【作者】 徐苏；

【导师】 孙怡宁；

【作者基本信息】 中国科学技术大学 ， 检测技术与自动化装置， 2010， 博士

【摘要】 通信技术和互联网的飞速发展从根本上影响了人类的生活方式,社会对于身份认证的准确性、安全性和实用性提出了更高的要求,生物特征识别技术受到了广泛的关注。基于步态触觉特征的生物特征识别方法是一种新兴的身份识别技术,与其它生物特征识别技术相比,基于步态触觉特征的身份识别具有独特的优点,第一,步态触觉信息的采集并不需要被测者的配合,可采取远距离非接触式的采集方式,不会涉及到隐私等问题；第二,步态主要由先天因素和后天因素影响,先天因素主要是指人体的脚型骨骼,后天因素则包括成长环境、习惯等,因此步态具有特定性和相对稳定性,不易伪装和模仿,非常适合用于生物特征识别。基于步态触觉特征的生物特征识别才刚刚起步,有关这方面的研究非常少,因此本文在介绍步态触觉信息的基础上,深入研究了如何利用步态触觉特征完成身份识别,全文的主要研究工作如下：首先简要介绍了步态触觉特征的生理本质、分类、模型和应用,随后主要针对三维地面反作用力,详细说明了实现生物特征识别的步骤以及需要解决的难点。在分析常用步态触觉信息获取装置的原理和不足的基础上,设计研制了一种包括压力测试板和三维测力台的测量平台,该测量平台能够测量包括足底压力分布和三维地面反作用力在内的完整步态触觉信息。利用该测量平台获取了完整步态触觉信息,建立了步态触觉信息数据库,随后针对步态触觉信息信号的特点,选取适当的方式对数据库中的数据进行去噪处理,有效地消除了噪声,又较好地保留了原始信号中的有效特征点,最后应用去噪后的信号,考察了三维地面反作用力的重复性和唯一性,证明了其满足进行生物特征识别的基本条件。接下来讨论了从原始信号中提取有效特征的方式,经过去噪处理后的信号中包含了大量的有用特征,如何提取这些特征对提高识别的准确性十分重要,通过分析常用特征提取方法的原理和适用范围,结合考虑GRF信号的特点,选取小波包分解作为特征提取方法,对原始信号进行大量不同的分解,提高信号的时频分辨率。通过实验对步态触觉信息数据库中的样本数据进行了分析,将原始信号分解到不同的频段上,选用能最清晰表现信号特征的小波包系数作为初始特征表征,为身份识别的下一步提供更精确的数据。随后讨论了特征选择和识别的方法,特征选择是模式识别中的关键步骤,它可以减少冗余特征,提供更快和更有效的模型,特征选择子集往往决定了最终的识别效率。本文提出了混合型的特征选择方法：采用基于模糊集的特征选择法进行首次特征选择,再利用遗传算法或蚁群算法进行二次特征挑选,在二次特征挑选中会用到分类器,支持向量机能够避免产生局部最小值,确保得到有效解,并且对小样本识别问题尤为有效,因此被用做分类。为了展示算法的有效性,采用它对步态触觉特征数据库中的数据进行分析,通过识别率比较和特征重构,比较了单纯采用一步特征选择和采用混合算法的效率,也对遗传算法和蚁群算法的效率进行了对比。前面的几章介绍了基于单步步态触觉特征的生物识别算法,在此研究的基础上,介绍了基于多步步态触觉信息的生物特征识别算法,构建了基于步态触觉特征的生物特征识别系统,详细说明了系统结构和算法流程,并利用该系统建立了一个步态触觉信息数据库,通过识别测验验证了所提出的识别算法的有效性和可行性,最后利用所提取的最优特征子集进行信号重构,说明了所提取的步态触觉特征的意义。更多还原

【Abstract】 The rapid development of telecommunication technology and internet has fundamentally changed people’s lifestyle, and the society proposes higher demand on the veracity, security and practicability of personal authentication, therefore, biometric authentication technology has drawn wide attention around the world.Plantar pressure and ground reaction force (GRF) based biometric authentication method is a novel personal identification technology. Compared with other methods, it has distinct advantages:first, the measurement of the information does not need the cooperation of the targeted subjects, for it can collect the information without privacy intrusion; second, plantar pressure and ground reaction force are determined by innate and postnatal factors:innate factors include the bone structure of human foot while postnatal factors refer to the environment, personal habits and so on. Therefore, these characteristics are distinctive, relatively stable and hard to disguise, making them perfect for using in biometric authentication.Research on biometric authentication based on plantar pressure and GRF is just started, and there’s little information on it, thus this paper will first introduce some basic knowledge about plantar pressure and GRF; then it will deeply study on how to accomplish personal identification with them, the main works of the dissertation are as follows:The first part gives brief introduction of the characteristics’physiological feature, type, model and application, while focusing on the steps and key problems in biometric authentication using GRF.Based on the analysis of the principle and deficiency of the popular information acquisition instruments of plantar pressure and ground reaction force, a platform combining pressure platform and force platform was designed and developed, and it can measure plantar pressure and ground reaction force at the same time. Using this platform, information was collected from different people, and a database was established. Considering the signal’s feature, a suitable method was used to denoising the original information, while maintaining some distinctive features. The repeatability of the signal for the same person and the uniqueness of the signal for different people were tested, proving that the signal is suitable to be used in biometric authentication.After that the paper discussed how to extract effective features from the denoised signals, which contain large amounts of useful features and how to extract them is very important for improving the efficiency of the authentication. Analyzing the principle of some popular feature extraction method and considering the characteristics of the GRF signals, wavelet packet decomposition (WPD), which can yield a large number of different decompositions and improve the time-frequency resolution, was used for this purpose. Experiments were carried to analyse subjects in the database, during which the original signals were decomposed to different frequencies, and the WP coefficients, which can best represent the characteristics of the signals were chosen as the original feature representation.Feature selection and classification method were discussed in the next chapter. Feature selection, which can reduce the dimensionality of the feature and provide better and more effective model, is a key step in authentication, for the feature subsets have a decisive role in determining the classification efficiency. A hybrid feature selection method was proposed:in the first step, a fuzzy set based feature selection criterion was used to drop unrelated features while genetic algorithm (GA) or ant colony optimization (ACO) was used in the second step to pick up pivotal ones. Support vector machine (SVM) was used for classification during the second selection process, for it can effectively avoid local optimum, guarantee best resolution and is very suitable in small sample classification problems. To show the availability of the algorithms, experiments were made to analyse the database, the classification efficiency and feature reconstruction of the pure one-step feature selection method and the proposed algorithms were compared, as well as the effect of using GA and ACO in the second selection process.Based on the above researches, the last part of the paper introduced the algorithm of biometric authentication based on multi-cycle GRF and a biometric authentication system was established. The system structure and proposed algorithm were explained in detail, and a database was established. The verification and practicability of the system was proved with authentication test, and signal reconstruction was carried out to discuss the essence of the selected feature.更多还原