物联网学报 ›› 2023, Vol. 7 ›› Issue (4): 153-167.doi: 10.11959/j.issn.2096-3750.2023.00360

• 理论与技术 • 上一篇    

基于CNN-BiGRU的复杂连续人体活动Wi-Fi感知方法

刘洋1,2, 董安明1,2,3, 禹继国3,4, 赵恺5, 周酉6   

  1. 1 齐鲁工业大学（山东省科学院）山东省计算中心（国家超级计算济南中心）算力互联网与信息安全教育部重点实验室，山东 济南 250353
    2 齐鲁工业大学（山东省科学院）计算机科学与技术学院，山东 济南 250353
    3 齐鲁工业大学（山东省科学院）大数据研究院，山东 济南 250353
    4 山东省基础科学研究中心（计算机科学）山东省计算机网络重点实验室，山东 济南 250353
    5 中国科学院自动化研究所，北京 100190
    6 山东海看新媒体研究院有限公司，山东 济南 250013
  • 修回日期:2023-07-01 出版日期:2023-12-01 发布日期:2023-12-01
  • 作者简介:刘洋（1996- ），男，齐鲁工业大学（山东省科学院）计算机科学与技术学院硕士生，主要研究方向为深度学习、Wi-Fi感知等
    董安明（1982- ），男，博士，齐鲁工业大学（山东省科学院）副教授，主要研究方向为通信信号处理、MIMO 无线通信、机器学习、智能物联网等
    禹继国（1972- ），男，博士，齐鲁工业大学（山东省科学院）教授，主要研究方向为智能感知、无线网络与通信、网络与数据安全及隐私保护、区块链、分布式计算等
    赵恺（1984- ），男，博士，中国科学院自动化研究所副研究员，主要研究方向为深度学习、决策智能、自主作业机器人等
    周酉（1988- ），男，博士，山东海看新媒体研究院有限公司高级工程师，主要研究方向为多媒体智能信息处理、机器学习、大数据分析等
  • 基金资助:
    国家重点研发计划(2019YFB2102600);国家自然科学基金资助项目(61701269);国家自然科学基金资助项目(62272256);山东省科技型中小企业创新能力提升工程(2022TSGC2180);山东省科技型中小企业创新能力提升工程(2022TSGC2123);济南市“高校20条”自主培养创新团队项目(202228093);齐鲁工业大学（山东省科学院）科教产融合试点工程项目（基础研究类）先导项目(2022XD001);齐鲁工业大学（山东省科学院）计算机科学与技术学科基础研究加强计划(2021JC02014)

A Wi-Fi sensing method for complex continuous human activities based on CNN-BiGRU

Yang LIU1,2, Anming DONG1,2,3, Jiguo YU3,4, Kai ZHAO5, You ZHOU6   

  1. 1 Key Laboratory of Computing Power Network and Information Security, Ministry of Education, Shandong Computer Science Center (National Supercomputer Center in Jinan), Qilu University of Technology (Shandong Academy of Sciences), Jinan 250353, China
    2 School of Computer Science and Technology, Qilu University of Technology (Shandong Academy of Sciences), Jinan 250353, China
    3 Big Data Research Institute, Qilu University of Technology (Shandong Academy of Sciences), Jinan 250353, China
    4 Shandong Provincial Key Laboratory of Computer Networks, Shandong Fundamental Research Center for Computer Science, Jinan 250353, China
    5 Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China
    6 Shandong HiCon New Media Research Institute Co., Ltd., Jinan 250013, China
  • Revised:2023-07-01 Online:2023-12-01 Published:2023-12-01
  • Supported by:
    The National Key Research and Development Program(2019YFB2102600);The National Natural Science Foundation of China(61701269);The National Natural Science Foundation of China(62272256);The Innovation Capability Enhancement Program for Small and Medium-sized Technological Enterprises of Shandong Province(2022TSGC2180);The Innovation Capability Enhancement Program for Small and Medium-sized Technological Enterprises of Shandong Province(2022TSGC2123);The Innovation Team Cultivating Program of Jinan(202228093);The Pilot Engineering Project of Science, Education and Industry Integration (Basic Research) of Qilu University of Technology (Shandong Academy of Sciences)(2022XD001);The Basic Research Strengthening Program of Computer Science and Technology Discipline of Qilu University of Technology(2021JC02014)

在线阅读

19

PDF下载

86

摘要：

基于Wi-Fi信道状态信息（CSI, channel state information）的人体活动感知在虚拟现实、智能游戏、元宇宙等未来智能交互场景具有重要的应用前景，复杂连续人体活动的精准感知是Wi-Fi感知的重要挑战。卷积神经网络（CNN, convolutional neural network）具备空间特征提取能力，但对数据的时序特征建模能力差。而适用于时间序列数据建模的长短期记忆（LSTM, long short-term memory）网络或门控循环单元（GRU, gated recurrent unit）网络忽视了对数据空间特征的学习。针对此问题，提出了一种融合双向门控循环单元（BiGRU, bidirectional gated recurrent unit）网络的改进型 CNN。所提网络利用 BiGRU的双向特征提取能力捕捉时序数据前后信息的关联和依赖性，实现时序CSI数据的时空特征提取，进而呈现动作与CSI数据的映射关系，从而提高对复杂连续动作的识别精度。以篮球动作为场景对所提网络结构进行了实验，结果表明，该方法在多种条件下识别准确率均高于95%，与传统多层感知机（MLP, multi-layer perceptron）、CNN、LSTM、GRU、具有注意力机制的双向长短期记忆（ABLSTM, attention based bidirectional long short-term memory）网络等基线方法相比，识别准确率提升了1％～20％。

关键词: 信道状态信息, 人体活动感知, 复杂连续活动, 卷积神经网络, 双向门控循环单元

Abstract:

Human activity sensing based on Wi-Fi channel state information (CSI) has an important application prospect in future intelligent interaction scenarios such as virtual reality, intelligent games, and the metaverse.Accurate sensing of complex and continuous human activities is an important challenge for Wi-Fi sensing.Convolutional neural network (CNN) has the ability of spatial feature extraction but is poor at modeling the temporal features of the data.While long short-term memory (LSTM) network or gated recurrent unit (GRU) network, which are suitable for modeling time-series data, neglect learning spatial features of data.In order to solve this problem, an improved CNN that integrates bidirectional gated recurrent unit (BiGRU) network was proposed.The bi-directional feature extraction ability of BiGRU was used to capture the correlation and dependence of the front and back information in the time series data.The extraction of the spatiotemporal features of the time series CSI data was realized, and then the mapping relationship between the action and the CSI data was present.Thus the recognition accuracy of the complex continuous action was improved.The proposed network structure was tested with basketball actions.The results show that the recognition accuracy of this method is above 95% under various conditions.Compared with the traditional multi-layer perceptron (MLP), CNN, LSTM, GRU, and attention based bidirectional long short-term memory (ABLSTM) baseline methods, the recognition accuracy has been improved by 1%～20%.

Key words: channel state information, human activity sensing, complex continuous action, convolutional neural network, bidirectional gated recurrent unit

中图分类号: 

No Suggested Reading articles found!