摘要
针对传统基于加速度传感器的运动识别方法依赖于传感设备的放置方式和位置的问题,提出一种基于深度学习的运动识别方法,且与放置方式和位置无关.使用栈式自动编码器构建深度网络,结合逐层无监督学习和全局有监督微调的方式,快速、有效地学习出原始数据的深层特征.设计不同放置方式和不同设备放置位置的学习策略,并利用所学特征对不同设备放置方式和位置下的运动进行识别.实验结果表明:基于深度学习的方法可以从原始数据中提取出与放置方式和位置无关的深度特征,相比传统方法,能够有效提高在非固定加速度传感设备放置方式和位置下的运动识别准确率.
An activity recognition method based on deep learning and independent of device orientation and placement was proposed,to address the problem that traditional acceleration activity recognition methods usually depend on the fixed device orientations and placements.Based on a layer-wise unsupervised learning and a global supervised fine-tuning,deep neural networks were constructed by stacked autoencoder.Then,the deep features of the proposed method were efficiently and effectively extracted from original acceleration data.Finally,a cross orientation and placement evaluation strategy was presented to recognize the activities under different device orientations and placements.Experimental results show that the proposed method can extract discriminative deep features from original data and achieve better performance than other methods under the condition of uncontrolled acceleration device orientation and placement.
出处
《浙江大学学报（工学版）》
EI
CAS
CSCD
北大核心
2016年第6期1141-1148,共8页
Journal of Zhejiang University：Engineering Science
基金
国家自然科学基金资助项目(60703040
61332017)
浙江省科技计划资助项目(2011C13042
2015C33002)
关键词
运动识别
传感设备位置
加速度传感设备
深度学习
普适计算
activity recognition
sensor device placement
acceleration sensor device
deep learning
pervasive computing
