基于两阶段的交通标志识别方法研究

发布时间：2018-01-25 07:00

本文关键词： 交通标志识别两阶段稀疏表示 HOG特征 LBP特征 SVM分类器特征融合　出处：《南京理工大学》2015年硕士论文　论文类型：学位论文

【摘要】：交通标志识别是智能交通系统(ITS)中的一个重要组成部分。经过国内外学者几十年的研究,交通标志识别领域的理论和实践体系逐渐形成,并取得了很多突破性的进展。然而自然场景具有复杂性以及交通标志种类繁多,使得交通标志的识别依然具备挑战性。本文主要针对交通标志的多类别,研究通过两阶段的方法识别交通标志。首先,从交通标志的功能类别出发,设计了基于PCA-LDA的两阶段交通标志识别方法。功能相近的交通标志一般具有相似的图案设计,因此,根据功能类别将交通标志划分为限速、警告、指示以及无规则四个相似的子类进行识别。首先采用结合PCA与LDA的方法对交通标志进行快速的预分类,即得到交通标志的所属子类；然后采用稀疏表示的方法进行子类内类别识别,得到交通标志具体所属的类别。实验表明,基于PCA-LDA的两阶段交通标志识别方法在识别准确率上要优于PCA与LDA结合的方法以及基于局部字典的两阶段稀疏表示等方法。然后,根据交通标志的形状以及颜色特征对交通标志的子类划分进行了调整,分为了红色圆形、红色正三角形、蓝色圆形、白色圆形以及无规则五个相似的子类。在第一阶段中采用HOG特征与SVM分类器对交通标志进行预分类；第二阶段中提出了相似类核心区域的提取方法,并采用稀疏表示的方法进行子类内类别识别。通过实验表明,在使用相同特征提取方法的基础上,基于颜色与形状的相似子类划分方案要优于基于功能类别的子类划分方案,而在同种相似子类划分的基础上,基于HOG特征方法的子类识别效果要优于基于PCA-LDA方法,同时在最终的两阶段识别上也取得了更好的效果。最后,提出了基于多特征融合的两阶段交通标志识别。在第一阶段识别过程中,可以分别以颜色特征与边缘特征作为依据进行子类的识别,然而单一特征难以全面描述交通标志,因此在第一阶段中采用融合颜色直方图以及HOG边缘特征的方法对交通标志进行预分类；第二阶段中采用融合LBP纹理特征与HOG边缘特征对交通标志各子类内的类别进行识别。实验表明,特征融合方法能够获得更高的识别率以及更好的鲁棒性,识别率达到96.9%。
[Abstract]:Traffic sign recognition is an important part of Intelligent Transportation system (its). After decades of research by domestic and foreign scholars, the theory and practice system of traffic sign recognition has gradually formed. A lot of breakthrough progress has been made. However, the complexity of natural scenes and the variety of traffic signs make the recognition of traffic signs still challenging. This paper mainly focuses on the multi-category of traffic signs. This paper studies the identification of traffic signs by a two-stage method. First of all, it starts from the functional categories of traffic signs. This paper designs a two-stage traffic sign recognition method based on PCA-LDA. Traffic signs with similar functions generally have similar pattern design. Therefore, traffic signs are divided into speed limit and warning according to the function category. First, the method of combining PCA and LDA is used to pre-classify traffic signs quickly, that is to say, the subclasses of traffic signs are obtained. Then the sparse representation method is used to identify the class within the subclass, and the specific category of traffic sign is obtained. The two-stage traffic sign recognition method based on PCA-LDA is superior to the combination of PCA and LDA in recognition accuracy and two-stage sparse representation based on local dictionary. Then. According to the shape and color characteristics of traffic signs, the subcategories of traffic signs are divided into red circle, red square triangle and blue circle. White circular and irregular five similar subclasses. In the first stage, HOG features and SVM classifier are used to pre-classify traffic signs. In the second stage, the method of extracting the core region of similar classes is proposed, and the sparse representation method is used to recognize the subclass. The experiments show that the same feature extraction method is used on the basis of the method. The similar subclass partition scheme based on color and shape is superior to the subclass partition scheme based on function category, but on the basis of the same similar subclass partition. The subclass recognition effect based on HOG feature method is better than that based on PCA-LDA method. At the same time, better results are obtained in the final two-stage recognition. Finally. A two-stage traffic sign recognition based on multi-feature fusion is proposed. In the first stage, the color feature and edge feature can be used as the basis for sub-class recognition. However, it is difficult to describe traffic signs in a single feature, so in the first stage, the method of combining color histogram and HOG edge features is used to pre-classify traffic signs. In the second stage, LBP texture features and HOG edge features are used to identify the subclasses of traffic signs. The feature fusion method can obtain higher recognition rate and better robustness, and the recognition rate reaches 96.9.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U495;TP391.41

【参考文献】