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 realtime 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 favourable 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, offering a practical solution for real-time HAR systems.

Radar-based Human Activity Recognition (HAR) has attracted much attention in various fields such as smart security, medical monitoring, and human computer interaction. Integrating Convolutional Neural Networks (CNNs) with radar spectrum techniques for HAR is becoming increasingly popular. However, traditional network models usually have a large number of parameters and require long training and inference times, making them less suitable for real-time applications. To address these issues, this study proposes a lightweight CNN model based on Frequency-Modulated Continuous Wave (FMCW) radar, designed for edge devices for efficient real-time monitoring. We compare three different 2D domain radar data preprocessing techniques - Time Range (TR), Short-Time Fourier Transform (STFT), and Smoothed Pseudo-Wigner-Ville Distribution (SPWVD) - along with four state-of-the-art neural networks. Our approach achieves high accuracy in HAR classification and effectively addresses the challenges posed by limited radar data through Transfer Learning (TL), demonstrating the potential for real-time applications. After evaluating 12 configurations of CNN models and preprocessing methods, we found that MobileNetV2 with STFT was the most efficient and lightweight, with STFT taking only 220 ms to generate a spectrogram sample. This combination achieved an inference time of only 2.57 ms per sample and a recognition accuracy of 96.30%, setting a new benchmark for real-time intelligent systems on edge devices.