基于特征融合和字典学习的交通标志识别

发布时间：2018-01-31 23:13

本文关键词： 交通标志识别 HOG GIST GCCA K-SVD 线性SVM　出处：《安徽大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着经济的发展,人们生活水平的提高和车辆的日益增多,引发的交通堵塞和生命安全问题也愈发严重,因此智能交通系统的研发应用受到工业界和学术界的高度重视。作为智能交通系统中核心技术之一,交通标志识别一直是研究的热点,然而研究成果还尚未达到成熟,这是由于在真实复杂的场景中,交通标志的识别易受天气状况、运动模糊、外物遮挡、颜色退化、旋转倾斜等因素的影响。因此,交通标志识别依然是一个具有挑战性的课题。本文的主要研究工作有以下四个方面:(1)由于外界场景各因素的影响,各交通标志图像的尺寸、亮度、形态等差异不一,因此在特征提取之前,需要对其进行预处理操作。本文主要从三个方面进行了处理:①图像进行灰度化以及灰度归一化;②感兴趣区域分割提取;③尺度归一化。(2)根据交通标志的局部边缘、全局轮廓、纹理等特点,本文深入研究了两种特征算子:HOG和GIST,并对各特征的参数进行了优化。由于单一特征难以全面地对交通标志的特点进行描述,因此本文提出采用多特征融合,但不适当的融合算法会使得融合后特征表示力更差。鉴于广义典型相关分析(Generalized Canonical Correlation Analysis,GCCA)已在人脸识别特征融合方面取得不错的效果,本文将其应用于交通标志识别领域中,实验结果表明,HOG与GIST的融合相对于单一特征更有利于分类识别。(3)从实验中发现,融合特征对于结构相似的交通标志在表达上还是存在一定的冗余性,特别是限速类交通标志。因此,本文提出采用字典学习稀疏编码方式对融合特征进行优化,即使用K-SVD算法进行字典学习,正交匹配追踪(Orthogonal matching pursuit,OMP)算法对其进行稀疏编码。实验结果表明,优化后的特征对结构相似的交通标志更有表达力。(4)在GTSRB(German Traffic Sign Benchmarks)上,通过对比实验分析了单个特征、融合特征以及融合-稀疏特征的分类识别效果。实验结果表明经过优化后的特征在线性SVM分类器上获得了更好的识别效果。
[Abstract]:With the development of economy, the improvement of people's living standard and the increasing number of vehicles, traffic jams and safety problems are becoming more and more serious. Therefore, the research and application of Intelligent Transportation system (its) has been attached great importance to by industry and academia. As one of the core technologies of its, traffic sign recognition has always been a hot research topic. However, the research results are not yet mature, this is because in the real complex scene, traffic sign recognition is vulnerable to weather conditions, motion blur, object occlusion, color degradation. Therefore, traffic sign recognition is still a challenging task. The main research work in this paper is as follows: 1) due to the influence of various factors in the external scene. The size, brightness and shape of each traffic sign image are different, so before feature extraction. It needs to be preprocessed. This paper mainly deals with the grayscale and grayscale normalization of the image of 1: 1 from three aspects. (2) Segmentation and extraction of region of interest; (3) based on the characteristics of local edge, global contour and texture of traffic signs, two kinds of feature operators: Hog and GIST are studied in this paper. And the parameters of each feature are optimized. Because it is difficult to describe the characteristics of traffic signs in a single feature, this paper proposes a multi-feature fusion. However, improper fusion algorithm will make the feature representation force worse after fusion. In view of the generalized canonical correlation analysis (. Generalized Canonical Correlation Analysis. GCCA has achieved good results in face recognition feature fusion. In this paper, it is applied to the traffic sign recognition field, experimental results show that. The fusion of HOG and GIST is more favorable to classification and recognition than a single feature. From the experiment, it is found that the fusion features have some redundancy in the expression of similar traffic signs. Especially speed limit traffic signs. Therefore, this paper proposes a dictionary learning sparse coding method to optimize the fusion features, even K-SVD algorithm for dictionary learning. Orthogonal matching tracking algorithm is used for sparse coding. The experimental results show that the proposed algorithm is very effective. The optimized features are more expressive for traffic signs with similar structure. The classification and recognition effects of single feature, fusion feature and fused sparse feature are analyzed by comparison experiments. The experimental results show that the optimized feature has better recognition effect on linear SVM classifier.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】