基于改进的DCNN人体行为识别  

Human behavior recognition based on improved DCNN

作　　者：周鹏 袁国良[1] 张颖[1] 孙莉 ZHOU Peng;YUAN Guoliang;ZHANG Ying;SUN Li(College of Information Engineering,Shanghai Maritime University,Shanghai 201306,China)

机构地区：[1]上海海事大学信息工程学院,上海201306

出　　处：《传感器与微系统》2021年第10期125-128,共4页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(61273068);国家自然科学基金青年科学基金资助项目(61901255)。

摘　　要：在基于可穿戴传感器的人体行为识别领域中,提取原始数据的有效特征和建立合适的分类模型是提高识别准确率的关键。针对上述问题,提出一种改进的深度卷积神经网络(DCNN)模型,在经典的DCNN模型中增加了信号融合单元,并提出一种将时间序列转换成单通道行为图片的方法,由加速度、角速度和俯仰角信号构成的行为图片在经过信号融合单元处理后,可实现跨通道的信息融合,然后提取行为图片的张量特征,实现对行走、奔跑、坐下、躺下、跌倒、跳跃共6种日常行为的识别。实验表明:该方法在UCI开源数据集上的识别率达到97.05%,高于传统分类模型的识别率。In the field of human behavior recognition based on wearable sensors,extracting effective features of original data and establishing suitable classification model are the keys to improve recognition accuracy.In view of the above problems,an improved deep convolutional neural network(DCNN)model is proposed.This method adds signal fusion unit to the classic DCNN model,and propose a method for converting time series into single-channel behavior picture.After being processed by signal fusion unit,the behavior pictures composed of the acceleration signal,angular velocity signal,and pitch angle signal can realize cross-channel information fusion,and then extract the tensor characteristics of the behavior picture to achieve the recognition of six kinds of daily behaviors like walking,running,sitting,lying,falling and jumping.Experiments show that the recognition rate of this method on the UCI open source dataset reaches 97.05%,which is higher than that of traditional classification models.

关 键 词：行为识别 信号融合 深度卷积神经网络(DCNN) 融合深度卷积神经网络(F-DCNN) 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP212[自动化与计算机技术—计算机科学与技术]