Hand gesture recognition method based on mmWave radar with MobileViT and knowledge distillation
Xiangqun Zhang, Zhizhou Ge, Kai Lü, Genyuan Du, Jiawen Shen, Xiangqian Gao · 2025
Human-hand gesture recognition using millimetre wave radar is attractive in human-computer interfaces, industrial Internet of Things, and smart home. However, the existing CNN or RNN model is so complex and large that it is hard to apply to mobile vision tasks and embedded devices. This paper proposes dual lightweight convolutional neural networks, MobileViT and knowledge distillation, to solve this problem. Firstly, we acquire the original radar echoes of frequencymodulated continuous wave (FWCW) signal and reshape the three-dimensional matrix by Chirps * Samples * Frames. Then, we perform the signal processing method to eliminate the noise and static background. Secondly, we employ the Fast Fourier Transform to extract the hand gestures feature map of distance and Doppler information. Finally, the feature map is input into the modified MobileViT to recognize dynamic hand gestures, and knowledge distillation models is used to simplify the model structure. The experimental results show that the parameter space complexity of the constructed model is reduced to 0.018 M, and the computational complexity is 0.082 GFLOPs after knowledge distillation model. The method is verified in 12 complex gesture. It has the potential to be applied in mobile tasks and embedded devices.