【作者】 张明恒；

【导师】 王荣本；

【作者基本信息】 吉林大学 ， 载运工具运用工程， 2007， 博士

【摘要】 行车安全是汽车交通发展的永恒主题,随着汽车保有量的增长和公路等级的不断提高,公路交通事故的发生越来越频繁,交通安全问题日益突出。在这种背景下智能交通应运而生,它是交通工程领域的研究前沿,体现了车辆工程、人工智能、自动控制、计算机科学等多学科领域理论技术的交叉和综合,是未来车辆发展的趋势,其中安全保障技术是必不可少的组成部分。近年来,国内外一部分较成熟的安全辅助驾驶技术正在逐步实用化和产品化,然而仍有相当多的技术难题亟需解决。本文从车辆的主动安全性角度出发,在利用机器视觉方法对驾驶员面部朝向估计方面进行了一些积极有益的探索,从而为我国安全辅助驾驶领域的相关实用产品研发和进一步深入研究提供理论和技术支撑。驾驶员面部及面部各特征区域的准确定位是进行朝向估计研究的基础。在对图像进行Gamma校正和白平衡预处理基础上,利用AdaBoost检测器与肤色模型相结合的方法实现了驾驶员面部区域的快速准确定位,从而有效提高了面部检测的实时性与稳定性。在面部区域定位基础上,根据眼睛、嘴巴与肤色之间存在的灰度及色彩差异,分别对眼睛与嘴部的定位方法进行了研究。依靠视觉特征对驾驶员面部朝向所进行的估计是一种多特征综合的非线性模式分类问题。由此,论文基于面部轮廓相似于椭圆的事实,在面部边缘点检测基础上利用边缘链码组对面部轮廓线进行拟合。以面部特征区域(眼睛、嘴巴等)相对于轮廓线的位置变化作为特征量,利用BP神经网络对面部朝向估计问题进行了深入研究。试验表明,上述驾驶员面部朝向估计方法是可行的。根据论文所研究问题的具体特点,对利用卡尔曼滤波器结合MeanShift算法进行面部特征区域的跟踪方法进行了较为深入的研究。更多还原

【Abstract】 Safety problem is a perpetual theme in the development of automobile traffic. With the rapid increase of vehicle conservation and the enhancement of highway level, traffic accidents occurred more and more frequently. Under this kind of status, the intelligent transportation occurred. Safety Driving Assist (SDA), mainly solving the problem of road safety, is a key component of the Intelligent Transportation System. Simultaneously, SDA also can relax the pressure of traffic jam and environmental pollution. At present, Europe, America and other developed countries have invested massive resources in this field and obtained many valuable researches.On the vehicle active safety consideration, this thesis carries on some positive beneficial study in the field of driver’s status monitoring based on machine vision. The purpose of the study is reducing the traffic accidents, through monitoring the driver’s face features real-timely and giving warning message in time, and providing theory and technology support for the SDA research of our country.According to the domestic and international research status in this field, the thesis mainly has carried on some studies about the following several topics. The face segmentation is the basis of locating face and other features such as eyes and mouth. In this thesis, the brightness correction and white balance have been conducted in the image pre-processing. As the environment illumination and the image collector system’s influence, the image used for face location possibly appears brightness and color disproportion. Therefore, the brightness Gamam correction method is used for solving the brightness influence and the white balance based on Grey World Model is introduced for solving the image color disproportion. Test results show that the method adopted can adjust the image quality very well.The face and features location is the precondition of face orientation estimating. At present, there are many methods for face location. These methods can be divided into two types roughly: method based on knowledge and on statistical characteristics. They have their own advantages and shortcomings: the former can detect face rapidly, but its precision is lower than the latter’s. Therefore, this paper uses AdaBoost classifier based on knowledge to detect the possible face ROI in image. In the ROI, the skin color model based on statistical characteristics is adopted to locate face region accurately. For the eye and mouth detection, methods based on grey projection and Fisher linear transformation are used to locate the regions accurately.The face’s contour detection is the key component of face orientation judgment in this thesis. On the basis of the fact that face’s contour looks like an ellipse, a method based on edge point restriction is proposed to fit the outline’s curve. In the process of ellipse fitting, three restraint factors (face geometry restraint, the curvature symmetrical phase restraint and the edge point coordinates restraint) are used. All of the above process steps provide an insurance of the ellipse fitting precision and the face orientation estimating result.Using monocular vision to analysis the face orientation is an estimating method of acquiring face 3D information. Research result shows that the eye and mouth’s position in face region could be changed when the driver’s face occurs deflecting. Based on this fact, with the help of face edge point, eyes and mouth region detection, BP neural net is used to estimate the face’s orientation.The effect of face and features tracking is the key component of the system developed. MeanShift tracking method based on target color characteristics has the merits of quick speed and strong robustness. But it is sensitive to the target moving speed and the object looks like the target in the background. Considering the Kalman filter’s advantages such as simple computation and quick speed, this thesis combines the Kalman filter with MeanShift algorithm to carry on the tracking task. Firstly, Kalman filter is used to forecast the possible target’s region in the image. Then the MeanShift algorithm is adopted to locate the target accurately in the possible area. On the one hand this method enhances the tracking speed; simultaneously it also provides an insurance of tracking accuracy.In summary, many systematic and scientific researches have carried on in this thesis, which are the key technologies in Vision-based Driver’s face orientation monitoring. The achievements not only can be adopted by the product research, but also can provide technical and theoretical support for deep research in SDA field.更多还原