Hand gesture recognition using mmWave radar has emerged as a promising technology for human–computer interaction (HCI), smart home systems, and the Internet of Things (IoT). However, the practical application of this technology is often constrained by the high computational complexity and significant storage demands of contemporary deep neural networks, which impede their deployment on resource‐limited embedded devices. To address this limitation, we present a novel approach that combines an improved MobileViT model with a knowledge distillation (KD) framework. The proposed method consists of three main stages. First, raw radar signals are captured and restructured into a three‐dimensional format (Chirps × Samples × Frames, a 3D tensor) and processed to generate range‐time maps (RTMs) and Doppler‐time maps (DTMs). Second, an improved MobileViT network is designed, incorporating fewer redundant blocks, a lower input resolution, and a dual‐branch input structure to effectively fuse features from the RTM and DTM. This enhanced architecture serves as a robust teacher model, excelling at extracting both local and global spatiotemporal features for accurate gesture recognition. Finally, KD is applied to transfer knowledge from the teacher model to a compact student network, thereby achieving model compression. Experimental results demonstrate that the final distilled student model, evaluated on the test set, has only 0.018 M parameters—~10% of the teacher model’s size—while still achieving a high recognition accuracy of 99.16%. Consequently, the resulting model is highly compact and accurate, demonstrating its suitability for real‐world embedded deployment.
Zhang et al. (Thu,) studied this question.