基于连续Adaboost算法的多视角人脸检测  

A Multi-View Face Detection Based on Real Adaboost Algorithm

作　　者：武勃[1] 黄畅[1] 艾海舟[1] 劳世竑 

机构地区：[1]清华大学计算机科学与技术系智能技术与系统国家重点实验室,北京100084 [2]日本欧姆龙公司传感技术研究所,京都6190283

出　　处：《计算机研究与发展》2005年第9期1612-1621,共10页Journal of Computer Research and Development

基　　金：国家自然科学基金重点项目(60332010);日本欧姆龙公司合作基金项目(0302J05)~~

摘　　要：提出了一种基于连续Adaboost算法的多视角人脸检测方法.人脸按其三维姿态被划分成若干个视点子类,针对每个子类使用Haar型特征设计了具有连续致信度输出的查找表型弱分类器形式,构造出弱分类器空间,采用连续Adaboost算法学习出基于视图的瀑布型人脸检测器.为了提高检测速度,使用了多分辨率搜索和姿态预估计策略.对于正面人脸检测,在CMU+MIT的正面人脸测试集合上检测的正确率为94.5%,误报57个;对于多视角人脸检测,在CMU侧面人脸测试集合上检测的正确率为89.8%,误报221个.在一台PentiumⅣ2.4GHz的PC上,处理一幅大小为320×240的图片平均需80ms.实验结果表明该方法十分有效,具有明显的应用价值.In this paper, a multi-view face detection method based on real Adaboost algorithm is presented. Human faces are divided into several viewpoint categories according to their poses in 3D, and for each of these categories a form of weak classifiers in look-up-table （LUT） type is designed using Haar-like features that have confidences in real values as their outputs, and correspondingly its space of weak classifiers is constructed, from which the cascade face detector is learnt by using real Adaboost algorithm. For speed up, multi-resolution searching and pose prediction strategies are introduced. For frontal face detection, the experiments on CMU ＋ MIT frontal face test set result in a correct rate of 94.5 % with 57 false alarms; for multi-view face detection, the experiments on CMU profile face test set result in a correct rate of 89.8% with 221 false alarms. The average processing time on a PⅣ 2.4GHz PC is about 80 ms for a 320×240- pixel image. It can be seen that the proposed method is very efficient and has significant value in application.

关 键 词：Haar型特征 查找表型弱分类器 姿态估计 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]