogram images. We evaluated our proposed architecture on both synthetic and real-world data. For synthetic data, we generated 1-D complex time-domain signals with multiple time-varying frequencies and evaluated and compared the ability of our network to generate high-resolution $\mu $ -DS and perform in different SNR levels. For real-world data, a challenging radar-based American Sign Language (ASL) dataset consisting of 100 words was used to evaluate the classification performance achieved using the $\mu $ -DS generated by the proposed approach. The results showed that the proposed approach outperforms the classification accuracy of traditional STFT-based $\mu $ -DS by 3.48%. Both synthetic and experimental $\mu $ -DSs show that the proposed approach learns to reconstruct higher resolution and sparser spectrograms.