【作者】 李伦波；

【导师】 马广富；

【作者基本信息】 哈尔滨工业大学 ， 控制科学与工程， 2008， 博士

【摘要】 随着社会的发展,汽车逐渐地普及,现有的道路通过能力逐渐难以满足交通量快速增长的需要。交通拥挤加剧,事故频发,公路交通安全和运输效率等问题变得日益突出。在此背景下,欧美发达国家已从修建更多的道路、扩展路网规模逐步转移到采用高新技术改造现有的道路交通系统和管理体系上,即开展了智能交通系统(ITS: Intelligent Transportation System)的研究。ITS是一个集通讯、检测、控制与计算机等技术为一体的综合信息管理系统,实现该系统的一个关键技术是发展具有主动安全技术的智能车辆。发展具有主动安全技术的智能车辆已成为各国政府、研究机构和汽车制造商的一个重要目标。基于计算机视觉的交通标志识别是智能车辆的关键技术和难点之一,它包括自然场景下交通标志的检测与较大类别交通标志的分类理解技术。尽管学者们对这些问题进行了多年的研究,但问题仍然没有得到很好的解决,主要表现为复杂环境下算法的鲁棒性较差。针对这种情况,本论文在交通标志检测、自然场景图像的色彩增强与交通标志分类三方面展开研究。为了快速、准确地从自然场景图像中检测出交通标志,提出了一种基于局部特征的标志检测算法。该算法先将RGB格式的交通标志图像转换到HSV彩色空间,通过固定阈值进行颜色分割,提取出目标区域;然后针对多边形和圆形交通标志,提出一种统一的对称局部特征检测模板来提取目标区域的特征,在此基础上根据交通标志的形状特征设计一组模糊推理规则来判定目标区域的形状,进而从场景图像中检测出交通标志。对晴天、多云与小雨天气状况下共3000幅自然场景图像进行交通标志检测实验,实验结果表明,该算法能够克服标志大小变化、旋转变化和视角变化等的影响,具有良好的鲁棒性,验证了该检测算法的有效性。针对光照变化、天气变化和颜色退化等因素造成的交通标志漏检问题,提出了一种改进的颜色分割算法。该算法先对RGB格式的交通标志图像采用色彩恒常性算法进行彩色增强,然后将其转换到HSV彩色空间进行颜色分割。标志检测实验结果表明,该算法能够有效地克服光照、天气、颜色退化对交通标志颜色的影响,更好地分割出交通标志的特征颜色,从而进一步提高交通标志正确检测率。针对交通标志这类大类别分类问题,提出一种由粗到精的分层决策分类算法。该算法先利用我国交通标志的特征颜色进行分类决策,然后再根据颜色和形状进一步分成子类,这样可简化分类系统设计,提高分类精度。在对子分类系统设计时,提出一种概率神经网络(PNN: Probabilistic Neural Network)优化设计算法。该算法在利用径向Tchebichef不变矩提取交通标志特征的基础上,采用Global K -means聚类算法优化PNN网络结构,然后利用粒子群算法(PSO: Particle Swarm Optimization)优化模式层核函数控制参数矩阵,以提高网络的泛化性能。实验结果表明,该分类系统不仅具有精简的结构,而且可获得较高的分类精度和较好的泛化性能。考虑到交通标志分类始终是有限样本分类问题,在研究基于支持向量机(SVM: Support Vector Machine)的多类别分类理论的基础上,对交通标志分类系统中子类标志设计了C -支持向量分类机和ν支持向量分类机进行分类。交通标志分类实验表明,基于支持向量机的分类器能够获得较高的分类精度和更好的泛化性能。更多还原

【Abstract】 With the development of the times, automotive vehicles gain popularity, the existing roads cannot meet the demand of traffic flux which increases gradually, and the problems such as traffic jams, accidents, road safeties and efficiencies in transportation are becoming more and more serious. Under this background, developed countries in Europe and America have been improving the existing transportation systems with high and new technique, rather than constructing more highways to expand the transportation capacity, which means that the developed countries have been developing Intelligent Transportation System (ITS). ITS is an integrated system combining the technology of communication, measurement, control and computer, and the research of intelligent vehicles based on computer vision is one of the key problem for the realization of ITS.Developing the intelligent vehicles with active safety technique has been one of the important motives of the governments, institutes and vehicle manufactures. Traffic sign recognition based on computer vision is one of the key techniques for the design of intelligent vehicles, and the system consists of two parts: one is traffic sign detection under natural scenes, the other is traffic signs classification. Although many years of studies, the robust performance of existing recognition algorithms is not good, and traffic sign recognition is still an open problem. According to this problem, the main contents of this dissertation are concentrated on the three parts: traffic sign detection, color constancy algorithm for the images of natural scenes, and traffic sign classification.A novel algorithm based on local features of region of interest is proposed for the fast detection of traffic signs in natural environments. In the algorithm, every RGB image was converted into HSV color space which was segmented by the hue and saturation thresholds, then the region of interests (ROIs) were extracted. A symmetrical detector of local binary features is proposed to extract the local features of the region of interests, and the shape of ROI is determined using a set of fuzzy rules with the local features, then the traffic signs are detected from the images under natural scenes. Experiments were conducted for the detection of traffic signs which involved 3000 images under sunny, cloudy and rainy weather conditions. The experimental results indicate that the proposed algorithm possess of high detection accuracy, and good robust performance.A modified color segmentation algorithm is presented with respect to the variations of the light conditions, weather conditions, and the paint on the signs which are the important factors resulting false detection of traffic signs. The RGB images were enhanced with color constancy algorithm, then were converted into HSV color space and segmented by the constant thresholds. Experimental results of traffic sign detection show that the segmentation algorithm is efficient with respect to the variations of the light conditions, weather conditions, and the paint on the signs, which improves the detection accuracy.A hierarchical coarse-to-fine classification scheme is proposed for traffic signs classification. In the classification module, a decision tree was designed using the features of color and shape for the coarse classification, and radial Tchebichef moments were adopted to extract the features of traffic signs, then probabilistic neural networks (PNN) were adopted for the further classification which incorporates the global K-means algorithm and Particle Swarm Optimization to improve the generalization. The experimental results demonstrate that the classification module is not only parsimonious but also has higher accuracy better generalization.After all, the traffic signs classification is a kind of classification with limited samples, Support Vector Machine (SVM) is the best tool for limited samples classification with perfect generalization. With the comprehension of SVM, C -Support Vector Classifiers andν-Support Vector Classifiers were designed for the fine classification of the hierarchical classification model. The experimental results demonstrate that the sub-classifiers based on SVM have achieved high accuracy and better generalization.更多还原