【摘要】 基于智能手机传感器的人体活动识别是普适计算领域的研究热点.为扩展可识别的活动种类,并提高准确率和实时性,提出了由智能手环和智能手机组建无线体域网通过深度神经网络在线识别人体活动的方法.首先,设计由智能手环和智能手机组成的无线体域网的总体框架;然后,对预处理后的传感信号,构造带有Inception结构的卷积神经网络和长短时记忆递归神经网络来分别提取时空域特征,并结合两类网络结构来融合多模态传感数据,离线进行神经网络模型训练;最后,对训练好的神经网络模型进行优化,并部署到智能手机上,在线实时识别人体活动.实验结果表明,本文方法无需手工设计特征,可自动融合各类异构传感数据,更加准确、高效地识别了更多种类的活动.更多还原

【Abstract】 Nowadays human activity recognition( HAR) based on smartphone-mounted sensors has become a hot topic in ubiquitous computing. In order to improve its ability to classify different activities,and to improve the accuracy and efficiency,deep neural networks were proposed to classify human activities based on wireless body area network consist of smartphone and smart wristband.First,the overall structure of the wireless body area network was designed; Second,after the rawsensor data being preprocessed,a convolutional neural network( CNN) was designed to extract its local spatial features,and a long short-term memory( LSTM) network to extract its temporal features. Then,the heterogeneous sensor data was fused by joint training of these two neural networks. Finally,the trained neural network model was optimized before deploying it to the smartphone and recognizing human activities in real time. Experimental results showthat the proposed scheme has improved efficiency and accuracy when classify more human activities,and can fuse heterogeneous sensors automatically,without need to handcraft features.更多还原