Enhancing FMCW Radar Gesture Classification With Physically Interpretable Data Augmentation

This study introduces a novel, physically interpretable data augmentation framework that improves the robustness and accuracy of hand gesture recognition using Frequency-Modulated Continuous Wave (FMCW) radar and Convolutional Neural Networks (CNN). The proposed reconfigurable and parametric method modifies specific characteristics of five time-series features, namely range, velocity, azimuth and elevation angles, and signal magnitude, to generate synthetic gesture samples with realistic variations. By simulating variations in hand gesture distance, angle, duration, and noise, the framework improves model generalization while reducing the need for extensive and costly data collection. The augmentation techniques employed in this research include time scaling, range and angle transformation, and noise injection, effectively simulating different gesture speeds, orientations, distances, and interference levels. Additionally, applying augmentation at the feature level, rather than on raw radar data, reduces data size significantly, leading to faster training, lower memory requirements, and improved scalability. Experimental results demonstrate significant performance gains in a 1D-CNN classifier deployed on an ARM Cortex-M4 microcontroller after applying the proposed augmentation techniques. Specifically, the combination of time scaling, range transformation, and noise injection improves accuracy by 16.67%, precision by 6.4%, recall by 15.07%, and F1 score by 13.12% in the extended range scenario, 1 meter to 1.2 meter, compared to the baseline model trained on unaugmented data.