A Lightweight Knowledge Distillation Framework for Multi-Target Action Recognition
作者
Z. Zhang,Z. Chen,Z. Xu,Z. Zhou,Z. Chen
出处
期刊:日期:2025-11-14卷期号:: 1-3
标识
DOI:10.1109/csrswtc67757.2025.11383988
摘要
Fast and accurate multi-target action recognition using Frequency Modulated Continuous Wave (FMCW) radar is a significant challenge for real-time applications. To address this, we propose a lightweight dual-input network named Quickly Identify Multi-Target Action Network (QIMTA). Built upon the MobileViT2 architecture, our model is trained via a knowledge distillation framework, utilizing a Swin Transformer as the teacher model. This design enables the parallel recognition of separated Range-Doppler maps, significantly enhancing computational efficiency. Experimental results demonstrate that QIMTA achieves 93.49% recognition accuracy with an inference speed of only 5.62 ms, demonstrating its high efficiency and effectiveness for deployment on edge devices.