摘要
基于惯性测量的人员行为识别(human activity recognition,HAR)在智能运动、智能家居、医疗健康领域有着较广泛的应用前景,是当前的研究热点。提出了一种基于空间变换网络(spatial transformer network,STN)的人员行为识别方法,该方法在传统的卷积神经网络(convolutional neural network,CNN)中加入空间变换单元,使得网络对输入数据的时间平移、频率变化等因素不敏感,从而提高识别稳健性。在方法实现中,根据人员惯性数据的特点,对仿射变换的参数模型进行了改进,使其能够适应相同行为类别中惯性数据的样本变换,包含时间的平移、频率的变化等。开源数据集的试验结果表明,相比于传统的卷积神经网络和循环神经网络(recurrent neural network,RNN)的方法,提出方法的识别率分别提高了约5.1%和3.4%,网络训练收敛时间降低了2.4%和30.8%。
Inertial measurement based human activity recognition(HAR)has a great prospect in smart sports,smart homes,medical and health care,and it is a hotspot in current research.In this paper,a HAR method based on spatial transformer network(STN)is proposed.In the method,a spatial transformer is added to the traditioanal convolutional neural network(CNN),which can render the network insensitive to time shift and frequency change in input samples.The robustness of the proposed method is thus improved.In the implementation of the method,the affine transformation parameter model is improved according to the nature of human inertial measurements,so that the model is adaptable to the sample transformation of inertia data in the same activity category,including time shift,frequency changes and so on.The experimental results of open-source data show that compared with traditional convolutional neural network and recurrent neural network(RNN),the recognition rate of the proposed method increases by 5.1%and 3.4%,respectively;and the training convergence time decreases by 2.4%and 30.8%,respectively.
作者
袁帆
YUAN Fan(School of Electrical Engineering,Shaanxi University of Technology,Hanzhong 723001,China)
出处
《武汉大学学报（工学版）》
CAS
CSCD
北大核心
2022年第7期740-746,共7页
Engineering Journal of Wuhan University
基金
陕西理工大学科研基金项目(编号:SLGKY2017-18)。
关键词
传感器数据
人员行为识别
深度学习
空间变换网络
sensor data
human activity recognition
deep learning
special transformer network
