摘要
针对人体动作识别问题,提出一种基于智能手机加速传感器数据并运用深度卷积神经网络进行分类识别的方法,可以有效地分类人体的走、坐、躺、跑、站五类动作.该方法模型由输入层、两层卷积层、两层池化层、一层全连接层和输出层组成,使用滑动窗口折叠法将传感器数据变换为类似于三通道的RGB图像格式,自动提取加速传感器数据的特征,对各个动作进行分类,免去了传统方法繁琐的特征提取工程.该方法在Actitracker开源数据库上达到了0.912 6的识别率,验证了该方法的可行性.
Focusing on issues of human activity recognition,a classification and recognition method was proposed,which was based on the data from accelerometer sensor of smart phone and utilizes the deep convolutional neural networks.The proposed method was able to effectively recognize 5types of human activities,including walking,sitting,lying down,jogging and standing.The model consisted of one input layer,two convolutional layers,two max-pooling layers,one fully connected layer and one output layer.The sliding-window folding method was used to transform accelerometer data into the format which was similar to three-channel RGB image.The features of the accelerometer data could be extracted automatically so as to classify each activity,avoiding tedious work of traditional feature extraction methods.The proposed method achieves accuracy of 0.9126 using Actitracker public dataset,which shows its feasibility.
作者
吴军
肖克聪
Wu Jun;Xiao Kecong(Institute of Network Technology,Beijing University of Posts and Telecommunications,Beijing 100876,China)
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2016年第S1期190-194,共5页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家国际科技合作与交流专项资助项目(2013DFE13130)
关键词
动作识别
卷积神经网络
深度学习
机器学习
加速传感器
activity recognition
convolutional neural networks
deep learning
machine learning
accelerometers
