【作者】 吴法；

【导师】 孔德兴；

【作者基本信息】 浙江大学 ， 应用数学， 2013， 博士

【摘要】 本文将图像处理和模式识别应用于智能交通领域,这对日益严峻的交通问题有着重大意义,随着新增的摄像头的增加,执法监控能力也会得到进一步的加强,目前车牌识别,闯红灯识别,超速驾驶识别等均已实现自动化处理,而更深入的识别如边开车边打手机,未系安全带驾驶,车型识别等才刚刚起步,而其中对于佩戴安全带是交通法规明文规定的,并且最有效的降低交通事故中的伤亡率的措施。本文设计了一套安全带佩戴识别系统,完整阐述了了安全带处理的整个流程,经过大量实验并投入使用中,得到非常满意的效果。全文的主要内容共分为三个部分：1)基于颜色和纹理的蓝色车牌检测和基于机器学习的车窗定位对安全带的监测,首先要定位车窗,而车窗与车牌有相对固定的位置关系,车牌定位又相对简单,因此本文采用逆向思维考虑,通过定位简单图像,进而最后辅助定位困难图像,也就是先定位车牌再定位车窗的方法。将车牌颜色投影到HSL色彩空间,对空间中车牌颜色可能出现的区域进行了细致的刻画,从而过滤掉图片的大多数区域,然后采取sobel算子的横向极大值计数的方法,抽取区域的纹理特征。对于车窗定位,选取车窗一角用统计学习的方法,得到车窗检测器,以车牌为基准偏移一定方位后对区域作adaboost判别,选取响应最大的位置坐标作为车窗角点坐标。通过以上步骤,实现了车牌监测和车窗定位,为后续的安全带的位置监测奠定了基础。2)基于Canny边缘检测及概率Hough变换(PPHT)的安全带边缘检测大部分的图片样本的安全带边缘是清晰可见,如何快速筛检出这些清晰的样本是检测的第一步,本文将Cann边缘检测的信息加以梯度方向的筛选,再根据检出的线段夹角,距离等信息判断是否为安全带边缘,然后更进一步的,本文改进了Hough圆检测,将其变为双上半圆弧检测,细化直线的方位信息,从而达到更快速度地对清晰可见安全带边缘的图片样本进行筛检的效果,大幅提升了计算速度和效率。3)基于adaboost统计学习算法的安全带检测非统计的算法对于纹理、遮挡等干扰极其敏感,而约有三分之一以上的图片或多或少有些于扰。例如前车窗的反光造成的区域屏蔽；衣服的纹理形成的视觉识别困难；手臂的遮挡影响机器的图形提取。因此本文引入了机器学习的算法,通过构造两种截然不同的样本集,通过同一种算法,得到两个分类器,具有不一样的测试结果,经过大量实验,证实两者共同使用可以达到更好的效果。更多还原

【Abstract】 Patten recognition applied in intelligent traffic system significantly help peo-ple soft traffic problem caused by increasingly large transport network and the large number of new camera.Recently, license plate recognition, through a red light, speeding, etc. have been identified to automate processing.The more in-depth identification such as a using cell phone while driving, not wearing a seat-belt driving detection, vehicle identification has just begun. This paper describes the entire process of handling not wearing a seatbelt driving detection. Paper is divided into three parts:1. The blue license plate detection based on the color and texture informa-tion and locating windows based on machine learning method.To detect belts, we must first locate the windows, and the windows and license plates have a relatively fixed position relationship.plate location is rela-tively simple, so we took a first positioning plate reorientation window approach. The color plates projected onto HSL color space. The area where color plates may sitting were detailed characterization, which filter out most of the picture area, and then take the Sobel operator horizontal maxima counting method to extract the region texture features. For the window locating process, using statistical learning methods to obtain the window detector.2.Seatbelt edge detection based on Canny edge detectionFor most of the image samples, Seatbelt edge is very clear, The problem is how fast we can screening of samples of these clear image.we filtered Canny edge detection result using gradient direction information, the angle between the line segment according to the detected, distance and other information to determine whether the seat belt edge, further, we have improved Hough circle detection, turn it into a pair of semi-circular arc on the detection,specify line location information, to achieve the effect of rapid screening.3.Seatbelt edge detection based on adaboost learning algorithm Non-statistical algorithm for texture, occluding and other disturbances are extremely sensitive, and approximately one-third or more of the picture more or less interference, and some from reflections of the front windows, some from clothes texture, some from the arm occluding, Therefore, we introduced a ma-chine learning algorithm, by constructing two distinct set of samples, by the same algorithm to obtain two classifiers, with different the test results, the two can be used together to achieve better results.更多还原