对灰度复杂背景下的正面多视角人脸检测与跟踪技术的研究计算机软件与理论专业论文.docxVIP

摘 摘 要 人脸检测由于它的学术价值和应用范围，越来越受到人们的关 注。其中自适应自举学习算法(AdaBoost)将人脸检测带入了实用领 域。人脸检测的任务是从静态图像或视频中定位人脸的位置。深入研 究了灰度图像复杂背景下的平面内多视角的人脸检测以及人脸跟踪 技术，具体工作如下： 改进了AdaBoost的训练算法。该算法改原分类器的单阈值为双 阈值，且对样本的一次遍历便可完成一个弱分类器的训练，降低训练 耗时。对其可行性给出了数学证明。 提出了一种改进的AdaBoost权值更新算法，该算法较原算法加 入了评价其整体分类性能参数，以缓解的过拟合特性。对于改进后算 法在训练集上的错误率上限给出了数学证明。 对Haar型特征进行了扩展，采用45。积分图像并放宽了特征的计 算条件以发掘具有更好分类性能的特征，从而保证多视角人脸检测的 需要。 提出了层级决策树分类器以满足检测多视角人脸需要。出于不同 层次的子分类器对其性能有不同的要求，在训练时，对于不同样本赋 予不同的分类错误代价，并参考错误代价，改进AdaBoost的权值更 新算法，使其满足不同分类器性能要求。 在人脸跟踪方面，改进了Mean—shift算法的采样方法，加入了灰 度几何特征，使算法在复杂背景下的人脸跟踪更具鲁棒性。数学证明 了加入核带宽后的自适应检测窗口大小的Mean—shift算法的收敛性。 关键字人脸检测，AdaBoost算法，过拟合，Haar型特征，Mean．shift 算法 ABSTRACTHuman ABSTRACT Human face detection draws more attention tor its academic value and the range of application．Face detection is described as to locate the human face in images or video．The technology of in—plane rotated face detection and tracking in gray and complex background is deeply discussed in this paper．The detail is as bellow： AdaBoost training algorithm is improved．This method rums the single threshold into double threshold，and finishes the week classifier training only by going through training samples once to cut the time consuming．And effectiveness is mathematically proved． AdaBoost samples weights updating algorithm is improved．Add a parameter which describes the performance of the classifier to reduce the over fit while the weights of samples are being updated．A mathematically prove is given for the up bound of the learning error． The Haar-like features are extended．And the 45。integral image is adopted to extract more powerful features and meet the needs of multi．view face detection． The hierarchical decision tree is proposed to detect multi—view face． In training，various error cost is given to different kinds of samples when updating the samples’weights to meet the different requirement of different classifiers’layer performance． In aspect of face tracking，gray geometric data is added when images are sampled by Mean—shift