【作者】 路玉峰；

【导师】 王增才；

【作者基本信息】 山东大学 ， 车辆工程， 2008， 博士

【摘要】 驾驶员视觉分散是众多交通事故的诱因,并且随着车载信息系统的增加,将会引起驾驶员越来越多的视觉分散行为,从而引发更多的交通事故。检测驾驶员视觉分散并警告驾驶员可减少类似原因造成的交通事故。本文从视觉分散对驾驶性能的影响研究入手,开展驾驶员视觉分散检测技术的研究,并重点研究基于视频图像分析的驾驶员视觉分散特征的提取方法。主要工作如下:首先研究视觉分散对驾驶性能的影响。设计实验让驾驶员阅读4处位置上的2类文本信息,使其产生8种不同的视觉分散。分析不同视觉分散时车辆的SDLP(车辆偏离道路中心距离的标准差),发现SDLP随着视线偏离道路中心角度的增大而增大。研究视觉分散影响驾驶员的机理,分析基于脑电、皮电、行为等检测驾驶员视觉分散的方法。根据驾驶过程中驾驶员视线变化的特点,建立基于驾驶员面部姿势估计与眼睛视线方向识别,并包含转向行为识别的视觉分散检测模型。然后对驾驶员视觉分散的特征进行提取研究。研究多姿势下驾驶员面部、面部特征点精确定位的方法。研究利用肤色混合高斯模型预定位人脸区域,然后根据眉毛、嘴唇位置精确定位驾驶员面部的方法。针对眉毛区域灰度值低、变化剧烈的特点,研究基于联合投影函数定位驾驶员眉毛上边缘的方法。研究背景滤除的方法,克服面部横摆角度较大时眉毛定位不准的缺点。研究利用唇色多项式模型及嘴唇比人脸肤色更红的特点定位驾驶员嘴唇区域。研究驾驶员面部图像归一化的方法。研究驾驶员面部姿势的提取方法,提出利用核主元分析估计驾驶员面部姿势的方法。分析核主元分析实现原理,研究利用核主元分析估计面部姿势的步骤。研究获取标准样本图像的方法,设计样本图像采集系统。利用核主元分析把高维面部图像存在的流形结构嵌入到二维空间,建立估计面部姿势的标准曲线,并根据姿势曲线拟合圆。提出利用拟合圆心及姿势曲线上距新投影点最近的两个点,来估计新投影图像对应角度的方法。该方法克服传统模式分类方法需要为不同人建立不同姿势曲线的缺点,并且估计精度可满足一定实际需要。研究不同核函数、核函数参数对估计精度的影响。研究驾驶员眼睛视线方向提取方法,提出基于Multi-PCA(多主元分析)的眼睛视线方向识别方法。研究PCA实现原理,分析常用PCA应用于识别时存在的问题。针对驾驶环境中精确提取视线方向的困难,把视线方向分为5类(上、下、左、右、前),为每类建立特征空间,通过主元分析提取每类视线的共同统计特征,然后根据测试样本在每类特征空间下的重构误差进行分类。该方法充分运用了PCA变换的最佳逼近性能,并提取了单类眼睛视线图像的独有特征,实验证明该方法可以获得比常用PCA方法更高的识别率。研究驾驶员转向行为识别,提出根据手部位置标准差来识别驾驶员转向行为的方法。驾驶员在十字路口处的转向过程中,视线方向偏离车辆前方的时间将超过2秒钟,检测系统会误认为是视觉分散,因此需要检测驾驶员的转向行为以减少这种误判。研究基于视频分析的驾驶员双手定位方法,并研究基于粒子滤波算法的驾驶员手部跟踪方法以提高双手定位的实时性。研究驾驶过程中驾驶员双手位置变化的特点,根据转向过程中驾驶员手部位置变化剧烈的特性,提出利用双手位置标准差识别驾驶员转向行为的方法。最后根据视觉分散检测模型,建立驾驶员视觉分散检测系统的软、硬件框架,并进行视觉分散检测的实验研究。摄取行驶过程中驾驶员观察仪表盘、调节收音机、十字路口转向时的手、面部视频图像,利用本文提出的识别算法提取驾驶员视觉分散特征。对本文提出的视觉分散检测算法进行验证,实验表明算法可行,并能有效防止检测系统在驾驶员转向时发出虚警的现象。更多还原

【Abstract】 Many studies reveal that visual distraction degrades driving performance and it is the main reason of traffic accidents. The distraction problem may be expanded in the near future, as many drivers will use an increasing number of electronic devices such as cell phones, navigation systems, and wireless Internet. A detection system that can predict distraction and alert the driver by monitoring driver faces’ features could reduce the number of distraction crashes.In this thesis we start with analyzing the effects of visual distraction on driving performance. Then the study of visual distraction detection is developed based on video analysis. And we focus on developing methods to extract features of visual distraction based on video analyzing. In this paper the main work is as follows:First, effects of visual distraction on driving performance have been analyzed. An experiment is designed to test the effects of visual distractions in different positions have on driver. The SDLPs (standard deviation of lane positions) are analyzed when driver read two kinds of texts in four different positions. We conclude that the effects degree is deeper as the visual distraction departure father. Methods of detecting driver visual distraction are studied. A driver visual distraction detection model is constructed based on visual line. And turn activity recognition is embraced in this model.Then face detection and face features location has been studied under variety face pose. We propose to predict face region using mixture-of-Gaussians modeling of face color first. And precise face region is located base on eyebrow and lips locating. We propose to locate eyebrow using combine projection function because intensity of eyebrow is low and change acutely. A method to get rid of background is proposed to enhance precision of eyebrow locating when face turn to one side. The lips’ region is located based on color quadratic polynomial model. The information that lips color is redder than face’s is used to locate lips’ region further. The normalize method which is fit for resizing face image has also been studied in this paper.Kernel Principal Component Analysis (KPCA) is proposed to estimate driver face pose. Getting the standard face images with exact pose is the first step, so we design an images collection system. The manifold in high dimension face images can be embodied into low dimension space. So a standard curve to estimate head pose is constructed. A circle is fit using face pose curve. A new sample’s pose can be calculated using the centre of the circle and the two points which are nearest to the new point in pose curve. The kernel functions and their parameters’ effects on pose angle estimation precision are also studied.A method based on Multi-PCA (Principal Component Analysis) to recognize eye gaze direction has been proposed in this paper. The principle of PCA is studied and its shortage is analyzed too. First, the eye gaze direction is classified into five kinds and feature space is constructed for every kind. Then, the common statistical features of each space are extracted. Finally, the test samples are classified based on their reconstructing errors in different feature spaces. The experiments show that Multi-PCA gets a higher recognition rate than PCA because the method takes full advantage of sole features.Driver gaze will departure the front of vehicle for more than two seconds when driver turns at the intersections of the ways. So the driver visual distraction detection system will take this as visual distraction wrongly. A method is proposed to recognize driver turn activities to reduce percents of negative alarm. The standard deviation of driver hands’ positions is used. Particle filter based tracking method is proposed to enhance the speed of locating driver hands.Last a visual distraction detection system comprising hardware and software has been constructed and detecting experiments are carried out. Videos of driver monitoring the panel, adjusting radio, turning in the intersection are getten when driver is on the road. The features of visual distraction are extracted using methods proposed in this thesis. The experiments are carried out to test visual distraction detecting method. The experiments result that the method is effective. And the method can avoid recognizing turn activities as visual distraction.更多还原