【作者】 李玲玲；

【导师】 陈阳舟；

【作者基本信息】 北京工业大学 ， 模式识别与智能系统， 2010， 硕士

【摘要】 驾驶员疲劳驾驶是造成交通死亡事故的重要原因之一,驾驶疲劳检测已成为智能运输系统(ITS)研究的热点之一。本文首先对基于计算机视觉的驾驶员疲劳检测的研究现状进行了总结和分析,并提出一种基于多视觉信息融合的疲劳检测方法。该方法采用双摄像机联合定位跟踪面部,有效地提高了面部信息采集的精度,然后通过一系列的脸部特征提取和跟踪的算法实时获取眼睛、嘴部、头部运动等多种与疲劳程度相关的重要视觉特征信息;最后提出一种改进的贝叶斯算法融合多视觉信息来估计驾驶员的疲劳程度。本文研究的核心内容包括:基于双摄像机的人脸定位跟踪算法研究与实现;眼睛和嘴部特征实时检测与跟踪算法研究与实现;多种视觉疲劳特征信息提取、贝叶斯多信息融合等算法的实现与改进。首先提出了一种基于双摄像机的人脸定位跟踪算法。本文通过两个摄像机实时地采集驾驶员视频图像,其中固定摄像机A用于拍摄驾驶员上半身,可控摄像机B用于跟踪拍摄驾驶员面部,并提出了一种基于肤色区域分割与人脸验证相结合的人脸初步定位方法和基于CAM Shift跟踪与人脸验证相结合的跟踪方法从摄像机A采集的图像中快速定位和跟踪脸部位置。系统根据摄像机A中定位的脸部位置信息,通过串口发送云台控制指令控制高速可控摄像机B实时转动,跟踪拍摄驾驶员头部。实验结果显示基于双摄像机的人脸定位跟踪算法相比单摄像机算法可以有效地提高脸部图像的采集精度,获得大分辨率的脸部图像,以便更精确地提取脸部疲劳特征信息。同时系统还可以在对摄像机A所得视频的处理过程中获取驾驶员的头部运动信息。除了获取头部运动信息,还改进和提出了一系列眼睛及嘴部特征实时检测与跟踪的算法,提取从摄像机B所得视频中实时地采集驾驶员的眼睛和嘴部状态信息作为疲劳程度估计的主要特征。本文提出了一种改进的粒子滤波人眼定位及跟踪方法,首先采用了基于Haar-Like特征级联分类器的检测方法,在人脸区域内按照不同尺度搜索存在的眼睛,然后通过检测到的眼睛位置初始化粒子滤波算法。为了提高算法的准确性,减小环境噪声和干扰的影响,本文提出了一种一阶滤波算法来对跟踪结果进行再次滤波修正,并根据人脸的位置信息和眼睛定位的历史信息来判断眼睛是否跟踪丢失。若跟踪丢失则采用Haar-Like特征级联分类器重新检测眼睛位置。嘴部定位采用三停五眼的方法。此方法简单,效率较高。其原理是根据眼睛的位置及大小确定嘴部的相对位置。本文选取了多种视觉特征作为疲劳估计特征,包括眼睛闭合程度、眨眼频率、打哈欠频率和点头频率。本文分析了相关的视觉疲劳特征计算方法,并提出一种判断眼睛及嘴部开合状态的新方法。其原理是根据眼睛或嘴部的外接矩形宽高比及面积判断眼睛及嘴部状态。与以往的算法相比该方法可以更精确地表示眼睛及嘴部的状态。该方法是获得眼睛闭合程度、眨眼频率、打哈欠频率的基础算法。点头频率特征是通过跟踪头部运动获得的。最后,采用贝叶斯算法对以上疲劳特征进行融合,计算驾驶员疲劳程度。由于动态贝叶斯网络中转移概率较难获取,本文提出一种概率更新的方法实现了对疲劳特征的动态贝叶斯融合,并在室内环境对算法进行仿真测试,取得了良好的实验结果。更多还原

【Abstract】 Fatigue driving is one of the major causes of traffic accident, and driver fatigue detection has become hot spots of Intelligent Transportation Systems (ITS). This paper first summarizes and analyzes the current reaearch of driver fatigue detection that based on computer vision and proposes a new method using multi-visual information fusion. This method first proposed a two-camera co-location and tracking facial approach that effectively improve the accuracy of facial information collection. Then obtain real-time eyes, mouth, head movement and fatigue-related visual feature information through a series of facial feature extraction and tracking algorithm. Finally, we propose an improved Bayesian algorithm to estimate driver’s fatigue over integration of visual information.The core study of this paper include: research of dual-camera-based human face location and tracking algorithm and implementation, eyes and mouth features real-time detection and tracking algorithm and implementation, a variety of visual fatigue characteristic information extraction, implementation and improvement of multi-information fusion algorithm for Bayesian.This paper first presents a two-camera-based face location and tracking algorithm. We use two cameras to get real-time video of the driver, including a fixed camera which is used to shoot driver’s upper-body and a controllable camera which is used to shoot driver’s face. Then we propose a face positioning method that combines skin color segmentation and face validation and a tracking method that bases on CAM Shift and face validation to quickly locate and track the face location from camera A.Then the system sends PTZ control commands through serial port according to the face position in camera A to make the high-speed controllable camera B rotate in real time and follow the driver’s head. The results showed that this method can effectively improve the accuracy of facial image collection, access to a large solution of facial image and also could more precisely extract facial fatigue information compare to the usual method. At the same time, we could obtain the driver’s head movement information during the processing of camera A video images.In the paper, not only get head movement information, we also propose and improve some real-time detection and tracking algorithms of eyes and mouth’s features which collect eyes and mouth’s status as the main features of the estimated fatigue from the real-time video of camera B. This paper presents a eye’s locating and tracking method based on particle filter. Firstly, we detect eyes based on haar-like features in face region, then initialize particle filter through the detected eyes. In order to improve the algorithm’s accuracy and reduce the impact of environmental noise and interference, this paper presents a first-order filtering algorithm to amend the track result and determine whether the eye tracking is lost according to the history face position and eye’s position. If it is lost, we will detect eyes with haar-like features and re-initialize particle filter.Mouth locating is according to face distribution. This method is fast and easy to realize. The principle is to locate mouth region according to eye’s size and position.This paper selects a variety of visual features as the fatigue characteristics, including Perclos, blinking frequency, yawning frequency and nodding frequency. We also analyze some related visual fatigue calculation method and propose a new method judging the opening and closing state of eyes and mouth. The principle is based on the external rectangle ratio and area of eyes and mouth. Comjpared with the other methods, this method could more accurately represent the state of the eyes and mouth. This method is the foundation method of obtaining Perclos, blinking frequency and yawning frequency. Nodding frequency is obtained by tracking head movements.Finally, we use Bayesian method to integrate the fatigue features that metioned above. As the transition probability of dynamic Bayesian networks is difficult to obtain, we propose a probability updating method to integrate the fatigue features based on Bayesian. And we achieved good results in the indoor environment simulation.更多还原