复杂环境下交通标志的检测

发布时间：2019-01-28 21:30

【摘要】：近几年来智能交通系统(Intelligent Transportation System, ITS)引起了人们越来越多的兴趣和关注。该系统可以减轻人们的驾驶压力,使人们的出行更加自由、安全、可靠。交通标志的检测与识别系统(Traffic Signs Detection and Recognition System)作为ITS的一个必要环节,是智能车或驾驶员获取外界路况信息的可靠保障。交通标志的检测是识别的关键,准确实时的检测出感兴趣区域才会为识别奠定良好的基础。针对复杂的外界环境和行车环境下拍摄到的图像,需要预处理才可以保证检测的准确率。本文首先针对去除运动模糊和浓雾两个预处理环节进行了研究,然后探究了标牌的检测方法。图像的运动模糊复原问题分为匀速直线运动模糊复原和非匀速直线运动模糊复原两类。针对前者,介绍了模糊核尺度和方向估计的普遍方法以及图像复原的基本算法；针对后者本文在强边缘估计模糊核方法上进行了改进,用导向滤波器进行保持边缘去噪,并且在R、G、B三通道进行运算,以增加较小计算量为代价,使该算法可直接处理彩色图像,复原结果保留了图像的颜色信息。针对图像去雾问题,本文采用了去雾领域最近研究成果“黑暗通道(Dark Channel Prior)法”。利用黑暗通道先验知识对雾化图像进行复原,并在透射函数优化方面提出了导向滤波与线性插值相结合的方法。自然拍摄的图像经过预处理后,再进行交通标志的检测,检测环节分为两个部分：(1)针对红、蓝、黄三种颜色在YIQ颜色空间内建立1分量和Q分量的二维正态分布模型,利用待检测像素对三个正态分布模型的符合度来进行颜色分割,然后进行形态学处理,得到二值化图像中的白色区域作为初步感兴趣区域。(2)选取对形状的大小、角度、旋转鲁棒性较强的Hu不变矩特征,针对初步分割得到的感兴趣区域,求得各区域Hu不变矩特征,利用支持向量机(Support Vector Machine, SVM)进行训练得到分类器。剔除不规则类别,保留交通标志特征形状圆形、矩形和三角形类别,作为最终确定的检测结果。本文的主要工作是对去除运动模糊和浓雾的算法进行了研究与改进,使之效果更佳、实时性更好、更符合交通标志预处理的现实要求。在检测阶段,本文算法效果较优,误检率和漏检率低,实时性好,对光照变化和少量遮挡等情况也有较好的表现。
[Abstract]:In recent years, intelligent transportation system (Intelligent Transportation System, ITS) has attracted more and more attention. The system can reduce driving pressure and make people travel more freely, safely and reliably. As a necessary link of ITS, the traffic sign detection and recognition system (Traffic Signs Detection and Recognition System) is a reliable guarantee for intelligent vehicle or driver to obtain information of road condition. Detection of traffic signs is the key to recognition, accurate and real-time detection of the region of interest will lay a good foundation for recognition. In order to ensure the accuracy of detection, preprocessing is needed for the images taken in complex environment and driving environment. In this paper, two preprocessing steps of removing motion blur and dense fog are studied firstly, and then the detection method of sign is explored. Image blur restoration is divided into two categories: uniform linear motion blur restoration and non-uniform linear motion blur restoration. Aiming at the former, the general methods of fuzzy kernel scale and direction estimation and the basic algorithm of image restoration are introduced. For the latter, the fuzzy kernel method of strong edge estimation is improved, the edge preserving denoising is carried out by the guide filter, and the operation is carried out in the three channels of RG GnB at the cost of increasing the computational cost. The algorithm can directly process color images, and the restoration results retain the color information of the images. Aiming at the problem of image de-fogging, this paper adopts the dark channel (Dark Channel Prior) method, which is a recent research result in the field of de-fogging. Based on the prior knowledge of dark channels, the atomization image is restored, and a method of combining guidance filtering with linear interpolation is proposed in the aspect of transmission function optimization. After pre-processing, the natural images are detected by traffic signs. The detection links are divided into two parts: (1) the 2-D normal distribution model of 1 component and Q component in YIQ color space is established for red, blue and yellow colors. The color of the three normal distribution models is segmented by the pixels to be detected, and then the white area in the binary image is obtained as the initial region of interest. (2) the size and angle of the shape are selected. The Hu moment invariant feature with strong rotation robustness is obtained. The Hu moment invariant feature of each region is obtained for the region of interest obtained by initial segmentation. The classifier is trained by support vector machine (SVM) (Support Vector Machine, SVM). Eliminating irregular categories and retaining traffic signs with circular, rectangular and triangular types as the final detection results. The main work of this paper is to study and improve the algorithm of removing motion blur and dense fog, which makes it more effective, more real-time and more in line with the practical requirements of traffic sign preprocessing. In the detection stage, the algorithm is effective, the false detection rate and the missing detection rate are low, the real-time performance is good, and it has good performance to the illumination variation and a small amount of occlusion and so on.
【学位授予单位】：山东大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U495;TP391.41

【相似文献】