摘要
人体活动识别(HAR)可为智慧生活、医疗监护、虚拟现实等上下文感知系统提供重要的基础信息,是模式识别领域的热门研究方向。针对现有基于惯性传感器的活动识别深度学习算法对于多维时间序列特征的提取效果欠佳的问题,提出了一种基于无阈值递归图(URP)和卷积神经网络—长短期记忆(CNN-LSTM)的活动识别算法。首先,使用SMOTE-ENN算法对惯性数据集进行增强,平衡各个类别样本数量比例;然后,使用URP方法将多维惯性传感时序波形构造为对应多个二维递归矩阵;最后,构建CNN-LSTM组合的分类模型。通过在UCI-HAR、WISDM公开数据集上的实验结果表明:所提算法在测试集上4种分类指标均得到提高,其中准确率分别达到98.32%和98.97%,性能优于现存的其他深度学习算法。
Human activity recognition(HAR)a hot research direction in the field of pattern recognition pattern recognition topic,since it can provide importanrt fundamental information in many context-awareness system such as smart life,medical surveillance and virtual reality.Aiming at the problem of poor effect in extracting multidimensional time series features based on deep learning algrithm for activity recognition of inertial sensors,an activity recognition algorithm based on the unthresholded recurrence plot(URP)and CNN-LSTM is proposed.Firstly,the inertial measurement unit(IMU)datasets are enhanced by using the SMOTE-ENN algorithm to balance sample quautities ratio of each category.Then,the multidimensional time series from IMU are constructed into multiple two-dimensional correlation matrices by using the URP approach.Finally,a combined CNN-LSTM classification model is constructed.Experimental results on two public datasets UCI-HAR and WISDM.The show that four classification indicators of the proposed algorithm are improved on the test set accuracy reaches 98.32%and 98.97%,respectively the performance is prior to other existing deep learning algorithm.
作者
史立宇
孙杨帆
谢溢翀
黄旭萍
周彪
SHI Liyu;SUN Yangfan;XIE Yichong;HUANG Xuping;ZHOU Biao(School of IoT,Jiangnan University,Wuxi 214122,China;The Affiliated Wuxi Mental Health Center of Jiangnan University,Wuxi Central Rehabilitation Hospital,Wuxi 214151,China;The 904th Hospital of the Joint Logistics Support Force of the People’s Liberation Army of China,Wuxi 214041,China)
出处
《传感器与微系统》
北大核心
2025年第3期130-133,共4页
Transducer and Microsystem Technologies
基金
国家自然科学基金青年科学基金资助项目(61901206,61703185)
无锡市太湖人才计划资助项目(WXTTP2020008,WXTTP2021)。
关键词
人体活动识别
数据增强
深度学习
无阈值递归图
卷积神经网络
长短期记忆神经网络
human activity recognition
data augmentation
deep learning
unthresholded recurrence plot
convolutional neural network
long short-term memory neural network
