【作者】 郭军；

【导师】 文玉梅；

【作者基本信息】 重庆大学 ， 光学工程， 2007， 硕士

【摘要】 步态识别根据人走路的姿势进行身份识别。和其它生物特征相比,捕捉步态特征无需身体接触,不具有侵犯性。在步态识别中,完整的人体轮廓是特征提取、特征表达、目标分类与目标识别等后期处理的前提条件。因此,本课题的主要目的是研究运用目标分割技术从步态图像序列中提取人体轮廓。主要的工作包括以下几点:①针对块匹配算法初始搜索点选择与实际情况存在差异的问题,在运用UCBDS法(unrestricted center-biased diamond search,非限制性中心倾向分布的钻石形搜索方法)进行运动估计时,采用中值法预测初始搜索点,并根据各候选块匹配度减少UCBDS法的搜索范围。②研究了一种新的基于运动信息和分水岭变换的运动目标分割算法。采用改进的块匹配算法对人体目标运动进行估计,运用分水岭算法将当前帧图像分割成封闭而不重叠的小区域,应用仿射参数模型进行基于运动的块区域合并,从而完成人体轮廓的提取。实验结果表明,该算法提高了步态识别率。③提出基于Snake模型的优化人体轮廓的方法。以基于运动信息和分水岭变换的运动目标分割算法得到的结果代替人工勾勒作为Snake模型的初始轮廓,实现了整个分割过程的自动化。Snake模型算法主要有变分法、动态规划法、Greedy算法。在本文中对Greedy算法进行了改进:通过修改外部能量函数区分真正的轮廓点和孤立的噪声点;构造了一个调节参数公式实现自动调节内部能量和外部能量的权值。实验结果表明,该方法能够在低对比度、局部动态背景下正确提取人体目标轮廓。更多还原

【Abstract】 Gait recognition is the process of identifying individuals by the way they walk. Compared with other biometrics, such as fingerprint, face and iris, the identification by gait is unobtrusive and does not require touching the human body. In gait recognition, effective extraction of human contour is a prerequisite of subsequent processing including feature extraction, feature expression, target classification and target recognition. The goal of this thesis is to research human contour extraction algorithm. The main contributions are listed as follows:①A new block-matching motion estimation algorithm based on UCBDS (unrestricted center-biased diamond search) is put forward in this thesis. For the original UCBDS, the motion vector of the current block is assumed to be zero, but this doesn’t match with the real condition. Aiming at a better prediction result, the median method is adopted to confirm the starting search position. In addition, according to the matching possibility of candidate blocks, the search region of UCBDS is decreased.②This thesis presents a new motion object segmentation algorithm based on motion information and the watershed algorithm. Firstly, the improved block-matching estimation algorithm is used to estimate the motion fields of the image. Secondly, we apply the watershed algorithm to divide the current image into a number of closed and non-overlapping regions. Finally, we utilize the affine parameter model to merge these regions. The experiments show that this new algorithm can effectively extract human contour and improve the performance of gait recognition.③An improved object contour method based on Snake model is investigated for an accurate object contour. The initial contour of Snake model is extracted automatically by the above algorithm rather than by manual work. In this paper, we use Greedy algorithm to compute the energy function of Snake model. The true points of contour and isolated noise points are distinguished automatically by applying a new external function. In addition, a formula is constructed to automatically adjust the weights of internal energy and external energy. The experiments show that the method can extract accurate contour under complex background.更多还原