【作者】 袁泉；

【导师】 杨杰；

【作者基本信息】 上海交通大学 ， 模式识别与智能系统， 2007， 硕士

【摘要】 人脸检测及跟踪属于模式识别与计算机视觉的研究领域,它作为人脸信息处理中的一项关键技术,在基于内容的图像与视频检索、视频监视与跟踪、视频会议以及智能人机交互等方面都有着重要的应用价值。本文主要研究了视频中的人脸检测及跟踪技术。在总结现有算法的基础上,针对视频这一应用背景的实时性要求,选取人脸的小波特征作为视频中人脸的主要特征,通过Adaboost训练算法架构人脸检测器。本文研究了基于直方图统计学习的人脸检测方法与基于haar-like特征的人脸检测方法两种方法。前者通过使用5/3小波变换,能有效地提取出空域、频域和方向场上的信息进行建模,同时反映目标各部分之间的几何关系,从而提取出完备的特征,能有效地检测正面和侧面人脸;后者先利用“积分图”快速计算特征,构造弱分类器,然后通过Adaboost学习算法从得到的大量弱分类器中产生一个高效的强分类器,最后采用级联方式将单个的强分类器再合成为一个更加复杂的层叠分类器,使图像背景区域快速地丢弃,保证了检测速度,满足视频的实时性需要。在人脸跟踪方法上,本文分别研究了两种侧重点不同的人脸跟踪方法。一种以基于直方图统计学习的人脸检测为基础,通过肤色预处理和视频中的运动信息获得人脸候选区域,再通过人脸检测算法精确定位人脸,实现了视频中基于人脸检测的人脸跟踪。另一种将Mean-Shift目标跟踪算法和Kalman滤波运用到人脸跟踪上,通过不断的进行均值偏移矢量的迭代和目标模版更新,可以快速有效的在视频中跟踪人脸。本文的创新点分别体现在以下3个方面:(1)高效的肤色分割预处理算法;对已有的光照补偿算法做了改进;提出了针对于肤色二值图像的区域分割与合并算法;(2)提出了求取小波系数量化参数的方法,并提出了分组量化的概念;改变了Adaboost训练过程中弱分类器的输出值,给出了弱分类器的阈值选取方法,减小了分类误差;提出了基于样本统计的最终人脸分类器阈值选取准则。(3)在Mean-Shift目标跟踪算法和Kalmam滤波的基础上,提出了新的实时人脸跟踪算法。提出了把Adaboost人脸检测算法和Mean-Shift人脸跟踪算法相结合实现人脸的实时检测与跟踪,并对跟踪人脸实现姿态估计的新思路。实验结果以及与其他算法的比较分析表明,本文算法在准确率、误检率和检测与跟踪速度等方面均可获得较理想的结果,是两个综合性能很强的完整、鲁棒、高效的人脸检测与跟踪算法。更多还原

【Abstract】 Face detection and tracking is an important research aspect in artificial intelligence and computer vision. As a key technology of face information processing, it has a broad application values in many fields such as video surveillance, content-based image retrieval, videoconference, etc.In this thesis, a study on face detection and tracking in video is presented. After a general review of existing schemes on this particular topic, we choose Wavelet feature as the main feature of human face and Adaboost training algorithm to construct the face detectore, due to the real-time requirement of video. We study two face detection approaches including histogram-based statistical learning approach and Haar-like feature-based face detection approach. By using 5/3 wavelet transformation, the former one could effectively decompose the image in frequency, orientation, space, and geometry, obtaining overcomplete feaure set and detecting frontal and profile faces effectively; While The latter utilize integral image to quickly calculate the feature, and construct weak classifier by the feature; then weak classifiers are combined to a strong classifier in a linear way.The final classifier is built in a cascade structure which could reject most non-face samples in the early layer.Two face tracking methods with different emphasis are also studied in this thesis. The first one accomplish face tracking in viedo by utilizing histogram-based statistical learning approach to detect faces in face-candidate regions, which was obtained by using skin pre-process and motion information in video. The second one applies Mean-shift object tracking algorithm and Kalman filtering to face tracking. The central computational module is based on the mean shift iterations and target model updating and finds the most probable face target position in current frame.The contributions of this paper could be expressed as the following 3 aspects:(1) Efficient skin segmentation pre-processing algorithm; improve the current lighting compensation algorithm;propose a new region segmentation and combination algorithm especially for skin binary image;(2) Present a method to quantize wavelet parameters and concept of quantization in different groups; change the output of weak classifier in Adaboost training and provide a method to set the threshold of the best weak classifier; propose a sample statistic-based criterion to set the threshold of face classifier;(3) A novel real-time face tracking algorithm was presented based on Mean-Shift target tracking algorithm and Kalman filtering algorithm; propose a new thinking to detect and track face in real-time by combining Adaboost face detection algorithm and Mean-Shift algorithm with pose estimation implemented to the tracked face.Experiment results and comparison with other published method show that algorithm in this thesis obtains almost ideal result in the field of detection accuracy rate, false alarm rate and detection and tracking speed, and both of these two algorithms are complete, robust and efficient face detection and tracking algorithms with great comprehensive performance.更多还原