摘 要

随着日益增长的物质文化水平以及对现在智能汽车的方向的探索，人们越来越关注于对汽车功能性的改善。从安全方面考虑，也可能是为年轻的驾驶员提供功能性的辅助，类似于辅助驾驶员更好的转向，以及提醒驾驶员车道上面的行人的数量和相对的位置；对于驾驶技艺高超的驾驶员可能能够提供更好的驾驶趣味性，甚至是代替驾驶员。这样的智能驾驶系统是现代汽车的研究方向，而在这个庞大的研究基础上，为驾驶员提供车道信息，类似于车道线所在的位置。这不仅能帮助智能驾驶系统更好的决策，也能更好的维护驾驶员的安全，可见，利用图像处理检测车道线是智能驾驶系统一个非常重要的研究方向。

论文主要研究了利用opencv程序库处理放置在车辆前部的摄像头传递回来的图像信息，对图像进行灰度化、二值化处理，去除传递回来的图像因为硬件或者外界环境过于复杂而存在的噪声，将处理之后的图像通过相关算子提取出图像的轮廓。通过hough变换检测图像轮廓中的直线,之后再进行透视变换，将摄像头畸变下的图像转变为正常的图像，对于二值化之后的图像通过膨胀和腐蚀消除图像中出现的噪点。同时在确定车道线的出现的区域之后，通过RANSAC算法检测在这个区域中出现的有效的车道线数据。之后，对于这些有效的车道线数据通过Bezier生成合理的曲线。当然为了减少车道线在检测之后出现跳动，后期处理还是用了车道线跟随算法：kalman算法，取得一定的效果。

本文的特色在于图像处理过程中的灰度化过程进行优化，提高车道线对比度的同时，消除图像中存在的噪声，在创建曲线拟合车道线的时候并非采用一般的直线，更多是采用双曲线来拟合，这样讲几条车道线联系起来的同时，也让车辆提起获取形式过程中下一步的曲率信息，方便车辆的智能系统进行决策判断。 同时本文还提出了一种车道线状态检测矩阵。通过创建一个斜置的高斯矩阵，计算二值化之后 的图像与斜置高斯矩阵的和，可以通过设置合适的阈值判断图像中是否检测到车道线，以及计算矩阵的和，计算二级梯度判断是否车辆是否在换道过程中。

关键词：hough变换；智能驾驶系统；双曲线模型

Abstract

With the increasing level of material culture and the exploration of the current direction of smart cars, people are paying more and more attention to improving the functionality of automobiles. There is a need to provide more assistance to the driver during the driving process. These assistances may be for driver's safety considerations. They may also provide young drivers with functional assistance, similar to assisting the driver in better steering and the number of lanes above the driver's remind pedestrians and position relative; for driving highly skilled driver may be able to provide better driving fun, even instead of the driver. This research direction is intelligent driving systems of modern cars, but in this huge study, based on information provided to the driver's lane, similar to the position of the lane line is located. This can not only help the smart driving system to make better decisions, but also better maintain the driver's safety. It can be seen that using image processing to detect lane lines is a very important research direction of smart driving systems.

The paper mainly studies the use of opencv library to process the image information transmitted from the security camera at the front of the vehicle. The image is grayed and binarized. The image is removed because the hardware or the external environment is too complex and there is noise. , The processed image is extracted by the correlation operator to outline the image. The hough transform is used to detect the straight line in the image outline, and then the perspective transformation is performed to convert the image under the distortion of the camera into a normal image. The image after binarization is used to eliminate the noise appearing in the image through expansion and erosion. At the same time after determining the area where the lane line appears, the valid lane line data appearing in this area is detected by the RANSAC algorithm. Afterwards, Bezier generates a reasonable curve for these valid lane line data. Of course, in order to reduce the occurrence of bouncing after the lane line is detected, the post-processing still uses the lane-line following algorithm: kalman algorithm to obtain a certain effect.

The research results show that the lane line after image processing can provide the lane line and related position information for the vehicle, assist the vehicle in making driving decisions and judgement, provide correct and effective data for the form, and provide path decision for the subsequent establishment of the smart driving system. The input variable has very significant meaning.

The feature of this paper is to optimize the graying process in the image processing process, to improve the contrast of the lane lines, and to eliminate the noise in the image. When the curve is fitted to the lane line, the general straight line is not used, and the double line is more often used. Curve to fit, so that talking about several lane lines at the same time, let the vehicle mention the next step in the process of obtaining curvature information, to facilitate the vehicle's intelligent system for decision making. At the same time, this paper also proposes a lane line state detection matrix. By creating an oblique Gaussian matrix and calculating the sum of the binarized image and the oblique Gaussian matrix, you can determine whether the lane line was detected in the image by setting an appropriate threshold, and calculate the sum of the matrices to determine the secondary gradient. Whether the vehicle is changing course.

Keywords:gough transform;smart driving system;hyperbolic model

目录

第1章.绪论 1

1.1 研究的背景及意义 1

1.2国外发展现状 3

1.3国内发展现状 4

1.4本文研究内容 5

第2章．车道线检测现有的解决方法 7

2.1常见方法分类 7

2.1.1基于模型的方法 7

2.1.2基于模型的方法 9

2.1.3本文提出的方法 10

第3章.道路图像预处理 12

3.1感兴趣区域的划分 12

3.1.1 动态ROI区域 12

3.1.2 固定ROI区域 12

3.2 去除图像中影响检测的像素 13

3.3图像畸变^[16] 13

3.4透视变换 15

3.5滤波函数 16

3.5.1方框滤波函数 16

3.5.2均值滤波函数 18

3.5.3高斯滤波函数 18

3.5.4中值滤波函数 20

3.5.5双边滤波算法 21

3.6图像的灰度化处理 22

3.6.1新的图像灰度化算法 24

3.7图像的二值化 24

3.7.1固定阈值的二值化 25

3.7.2全局灰度平均值法 25

3.7.3百分比阈值法 26

3.7.4大津法 27

第4章 车道线检测 29

4.1车道线轮廓的提取 29

4.1.1 sobel算子 29

4.1.2 laplace算子 31

4.1.3 canny算子 32

4.2霍夫变换 34

4.2.1hough变换的原理介绍 34

4.2.2标准hough变换 36

第5章 曲线拟合和跟踪 38

5.1车道线检测范围 38

5.2曲线拟合 40

5.3 Bezier3次曲线拟合 42

5.3.1逆透视变换 43

5.4车道线跟踪算法 43

5.4.1 kalman滤波算法 44

第6章 总结和展望 47

参考文献 49

第1章.绪论

1.1 研究的背景及意义

2017年，中国死于交通事故的人数虽然相对往年有了大幅度的降低，但是死亡的人数仍在全世界排名第二位。交通事故频发，伤亡惨重。而交通事故发生的主要原因有以下几点：1.由于天气、光照的原因，驾驶人员不能清楚辨识路况信息；2.驾驶人员由于疲劳以及失去正确的驾驶判断；3.驾驶人员存在不良的驾驶习惯等方面的原因。交通上面的问题引发学长的思考，通过智能汽车和智能驾驶网络两个方面解决上述三种原因造成的交通安全隐患以及合理的避免交通事故，现如今智能汽车和智能交通网络开始成为降低交通死亡率的重要途径。

对于天气、光照方面的原因，智能驾驶汽车有众多的检测路况信息的传感器：激光雷达、毫米波激光雷达、摄像头、超声波。这些传感器能够提供人眼所不能辨识到的路况信息，同时受天气、光照的影响相对于人眼要小得多。