【作者】 童兵亮；

【导师】 王荣本；

【作者基本信息】 吉林大学 ， 载运工具运用工程， 2004， 硕士

【摘要】 安全是汽车交通发展的永恒主题，随着汽车保有量的迅速增加，公路上的交通事故，特别是恶性交通事故发生率居高不下，交通安全问题日益突出。在这种情况下，作为智能车辆的关键技术之一，安全辅助驾驶技术受到人们日益关注。安全辅助驾驶的相关技术为解决常规车辆因驾驶员主观因素产生的交通事故提供了有力的技术支持。美、欧、日等西方发达国家已经抢先一步，在汽车安全辅助驾驶技术的研究方面相继投入大量人力、物力，取得了许多有价值的研究成果，在乘用车、重型卡车、公共交通汽车及特殊车辆上进行了汽车安全辅助驾驶系统的研究和应用。驾驶员人为因素已经成为交通事故发生的主要因素之一。驾驶员监测已成为当前安全辅助驾驶技术的一个研究热点。机器视觉在实时性、准确性、适用性及经济性等方面具有比其他监测方法更大的优势。许多研究者利用车载摄像机系统进行驾驶员视觉监测技术的研究。目前许多研究者集中于通过跟踪驾驶员的人脸，眼睛，瞳孔等，得到头部转动和方向，眼睑运动，眨眼频率，注意力方向等监测驾驶员疲劳驾驶或精神分散状态。然而，驾驶员打哈欠，长时间与他人说话或打手机通话等疲劳驾驶或驾驶精神分散状态没有得到人们的重视。本文通过一个车载CCD摄像机对驾驶员面部状态进行实时监测，在国内首次提出了利用机器视觉识别驾驶员嘴部状态进而判别驾驶员打哈欠疲劳驾驶和与他人说话或打手机通话等驾驶精神分散状态的方法，进一步拓展了驾驶员监测技术的涵盖范围，也为驾驶员监测系统的综合信息监测技术提供参考和支持。很显然，疲劳驾驶和驾驶精神分散状态监测系统对于降低交通事故率有着重要的作用。论文的研究内容包括四个方面，即驾驶员人脸检测、驾驶员嘴部定位与跟踪、驾驶员嘴部状态识别、驾驶员疲劳驾驶和驾驶精神分散状态判别。在驾驶员人脸检测方面，本文利用了人脸皮肤颜色模型的驾驶员人脸检测方法。不同人的皮肤颜色在YCrCb颜色空间中色度信息Cr、Cb具有一定的分布特性，虽然不同人的皮肤颜色有差异，但它们在Cr、Cb色度上的差异远<WP=94>小于亮度上的差异，也就是说，不同人的肤色在色度上往往很相近，只是在亮度上差异较大。实验证明：不同的肤色Cr、Cb具有相同的二维高斯模型。本文利用人脸的皮肤颜色在YCrCb颜色空间的分布特点，得到了一种快速、有效的在复杂背景下人脸检测方法。实验结果表明，该种驾驶员人脸检测方法可靠性高，具有良好的动态定位能力，而且对驾驶员坐姿的变化有较好的适应能力。在驾驶员嘴部定位和跟踪方面，嘴唇的主要颜色特征是唇色相对肤色颜色较红，而且归一化RGB颜色向量对光照和人脸运动和旋转具有不变性，利用Fisher线性变换得到肤色和唇色rgb颜色向量的最佳投影方向。将肤色和唇色颜色向量投影到该方向上后，肤色和唇色能够很好地区分出来。采用该方法不仅使唇色和肤色能够区分开而且嘴唇轮廓有较明显的边界，提高嘴唇检测、定位的准确性。在嘴唇唇色分割的基础上，本文利用连通成分标示算法和嘴部区域几何约束进行驾驶员嘴部定位，利用卡尔曼滤波跟踪算法和假设约束对驾驶员嘴部进行跟踪。连通成分标示算法可将驾驶员嘴部感兴趣区域的图像与唇色相似的各个独立区域标示出来，结合脸部下半部分各器官的几何特征即可将驾驶员嘴部定位，首先对获取的人脸下半部分区域即嘴部感兴趣区域的图象进行嘴唇分割处理，得到嘴部感兴趣区域内与唇色相似且相互孤立的几个区域。然后利用连通成分标示算法对这几个孤立的区域进行区域连通标示，进而获取各孤立区域的各种参数，如坐标位置、区域面积等。然后利用前文所讲到的几何约束条件对这些孤立区域进行判断，从中选取最符合约束条件的区域作为驾驶员嘴部区域。为了持续不断地监测嘴部，实时每帧地跟踪嘴部是重要的。这可以通过使用预测和定位的方案有效地解决这种问题。预测包括根据嘴部当前的位置信息确定下一帧图像嘴的近似位置。定位即通过局部搜索精确地定位嘴部的位置。根据卡尔曼滤波理论，本文建立嘴部状态方程和观测方程，第帧图像嘴部中心点的当前估计值决定和嘴部区域搜索窗口大小由协方差矩阵值决定，这样在此范围内经过嘴唇分割得到嘴部区域。在驾驶员嘴部状态识别方面，本文提出了利用图像投影定位驾驶员嘴唇特征点的算法，根据嘴部区域几何特征值进行BP神经网络识别驾驶员不说话时的嘴闭合、说话时的普通张嘴、打呵欠时的大张嘴的三种不同的嘴部状态。在嘴唇分割和定位的基础上，利用垂直投影得到左右嘴角，利用水平投影得到上嘴唇中心最上点、上嘴唇中心最下点、下嘴唇中心最上点、下嘴唇中<WP=95>心最下点。本文根据驾驶员三种不同的嘴部状态，嘴部区域的最大宽度、嘴部区域的最大高度、 上、下嘴唇之间的高度值不同，将这三个嘴部区域的几何特征值作为BP神经网络的输入向量。BP神经网络是应用最普遍的一种人工神经网络，目前其应用实例约占神经网络应用实例的80%，它已成为人工神经网络的经典代表。本文将采用BP神经网络来识别驾驶员的嘴部状态。该BP神经网络为三层结构，输入层有3个神经元，隐层有14个神经元，输出层有3个神经元，代表驾驶员三种更多还原

