【作者】 徐俊红；

【导师】 丛望；

【作者基本信息】 哈尔滨工程大学 ， 控制理论与控制工程， 2012， 博士

【摘要】 步态(Gait)是指人行走时协调的、周期的姿态,是人体生物特征中的一种行为特征。近年来,基于步态的身份识别(Identification Recognition)方式,作为一种新兴的生物识别技术,以其独特性质如可远距离识别性、无接触性、容易采集、难于模仿等,引起了众多研究人员的重视。就现有的生物特征所处理的对象以及所面对的环境而言,任何一种生物特征的识别方法都不是完美无缺的,步态识别也无例外地遇到一些难题,如视点的变化、行人的衣着穿戴、行人是否拎包、较长时间上步态变化、有效的特征表征等,论文重点对步态运动目标提取,基于椭圆模型识别,基于步态轮廓和光流识别和基于局部线性嵌入流形降维步态识别方面进行了深入的研究。在提取运动目标时,针对快速性计算,并较为准确的提取运动目标轮廓,提出了一种基于改进GVF和二阶差分阈值函数预测的模型,即根据二阶差分阈值函数预测初始化轮廓,改进的GVF模型使用线性映射规则进行二次初始化,迭代计算提取出运动目标,该计算模型同其他的GVF模型跟踪算法相比较,迭代次数少,速度更快；同时,为了提取出精确的运动目标轮廓,提出了基于模板匹配的运动目标提取模型,该方法可以更为精确的分割出人体目标,而且有利于消除伴随人体移动的阴影,为进一步更为有效地进行.步态识别提供了坚实的基础。对于低分辨率运动目标轮廓的不连续性,提出了一种基于正交联系点改进的椭圆拟合方法,不需要提取目标的闭合轮廓。该方法使用关键帧而不是动态时间规划方式,降低了计算复杂度,拟合的效果良好,识别率比代数拟合方法更高。为了更为有效地利用独特个性中的静态和动态的步态身份信息,提高正确识别率,提出了基于轮廓上八邻域像素光流进行步态识别算法。使用该模型研究发现：沿着目标的运动方向分量和综合指标分量正确识别率最高,几乎达到100%,而沿运动方向相反的分量部分识别率最低；基于轮廓上区域像素光流进行步态识别算法优于仅轮廓上像素进行识别的光流算法。为了降低步态序列数据的复杂度,采用符合认知规律的流形算法,把高维数据本征结构映射到低维的空间进行识别,论文对基于局部线性嵌入(Locally Linear Embedding, LLE)非线性流形降维方法进行了详细研究,提出了一种基于加权距离进行识别的LLE模型。从正确识别率的稳定度上,从参数的选择容易度上,从正确识别率的高低上,该计算模型都要高于加权LLE和LLE算法,而且从计算时间考虑,提出的算法与LLE算法相当,但是低于加权LLE算法。更多还原

【Abstract】 Gait is defined as the coordinated, cyclic combination of movements that result in human locomotion, and it is a behavior characteristic on human walking posture which belongs to one of human biological features. Recently, the technology for human identification recognition based on gait is attractive since it has the characteristics for example, no subject contact, easy collection at a distance and hard to disguise etc al. Existing biometric system are never perfect for all required performances and environments, therefore, gait faces some inevitable difficulties for data changes in different viewpoints, data changes for wearing clothing and package, data change for a long time, etc al. The paper aims to the human object abstraction, elliptical model recognition, optical flow effect on gait and dimensional reduction and the main contributions in detail as follows:Aiming to rapidity for moving objection and accuracy contour, the algorithm of threshold function on second order difference and improved GVF is raised. The initial contour is established by threshold function, which is built by experience and displacement equation. Subsequently, the reinitial contour for improved GVF according to rectangle range is obtained by linear rules, and is used to extract moving object. The method compared to other GVF algorithm is not only simple, but also has little iterations and more rapidity. Aiming to the more accurate contour extraction in single view-point, a new method based on human template matching and improved GVF snake is proposed in the paper. It can segment object accuratly, eliminate the effect of shadow more effectively and make solid foundation for gait recognition.Aiming to discontinuous moving object contour in low resolution images, a robust recognition algorithm based on improving ellipse model of orthogonal contacting points is raised, which need not closed object contour. A robust key fame, not dynatic time programming, makes detecting key fames easily and accurately. According to elliptic parameters in the key frames, correct classification rate(CCR) for gait recogntion is higher, and it reduces calculation complexity, and fitting result is better than algebraic one.In order to benefit from static and dynamic information in walking posture and improve correct classification rate, an optical flow method based on pixels in neighbor area is proposed. The results display that the correct recognition rate(CCR) for compontes on syntheses and moving direction is higher to100percents almostly, and contrary moving direction along human walking has lowest CCR. On total, the raised optical flow algorithm in neighbor pixels is better than simple contour pixels.Aiming to much information in gait sequence, manifold algorithm according with congnitive law is used to map the intrinsic dimension data in high dimension into lower dimension, in which is recognized. Therefore, locally linear embedding(LLE) manifild to reduce dimensionality is researched in detail, and an algorithm based on weighted distance testing and LLE train is raised. The expermental results show that the CCR of raised algorithm is higher than LLE and weighted LLE method, parameters for raised method are chosen easily, it runs rapidly than weighted LLE and is alike LLE in time.更多还原