【作者】 杨萍；

【导师】 蔡坤宝；

【作者基本信息】 重庆大学 ， 信号与信息处理， 2016， 硕士

【摘要】 随着目前一些不良驾驶习惯的呈现,如开车时接电话,与车里其他人聊天等,使得驾驶员的注意力被转移,以至驾驶员忽略交通标志的现象越来越普遍。恶劣的天气和其它不良的驾驶环境有时也会使驾驶员不能一直保持警惕,而忽略路上的重要交通标志。众所周知,当驾驶员错过或者忽略一些重要的交通标志时,可能会导致交通事故的发生。并且目前交通越来越拥挤,交通标志也越来越多,因此设计一个能够帮助驾驶员识别交通标志以便减少交通事故发生的系统是很有必要的。本文主要在研究分析现有的交通标志识别相关算法的基础上,将低照度环境现实场景下的交通标志作为主要研究对象,提出适合全天候全路况的交通标志检测与识别算法,并在算法设计验证完成后,将其移植到硬件平台上实现。针对能见度较低的道路环境,在进行交通标志检测前,本文恰当地采用了带色彩恢复的Retinex算法对图像进行预处理增强,为提高系统识别的准确率作了较好的前期处理。在交通标志检测阶段采用基于颜色空间和形状特征相结合的方法,对比了基于归一化RGB和基于HSV颜色空间的分割效果,并将分割后的感兴趣颜色分量分别存储以避免相互干扰。此外,颜色分量低于某阈值的图像没有后随的处理,可提高系统的效率。在形状检测方面,为了提高准确率,对交通标志的不同形状采用了不同的检测方法,最后提取出交通标志候选区域。针对交通标志识别阶段,为了得到良好的识别效果和实现较好的推广能力,本文采用Hu不变矩作为特征提取对象,设计基于支持向量机的多类分类器,并通过实验选择了合适的分类器参数。最后,本文选择Cubieboard4作为系统实现的硬件平台,对于Cubieboard4较高的主频以及足够的内存,本文通过使用GCC直接编译将软件系统移植到嵌入式平台。根据实验结果,本系统能够达到实时性要求,并且在低能见度下具有较高的识别精度,体现了系统的可行性。更多还原

【Abstract】 With the presence of some bad driving habits, such as answering the phone while driving, and chatting with other people in the car, the driver’s attention is easy to be scattered. Thus, the phenomenon of driver’s ignoring traffic signs is becoming more and more common. Bad weather and other bad driving conditions sometimes cause drivers not to be vigilant all the time, and thus, they often ignore important traffic signs along the road. It is well known that traffic accidents can occur when drivers miss or ignore some important traffic signs. As the traffic becomes more and more crowded and there are more and more traffic signs, it is necessary to design a system that can help drivers to identify traffic signs in order to reduce traffic accident.On the basis of researching and analyzing existing traffic sign recognition algorithms, this thesis takes such traffic signs lying in a real scenarios with low illumination as the main research object, and presents such an algorithm that is suitable for all-weather road traffic sign detection and recognition. After the algorithm design verification is completed, the identification system is transplanted to and implemented on the hardware platform.In view of the low visibility road environment, the thesis, in the stage before the traffic sign detection, properly uses the Retinex algorithm with color restoration to preprocess the images for enhancement. This is really a good preprocessing algorithm for improving system identification accuracy. In the stage of traffic sign detection, a method based on the combination of color space and shape characteristics is adopted. The segmentation results based on the normalized RGB and the HSV color space are compared. In order to avoid mutual interference, the thesis proposes that the interesting color components after segmentation are stored separately. Additionally, the image with color components below a threshold has no subsequent processing, which can improve the system efficiency. In the stage of shape detection, in order to improve the recognition accuracy, different methods are used to detect traffic signs with different shapes, and finally candidate regions of traffic signs are extracted. In the stage of road traffic sign recognition, in order to achieve good recognition effect and generalization ability, Hu invariant moments are used for feature extraction object in this thesis. Multi class classifier with appropriate classifier parameters based on support vector machine is successfully designed.Finally, Cubieboard4 is chosen as the hardware platform of system implementation. Because of Cubieboard4’s higher frequency and plenty of memory, software system can be transplanted into the embedded platform using GCC by direct compiling for this system. According to the experimental results, the system can achieve the real-time requirements, and has high recognition accuracy in low visibility, which shows the feasibility of the system.更多还原