人体行为多尺度卷积双向门控循环识别模型  

uman Activity Recognition Model Based on Multi-Scale Convolutional-Bidirectional Gated Recurrent Unit

作　　者：马鹏飞 卫芬 沈意平 吕泽强 MA Pengfei;WEI Fen;SHEN Yiping;LYU Zeqiang(Hunan Provincial Key Laboratory of Health Maintenance for Mechanical Equipment,Hunan University of Science and Technology,Xiangtan 411201,China;Laboratory of Science and Technology on Integrated Logistics Support,National University of Defense Technology,Changsha 410073,China;School of Mechanical and Electrical Engineering,Hunan University of Science and Technology,Xiangtan 411201,China)

机构地区：[1]湖南科技大学机械设备健康维护湖南省重点实验室,湖南湘潭411201 [2]国防科技大学装备综合保障技术重点实验室,湖南长沙410073 [3]湖南科技大学机电工程学院,湖南湘潭411201

出　　处：《湖南科技大学学报(自然科学版)》2024年第4期98-108,共11页Journal of Hunan University of Science And Technology:Natural Science Edition

基　　金：湖南创新型省份建设专项经费资助(2020RC3049);湖南省教育厅科学研究项目资助(20C0775,21A0310);湖南省自然科学基金青年项目资助(2023JJ40289)。

摘　　要：为实现深度挖掘人体行为传感数据中的前后关联信息,进而达到精确识别人体行为模式的目的,针对传统卷积神经网络无法充分利用行为特征之间蕴藏的时序信息,提出一种基于注意力机制的多尺度卷积神经网络(Multi-scale Attention Convolutional Neural Network,MACNN)与双向门控循环单元(Bidirectional Gated Recurrent Unit,BiGRU)的人体行为识别模型.提出一种多尺度卷积结构,拓宽网络的宽度并实现不同维度特征的提取和降维;引入双向门控循环单元学习信号在时间维度上的关联特征;通过在多尺度CNN中加入卷积注意力模块,深度挖掘人体行为信号中的关键信息特征,实现人体行为的高准确度识别.采用UCI人体行为识别公共数据集对所提方法进行试验验证,结果表明:所提方法能够准确地实现对6种典型人体日常行为的分类识别,且相比于传统的单尺度CNN识别模型,所提出的MACNN-BiGRU模型的准确率提高5%以上,达到98.40%.In order to realize the deep mining of the backward and forward relevant information in human activity sensing data,and then achieve the purpose of accurately identifying human behavior patterns,meanwhile aiming at the issue that the traditional convolutional neural networkcannot make full use of the time-series information contained in the behavioral features,a human activity recognition model based on Multi-scale Attention Convolutional Neural Network with attention mechanism(MACNN)and Bidirectional Gated Recurrent Unit(BiGRU)is proposed.Firstly,a multi-scale convolution structure is proposed to broaden the width of the network and realize the extraction and dimensionality reduction of different dimensional features.Secondly,the bi-directional gated recurrent unit is introduced to learn the correlation characteristics in the time dimension of the signal,and thirdly,by adding convolution attention module into the aforementioned MCNN,the key information features in human activity signals are deeply mined,and ultimately high-precision human activity recognition is realized.The public data set of UCI-HAR is utilized to verify the validity of the proposed recognition model.The results indicate that the proposed model can accurately classify and recognize six kinds of typical human activity.Compared with the traditional single-scale CNN,the accuracy of the proposed model is improved by more than 5%,and the recognition accuracy reaches 98.4%.

关 键 词：人体行为识别 多尺度卷积神经网络 注意力机制 双向门控循环单元 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]