无人驾驶车辆的行道线检测方法研究

发布时间：2018-05-02 15:17

本文选题：无人驾驶汽车 + 行道线检测　；参考：《南京理工大学》2017年硕士论文

【摘要】：作为无人驾驶系统和辅助驾驶系统的关键技术,行道线检测方法研究备受关注。本文利用传统的检测方法提出并实现了一种新的直道检测和弯道检测方法,同时从机器学习的角度实现了较为鲁棒的行道线区域粗定位算法,同时也介绍了本文实验所用平台。本文工作主要分为以下四个部分:(1)提出并实现了一种基于投影变换的快速行道线直道检测方法,并研究了三种行道线过滤机制。当行道线磨损时,传统方法中的边缘检测会无能为力,当出现强光、阴影遮挡时,传统方法中的二值化方法漏检率也会升高,因此本文提出利用图像中的投影信息来提取行道线的候选点。针对行道线图像中的"近大远小"现象,本文利用积分图实现了分段投影变换,也相当于实现了行道线的变分辨率检测。直线提取阶段,本文采用了基于角度估计的霍夫变换方法。针对霍夫变换后直线较多的情况,本文采用对比度判断、消失点估计以及直线聚类的方法对结果进行处理。实验结果表明,本方法对强光、阴影遮挡以及磨损等情况具有较强的鲁棒性,同时能够在保证精度的同时满足无人车的实时性要求。(2)提出并实现了一种基于分段式消失点估计的行道线弯道检测方法。根据弯道的多消失点特性,本方法采用分段式消失点估计的方法进行弯道判别。在弯道检测过程中,本方法在前文工作的行道线候选点的基础上采用分段式霍夫变换的方法提取直线,并加以抽样获得曲线候选点,然后通过最小二乘法和三次样条插值法分别对候选点进行曲线拟合。为了对一些拟合后不合理的弯道结果进行拒识,本方法采用通过相机标定的结果对拟合后的曲线进行后验。实验结果表明,本方法对于弯道能够较好的判别并加以检测,同时发现最小二乘法较为鲁棒,而三次样条插值法的误差更小。(3)实现了一种基于机器学习的行道线候选区域粗定位算法,该方法能够对以上方法的鲁棒性进行进一步加强。由于行道线具有明显的梯度和边缘特征,本文采用了HOG和Haar特征分别对行道线图像进行处理。本文首先对行道线的HOG梯度方向直方图特征进行了提取,并配合支持向量机SVM进行了训练分类。同时本文又提取了行道线的Haar特征,并配合级联分类器Ababoost加以实现。实验数据表明,Haar特征配合Adaboost级联分类器具有较高的检测率。(4)最后本文介绍了本课题项目所处的无人驾驶汽车实验平台的硬件系统和软件系统设计,本文算法在无人车系统中得到了实验验证,并取得了很好的应用效果。
[Abstract]:As the key technology of driverless system and auxiliary driving system, the research of lane detection method has attracted much attention. In this paper, a new method of straight track detection and curve detection is proposed and implemented by using the traditional detection method. At the same time, a more robust algorithm of rough location of line region is realized from the point of view of machine learning, and the platform used in this experiment is also introduced. The work of this paper is divided into four parts as follows: (1) this paper proposes and implements a fast line line detection method based on projection transformation, and studies three kinds of line line filtering mechanisms. When the track line is worn, the edge detection in the traditional method will be powerless, and when there is strong light and shadow occlusion, the miss rate of the traditional binary method will also increase. Therefore, the projection information in the image is used to extract the candidate points of the line path. In view of the phenomenon of "near, far and small" in the image of line trace, this paper realizes the piecewise projection transformation by using integral graph, which is equivalent to the variable resolution detection of line trace. In the phase of line extraction, the Hough transform method based on angle estimation is used in this paper. In this paper, contrast judgment, vanishing point estimation and linear clustering are used to deal with the problem of more lines after Hough transform. The experimental results show that the proposed method is robust to strong light, shadow occlusion and wear. At the same time, it can satisfy the real-time requirement of the unmanned vehicle while ensuring the accuracy.) A new detection method based on segmented vanishing point estimation is proposed and implemented. According to the characteristics of multiple vanishing points, the method of segmental vanishing point estimation is used to distinguish the curve. In the process of curve detection, this method uses segmented Hough transform method to extract straight lines and get curve candidate points by sampling. Then the candidate points were fitted by least square method and cubic spline interpolation method. In order to reject some unreasonable curve results after fitting, this method uses the camera calibration results to carry out a posteriori on the fitted curve. The experimental results show that the method can be used to distinguish and detect the curve, and the least square method is more robust. The error of cubic spline interpolation method is smaller than that of cubic spline interpolation method. It implements a rough location algorithm based on machine learning, which can further enhance the robustness of the above methods. Because of the obvious gradient and edge characteristics of the track, the HOG and Haar features are used to process the line image respectively. Firstly, the feature of HOG gradient direction histogram is extracted, and the training classification is carried out with support vector machine (SVM) SVM. At the same time, the Haar features of the line are extracted and implemented with cascaded classifier Ababoost. Experimental data show that Haar feature and Adaboost cascade classifier have high detection rate. Finally, this paper introduces the hardware system and software system design of the experiment platform of driverless vehicle. The algorithm of this paper has been verified by experiment in the unmanned vehicle system, and has obtained the very good application effect.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U463.6;TP391.41

【参考文献】