【作者】 张伟；

【导师】 成波；

【作者基本信息】 清华大学 ， 机械工程， 2011， 博士

【摘要】 疲劳驾驶是造成交通事故的重要原因之一。基于机器视觉技术通过对驾驶人面部表情特征的分析可实现疲劳状态的有效估计。由于该方法具有非侵入、准确、实时的特点而成为疲劳驾驶在线辨识中最具潜力的技术手段。然而，受实际行车环境中光照条件的复杂性、驾驶人面部姿态的不确定性、疲劳表征的隐匿性、驾驶人的个体差异性等诸因素影响，高鲁棒、全天候的的驾驶人疲劳状态在线辨识仍存在众多技术瓶颈。本文围绕光照与姿态变化条件下眼部特征的定位提取、驾驶人姿态估计与校准、疲劳特征空间建模及疲劳模式推断等核心问题展开研究，开发了可适用于实际交通环境的疲劳驾驶实时辨识系统并进行了实验验证。论文深入分析了实际行车环境中光照条件、驾驶人姿态变化对眼睛定位算法适应性的影响，建立了基于层叠式形状模型和自商图局部纹理模型的主动形状模型算法，实现了眼睛局部邻域的有效分割和可靠跟踪。在此基础上，充分利用自商图、色度以及梯度的统计分布等光照不变特征，建立了光照不变量约束下的参数化模板算法，实现了眼睛轮廓的精确定位。另外，设计了采用偏振光照明的双光谱互补照明光路，有效解决了夜晚在辅助光源照明下眼镜片的反光问题。充分考虑了行车过程中驾驶环境的时域稳定性，提出了一种综合利用机器学习、在线自适应肤色建模、纯背景建模技术的面部区域分割算法，并通过对面部区域内角点的跟踪，基于外极线约束方程建立了驾驶人相对姿态角解算模型。同时，基于Candide模型实现了驾驶人头部的个体三维重建，并通过三维模型配准完成了驾驶人初始姿态角的确定。采用统计学方法分析论证了不同疲劳水平下眼睛动作参数差异的显著性，建立了基于眼睛动作特征的疲劳特征空间，并模拟人的认知过程，提出了在驾驶任务初期采用基于训练样本得到的先验知识对疲劳模式进行分类，并在自学习基础上基于贝叶斯置信网络对驾驶人疲劳状态进行推断的辨识方法。更多还原

【Abstract】 Decreased vehicle control due to driver drowsiness is one of the major causes ofroad accidents. Computer vision based methods have shown the possibility ofdrowsiness detection through driver eye movement analysis. Camera monitoring of agiven driver’s eye status has proved to be the most promising technology due to goodaccuracy, real-time performance and non-intrusiveness. However, there are still manychallenges posed by illumination, driver postures and unapparent facial appearancechanges when a driver becomes drowsy. This paper focuses on the key issues in highaccuracy contour extraction methods across illumination and face orientation,orientation estimation and registration, drowsiness feature space modeling anddrowsiness state inferring. Real on-road experiments are performed to testify theaccuracy and robustness of the proposed methods as well.According to a thorough analysis on the influence of dynamic illumination andorientation in eye location, an improved Active Shape Model (ASM) involving twocontributions is introduced for face alignment. First, a novel local texture modelmaximizes the ASM tolerance to illumination changes by learning from theSelf-Quotient image instead of the original image. Second, a cascading overall shapemodel is proposed to enhance ASM orientation adaptability. On the basis of facialfeature alignment using ASM, a more precise eye contour location parameterizationmodel is performed by introducing some illumination insensitive features such aschromaticity information and gradient distribution characteristics. Furthermore, a twolayer cascaded illumination system is presented to eliminate reflections ofglasses. A polarized lighting method is adopted in the first layer, and a doublechannel narrow band multi-spectral imaging system is set up in the secondlayer.Based on the assumption that the features extracted from sequential imagesresemble each other, this paper presents a face detection algorithm which combines alearning-based approach with adaptive skin color segmentation and background modeling methods. The driver’s attitude angles are calculated by tracking corners inthe facial region. Moreover, the initial attitude angles are determined by matching thethree dimensional model of the driver’s head, which is reconstructed from acombination of the Candide model and the driver’s facial image.The changes of each measure with varying drowsiness levels were comparedwith the analysis of variance (ANOVA), and those with statistically significantdifferences are introduced into the drowsiness feature space. Imitating cognitivebehavior of human beings, an identification method is proposed where prioriknowledge obtained from train data sets are used to classify different drowsinessstates during the initial phase of a driving task and then Bayesian networks areintroduced to drowsiness assessment after a period of real-time learning.更多还原