基于感兴趣区域的道路交通标志检测算法研究

发布时间：2018-02-24 10:21

本文关键词： PHOG特征自适应谱聚类 ICF特征形状分类器感兴趣区域　出处：《安徽大学》2017年硕士论文　论文类型：学位论文

【摘要】：面对多样化的道路环境以及日趋复杂的交通安全问题,智能交通系统(Intelligent Transportation System,简称ITS)和先进辅助驾驶系统应运而生并得到了广泛关注。与此同时,交通标志识别系统作为ITS系统中的核心技术和前沿领域,现已逐渐成为国内外众多科研学者的热门研究课题。在复杂道路环境下的交通标志容易受到光照、天气、颜色退化、形状失真及动态模糊等因素的影响,因此有效的道路交通标志检测方法就成为了研究热点,本文的主要工作内容包括以下几点:(1)由于自然场景下的道路环境复杂多变,因此采集到的交通标志图像需经过预处理环节才能更好地进行检测与识别。本文采用了 Gamma校正以及梯度锐化方法来减少光照强度变化的影响以达到增强图像的目的;同时也比较了中值滤波与自适应多级中值滤波以及高斯滤波与自适应高斯滤波的噪声抑制能力;最后对比了 RGB、HSV、Lab三种不同颜色空间的阈值分割效果。(2)提出了一种基于感兴趣区域和随机森林分类器的交通标志检测方法,该检测方法首先使用线性SVM分类器对交通标志图像进行颜色转换,同时结合形状模板匹配思想计算其对应尺度和位置上的交通标志匹配分数,并根据其匹配分数与事先设定的阈值进行比较判断,可获得最终的候选ROIs区域;然后提取候选ROIs区域的PHOG特征描述子,结合结构模型学习策略生成随机森林分类器以完成干扰ROIs区域的剔除,提高了目标区域检测的准确性;最后在学校周边不同路段环境下拍摄到的380张交通标志图像测试数据库中进行了三组不同的对比实验,结果表明基于ROIs区域和随机森林分类器方法有着较好的检测效果。(3)针对传统的基于ICF特征和Adaboost交通标志检测算法,召回率过低和误检率过高的问题,提出了一种基于积分通道特征和多级分类器的交通标志检测方法,该检测方法首先对ICF特征进行自适应谱聚类并结合Adaboost算法学习得到目标的ROIs区域;然后对所获得的ROIs区域进行直方图均衡化,并利用LLC编码后的SIFT特征与线性SVM相结合生成四种不同的形状分类器;最后通过GTSDB数据库的验证,结果表明采用SICF-Adaboost+LSIFT-SVM构建的交通标志级联分类器检测方法对光照变化、形状失真、颜色退化、动态模糊等情况具有一定的鲁棒性。
[Abstract]:In the face of diversified road environment and increasingly complex traffic safety problems, Intelligent Transportation system (ITSs) and Advanced Auxiliary driving system (ACIS) have emerged as the times require and received extensive attention. Traffic sign recognition system, as the core technology and front field of ITS system, has become a hot research topic for many researchers at home and abroad. Traffic signs in complex road environment are vulnerable to light, weather and color degradation. Because of the influence of shape distortion and dynamic ambiguity, effective road traffic sign detection method has become a research hotspot. The main work of this paper includes the following points: 1) because of the complex and changeable road environment in natural scene, Therefore, the collected traffic sign images can be detected and recognized better only by pre-processing. In this paper, Gamma correction and gradient sharpening methods are used to reduce the influence of light intensity change to achieve the purpose of image enhancement. At the same time, the noise suppression ability of median filter and adaptive multistage median filter, Gao Si filter and adaptive Gao Si filter are compared. Finally, a method of traffic sign detection based on region of interest and random forest classifier is proposed by comparing the threshold segmentation effect of three different color spaces. The detection method firstly uses the linear SVM classifier to transform the color of the traffic sign image and calculates the traffic sign matching score on the corresponding scale and position with the shape template matching idea. The final candidate ROIs region can be obtained by comparing the matching score with the pre-set threshold, and then the PHOG feature descriptor of candidate ROIs region is extracted. The random forest classifier is generated by structural model learning strategy to eliminate the interference ROIs region and improve the accuracy of target area detection. Finally, three groups of different comparative experiments were conducted in 380 traffic sign image test databases taken in different road sections around the school. The results show that the method based on ROIs region and stochastic forest classifier has better detection effect. Aiming at the problems of low recall rate and high false detection rate based on traditional ICF feature and Adaboost traffic sign detection algorithm, the proposed method has the following advantages: 1. A traffic sign detection method based on integral channel feature and multilevel classifier is proposed. Firstly, the ICF features are clustered by adaptive spectral clustering and the target ROIs region is obtained by combining with Adaboost algorithm. Then, the obtained ROIs region is equalized by histogram, and four different shape classifiers are generated by combining the SIFT features of LLC coding with linear SVM. Finally, four different shape classifiers are generated through the verification of GTSDB database. The results show that the traffic sign cascade classifier based on SICF-Adaboost LSIFT-SVM is robust to illumination change, shape distortion, color degradation and dynamic blur.
【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】：U463.6;TP391.41

【相似文献】