行车环境下多特征融合的交通标识检测与识别研究

发布时间：2018-01-12 14:41

本文关键词：行车环境下多特征融合的交通标识检测与识别研究　出处：《山东大学》2016年博士论文　论文类型：学位论文

【摘要】：随着社会经济的发展,车辆与日俱增,智能交通系统的应用受到人们的高度重视。作为智能交通系统的核心关键技术,交通标识自动检测和识别获得越来越多学者的关注和研究,并在驾驶员辅助系统、无人驾驶车辆及道路标识的维护等方面获得广泛应用。然而,在真实的复杂场景中,交通标识会出现褪色、破损、阴影、遮挡、运动模糊及颜色与形状相似物体的干扰,面对这些问题,很多学者进行了深入的研究,但是研究成果还远未达到成熟。尤其是在我国人口众多、私家车日益普及的情况下,交通堵塞和生命安全问题愈发严重,因此对交通标识自动检测与识别的研究具有非常重要的理论与现实意义。论文围绕智能交通系统中交通标识自动检测与识别关键技术,重点研究了行车环境下由于车辆加速或者摄像头抖动造成交通标识模糊的问题,图像的底层特征融合、交通标识的颜色分割及感兴趣区域提取问题,以及支持向量机、极限学习机和其它分类器在交通标识检测与识别中的应用问题。论文的具体研究工作及成果如下：(1)针对行车环境下摄像头获取的视觉图像产生运动模糊的问题,研究了一种基于稀疏表示和Weber定律的图像盲复原算法。该方法首先通过冲击滤波器来预测模糊图像的显著边缘,并用多尺度策略由粗到细进行模糊核的估计。然后对图像盲复原模型进行稀疏正则化约束,并结合反映人类视觉特性的Weber定律对图像进行盲复原。实验结果表明,提出的盲复原算法能获得较优的性能,在图像纹理上能取得较好的复原效果,并且该方法降低了复原图像的边界伪影,符合人的视觉感知特性。(2)针对交通标识检测中样本类别间的不平衡常常导致分类器的检测性能弱化的问题,研究了一种基于感兴趣区域和HOG-MBLBP融合特征的交通标识检测方法。根据交通标识鲜亮的颜色特点,采用颜色增强技术分割并提取出自然背景中交通标识所在的感兴趣区域；研究了HOG-MBLBP图像底层融合特征,并对交通标识样本库提取该融合特征,利用遗传算法对SVM交叉验证进行参数的优化选取,以此来训练和提升SVM分类器性能；最后将提取的感兴趣区域图像的HOG-MBLBP特征送入训练好的SVM多分类器,进行进一步的精确检测和定位,剔除误检区域。在自建的SDU_CVPR_A交通标识样本库及GTS*德国交通标识库上分别进行了实验,结果验证了所提方法的优越性。(3)为了准确快速识别出检测到的交通标识,研究了一种基于HOG-MBLBP融合特征和极限学习机的交通标识识别方法。首先针对中国交通标识的特点建立了23类SDU_CVPR_B交通标识识别样本库,然后对交通标识样本库分别提取HOG特征、BLBP特征、MBLBP特征以及HOG-MBLBP融合特征,并将提取特征分别输入ELM分类器、SVM分类器、KNN分类器以及随机森林分类器进行分类训练。通过在自建SDU_CVPR_B交通标识识别库和GTSRB德国交通标识识别库上进行的实验表明,融合特征结合ELM分类器可以取得较优的识别效果。(4)鉴于语义特征BoF模型在图像分类任务中的广泛应用,为了更好地表达图像,建立底层视觉特征与高层语义特征间的关系,研究了一种基于融合特征BoF模型的金字塔匹配交通标识识别方法。首先利用K均值聚类方法对各种局部不变特征进行聚类,根据聚类中心构建各自的词典,然后进行BoF模型的图像直方图表示,并采用空间金字塔策略以充分利用局部不变特征的空间结构信息,最后进行SVM分类器训练。在自建SDU_CVPR_B交通标识库和GTSRB德国库上的实验结果表明HOG-MBLBP融合特征的分类效果较优,且HOG-MBLBP融合特征的BoF模型表示进行分类识别的效果优于HOG-MBLBP融合特征进行ELM分类识别的效果。综上所述,本文主要针对行车环境下交通标识图像的检测与识别环节所涉及的一些关键问题进行了探索和研究,旨在提高交通标识检测和识别的准确性与快速性,丰富智能交通系统的理论体系,能够最大程度的解决我国现有的交通问题。
[Abstract]:With the development of social economy, the application of intelligent transportation system vehicle grow with each passing day, the affected people's attention. As the key technology of the intelligent transportation system, traffic signs more and more attention and research for automatic detection and recognition, and in driver assistant system, application maintenance and other aspects of the unmanned vehicle and road signs. However, in the complex scene, traffic signs will fade, breakage, shadows, occlusion, motion blur and color and shape similar to object interference, in the face of these problems, many scholars have conducted in-depth research, but the results are far from mature. Especially in our country has a large population, a private car the growing popularity of the situation, traffic congestion and safety problems become more serious, so it has very important theory and the identification of automatic detection and recognition of traffic Practical significance. This dissertation focuses on Key Technologies of automatic detection and recognition of traffic signs in intelligent traffic system, focusing on the traffic environment due to vehicle acceleration or camera jitter caused by traffic identification Fuzzy problem, the fusion of the low-level features of image segmentation and extracting regions of interest in traffic sign colors, and support vector machine, application of extreme learning machine and other classifier in traffic sign detection and recognition. The specific research work and achievements of the thesis are as follows: (1) motion blur problem for visual image traffic environment to access the camera, on a blind image restoration algorithm based on sparse representation and Weber law. The first method to predict significant edge blur the image through the shock filter, and multi-scale strategy from coarse to fine estimation kernel. Then the image blind restoration model