基于AdaBoost的视频人脸检测

发布时间：2018-04-14 13:11

本文选题：人脸检测 + AdaBoost算法　；参考：《杭州电子科技大学》2017年硕士论文

【摘要】：人脸检测是人脸识别、表情分析、人脸跟踪等领域的前提和基础,在视频监控领域有着广泛的应用价值。人脸检测具有的不易被检测目标发现的显著优势,可以目标不配合的情况下进行监控,为智能化的视频监控提供了有力保障。本文研究的内容是基于AdaBoost算法的视频人脸检测,由于在视频人脸检测时需要考虑众多非理想因素(如复杂环境、多人脸、旋转人脸等问题),而现有的检测算法往往仅针对其中某种情况有较好的检测效果。因此,对于视频人脸检测时存在的众多干扰因素,建立一个检测率高、误检率低、鲁棒性较强的人脸检测算法仍然是一个具有挑战性的问题。由于AdaBoost人脸检测算法具有法具有检测效果好,并且基本可以达到实时人脸检测的目的,因此本文采用该算法进行视频检测,以下是本文的主要研究内容:首先,以AdaBoost算法为基础,并针对于人脸检测分类器训练过程中训练特征数目过多导致训练过程过于耗时的问题,提出了一种基于“大T”型区域的AdaBoost人脸检测算法。通过随机抽取500张人脸样本并提取出面部的主要特征区域,投影到一个大小为20?20的模板中,然后对投影区域的重叠部分取其并集求得“大T”型Haar特征筛选模板,最后采用该模板对所有的Haar特征进行筛选,使得用于AdaBoost算法的训练特征全部集中于人脸面部的关键区域,而且“大T”型特征筛选模板只限于AdaBoost人脸检测算法训练特征的优化与不同的训练样本库无关。实验结果表明,采用“大T”型特征筛选模板对Haar特征进行筛选不仅降低训练特征的数目起到了优化训练时间的目的,而且在LFW、PKU-SVD-B数据库中的检测结果也表明本文的改进方法在不降低AdaBoost算法检测率的同时,对PKU数据库中的多人脸检测时算法的漏检率有所改善,因此对于AdaBoost算法训练性能的提高有一定的作用。其次,针对于监控设备采集到的视频图像中不可避免的存在多人脸、复杂背景、光照干扰、多姿态人脸等干扰因素,使得单独采用AdaBoost算法进行视频人脸检测时,算法的误检率和漏检率都比较高的问题,提出了采用AdaBoost算法与YCgCr混合高斯肤色模型相结合的方式进行视频人脸检测。通过对肤色样本在YCgCr色彩空间进行建模,发现亮度分量同样具有近似高斯分布的特性,因此采用线性加权的方式构建Y分量与CgCr分量的混合高斯肤色模型,并采用该混合高斯模型对视频图像进行肤色分割。实验结果表明,采用AdaBoost算法与新的肤色模型进行视频人脸检测时,可以较好的避免视频图像中复杂背景对人脸检测的影响,进而起到了降低算法的误检率和漏检率的目的。
[Abstract]:Face detection is the premise and foundation of face recognition, facial expression analysis, face tracking and so on. It has wide application value in the field of video surveillance.Face detection has the obvious advantage that it is not easy to be detected by the target detection, and can be monitored without matching the target, which provides a strong guarantee for intelligent video surveillance.The content of this paper is video face detection based on AdaBoost algorithm, because many non-ideal factors (such as complex environment, multi-face) need to be considered in video face detection.However, the existing detection algorithms usually have better detection effect only for some cases.Therefore, it is still a challenging problem to establish a face detection algorithm with high detection rate, low false detection rate and strong robustness.Because the AdaBoost face detection algorithm has good detection effect, and can basically achieve the purpose of real-time face detection, this paper uses the algorithm to carry out video detection. The following are the main contents of this paper: first,Based on the AdaBoost algorithm, and aiming at the problem that the excessive number of training features in the training process of face detection classifier leads to the time-consuming training process, a AdaBoost face detection algorithm based on "large T" region is proposed.By randomly extracting 500 face samples and extracting the main feature region of the face, the main feature region of the face is projected into a template of 20 ~ 20, and then the "large T" type Haar feature screening template is obtained by the union of the overlapping parts of the projection region.Finally, the template is used to filter all the Haar features, so that the training features used in the AdaBoost algorithm concentrate on the key areas of the face.Moreover, the "large T" feature selection template is limited to the optimization of training features of AdaBoost face detection algorithm and is independent of different training sample bases.The experimental results show that "large T" feature screening template can not only reduce the number of training features, but also optimize the training time.The detection results in LFWU PKU-SVD-B database also show that the improved method does not reduce the detection rate of AdaBoost algorithm, but also improves the detection rate of multi-face detection algorithm in PKU database.Therefore, it can improve the training performance of AdaBoost algorithm.Secondly, in view of the inevitable interference factors such as multi-face, complex background, illumination interference, multi-pose face and so on in the video image collected by the monitoring equipment, the AdaBoost algorithm is used to detect the video face alone.The false detection rate and missed detection rate of the algorithm are both high. This paper proposes a method of video face detection based on the combination of AdaBoost algorithm and YCgCr mixed Gao Si skin color model.By modeling the skin color sample in the YCgCr color space, it is found that the luminance component also has the property of similar Gao Si distribution, so we use linear weighting method to construct the mixed skin color model of Y component and CgCr component.And the mixed Gao Si model is used to segment the color of the video image.The experimental results show that the AdaBoost algorithm and the new skin color model can avoid the influence of the complex background on the face detection and thus reduce the false detection rate and the missed detection rate of the algorithm.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】