摘要
封闭浴室场景中的跌倒行为危害性大,救助及时性差,对其高效便捷的监测研究有重要意义.针对现有基于Wi-Fi信号感知的跌倒算法中存在的特征提取不足、识别精度有限问题,本文提出了一种基于深度学习的非接触式浴室跌倒监测模型WiSFall.首先,WiSFall将一维时序CSI数据流重构为二维的频率能量图形式,使得感知数据的特征容纳能力得以增强;其次,对重构后的感知数据进行巴特沃斯滤波,以去除环境噪声;最后,将滤波后的感知数据以频率能量图形式输入到构建好的深度学习模型中,通过对感知特征的有效提取和分类实现Wi-Fi环境下高精度的浴室跌倒监测.实验结果表明,WiSFall在居家浴室环境下对跌倒行为监测的准确率达到99.63%,相比同类模型表现更好,且具有较强泛化能力.
Falling behavior in the closed bathroom scene is very harmful,and the timeliness of rescue is poor,which is of great significance to its efficient and convenient monitoring research.Aiming at the problems of insufficient feature extraction and limited recognition accuracy in existing fall algorithms based on Wi-Fi signal perception,this paper proposes a non-contact bathroom fall monitoring model WiSFall based on deep learning.First,WiSFall reconstructs the one-dimensional time series CSI data stream into a two-dimensional frequency energy graph form,which enhances the feature capacity of the sensing data;secondly,it performs Butterworth filtering on the reconstructed sensing data to remove environmental noise;Finally,the filtered perception data is input into the built deep learning model in the form of frequency energy graph,and high-precision bathroom fall monitoring in the Wi-Fi environment is realized through the effective extraction and classification of the perception features.The experimental results show that the accuracy of WiSFall′s fall behavior monitoring in the home bathroom environment reaches 99.63%,which is better than similar models and has a strong generalization ability.
作者
段鹏松
李婧馨
王超
孔金生
DUAN Peng-song;LI Jing-xin;WANG Chao;KONG Jin-sheng(School of Information Engineering,Zhengzhou University,Zhengzhou 450001,China;School of Software Engineering,Zhengzhou University,Zhengzhou 450002,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2023年第2期232-238,共7页
Journal of Chinese Computer Systems
基金
国家自然科学基金面上项目(61972092)资助
郑州市协同创新重大专项项目(20XTZX06013)资助
河南省高等学校重点科研项目(21A520043)资助.
关键词
浴室跌倒监测
信道状态信息
数据重构滤波
深度神经网络
shower room fall detection
channel state information
data reconstruction filtering
deep neural network