【作者】 吴璇；

【导师】 李见为；

【作者基本信息】 重庆大学 ， 仪器科学与技术， 2010， 硕士

【摘要】 人脸检测与人脸跟踪技术是各种人脸图像处理算法的关键技术。人脸图像处理领域包含有人脸识别、姿态估计、表情识别、视频监控等多个研究方向,而几乎所有这些方向都涉及到人脸的检测与跟踪问题。本文在收集和分析了近年来国内外关于人脸检测与跟踪的学术论文及研究报告的基础上,对人脸检测与跟踪算法进行了深入研究,并根据国际、国内人脸检测与人脸跟踪技术的研究成果,设计并实现了基于Adaboost和CamShift算法的人脸检测和跟踪方法。该方法借鉴了前人的一些经验,并针对实验条件,作了一些关键性的改进。本文主要研究工作如下:1、针对传统Adaboost检测算法对于侧面人脸检测率不高的情况,对Adaboost人脸检测算法进行相关改进。首先从一个较大的特征集中选择少量关键Haar-Like特征完成高效强分类器的构建;其次将每个强分类器级联成为更加复杂的级联分类器,在此基础上,分别训练正面人脸分类器和侧面人脸分类器,将正面、侧面人脸检测结果进行融合得出人脸区域,并对其进行肤色验证,提高算法的鲁棒性、降低了虚警率。实验证明该算法检测速度快,实时性好,在CMU人脸测试库上进行了实验,检测率达到82.7%。2、分析对比基于肤色和基于Adaboost的两种人脸检测算法。针对Adaboost算法对人脸偏转和遮挡时检测效果差的问题,为了解决Adaboost人脸检测的这一缺点,本文将其与人脸跟踪算法相融合实现人脸的多角度检测。3、在软件开发平台Visual C++环境中完成对Adaboost算法与Camshift算法的融合。利用检测过程中人脸区域初始化跟踪窗口,建立肤色的色调信息模型对后续帧进行跟踪。通过实验得知该算法计算量小、跟踪速度快,并且不受人脸不规则运动和随意转动的影响,对场景中出现其他人脸不会出现误跟踪,对人脸被部分遮挡也不会出现跟踪丢失现象,大大提高了跟踪的匹配度和精度,证明了该算法的可行性和有效性。更多还原

【Abstract】 The face detection and tracking are one of the key techniques in various facial processing algorithms. The research fields of facial image processing include face recognition, gesture estimation, facial expression recognition, video monitoring and so on. And nearly all of these are involved in the area of face detection and face tracking. Based on the collection and analysis of numerous the domestic and international academic thesis and research reports on face detection and tracking in recent years, face detection and tracking algorithms is lucubrated. According to the previous research achievements on face detection and tracking, a face detection algorithm based on Adaboost and CamShift has been realized. Aiming at the experiment conditions, crucial improvements have been achieved by learning from some traditional and classical algorithms of face detection and tracking. The study of this paper focused on the following aspects:1. Aiming at the poor rate of side face detection by traditional Adaboost algorihtm, some related improvements to Adaboost algorihtm were proposed. Firstly, few key Haar-Like features from a large feature set were selected to generate an effective strong classifier. Secondly, cascades these single strong classifier into a more complex cascade classifier. On that basis, the frontal face classifier and side face classifier were trained individually and the detection results of frontal face with side face were inosculated to obtain face region. Afterwards, skin color was further validated on the face region to enhance the robustness of the algorithm and reduce the false alarm rate. The results from experiments demonstrated the high speed of test and good real-time. With the CMU face test library on this algorithm, the detection rate can reach 82.7%.2. In this paper, the face detection algorithm on skin color was comparative analyzed with the algorithm based on Adaboost. Aiming at the issue that the Adaboost algorithm is not effective when faces rotated or being blocked out. In order to overcome this shortcoming, a method to combine the Adaboost algorithm with face detection algorithm was presented in this paper.3. The Adaboost algorithm and CamShift algorithm were combined well in software development tool such as Visual C++. Firstly, the tracking window based on face region was initialized in detecting process. Then, color hue information model was established to track the follow-up frame. Presented the experiment results, it is showed that the algorithm needs less calculating amounts with fast tracking speed, and without influenced by irregular movement of face. And the phenomenon of face tracking error does not occur when other people appeared on the scene. Furthermore, track loosing does not happen when the face kept out partly. Therefore, the matching and accuracy of the tracking algorithm has been improved greatly. The feasibility and effectiveness of this improved algorithm was verified.更多还原