一种提取人脸特征3D信息的方法  

Method for Extracting 3D Facial Feature

作　　者：曾阳[1] 邹北骥[1] 王磊[1] 彭小宁[2] 

机构地区：[1]中南大学信息科学与工程学院,湖南长沙410083 [2]怀化学院计算机科学与技术系,湖南怀化418008

出　　处：《小型微型计算机系统》2009年第2期332-335,共4页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(60673093)资助;怀化学院重点建设学科项目资助

摘　　要：针对头部姿态大角度偏转时,传统的2D人脸模型(主动表观模型,Active Appearance Model)定位人脸特征点的精度会急剧下降.引入与MPEG-4兼容的3D人脸形状模型Candide,并为其建立2D的形状模型和表观模型.模型拟合人脸图像的过程中,在保持3D形状模型与2D形状模型形变一致的基础之上,应用反向组合算法,提取人脸特征点的位置和头部姿态等3D信息.实验结果表明:在C环境下,算法的处理速度为250帧/秒以上,能达到实时处理;当头部姿态的偏转角度为0°-60°时,算法的标准误差值介于0.01-0.1之间.When head pose rotates in a large scale, the accuracy of facial features localization becomes worse for traditional 2D facial model （Active Appearance Model）. In this paper, 2D shape model and 2D appearance model are established, based on 3D facial wire-frame model Candide compatible with MPEG-4. In the process of fitting, the algorithm keeps the deformation of 3D shape model consistent with 2D shape model and applies Inverse Compositional Algorithm to extract 3D information about the position of facial features and head pose. The experiments show that the C implementation of the real-time algorithm can run at well over 250 frames per second. When the head pose rotates in the scale of 0°-60°, the RMS point error maintains between 0. 01-0. 1.

关 键 词：人脸特征 2D+3D Candide 主动表观模型 反向组合算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]