自然场景下交通标志检测和分类算法研究
发布时间：2018-08-12 09:32
【摘要】：交通标志作为重要的道路安全附属设施,在规范交通行为、指示道路状况、保障道路功效、引导行人和安全驾驶等方面起到重要作用。为了确保交通标志的信息能够及时、准确的传达,交通标志自动识别系统(TSR)受到各国学者的重视。基于影像的道路交通标志检测和分类是交通标志自动识别系统的两个关键技术环节,经过多年的发展,在理论研究和实用系统方面均取得了 一定的成果。交通标志长期暴露在户外,标志本身会发生颜色退化、污损以及形变;在视频和图像的采集过程中,光照变化会导致交通标志颜色失真,视角倾斜会引起交通标志形状变化;在复杂环境下交通标志会被其它物体遮挡而形成不完整的边缘,这些都给交通标志的检测带来挑战。交通标志的分类是TSR中另一个关键技术,交通标志类别众多,是典型的多分类问题,追求分类算法的鲁棒性和有效性仍然是尚未有效解决的热点问题。针对上述不足,本文在交通标志颜色处理、交通标志检测和分类等方面进行了研究,主要研究工作如下:(1)针对交通标志的颜色特征,提出颜色分布模型和改进的颜色对比度模型,以突出图像中的交通标志区域。颜色分布模型通过计算交通标志主体颜色在Lab空间中的分布概率,得到输入图像相对于每种颜色的特征图,在特征图中突出相应颜色的区域。改进的颜色对比度模型则根据人眼视觉机制中存在的颜色对抗性,突出红色、蓝色和黄色区域。在实验阶段对比了这两种模型和其他常用的颜色处理算法在交通标志数据集上的运行结果。实验证明,本文提出的改进的颜色对比度模型在保证最短运行时间的同时取得了最高的检测率。(2)在研究交通标志快速检测算法的基础上,提出一种基于旋转对称投影的快速多边形检测算法。该方法以图像的边缘梯度为特征,选择满足特定旋转对称角度的点进行投影,得到图像中可能存在的多边形,之后采用多边形分类方法得到其具体类别。该算法时间复杂度较低,实验表明每幅图像的平均处理时间为55ms,能够满足交通标志检测的实时需求。(3)在分析现有交通标志形状检测算法的基础上,针对部分遮挡和视角倾斜的交通标志提出了基于连接分布(LD)模型的多边形检测方法。LD模型将多边形看作中心到边界点连接的集合,每个连接可用连接的长度、连接与水平线的夹角、边界点的边缘方向表示。在交通标志发生视角倾斜和边缘部分缺失的情况下,连接间的顺序和邻接关系不变,因此该算法能够有效检测视角倾斜和边缘不完整的多边形。在公开数据集上的实验表明,该算法在禁令、警告和指示标志上的检测率分别为98.63%、95.24%和94.40%,优于大多数国际先进算法。(4)针对复杂环境下交通标志检测结果不理想的情况,提出一种基于视觉显著性的交通标志检测算法。该算法将自底向上和自上而下的显著性结合在一起,完成对交通标志的检测。自底向上的显著性算法在颜色聚簇划分和区域分割的基础上,计算区域之间的对比度和区域所属聚簇的颜色分布紧致性,以此作为该区域的显著性度量。每一类交通标志均有特定的颜色特征,可形成类别相关的显著图,即自上而下的显著性检测。和自底向上的显著图相结合,完成复杂环境中交通标志的检测。实验证明该方法可以有效检测禁令、指示和警告标志,在公开数据集上取得了较高的检测率。此外,该方法能够检测白色交通标志,不需要额外的处理。(5)在分析我国交通标志特点的基础上,提出一种逐级细化的交通标志分类方法。首先根据颜色和形状特征将交通标志分为五个大类,即禁令标志、警告标志、指示标志、解除禁令标志和其它标志。在粗分类中,分别利用HOG和CN描述子表示每类标志的形状和颜色特征,采用线性SVM分类器得到感兴趣区域所属的大类。然后分析词袋模型中颜色和形状特征的融合方式,采用CN和SIFT特征早融合的方式表示感兴趣区域,最后利用高斯核SVM分类器得到每个感兴趣区域的最终类别标记。该算法在公开数据集上的交通标志分类正确率为99.15%,优于人工分类结果,在所有公开的分类结果中排名第二。
[Abstract]:Traffic signs, as an important accessory facility of road safety, play an important role in regulating traffic behavior, indicating road conditions, ensuring road efficacy, guiding pedestrians and driving safely. Detection and classification of road traffic signs based on images are two key technologies of automatic traffic sign recognition system. After years of development, some achievements have been made in both theoretical research and practical systems. In the process of acquisition, illumination changes will lead to color distortion of traffic signs, and angle tilt will cause shape changes of traffic signs; in complex environment, traffic signs will be blocked by other objects and form incomplete edges, which bring challenges to traffic signs detection. Traffic signs classification is another key technology in TSR, traffic signs. It is a typical multi-classification problem that there are many classifications. The pursuit of robustness and effectiveness of classification algorithm is still a hot issue that has not been effectively solved. A color distribution model and an improved color contrast model are proposed to highlight the traffic sign area in the image. The color distribution model calculates the probability of the main color distribution in the Lab space and obtains the feature map of the input image relative to each color. The corresponding color region is highlighted in the feature map. The model highlights the red, blue and yellow regions