摘要
针对传统基于特征提取的运动识别方法很大程度上依赖研究者的领域知识和训练样本的规模问题,提出一种基于主题的特征构建方法,使用基于符号化聚合近似(SAX)的主题模型对运动模式进行建模.使用降维的方法对加速度原始信号进行预处理,结合时序数据符号化聚合近似化的方法,将SAX化后的时序数据集作为主题分析的文档集.通过主题模型进行模式挖掘,实现文档数据的降维效果,构造隐主题相关的向量,并通过建立空间向量模型(VSM)进行运动识别.实验结果表明:基于符号化聚合近似的主题分析方法可以很好地应用于运动识别,并且与传统基于特征提取的方法和基于模体发现的方法相比,活动识别率明显提升.
A topic based feature construction method was proposed to address the problem that traditional activity recognition methods based on feature extraction heavily depend on the domain knowledge of researchers and the quantity of the training dataset.Based on Symbolic Aggregate approXimation（SAX）,the proposed method employed a topic model to discover activity patterns.After preprocessed by dimensionality reduction techniques and symbolic aggregate approximation,the acceleration data were used as document set for the topic model.Pattern mining was completed through topical model to reduce the dimensions of the document data and construct the latent topic related vectors,then vector space model（VSM）was utilized to classify different activities.Results show that SAX based topic model can be well applied on activity recognition,and the proposed method is more effective to improve the recognition accuracy than feature extraction based method and motif discovery based method.
出处
《浙江大学学报（工学版）》
EI
CAS
CSCD
北大核心
2016年第6期1149-1154,共6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
传感器
符号化聚合近似
模体发现
隐含狄利克雷分配
activity recognition
sensor
Symbolic Aggregate approXimation（SAX）
motif discovery
latent Dirichlet allocation