The sparse regularization constraint, and combined with the law reflect Weber human visual characteristics of image restoration. The experimental results show that the proposed blind restoration algorithm can obtain better performance, can obtain the good restoration effect on the image texture, and this method reduces the boundary artifact image, in line with the characteristics of human vision perception. (2) for the unbalanced data traffic sign detection in between categories often leads to weakening of the detection performance of classifier problem, studies a method of traffic sign detection region of interest and HOG-MBLBP fusion based on feature. According to the characteristics of the traffic signs color bright, the color segmentation enhancement technology and extract traffic signs where the interest in natural background areas; study of HOG-MBLBP image fusion feature, and the traffic signs sample extract the feature fusion, using the genetic algorithm to the SVM cross Optimal selection of validation parameters, in order to train and enhance the SVM performance of the classifier; HOG-MBLBP feature extraction finally the ROI image into the trained SVM classifier, accurate detection and positioning further, to eliminate the error regions. The experiments were conducted in a self built SDU_CVPR_A traffic signs and traffic in Germany GTS* sample database the logo, results show the superiority of the proposed method. (3) in order to quickly and accurately identify the detected traffic signs, on a pass identification method of HOG-MBLBP feature fusion and extreme learning machine. Based on the first established 23 types of SDU_CVPR_B traffic identification database according to the characteristics of China traffic signs then, HOG feature extraction of traffic sign sample library BLBP feature, MBLBP feature and HOG-MBLBP feature fusion, feature extraction and input of ELM classifier, SVM Classifier, KNN classifier and random forest classifier training. Through the self SDU_CVPR_B traffic identification library and GTSRB library on the German traffic sign recognition experiments show that the combination of the ELM classifier can achieve better recognition effect of feature fusion. (4) in view of wide application of semantic features of BoF model in image classification task, in order to better express the image, establish relationship between low-level visual features and high-level semantic features between the study of a new matching traffic sign recognition feature fusion method of BoF model based on Pyramid. The first use of K mean value clustering method of local invariant feature clustering, according to the dictionary to construct a cluster center, and then the image histogram BoF model said the Pyramid strategy to make full use of the information of spatial structure of local invariant features. Finally, SVM classifier training In traffic sign practice. Self SDU_CVPR_B library and GTSRB Library in Germany. The experimental results show that the classification results of HOG-MBLBP fusion feature is better than the BoF model and the HOG-MBLBP fusion feature representation feature fusion classification is better than the HOG-MBLBP ELM classification results. To sum up, this paper mainly focuses on some key issues of detection and recognition link traffic sign image traffic environment to explore and research, in order to improve the accuracy of traffic sign detection and recognition and speediness, enrich the theoretical system of the intelligent transportation system, can solve the maximum traffic problems existing in our country.

【学位授予单位】：山东大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】