Comparative Analysis of Time-Frequency Transformations and Deep Learning Architectures in the Classification of FMCW Radar Micro-Doppler Signatures

Frequency Modulated Continuous Wave (FMCW) radars present a robust and privacy-friendly alternative for human activity recognition (HAR) systems. In the classification of micro-Doppler signatures obtained from these radars via deep learning algorithms, the structural quality of the time-frequency (TF) transformation used as input directly affects the model's performance. Therefore, the explicit objective of this study is to design and optimize lightweight Convolutional Neural Network (CNN) architectures capable of efficiently extracting micro-Doppler features from various TF representations under the constraints of limited radar data. To achieve this, signals belonging to 6 different human activities collected by an FMCW radar were converted into spectrograms using Short-Time Fourier Transform (STFT), Continuous Wavelet Transform (CWT), and Wigner-Ville Distribution (WVD). Three novel convolutional neural network architectures (CNN–Base, CNN–Wide, CNN–LSTM), capable of providing different responses to the spatial features of these transformations, were designed and analyzed using a 6-fold cross-validation strategy. Experimental results demonstrated that the logarithmic scaling of CWT misled CNN filters, while the standard STFT provided a strong baseline with 80.00% accuracy. The highest performance of the study was achieved at 80.42% with the CNN–Wide architecture, which successfully processed the high-resolution texture of WVD containing cross-terms utilizing its wide receptive field. The WVD-based model identified life-critical falling activities with 100% precision, offering a promising approach for elderly care and autonomous health monitoring systems.