驾驶员疲劳特征提取方法的研究及检测系统的设计

发布时间：2018-07-27 12:06

【摘要】：在经济与科技发展的推动作用下,越来越多的上班族以汽车作为交通工具,这样既方便了出行又节省了时间。但是,道路上出现的事故伴随着汽车的普遍应用也出现了井喷式的增长。根据调查研究与科学统计,在发生交通事故的各种原因中,疲劳驾驶成为主要原因。尽管如此,由于检测技术及软件处理的不成熟与不完善,使得检测驾驶员疲劳程度的产品还不能得到广泛的引用,从而没能减轻甚至杜绝类似交通事故的发生,造成了不可估量的生命与财产的损失。本文以此为背景,并对行车过程中出现疲劳状况时的面部表现进行观察和分析,提取出疲劳面部信息,结合面部检测及多特征处理技术,设计了能检测疲劳状态的系统,并在系统上完成了算法验证。此系统包括人脸检测、疲劳特征提取和疲劳状态判断三个部分。其中,提取的面部特征有三个:眼睛特征、头部特征和嘴部特征。文中首先对系统的需求和工作流程进行了分析,然后对各部分采用的算法进行了论述。在检测之初,本文先采用了Adaboost算法,在计算出人脸Haar特征之后用于人脸的检测。在算法的使用中发现,其速度和对倾斜时人脸的检测有很多不足。因此,在后期引入了KCF跟踪算法,并把这两者进行结合,在应用对比后发现,此改进无论在时间上还是对倾斜人脸的检测上,都有相当大的提升。然后,论述了眼睛特征提取、头部特征提取和嘴部特征提取的方法:在提取眼睛状态的时候,首先根据先验知识确定眼睛子窗口,进行二值化处理之后利用灰度积分投影分别提取出单个眼睛窗口,并求出各眼睛区域的最小的外接矩形,以矩形面积大小确定眼睛状态;提取嘴部状态时,根据先验知识提取出鼻孔和嘴部区域,并针对传统Canny边缘检测存在的不足采用自适应边缘检测算法得到鼻孔和嘴部轮廓,根据鼻孔到上下嘴唇间距离的比值确定出嘴巴的状态;对头部特征的确定则是根据人脸矩形框坐标位置的变化得出的。在对疲劳特征提取算法进行论述之后,介绍了疲劳状态判定的方法,针对利用单特征进行疲劳检测存在的不足,提出融合眼睛特征、嘴部特征和头部特征进行检测的方法,该方法提高了检测的正确率。本文在实验室通过VS开发环境设计了疲劳驾驶检测系统,并在此系统上分别对采用单特征的疲劳检测和多特征融合的疲劳检测方法进行实验验证,结果表明:本文提出的融合多特征的疲劳检测方法在准确率上有了明显的提高。
[Abstract]:Driven by the development of economy and technology, more and more commuters use cars as means of transportation, which not only facilitates travel but also saves time. However, accidents on the road accompanied by the widespread use of cars also appeared blowout growth. According to investigation and scientific statistics, fatigue driving is the main cause of traffic accidents. However, due to the immaturity and imperfection of detection technology and software processing, the products used to detect drivers' fatigue degree can not be widely used, which can not reduce or even eliminate the occurrence of similar traffic accidents. Resulting in incalculable loss of life and property. Based on this background, this paper observed and analyzed the facial performance of fatigue state in the course of driving, extracted the fatigue facial information, combined with facial detection and multi-feature processing technology, designed a system to detect fatigue state. The algorithm is verified on the system. The system includes face detection, fatigue feature extraction and fatigue state judgment. Among them, there are three facial features extracted: eye features, head features and mouth features. In this paper, the requirements and workflow of the system are analyzed firstly, and then the algorithms used in each part are discussed. At the beginning of detection, Adaboost algorithm is used to detect face after Haar feature is calculated. In the use of the algorithm, it is found that there are many shortcomings in the speed and face detection of tilt. Therefore, the KCF tracking algorithm is introduced in the later period, and the two algorithms are combined. After application comparison, it is found that the improvement has a considerable improvement in both the time and the tilt face detection. Then, the methods of eye feature extraction, head feature extraction and mouth feature extraction are discussed. After binary processing, the single eye window is extracted by gray integral projection, and the smallest external rectangle of each eye region is obtained. The state of the eye is determined by the size of the rectangle area, and the state of the mouth is extracted. According to the prior knowledge, the nostril and mouth regions are extracted, and the adaptive edge detection algorithm is adopted to get the nostril and mouth contours according to the shortcomings of traditional Canny edge detection, and the state of the mouth is determined according to the ratio of the distance between the nostrils and the upper and lower lips. The determination of the head feature is based on the change of the coordinate position of the face rectangular frame. After discussing the algorithm of fatigue feature extraction, the method of fatigue state determination is introduced. Aiming at the shortcomings of fatigue detection using single feature, the method of combining eye feature, mouth feature and head feature is put forward. This method improves the accuracy of detection. In this paper, the fatigue driving detection system is designed in the lab by using vs development environment, and the fatigue detection method using single feature and multi-feature fusion is verified by experiments on this system. The results show that the proposed fatigue detection method with multiple features has a better accuracy.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U463.6;U495;TP391.41

【参考文献】