视频中运动人脸的检测与特征定位方法  

Face Detection and Feature Location of Moving Men in Video

作　　者：雷蕴奇[1] 柳秀霞[1] 宋晓冰[1] 袁美玲[1] 欧阳江帆[1] 

机构地区：[1]厦门大学计算机科学系,福建厦门361005

出　　处：《华南理工大学学报（自然科学版）》2009年第5期31-37,共7页Journal of South China University of Technology(Natural Science Edition)

基　　金：国家"863"计划项目(2006AA01Z129)

摘　　要：针对现有人脸检测方法存在的检测质量与速度不平衡的问题,提出了视频序列中运动人脸的检测与特征定位方法.首先利用Adaboost方法检测出人脸的大致范围,根据肤色模型确定人脸的具体位置,并从图像中提取出人脸部分;然后利用基于帧间亮度差的人脸区域的PSNR判断图像清晰度,从而找出人脸区域清晰度高且尽可能大的视频帧;最后对该视频帧进行人脸检测和特征定位.实验结果表明,与现有人脸检测方法相比,文中方法速度快、人脸检测率约为94.8%,眼角、口唇角定位结果更为准确.In order to balance the detection quality and the computing speed of the existing human face-detecting methods, an algorithm for face detection and feature location of moving men in a video is proposed. In this algo- rithm, first, the approximate face region is detected using the Adaboost method. ＇Next, the specific face region is determined using the skin color model, and the face part is picked up from the frame. Then, the video frame with a high face-region definition and a region as large as possible is selected by judging the image definition from the Peak Signal-to-Noise Ratio （PSNR） based on the difference between two neighbor frames in face region. Finally, the face detection and feature location of the video frame are performed. Experimental results indicate that, as com- pared with the existing face-detecting methods, the proposed method helps to perform more accurate feature location for corners of eye and mouth with higher calculating speed, the face-detecting rate being about 94.8%.

关 键 词：视频处理 人脸检测 特征定位 图像清晰度 峰值信噪比 肤色模型 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]