【作者】 徐翠；

【导师】 汪增福；

【作者基本信息】 中国科学技术大学 ， 模式识别与智能系统， 2009， 博士

【摘要】 行车安全是交通运输业的永恒主题。近年来,随着汽车保有量的迅速增长,交通事故的发生越来越频繁,它给人类社会带来的危害也日趋严重。在这种背景下,汽车安全辅助驾驶技术受到广泛的关注。作为减少交通事故、降低事故损失的一种有效手段,它成为交通工程领域的研究前沿,代表了未来车辆发展的趋势。随着国内外汽车安全辅助驾驶研究的深入开展,一部分相关技术正逐步成熟,并开始走向产品化,但仍然有相当多的技术难题有待进一步解决。本文从中国的交通安全现状出发,利用计算机视觉方法对驾驶员疲劳检测和车外行人检测两个关键问题进行了一些有益的探索,以期为我国安全辅助驾驶技术的研发做出积极的贡献.首先对驾驶员眼睛及嘴巴状态检测问题进行了探讨。该研究的目的在于实时捕捉驾驶员长时间的闭眼及打哈欠等与瞌睡行为密切相关的动作。我们以眼睛和嘴巴的外在图像特征为切入点,提出了基于统计学习的疲劳状态检测算法。该算法将状态检测问题看成一个两类分类问题,通过提取感兴趣区域的LBPH-BIN-DIST(LBP直方图的位距离)特征,利用级联AdaBoost分类算法,实现了对闭眼及打哈欠两种状态的监测和区分,为后续疲劳状态的判断打下了坚实的基础。其次研究了驾驶员头部姿态的估计问题。该研究的目的在于捕捉驾驶员长时间的低头、抬头或者频繁点头这些反常的动作。该研究由两部分组成:(1)特征点(眼角及嘴角)的定位(2)基于特征点的头部姿态估计。首先用Harris算子得到候选的特征点,然后根据候选特征点邻域的灰度分布进一步提取特征,并用逻辑回归的方法进行特征融合以使定位更加准确。最后,基于定位的特征点信息,利用POSIT算法对头部姿态进行估计。该算法仅需四个以上的特征点就能得到准确的姿态估计。实验结果表明,只要人脸不发生过大角度的运动(约大于60度),所提出的算法均能正确估计出相应人脸的姿态,并对其运动趋势做出正确描述。最后是基于贝叶斯网络的疲劳状态估计。该研究的目的是根据疲劳所伴随的各种视觉表现对疲劳进行综合建模,并据此对驾驶员的疲劳程度进行判断。为此,选用睡眠质量、驾驶时段及天气状况等常见且重要的因素作为参数,对一段时间内驾驶员异常的面部动作发生的时长或频率进行观测和统计,并利用贝叶斯网络为工具,建立驾驶疲劳模型。该模型充分考虑了各个参数之间的相互依赖关系,保证了疲劳状态估计结果的准确性。此外,采用基于人体配置模型的方法对行人检测问题进行了研究。该研究以监测车辆外部特别是车辆前方的行人为主要目标。首先,用矩形近似人体的四肢和躯干,构建了一个处于行走状态的人体模型。其次,利用形状匹配算法寻找图像中近似矩形的区域作为四肢及躯干的候选,最后用隐马尔可夫解码算法从中寻找满足人体模型的候选区域作为最终检测到的人体。为了加快运算速度和提高匹配准确率,在进行优化处理时,加入了反映人体配置关系和有关人体外表先验知识的一些全局性约束.实验结果表明,我们的方法能够在有效抑制虚警的前提下,实现对车外行人的实时检测和报警。更多还原

【Abstract】 Vehicle driving safety is a persistent issue in transportation industry.In recent years,along with the growth of the amount of cars,the traffic accidents arise more and more frequently,bringing huge damages to the society.Under this circumstance,Vehicle safety driving assistant technique has drawn great attentions.It is a frontier research topic,and also a effective measurement for preventing the happening and damage of traffic accident.Although a few of the vehicle driving-assistant techniques have been used in practical systems,most of the techniques remain immature.This dissertation researches the vehicle driving assistant techniques,mainly on the following two topics:driver fatigue detection,and outside-car pedestrian detection.The purpose of eye and mouth detection is to detect the closed eye and yawning mouth which are two major features of fatigue.According to the different appearance of different eye and mouth statuses,we propose a statical learning based method.We treat the status detection problem as a classification problem which is to recognize the abnormal status(closed eye or yawning) from eye and mouth images.We propose to use LBP Histogram Bin Distance(LBPH-BIN-DIST) as an effective feature extraction method.Based on this feature,we construct a cascade AdaBoost classifier to detect the abnormal status.The proposed method is highly accurate and fast.The purpose of head pose estimation is to detect the long-time lower or higher head poses,or frequent nodding.These affairs are also an important clue of fatigue.The research consists of two steps.The first step is to accurately locate several facial feature points,such as eye and mouth corners.Then,the second step uses these points for pose estimation.The detection of feature points is based on the grayscale distribution of the neighborhood around these points.We utilize the distribution to extract features, and use logistic regression to fuse several features,and give the final estimation of the feature points.Once we have achieved the location of these feature points,we can use POSIT algorithm for pose estimation.To fuse the available fatigue clues,such as eye and mouth status the head pose, also to integrate other external factors that will lead to driver fatigue,such as sleeping quality,time and weather,we use Bayes network to fuse all of these information and to get a final estimation of the probability of fatigue.The fatigue estimation model built by Bayes network considers the correlation and the conditional probabilities of different factors,thus can give more reasonable and overall judgement of fatigue.Besides the fatigue detection,the outside-car pedestrian detection is another important issue in driving-assistant research.The aim is to monitor the pedestrian activity, and prevent the potential accidents against pedestrians.Regarding this problem,we use a method based on the model of human body configuration.We first use rectangular to simulate the torso and limbs,and construct a walking human body model.Then,we use chamfer matching algorithm to find the potential human body regions in the images. The algorithm will give several candidates for human body.To identify the real human body,we use hidden Markov model(HMM) decoding algorithm.Some priors about human body are also added into the model as global constraints.The experiment shows that our method can effective detect pedestrian with low false-alarm rate.更多还原