Structured Time-Frequency Feature Driven Meta Learning for Fall Detection with mmWave Radar

Millimeter-wave radar has emerged as a preferred modality for privacy-preserving human activity recognition. However, the effectiveness of current deep learning approaches is often severely compromised by data scarcity and incomplete feature representation in complex, real-world environments. To address these challenges, this paper presents a novel meta-learning framework tailored for few-shot radar recognition. Structured time-frequency feature strategy is proposed, representing the systematic extraction of multi-spectral joint temporal-spectral features from raw IQ data. Unlike traditional passive mapping, this strategy utilizes structured temporal alignment to construct distinctive two-dimensional maps that actively compensate for information loss. Synergizing with this, we propose the SAMSCNN architecture, which departs from conventional layer stacking to employ a customized hierarchical multi-scale topology. This architecture innovatively integrates spatially-aware self-attention with meta-learning optimization, enabling the network to rapidly adapt to new tasks while preserving fine-grained signal structures. Experimental results demonstrate state-of-the-art performance, achieving 99.21% accuracy specifically in fall detection. Notably, under severe few-shot conditions with only five training samples, the method sustains an average accuracy of 90.24% across six distinct human activities, significantly outperforming existing methods. Extensive comparative analyses further confirm the framework’s superior robustness across varying scene configurations and environmental interference.