Human activity recognition (HAR) using frequency-modulated continuous wave (FMCW) millimeter-wave radar is a promising alternative to wearable and vision-based systems due to its unobtrusive and privacy-preserving nature. However, modeling multi-dimensional radar data under limited training samples while remaining robust to user and environmental variations is challenging, particularly for edge-based applications. To address this challenge, we propose a lightweight artificial intelligence-based framework for FMCW radar-based HAR that enables accurate and computationally efficient activity recognition on edge devices. The framework processes radar-derived Range-Doppler, Range-Azimuth, and Range-Elevation feature maps as structured multi-dimensional data vectors rather than conventional two-dimensional images, allowing compact representation of motion dynamics and spatial relationships. A lightweight deep learning architecture combining a modified ResNet-18 with depthwise separable convolutions and a bidirectional long short-term memory module is employed to extract spatial–temporal features with reduced complexity. To improve generalization under limited data conditions, we used data augmentation strategies including spatial shifting, intensity scaling with bias shift, horizontal Doppler flipping, and additive Gaussian noise. The framework is evaluated on a newly collected 60 GHz FMCW radar dataset covering seven daily activities in a realistic home-like environment. Experiments using cross-scene and leave-one-person-out validation demonstrate superior performance over baseline methods, achieving up to 91.98% accuracy and 89.82% F1-score.
Fard et al. (Sun,) studied this question.