【作者】 刘宇；

【导师】 邢永康；

【作者基本信息】 重庆大学 ， 计算机软件与理论， 2010， 硕士

【摘要】 当今世界纷繁复杂,各种场合对安全的需求也越来越高。安全、有效、唯一等等得天独厚的优势使得生物特征识别技术逐渐被人们所重视。而根据人行走方式的不同来进行身份识别的步态识别则是一种新兴生物特征识别技术。由于具有其它生物特征识别所无可比拟的隐蔽性、非侵犯性和对图像分辨率要求不高等优点,在视觉监控领域有着广泛的应用前景。步态识别过程主要由步态图像预处理、特征提取和分类识别三个部分组成。其中步态图像预处理就是通过背景建模、前景检测和形态学后处理等工作,从输入序列图像中检测运动目标,这是后续特征提取和目标分类工作的基础,有很重要的意义。而特征提取是决定识别效果的决定因素,是重中之重,也是本文研究的重点。本文针对侧面视角条件下提出了两种提取特征的步态识别方法。其中之一基于Radon变换,将用于直线检测的Radon变换推广到特征空间的建立上。根据人体运动特性,将运动人体时空不变的身体结构参数和随时空变化的动态参数结合起来,构造一个周期的特征向量模板。其后运用主成分分析(PCA)降降低特征空间维数,提取特征主向量。另一种方法利用反映动态信息的各个对象步态能量图(GEI )的标准差为动态权值掩模(DWM),通过校准后与GEI做Hadamard积,从而得到动态信息增强的步态能量图(EGEI),这幅图像不仅保留了轮廓、频率、相位等步态信息,而其一定程度上解决了遮挡问题。再运用结合了行方向与列方向的二维主成分分析((2D)2PCA)来计算训练和测试样本的主元分量,得到对识别贡献最大的特征向量矩阵。单一的步态特征对步态的描述存在局限性,本文抓住不同步态特征在识别时能够提供互补信息这一关键,采用基于计分规则的方法,在决策层对两种特征的信息进行了融合。实验采用CASIA步态数据库中包含124个对象、3种行走条件的Dataset B,从以上算法的识别性能进行测试评估。结果表明本文提出的通过增强能量图来提取特征的算法对衣着变化和携带物的影响具有很好的鲁棒性。将两种特征信息通过合理的规则进行融合对提高识别性能更具效果。更多还原

【Abstract】 The world we live in is so diverse and complicated that we need higher and higher security guarantee in every situation. Its advantageous characteristics such as safety, efficiency and unique etc. make biometrics identification technology (biometrics) valued by more and more people gradually. Gait recognition, which is to identify one’s identity according to the different ways one walks, is a rising biometrics. And because of its incomparable crypticity, noninvasive and low demand of image resolution ratio which other biometrics doesn’t have, it has an extensive prospect of application in the vision monitoring area.The process of gait recognition is mainly made up of three parts, that is, pretreatment of gait image , feature extraction and discrimination. Pretreatment of gait image is to input detected moving targets of the image sequence via background modeling, future check and the morphology reprocessing. It is the foundation of following extracting features and classifying objective subjects, which is significant. Of all the work, extracting features is the most important, and it is the decisive factor in an effect of discrimination. And because of this, it is a key point this paper deals with. In view of side vision, this paper comes up with two ways of extracting features.One is based on Radon transformation. That is, Radon transformation used in line detection is extended to the foundation of feature space. In line with characteristics of human motion, we can combine space-time invariant physical structure parameters of moving human with Time-varying motion parameters, making up a periodic eigen vector template. Then we can use Principal Component Analysis(PCA) to reduce the feature space dimension to extract main vctor of characteristics. The other way is to use standard deviation of each object’s gait energy image(GEI) which can reflect the dynamic information as dynamic weight mask(DWM). After correcting, we can use Hadamard product with GEI to get enhanced gait energy image(EGEI) whose dynamic information is strengthen. Not only does it continue to have the gait information of contour, frequency, phase etc. but also it solves occlusion issue to a degree. Then, we can use Two Directional Two Dimensional Princdipal Component Analysis((2D)2PCA) to calculate prineipal components of training and testing samples, and we get eigenvector matrix which contributes most to identification.There are limitations in single gait description to gait characteristics. The text keeps to the key point that different gait characteristics can provide complementary information when identifying, adopts methods based on score and fuses together information of two kinds of character at decision layer.The experiment adopts CASIA which contains 124 objects and Dataset B of 3 walking conditions. Using above ways of algorithm, we test and assess its distinctions. The results show that the algorithm referred to in the paper which extracts feature by enhanced gait energy image, has a perfect robustness to clothes varying and carried materials. And it’s better to merge together two sorts of characteristic information by reasonable rules to improve its distincions.更多还原