为了利用HMM抽取的步态序列的动态特征来进行身份确认，首先提出一种改进的角度向量用来表征二值化的步态序列图像，以便将每幅图像转化为1维向量，然后再以此作为特征向量，对每个人物建立并训练HMM模型，用于确定人物身份。这种改进的角度向量由于具有较强的抗噪性和方便的尺度伸缩性能，因此既适用于分割质量较差的图像，又能减小行走方向和距离的影响。实验表明，这种HMM不仅能较好地模拟步态的动态特征，还能描述序列图像间的联系，而且算法执行速度快，从输入原始数据到输出识别结果所需时间不超过2min，能满足实时要求。在Soton和NLPR数据库上进行的实验，分别获得了100％和85％的识别率，证明该方法是有效的。

This paper explores the dynamic feature of human gait extracted by hidden Markov model(HMM),which is used for identifying people. At first,an improved angular vector representation is proposed for binarized human images in a gait sequence so that every image is turned into a one-dimension vector. Then these vectors act as feature vectors to build and train HMMs which are the final identifying tools for each person based on input gait sequences. The improved angular vector is equipped with better robustness against segment errors,so it is suitable for imperfectly segmented silhouettes. It is also easy to scale up or down,thus scarcely vulnerable to the change of walking direction and distance from data-collecting camera. HMM models not only the dynamic characteristic of gait but also the relation between images in the same sequence. Besides,it can guarantee a high-speed operation which carries out the whole process within 2min. The experiments on Soton and NLPR database yield encouraging correct identifying rate of 100% and 85%,which demonstrates the effectiveness of this method.