CASIA OpenIR  > 毕业生  > 硕士学位论文
基于肌电信号的身份识别技术研究
卢立静
2022-05-17
页数94
学位类型硕士
中文摘要

伴随着互联网的日益发展,个人信息的安全受到人们越来越多的关注。身份识别在信息安全领域至关重要。为了确保个人的身份信息安全,研究者提出了各式各样的身份识别方法。早期的传统身份识别方法如身份识别标签(ID)或者身份识别号码(PIN)等,容易遭到黑客的攻击。接着,研究者们又提出了基于个体生物特征的生物识别技术,该项技术通过人体具有的独特生物特性进行身份识别。目前,人脸识别、虹膜识别、指纹识别等识别技术在日常生活中受到广泛应用。然而人脸、指纹等生物特征识别技术也存在着可以被伪造的问题。因此,基于人体生理信号的生物特征识别受到特别关注,比如脑电信号,心电信号,肌电信号等。这些信号具有活体生物特性,有利于活体检测以及在防止欺骗攻击上更加健壮。在这些生理信号中,肌电信号(Electromyography, EMG)由于其独特的优势吸引了人们的关注,使其在其他生物特征中突出。因此,利用肌电信号进行身份识别技术具备良好的应用前景。目前,基于肌电信号的身份识别研究却很少。本文主要研究利用具有活体特性且易采集的肌电信号进行身份识别的系统,包含从离线身份识别系统递进到实时的身份识别系统应用研究。本文取得的主要研究成果如下:


(1)针对离线身份识别系统,本文提出了基于肌电信号的一对多的个人识别系统和基于肌电信号的一对一身份识别系统。离线身份识别系统由采集模块,数据预处理模块、特征提取模块以及身份识别模块组成。在一对多的身份识别系统中,系统采用连续小波变换方法挖掘肌电信号内部的时频特征,结合卷积神经网络算法对数据进行训练,最后达到了99.203%的识别率;在一对一的身份识别系统中,采用孪生网络算法对比系统提取出的样本特征之间的相似度,实现了99.285%测试准确率。

(2)在离线身份识别系统的基础上,本文进一步提出了一套基于肌电信号的实时身份识别系统。与离线身份识别系统不同的是,实时系统不仅包括采集模块、数据预处理模块、特征提取模块以及身份识别模块,还额外多了手势信号检测模块。在实时的系统中,系统需要着重考虑响应时间,使得系统具备实时、高效、鲁棒性强的特性。针对实时性要求,在事件信号检测环节,本文提出了一种峰值检测算法,用以实时的将手势信号从信号流中检测出。针对计算效率问题,本文采用小波变换结合一维卷积神经网络的方式,即采用离散小波变换方法结合统计学的方式对肌电信号进行处理及特征提取,最后,根据提取到的特征定制化地搭建一维卷积神经网络算法框架,实现了98.41%的识别准确率以及整个系统的处理时间为37ms。

英文摘要

With the rapid development of the internet, security of personal information becomes more and more important. Personal identification plays an important role in the information security field. To ensure the security of personal information, varieties of personal identification methods have been proposed. The traditional identification methods, such as Identification tag (ID) or Personal Identification Number (PIN), are vulnerable to hackers. Then, the biometrics which use the unique physiological characteristics of human body to identify personal information come into being. But the widely used biometrics at present, such as human face, iris , fingerprint, and voice, also can be falsified and forged. Thus, special attention has been paid to personal identification based on human physiological signals, such as electroencephalograph (EEG) signals, electrocardiograph (ECG) signals and electromyography (EMG) signals. These signals are living body features, which can be used to achieve aliveness detection and prevent the spoofing attacks. Electromyography (EMG) has attracted people's attention because of its unique advantages, which makes it outstanding among other biological characteristics. Thus, using EMG signal for personal identification has broad applications. However, there are few studies on personal identification based on EMG signal. This paper mainly studies the personal identification models based EMG signal which has characteristics of living body and is easy to collect, including offline identification system and real-time identification system. The latter is advanced for the former. The main contributions in this paper are summarized as follows: 

(1) For offline identification system, two EMG-based identification systems, which are one-to-many personal identification system and one-to-one identification system, are proposed in this paper. Both systems are consist of data acquisition module, data preprocessing module, feature extraction module, and personal identification module. For the one-to-many personal identification system, the continuous wavelet transform method is adopted to mine the time-frequency features of EMG signals captured from different subjects and convolutional neural network is used to identify the subjects. The identification accuracy of this method can achieve 99.206%. For the one-to-one personal identification system, the siamese network is adopted to learn a similarity metric from data, achieving the identification accuracy of 99.285%.

(2) Based on the offline identification system, this paper further proposes a real-time identification system based on EMG signal. Different from the offline system, the real-time system includes data acquisition module, data preprocessing module, feature extraction module, personal identification module, and additional gesture detection module. The real-time system focuses on the requirement of response time, so that the system has the characteristics of real-time, high efficiency, and strong robustness. To meet the requirement of real-time, a peak detection algorithm which can detect gesture signals from the signal stream in real time is adopted in the system. Then, to improve the computational efficiency, this paper adopts wavelet transform method combined with one-dimensional convolutional neural networks. Specifically, a discrete wavelet transform method combined with statistical methods is used to process EMG signals and extract features. Then, based on the extracted features, one-dimensional convolutional neural network architecture is delicately designed to identify subjects. The result shows that the identification accuracy for 21 subjects under the hand-open gesture could achieve 98.41% and the processing time between gesture event and identification is 37ms.

关键词身份识别 肌电信号 小波变换 卷积神经网络 峰值检测
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/48492
专题毕业生_硕士学位论文
推荐引用方式
GB/T 7714
卢立静. 基于肌电信号的身份识别技术研究[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2022.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
明版-卢立静-签名.pdf(7880KB)学位论文 限制开放CC BY-NC-SA
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[卢立静]的文章
百度学术
百度学术中相似的文章
[卢立静]的文章
必应学术
必应学术中相似的文章
[卢立静]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。