基于图模型与卷积神经网络的交通标志识别方法  

Traffic sign recognition method based on graphical model and convolutional neural network

作　　者：刘占文[1] 赵祥模[1] 李强[1] 沈超[1] 王姣姣[2] LIU Zhan-wen ZHAO Xiang-mo LI Qiang SHEN Chao WANG Jiao-jiao(School of Information Engineering, Chang＇an University, Xi＇an 710064, Shaanxi, China School of Electronic and Control Engineering, Chang＇an University, Xi＇an 710064, Shaanxi, China)

机构地区：[1]长安大学信息工程学院,陕西西安710064 [2]长安大学电子与控制工程学院,陕西西安710064

出　　处：《交通运输工程学报》2016年第5期122-131,共10页Journal of Traffic and Transportation Engineering

基　　金：高等学校学科创新引智计划项目(B14043);国家自然科学基金项目(51278058;61302150);陕西省自然科学基金项目(2012JM8011);中央高校基本科研业务费专项资金项目(2013G1241111;2014G1321035)

摘　　要：为了提高交通标志识别的鲁棒性,提出了一种基于图模型与卷积神经网络(CNN)的交通标志识别方法,建立了一个面向应用的基于区域的卷积神经网络(R-CNN)交通标志识别系统。构造了基于超轮廓图(UCM)超像素区域的图模型,有效利用自底向上的多级信息,提出了一种基于图模型的层次显著性检测方法,以提取交通标志感兴趣区域,并利用卷积神经网络对感兴趣候选区进行特征提取与分类。检测结果表明:针对限速标志,基于UCM超像素区域的图模型比基于简单线性迭代聚类(SLIC)超像素的图模型更有利于获取上层显著度图的大尺度结构信息;基于先验位置约束与局部特征(颜色与边界)的层次显著性模型有效地融合了局部区域的细节信息与结构信息,检测结果更加准确,检测目标更加完整、均匀,查准率为0.65,查全率为0.8,F指数为0.73,均高于其他同类基于超像素的显著性检测算法;基于具体检测任务的CNN预训练策略扩展了德国交通标志识别库(GTSRB)的样本集,充分利用了CNN的海量学习能力,更好地学习目标内部的局部精细特征,提高了学习与识别能力,总识别率为98.85%,高于SVM分类器的95.73%。In order to improve the recognition robustness of tra{fic sign, a recognition method ot traffic sign based on graphical model and convolutional neural network （CNN） was proposed, and a application-oriented recognition system of traffic sign based on the regions with convolutional neural network （R-CNN） was set up. A graphical model based on UCM superpixel region was constructed, the multi-scales information from bottom to up was used efficiently, and a hierarchical saliency detection method based on graphical model was proposed to extract the interest regions of traffic sign. The candidate interest regions were processed for the feature extraction and classification of traffic sign with CNN. Detecting result indicates that aiming at the speed-limited signs, the graphical model based on UCM superpixel can get more larger-scale construction information of upper saliency map contrasting to the graphical model based on simple linear iterative cluster （SLIC） superpixel. Because the hierarchical saliency detection model based on the prior location restriction and the local properties combines the detail information of local regions and the constructional information of whole image, the detected results are more precise, and the detected targets are more complete and homogeneous. The detecting precision of detection model is 0.65, the recall ratio is 0. 8, F index is 0.73, and the indexes are higher than the indexes of other saliency detection methods based on superpixel. The CNN pre-training strategy for the specific detection task expends the data base of German traffic sign recognition benchmark （GTSRB）, and fully uses the learning skills of CNN to learn the local fine detail features of object, so the recognition precision of CNN rises, and the recognition rate of CNN is 98.85% beyond the rate of SVM with 95.73%. 19 figs, 31 refs.

关 键 词：交通控制 交通标志 显著性检测 卷积神经网络 预训练策略 

分 类 号：U491.52[交通运输工程—交通运输规划与管理]