【作者】 王依默；

【导师】 姜琳颖； 高晓兴；

【作者基本信息】 东北大学 ， 计算机科学与技术， 2009， 硕士

【摘要】 随着计算机科学的飞速发展,计算机相关领域的应用已经深入到人们生活的各个方面,并起着十分巨大的作用。其中,人工智能与模式识别是计算机科学研究中比较前沿的研究课题,是计算机智能化的基础。让计算机拥有人类的视觉辨识能力一直都是人工智能领域的一个重要课题,也是人工智能领域中使计算机认识世界的第一步。本文主要研究实时视频流中的人脸检测技术。在人脸检测的研究方面,主要侧重于在视频流中人脸检测算法的实时性问题。针对现有人脸检测方法存在的问题及人脸检测技术的主要难点,结合目前人脸检测领域的实际应用情况,提出了一种适合应用于视频流中的人脸检测方法。具体工作如下:(1)实时视频流与静态图像人脸检测的不同在于前者含有运动信息。针对视频流的这个特点,研究几种主要的运动目标分割的方法,针对背景差分法对于光线干扰较大的缺点,提出一种新的参考背景生成方法来进行目标分割,并设计可以实时更新参考背景的算法。(2)通过对图像分割时使用不同阈值产生不同分割效果的分析,提出一种新的自适应阈值的分割方法来进行运动目标分割。(3)针对肤色在YCrCb色彩空间下的聚类特性,对在该空间下的简单区域模型和高斯分布模型进行了研究,提出区域阈值与简单高斯模型相结合的肤色建模方法,并在分割时使用最优阈值的方法使分割的效果达到最佳。(4)利用分割肤色后的轮廓,使用最小二乘法进行椭圆拟合来精确定位出人脸在视频流中存在的位置,并进行筛选和验证。更多还原

【Abstract】 With the development of computer science and technology, the computer applications have entered into every aspect of human life, and play a very important role. In which, artificial intelligence is very hot research field, and also the foundation of the artificial computer. It is always a very important subject in the research of artificial intelligence that the computer can recognize the object as a man, and it is also the first step the computer understanding the world.This paper focuses on face detection technology in real-time videos. The research of face detection focuses on real-time problem in videos. Aiming at the existing disadvantages of face detection method and the main difficulties of face detection technologies, combination of application of face detection, a method for face detection in real-time videos is proposed. Specific tasks are as follows:(1) The difference of face detection between real-time videos and static image is the former with motion information. Aiming at the characteristic of real-time videos, this paper studies on several major methods of moving object segmentation. Against the disadvantage which is quite sensitive to light in the background subtraction method, the paper has put forward a new method on generation of reference background and algorithm of real-time updating reference background.(2) When image segmentation, through the analysis of different segmentation result with the use of different thresholds, a new adaptive threshold segmentation approach to segmentation of moving object is put forward.(3) Aiming at the cluster characteristics of face color in the YCrCb color space, this paper studies on simple regional model and Gaussian distribution model in this color space, and put forward a color modeling method which is combination of regional threshold and Gaussian distribution model, and then segment the skin color with the optimal threshold to make the result best.(4) Based on the contour after skin color segmentation, the position of face in video streaming is accurately located by ellipse fitting least squares method, and then selected and identified.更多还原