车牌及车标识别技术的研究

发布时间：2019-04-02 08:59

【摘要】：车辆识别技术是智能交通系统(Intelligent Transportation System,ITS)的关键技术之一,车牌及车标识别技术是车辆识别技术的重要研究领域。车牌与车标是车辆信息的重要组成部分,其在车库管理、路口收费、交通违章抓拍等场景都发挥了极其重要的作用,具有很大的经济价值和现实意义。本课题对近年来车牌识别及车标识别技术相关的理论和算法做了较为全面的了解和分析,系统地阐述了车牌识别技术和车标识别技术的难点。在PC平台上完成车牌识别系统和车标识别系统的开发和优化,其中重点对车牌识别技术中涉及的字符分割和字符识别、车标识别技术中涉及的车标定位、识别部分进行深入研究和改进。本文主要工作如下:(1)在车牌字符分割方面,应用了一组基本图像处理方法对车牌分割前图像进行预处理,并基于此提出了一种动态模板结合非零像素点的字符分割方法。首先根据车辆牌照字符按比例排列分布的特性设置车牌模板,使用模板在预处理后的车牌图像中滑动,并动态地改变模板的宽度,每滑动一次计算模板中七个字符区域内非零像素点的数目,以包含最大的非零像素点数目的模板位置作为最终车牌字符的分割位置,实现字符分割。对实验数据库进行测试,包含预处理在内的字符分割模块的正确率为95.62%,平均耗时14.75ms。(2)在字符识别方面,实现了基于支持向量机(Support Vector Machine,SVM)结合局部二值模式(Local Binary Pattern,LBP)特征的字符识别,经测试检验,该方法有较好的准确率且耗时较少。对识别错误的字符进行统计分析,归类观察到错误字符多为图像质量较差(如倾斜、残缺、模糊等)的字符。基于此,本文整理收集了大量低质量字符样本,加入原字符训练集中重新训练得到分类器,将字符识别正确率从94.80%提高至97.73%,实验结果表明训练集对样本的覆盖程度对分类器的性能有很大影响。(3)在车标定位方面,针对已有车牌检测技术实现车标的定位。该定位方法首先对车标周围水平或竖直纹理进行抑制以突显车标区域,然后消除车标周围的噪声点,利用矩形结构元素实现车标精确定位。对实验数据库进行测试,包含车牌定位在内的车标定位模块的准确率为94.19%。(4)在车标识别方面,实现了基于SVM结合方向梯度直方图(Histogram ofOriented Gradient)特征的车标识别。对实验数据库中的车标进行识别,准确率达到了 97.74%,平均耗时3.60ms,满足实时要求。
[Abstract]:Vehicle recognition technology is one of the key technologies of Intelligent Transportation system (Intelligent Transportation System,ITS). License plate recognition and vehicle mark recognition are important research fields of vehicle recognition technology. License plate and vehicle mark are important parts of vehicle information. They play a very important role in garage management, intersection charging, traffic violation capture and other scenes, which have great economic value and practical significance. This paper makes a comprehensive understanding and analysis of the theories and algorithms related to license plate recognition and vehicle mark recognition technology in recent years, and systematically expounds the difficulties of license plate recognition technology and vehicle mark recognition technology. The development and optimization of the license plate recognition system and the vehicle mark recognition system on the PC platform are completed, in which the character segmentation and character recognition involved in the license plate recognition technology and the car mark location involved in the car mark recognition technology are emphasized. In-depth research and improvement are carried out in the identification section. The main work of this paper is as follows: (1) in the aspect of license plate character segmentation, a group of basic image processing methods are used to pre-process the image before license plate segmentation, and a character segmentation method based on dynamic template combined with non-zero pixel is proposed. Firstly, the license plate template is set according to the character of the vehicle license plate character arranged proportionally, and the template is used to slide in the preprocessed license plate image, and the width of the template is changed dynamically. The number of non-zero pixels in the seven character regions of the template is calculated each time, and the position of the template containing the maximum number of non-zero pixels is used as the segmentation position of the final license plate character to achieve character segmentation. The experiment database is tested. The correct rate of character segmentation module including preprocessing is 95.62%, and the average time consuming is 14.75 Ms. (2) in the aspect of character recognition, the (Support Vector Machine, based on support vector machine is realized. SVM) combined with local binary pattern (Local Binary Pattern,LBP (local binary pattern) character recognition, the test results show that the proposed method has good accuracy and less time-consuming. By statistical analysis of the wrong characters, it is found that most of the wrong characters are the characters with poor image quality (such as slant, incomplete, fuzzy, etc.). Based on this, this paper collects a large number of low-quality character samples, adds the original character training set to re-train to get a classifier, and improves the correct rate of character recognition from 94.80% to 97.73%. The experimental results show that the coverage of the training set has a great impact on the performance of the classifier. (3) in the aspect of vehicle mark location, the vehicle mark location based on the existing license plate detection technology is realized. Firstly, the horizontal or vertical texture around the car mark is suppressed to highlight the area of the car mark, then the noise around the car mark is eliminated, and the precise positioning of the car mark is realized by using the rectangular structure element. The experimental database is tested, and the accuracy of the vehicle mark location module including license plate location is 94.19%. (4) in the aspect of vehicle mark recognition, the vehicle mark recognition based on SVM combined with direction gradient histogram (Histogram ofOriented Gradient) feature is realized. The recognition accuracy is 97.74% and the average time consuming is 3.60ms, which meets the real-time requirement.
【学位授予单位】：广西师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U495;TP391.41

【参考文献】