%0 Journal Article
%T 采用运动传感器的人体运动识别深度模型<br>A Deep Learning Model for Human Activity Recognition Using Motion Sensor Data
%A 滕千礼
%A A.ESMAEILI KELISHOMI
%A 蔡忠闽
%J 西安交通大学学报
%D 2018
%R 10.7652/xjtuxb201808010
%X 针对传统机器学习方法在采用运动传感器数据的人体运动识别领域中识别效果严重依赖人工特征且准确率受限的问题,提出一种改进的卷积网络与双层长短期记忆网络的深层混合(VGG？？LSTM)模型以实现特征自提取并进行运动识别。该模型结合传感器数据层状、时序的结构特点,将多维传感器数据类比于图像的RGB矩阵进行适应性处理;由一维串联卷积网络与双层长短期记忆网络复合而成。实验结果表明,在开源的人体运动识别(HAR)数据集和无线传感器信息控掘(WISDM)数据集上采用该模型的人体运动识别方法的平均准确率分别达到了97.17％和96.53％,该模型可以有效避免复杂的特征工程,在人体运动识别问题中具有很好的准确性和适应性。<br>An improved deep hybrid HAR model (VGG？？LSTM) is proposed to solve the problem that the performances of traditional machine learning methods for human activity recognition (HAR) using motion sensor data are heavily dependent on artificial features, and their accuracies are limited. The model combines a convolutional neural network and a long？？short？？term？？memory network to automatically extract features and to effectively recognize different activities. The proposed method combines multiple modified 1D？？convolutional neural networks and two layers of long？？short？？term？？memory networks by comparing the stratified and time？？series structure of sensor data with RGB matrix of image. Experiments show that the model achieves average accuracies of 97.17％ on the benchmark dataset HAR and 96.53％ on the dataset WISDM, and that it effectively avoids complex feature engineering and has a good accuracy in the human activity recognition
%K 运动传感器
%K 人体运动识别
%K 特征自提取
%K 深层混合模型&lt
%K br&gt
%K motion sensor
%K human activity recognition
%K automatic features extraction
%K deep hybrid model
%U http://zkxb.xjtu.edu.cn/oa/DArticle.aspx?type=view&id=201808010