Benchmarking Radar Preprocessing Techniques and Transfer Learning Models for FMCW-based Human Activity Recognition

Human Activity Recognition (HAR) using radar signals has gained significant attention due to its non-intrusive nature and robustness in various environments. However, the impact of radar signal preprocessing techniques on the performance of deep learning (DL) models remains an active area of research. This study investigates how different radar domain representations affect HAR accuracy by evaluating four preprocessing methods: Time-Range (TR) maps generated via Range-Fast Fourier Transform (FFT), Range-Doppler (RD) maps obtained through sequential FFTs, and Time-Doppler (TD) features extracted using Short Time Fourier Transform (STFT) and Smoothed Pseudo Wigner Ville Distribution (SPWVD). We employ a baseline Convolutional Neural Network (CNN) and state-of-the-art Transfer Learning (TL) models to assess whether advanced preprocessing or increased model complexity yields greater performance gains. The results reveal that high-resolution TD analysis using SPWVD does not significantly enhance classification performance and incurs substantial computational overhead, limiting its real-time applicability. Conversely, the TR representation offers computational efficiency but struggles to classify complex activities with the baseline CNN accurately. RD and STFT methods provide a favorable balance between classification accuracy and computational efficiency. Notably, transitioning from the baseline CNN to TL models leads to substantial improvements in recognition accuracy: up to 29.36% for TR, 21.42% for RD, 16.66% for STFT, and 11.11% for SPWVD representations. Overall, our findings demonstrate that TL models, when combined with computationally efficient radar preprocessing techniques like RD or STFT, significantly improve recognition accuracy and generalize well across datasets, as confirmed by evaluation on two publicly available radar-based HAR datasets. Among these, the RD representation combined with VGG-19 yielded the best trade-off between accuracy and latency, achieving a total processing time of 0.91 s per sample for a 10 s activity duration, making it highly suitable for latency-sensitive HAR applications.