according to the color antagonism in the human visual mechanism. The experimental results of the two models and other commonly used color processing algorithms on the traffic sign data sets are compared. The experimental results show that the improved color contrast model proposed in this paper guarantees the shortest running time. (2) On the basis of studying the fast detection algorithm of traffic signs, a fast polygon detection algorithm based on rotational symmetry projection is proposed, which is characterized by the edge gradient of the image and selects the points satisfying the specific rotational symmetry angle for projection to obtain the possible polygons in the image. The algorithm has a low time complexity, and the average processing time of each image is 55ms, which can satisfy the real-time requirement of traffic sign detection. (3) Based on the analysis of existing traffic sign shape detection algorithms, traffic signs with partial occlusion and oblique view angle are proposed. A polygon detection method based on connection distribution (LD) model is proposed. LD model regards polygons as a set of connections from center to boundary point. Each connection can be represented by the length of the connection, the angle between the connection and the horizontal line, and the edge direction of the boundary point. Experiments on open datasets show that the detection rates of prohibition, warning and indication signs are 98.63%, 95.24% and 94.40%, respectively, which are superior to most international advanced algorithms. (4) Traffic signs detection in complex environments. A traffic sign detection algorithm based on visual saliency is proposed, which combines bottom-up saliency with top-down saliency to detect traffic signs. Each type of traffic signs has a specific color feature, which can form a class-related saliency map, i.e. top-down saliency detection. Combining with bottom-up saliency map, traffic signs detection in complex environments is completed. Experiments show that the method is effective. In addition, the method can detect white traffic signs without additional processing. (5) Based on the analysis of the characteristics of traffic signs in China, a new classification method of traffic signs is proposed. First, according to the color and shape characteristics. Traffic signs are classified into five categories: prohibition signs, warning signs, indication signs, lifting prohibition signs and other signs. In rough classification, the shape and color features of each type of signs are represented by HOG and CN descriptors, and the regions of interest are classified by linear SVM classifier. In the fusion method of shape features, the region of interest is represented by the early fusion of CN and SIFT features. Finally, the final class markers of each region of interest are obtained by using Gaussian kernel SVM classifier. The classification accuracy of traffic signs on public data sets is 99.15%, which is better than that of manual classification. Among all the public classification results, the proposed algorithm is superior to manual classification. Ranked second.
