无线电工程2025，Vol.55Issue(8)：1598-1606,9.DOI:10.3969/j.issn.1003-3106.2025.08.006
基于时频多尺度融合的CSI双人动作识别方法
CSI Two-person Motion Recognition Method Based on Time-Frequency Multi-scale Fusion
摘要
Abstract
To address the challenge of Human Activity Recognition(HAR)based on WiFi Channel State Information(CSI)in multi-person environments,where motion occlusion and overlapping lead to recognition difficulties,methods to enhance recognition accuracy and robustness in complex scenarios are explored.A Time-Frequency Multi-Scale Pyramid Network(TFMS-Net)model is proposed,which employs a frequency-domain enhancement module to perform multi-channel filtering in the FFT domain and introduces a gating mechanism to enhance signal anti-interference capabilities.It utilizes a dynamic spatial-temporal convolution module combined with a dual gating mechanism to capture motion features,while adopting multi-scale dilated convolutions to mine interactive information.Additionally,it uses a channel segmentation strategy to process sub-signals in parallel and achieves individual motion feature separation through hierarchical down-sampling.Experiments in indoor multi-person activity scenarios show that the model achieves a recognition accuracy of over 92%for complex motions,significantly enhancing the robustness of WiFi CSI-based HAR technology in multi-person scenarios and providing an effective solution for non-contact multi-person motion recognition.关键词
深度学习/多人行为识别/时域卷积/信道状态信息Key words
deep learning/multi-person behavior recognition/time-domain convolution/CSI分类
信息技术与安全科学引用本文复制引用
张亚军,张涛,李峰,宋长军..基于时频多尺度融合的CSI双人动作识别方法[J].无线电工程,2025,55(8):1598-1606,9.基金项目
新疆维吾尔自治区自然科学基金(2022D01C54) （2022D01C54）
新疆大学博士研究启动基金(202212120001) Xinjiang Uygur Autonomous Region Natural Science Foundation of China(2022D01C54) （202212120001）
Doctoral Research Initiation Foundation of Xinjiang University(202212120001) （202212120001）
