STGR: Self-Training Cross-Domain Air-Writing Gesture Recognition Using FMCW Radar
Jiake Tian, Yi Zou, Jiale Lai, Dacheng Li · 2025
In the field of human-computer interaction within the Internet of Things (IoT), radar-based gesture recognition has garnered significant attention due to its non-contact nature and inherent robustness to variations in lighting, privacy concerns, and health risks. However, its cross-domain performance is often hindered by challenges such as inconsistencies in radar parameters, environmental noises, and gesture heterogeneity. To address these limitations, we propose STGR, a self-training framework for cross-domain recognition of air-writing gestures using Frequency-Modulated Continuous Wave (FMCW) radar. STGR begins by capturing raw signal data via millimeter wave (mmWave) FMCW radar and constructing an interference-resilient dataset of air-writing trajectories, which effectively mitigates the distance-angle parameter mismatches commonly encountered in conventional gesture recognition systems. We then design a high-precision recognition model by combining depthwise separable convolutions with a multi-head attention mechanism to enhance spatial-temporal feature extraction. To improve cross-domain generalization, a self-training optimization strategy is introduced, which leverages a small amount of labeled data to reduce dependency on fully annotated cross-domain datasets. Experimental evaluations on both the MNIST benchmark and our proprietary dataset demonstrate that STGR achieves a recognition accuracy of 98.97% and exhibits strong cross-domain robustness with an accuracy of 98.25%.