基于WGAN-GP和Mean Teacher的WiFi使能跨域人体行为识别  

WiFi-enabled Cross-domain Human Activity Recognition Based onWGAN-GP and Mean Teacher

作　　者：史心玥 夏文超 赵海涛[1] 杨丽花 阮欣雨 常天水 SHI Xinyue;XIA Wenchao;ZHAO Haitao;YANG Lihua;RUAN Xinyu;CHANG Tianshui(School of Communications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;School ofInternet of Things,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003 [2]南京邮电大学物联网学院,江苏南京210003

出　　处：《无线电通信技术》2024年第6期1192-1199,共8页Radio Communications Technology

基　　金：国家自然科学基金面上项目(62371250);国家自然科学基金青年基金项目(62201285);江苏省基础研究计划(自然科学基金)前沿引领技术基础研究专项(BK20212001);江苏省自然科学基金-杰出青年基金项目(BK20220054)。

摘　　要：人体行为识别(Human Activity Recognition,HAR)是当前众多研究工作的基石,对于推动人机交互和智能数字化转型具有巨大潜力。由于目标域样本较难采集,现有方法在跨域识别方面表现不佳。为解决这一问题,提出一种新的WiFi使能跨域HAR方法,从WiFi信号中获取信道状态信息(Channel State Information,CSI)并转化为图像,在基于Wasserstein距离和梯度的生成对抗网络(Wasserstein Generative Adversarial Network with Gradient Penalty,WGAN-GP)中引入双判别器,通过与源域样本和单目标域样本特征联合对抗,生成同时带有双域特征的虚拟样本。该方法还结合基于Mean Teacher的半监督学习设计识别分类(Recognition and Classification,RC)模块,通过对有标记样本与无标记样本分别构造损失函数,进行整体一致性损失的评估,实现对目标域样本的识别。实验结果证明了所提方法能够在减轻目标域样本采集压力的同时,实现较高的检测精度,在手势与动作的数据集上测试准确率分别达到92.71%和86.65%。Human Activity Recognition(HAR)serves as a cornerstone in numerous ongoing research endeavors,holding substantial promise in human-computer interaction and the realm of intelligent digital transformation.However,existing methods perform poorly in cross-domain recognition due to the fact that target domain samples are more difficult to collect.To address this problem,this research proposes a new WiFi-enabled cross-domain human activity recognition method.This method obtains Channel State Information(CSI)from WiFi signals and transforms it into images,and introduces a dual discriminator based on the Wasserstein Generative Adversarial Network with Gradient Penalty(WGAN-GP),which generates virtual samples with simultaneous dual-domain features by jointly competing with source domain samples and single target domain sample features.The method also incorporates Mean Teacher-based semi-supervised learning to design the Recognition and Classification(RC)module,which realizes the recognition of target domain samples by constructing the loss function for labeled and unlabeled samples separately and performing the evaluation of overall consistency loss.Experimental results demonstrate that the proposed method is able to achieve high detection accuracy while alleviating the pressure of target domain sample collection,and the test accuracy can reach 92.71% and 86.65% on the datasets of gestures and actions,respectively.

关 键 词：人体行为识别 生成对抗网络 Mean Teacher模型 跨域识别 

分 类 号：TN92[电子电信—通信与信息系统]