Deep Learning-Based High-Resolution Radar Micro-Doppler Signature Reconstruction for Enhanced Activity Recognition

Micro-Doppler signatures
$(\mu-\mathbf{DS})$
play a crucial role in activity classification using radar. However, conventional methods for
$\mu-\mathbf{DS}$
generation, such as the Short-Time Fourier Transform (STFT), suffer from several limitations, such as the resolution limit, sensitivity to noise, and the need for parameter tuning. To overcome these challenges, we introduce a novel deep learning (DL) based approach that directly reconstructs high-resolution
$\mu-\mathbf{DS}$
from 1D complex time-domain signals. Our deep learning architecture comprises three key components: an autoencoder block to enhance the signal-to-noise ratio (SNR), a Convolutional STFT block to acquire the knowledge of frequency transformations necessary for generating pseudo-spectrograms, and a UNET block for the reconstruction of high-resolution spectrogram images. We conducted evaluations of the proposed method using both synthetic and real-world datasets. In the case of synthetic data, we generated 1D complex time-domain signals with multiple time-varying frequencies and assessed the network's performance in generating high-resolution
$\mu-\mathbf{DS}$
under different SNR levels. For real-world data, A radar-based American Sign Language (ASL) dataset, consisting of 20 ASL signs are used, to assess the classification performance achieved with
$\mu-\mathbf{DS}$
generated by the proposed approach. Our results demonstrated a 3.34% increase in classification accuracy compared to traditional STFT-based
$\mu-\mathbf{DS}$
. Both synthetic and experimental
$\mu-\mathbf{DS}$
revealed that our approach effectively learns to reconstruct higher-resolution and sparser spectrograms, showcasing its potential for improving radar-based activity recognition applications.