【作者】 黎云汉；

【导师】 朱善安；

【作者基本信息】 浙江大学 ， 控制理论与控制工程， 2008， 博士

【摘要】 本文研究的主要内容为人脸特征的提取与跟踪。在机器视觉中,人脸特征的提取与跟踪起着十分重要的作用,如在虹膜识别、表情分析、人机交互等具体应用中,都需要提取人脸特征作为后续处理的基础。本文引入了一种基于改进色彩空间的人脸检测算法,该算法能较快的检测肤色区域并定位眼睛。针对眼睛特征的提取,本文提出了三种算法。第一种是基于Snake的眼睛特征提取算法,Snake是一种用于轮廓提取的算法,本文提出的改进算法首先需要建立眼睛区域的距离势能力场,然后用Snake方法眼睑轮廓进行匹配。在匹配过程中加入了修正算法,两段首尾相连的抛物线被用于Snake的初始化和匹配;第二种是基于改进Hough变换的算法,该算法采用改进的圆Hough变换来提取虹膜、眼睑轮廓和眼睛角点等眼睛特征,由于利用了梯度信息,计算复杂度得到了降低;第三种算法直接使用梯度信息来提取眼睛特征。我们在一个较大的数据库中做了实验,证明了这三种算法都能较好地提取眼睛特征。针对人脸跟踪,在第四章中提出了一种改进的Mean Shift跟踪算法,该算法采用了混合目标模型。与固定目标模型的Mean Shift算法相比,改进算法能较好地对复杂背景下的旋转人脸进行跟踪。更多还原

【Abstract】 The contribution of this thesis is facial feature extraction and tracking.Facial feature extraction and tracking is a key step for many applications such as iris recognition, facial expression recognition, human machine interface and biometric identification.In chapter II, a face detection algorithm based on modified color space is introduced. The algorithm can detect skin region rapidly. With the modified color space, eye region can also be extracted accurately.To extract eye features, three flexible, reliable and less complex algorithms were proposed. The first is based on Snake. Snake is a successful model for contour extraction. The algorithm first builds a distance force field of eye region, then a Snake deformation approach is used to extract eye contours. A correction procedure is constructed during deformation, and two end-to-end parabolas outside the candidate are used for initialization and Snake deformation. The second is based on Hough Transform. A modified circle Hough Transform is used to extract eye features including iris contours, eyelid contours and eye corners. Due to the use of the gradient information, the algorithm can reduce the combinatorial complexity. The third algorithm for eye feature extraction is directly based on the gradient information. We have tested the three algorithms on a large image database and achieved good performances.In chapter IV, a modified Mean Shift algorithm is proposed for face tracking. This algorithm uses combinative target model. Compared with invariable target model, this algorithm can track rotate target with complex background.更多还原