【学位授予单位】：南京理工大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：U495;TP391.41
本文编号：2178650
[Abstract]:Traffic signs, as an important accessory facility of road safety, play an important role in regulating traffic behavior, indicating road conditions, ensuring road efficacy, guiding pedestrians and driving safely. Detection and classification of road traffic signs based on images are two key technologies of automatic traffic sign recognition system. After years of development, some achievements have been made in both theoretical research and practical systems. In the process of acquisition, illumination changes will lead to color distortion of traffic signs, and angle tilt will cause shape changes of traffic signs; in complex environment, traffic signs will be blocked by other objects and form incomplete edges, which bring challenges to traffic signs detection. Traffic signs classification is another key technology in TSR, traffic signs. It is a typical multi-classification problem that there are many classifications. The pursuit of robustness and effectiveness of classification algorithm is still a hot issue that has not been effectively solved. A color distribution model and an improved color contrast model are proposed to highlight the traffic sign area in the image. The color distribution model calculates the probability of the main color distribution in the Lab space and obtains the feature map of the input image relative to each color. The corresponding color region is highlighted in the feature map. The model highlights the red, blue and yellow regions according to the color antagonism in the human visual mechanism. The experimental results of the two models and other commonly used color processing algorithms on the traffic sign data sets are compared. The experimental results show that the improved color contrast model proposed in this paper guarantees the shortest running time. (2) On the basis of studying the fast detection algorithm of traffic signs, a fast polygon detection algorithm based on rotational symmetry projection is proposed, which is characterized by the edge gradient of the image and selects the points satisfying the specific rotational symmetry angle for projection to obtain the possible polygons in the image. The algorithm has a low time complexity, and the average processing time of each image is 55ms, which can satisfy the real-time requirement of traffic sign detection. (3) Based on the analysis of existing traffic sign shape detection algorithms, traffic signs with partial occlusion and oblique view angle are proposed. A polygon detection method based on connection distribution (LD) model is proposed. LD model regards polygons as a set of connections from center to boundary point. Each connection can be represented by the length of the connection, the angle between the connection and the horizontal line, and the edge direction of the boundary point. Experiments on open datasets show that the detection rates of prohibition, warning and indication signs are 98.63%, 95.24% and 94.40%, respectively, which are superior to most international advanced algorithms. (4) Traffic signs detection in complex environments. A traffic sign detection algorithm based on visual saliency is proposed, which combines bottom-up saliency with top-down saliency to detect traffic signs. Each type of traffic signs has a specific color feature, which can form a class-related saliency map, i.e. top-down saliency detection. Combining with bottom-up saliency map, traffic signs detection in complex environments is completed. Experiments show that the method is effective. In addition, the method can detect white traffic signs without additional processing. (5) Based on the analysis of the characteristics of traffic signs in China, a new classification method of traffic signs is proposed. First, according to the color and shape characteristics. Traffic signs are classified into five categories: prohibition signs, warning signs, indication signs, lifting prohibition signs and other signs. In rough classification, the shape and color features of each type of signs are represented by HOG and CN descriptors, and the regions of interest are classified by linear SVM classifier. In the fusion method of shape features, the region of interest is represented by the early fusion of CN and SIFT features. Finally, the final class markers of each region of interest are obtained by using Gaussian kernel SVM classifier. The classification accuracy of traffic signs on public data sets is 99.15%, which is better than that of manual classification. Among all the public classification results, the proposed algorithm is superior to manual classification. Ranked second.
【学位授予单位】：南京理工大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：U495;TP391.41
【相似文献】
相关期刊论文 前8条
1 ;产品介绍[J];电气开关;1997年03期
2 ;企业之窗[J];电气时代;1997年03期
3 江厥中 ,杨公侠;略论室内建筑色彩[J];建筑学报;1983年07期
4 郝同枝;;为视障学生创设无障碍的校园环境[J];内蒙古教育;2010年02期
5 方荣国;;电影特效蓝/绿屏拍摄解析[J];现代电影技术;2011年01期
6 范树奇;有机玻璃瓶装寄生虫标本制作的改进[J];哈尔滨医科大学学报;1987年05期
7 徐和德;利用背景的几种技巧[J];照相机;2005年04期
8 ;[J];;年期
相关会议论文 前1条
1 许宗惠;林仲贤;潘广钺;;不同色光照明后相同白光色貌的色位移——相继颜色对比[A];全国第六届心理学学术会议文摘选集[C];1987年
相关重要报纸文章 前3条
1 德先;拍银杏的“非常规”技法[N];中国摄影报;2013年
2 戴定国;爱尔兰狗年发行导盲犬盲文邮票[N];中国集邮报;2006年
3 记者 聂根鹏;盘活万亩森林 开启农民致富之门[N];昆明日报;2013年
相关博士学位论文 前1条
1 徐丹;自然场景下交通标志检测和分类算法研究[D];南京理工大学;2015年
相关硕士学位论文 前1条
1 谭珊;基于颜色对比和各向异性扩散分割的显著性区域检测[D];大连理工大学;2013年
，本文编号：2178650
本文链接：https://www.wllwen.com/kejilunwen/daoluqiaoliang/2178650.html
