【作者】 杨倩倩；

【导师】 孙伟平；

【作者基本信息】 华中科技大学 ， 计算机系统结构， 2009， 硕士

【摘要】 人脸姿态估计是判断静态图像或者视频序列中的人脸在三维空间中的姿态的过程。人脸姿态估计作为计算机视觉领域的一个重要研究课题,在人机交互、智能视频监控、虚拟现实以及人脸识别领域有着广阔的应用前景和巨大的应用价值,是近年来的研究热点。目前大多数人脸姿态估计方法针对静态图像,方法复杂,计算复杂度高,不适合用于对视频图像中人脸姿态进行估计。提出了一种基于嘴巴在人脸中相对位置的人脸姿态估计方法,该方法简单有效,速度较快,完全能满足视频序列中人脸姿态估计的实时性要求。对视频图像中的人脸姿态估计涉及到人脸检测、人脸跟踪、人脸姿态估计等相关课题。在比较和总结了现有的人脸检测方法的基础上,选择了对姿态变化不敏感的肤色模型在视频序列的前几帧中检测出人脸区域,然后利用CAMShift(ContinuouslyAdaptive Mean Shift,连续自适应均值移动)算法对检测出的人脸区域进行跟踪,并用椭圆模型对人脸区域进行拟合,计算出人脸区域在视频帧中的位置。随后在检测到的人脸区域中利用唇色信息定位出嘴巴,根据嘴巴在人脸区域中的相对位置以及拟合人脸的椭圆模型的方向角来对视频图像中的人脸姿态进行估计。最后实现了一个人脸姿态估计原型系统,用录制的视频文件进行了实验,验证了该方法的可行性。更多还原

【Abstract】 Face Pose Estimation is a process that is through dealing with the static images orvideo sequences to determine the posture of human face in three dimensional space. As animportant research topic in the field of CV(computer vision), Face Pose Estimation inhuman computer interaction, intelligent video surveillance, virtual reality, as well as facerecognition has a wide application prospect and great value, and it is the research hotspot inrecent years.At present, a majority of Face Pose Estimation methods is for static images, which iscomplex, of high computational complexity, and is not suitable for estimating dynamicvideo images. In this thesis, a novel Face Pose Estimation method is proposed, which isbased on estimating the relative location of mouth in whole face. The method is simple,effective, fast and fully able to meet the real time requirement for estimating face pose invideo sequences.Estimating face pose in video sequences relates to Face Detection, Face Tracking,Face Pose Estimation and other related topics. To begin with, a comparison and summary ofthe existing face detection methods is given, color model which is non sensitive for posturechange is chosen for obtaining the detected face region in first few frames in the videosequences, and then the detected face region is being tracked by the CAMShift(Continuously Adaptive Mean Shift) algorithm, and at the same time which is fitted byellipse model, which is in order to locate human face region in video sequences.Subsequently, using the information about lip color, mouth is located in the detected faceregion and according to the relative location of mouth in whole face and the direction angleof ellipse model for fitted face regions, estimating human face posture in video sequences ismade, and then a prototype system for estimating the face posture is implemented, andsome experiments are made using the recorded video files, which has verified the feasibilityof the method.更多还原