【Abstract】 Safety is always one of the eternal topics in vehicle Transportation. With the rapid development of transportation, traffic accidents are greatly increasing, especially including traffic fatality. Safety problem in transportation is paid more and more attention worldwide. Under this circumstance, Safety Driving Assist technologies, as a part of Intelligent Vehicle technologies, are paid more and more attention and it can support greatly for reducing the road accidents due to drivers’ human factors. The developed countries, such as the U.S., U.K., Japan, Germany etc., have made their research programs to develop Safety Driving Assist technologies and have already achieved great progress. Some systems about Safety Driving Assist technologies have been applied in the passenger car, heavy truck, bus and public transport and special vehicle.Driver’s human factors have been one of the most important causes of road accidents. Driver monitoring has been a focus of Safety Driving Assist technologies research. The Driver monitoring method of Machine vision has better advantage than other method on time, accuracy, adaptability and economical respects. Therefore, most of researchers have studied the driver’s monitoring method based on machine vision by vehicle-mounted cameras. Until now many research have focused on monitoring the driver’s face, eye, pupil and so on to obtain his/her face rotation and orientation, eye activities, eye blinking rate, gaze direction, finally to determine his/her fatigue or distraction state. However, Most of researchers have neglected driver’s fatigue state such as driver’s yawning and his/her distraction like conservation and talking on a cellular phone while his/her driving. This paper presents a method for real-time monitoring a driver’s mouth <WP=97>state by one vehicle-mounted camera, and monitoring driver’s fatigue state such as driver’s yawning and his/her distraction like conservation and talking on a cellular phone while his/her driving by recognizing his/her mouth state based on machine vision first in our country, extends the Safety Driving Assist technologies, and provides the reference and support for the integrated driver monitoring technologies. Obviously, driver’s fatigue and distraction warning system take important role on reducing accident rate.The research work in this paper include the four parts, i.e. driver’s face detection, driver’s mouth detection and tracking, driver’s mouth state recognition and driver’s fatigue and distraction state identification.Driver’s face detection applies the human face skin color model. The color distribution of skin colors of different people was found to be clustered in a small area of the chromatic YCrCb color space. Although skin colors of different people appear to vary over a wide range, they differ much less in Cr，Cb chroma than in brightness. In other words, skin colors of different people are very close, but they differ mainly in intensities. Results showed that skin color Cr，Cb chroma distribution of different people can be represented by a Gaussian model. This paper proposed a fast, available face detection method under different background using the human skin color property in YCrCb color space. Experiment results showed that this method is much reliable and adaptable to driver’s different pose. Driver’s mouth detection starts with the driver’s lip pixels segmentation. Lip pixels are prevalently redder than skin pixels, not pure red. And normalized RGB is invariant to changes of the light source, face orientation and rotation. In normalized RGB color space Fisher transformation is used to determine an optimal projection vector , onto which rgb color vector data of lip and skin pixels are distinguished. This method segments lip and skin pixels, keeps the distinct lip contour boundary and increases lip detection accuracy. This paper locates driver’s mouth by connected component analysis and region geometric constraint. Kalman filter is used to track Driver’s mouth.Connected component analysis algorith更多还原