复杂场景下的车牌检测算法研究及其工程实现

【作者】 刘攀桂；

【导师】 茅耀斌；

【作者基本信息】 南京理工大学 ， 系统工程， 2008， 硕士

【摘要】 车牌识别系统(LPR)是智能交通的重要组成部分,在现代交通管理、社会治安等方面发挥着重要作用,有着广泛的应用前景。车牌检测是车牌识别系统的关键技术之一。随着车牌识别系统的广泛使用,过去在受控环境和简单场景下的车牌检测算法已经不能满足当前的需要。因此,在复杂场景和成像条件情况下的车牌检测逐渐成为当前研究的热点。本文对复杂场景和成像条件下的车牌检测算法进行了研究,主要完成了以下工作。首先,本文实现并改进了一种启发式的车牌检测算法。该算法根据车牌区域垂直边缘丰富的特点,提取图像中的垂直边缘,并通过连接边缘,获得车牌候选区域。最后根据车牌的几何、边缘分布以及颜色等特征对候选区域进行筛选得到车牌区域。其次,本文实现了一种基于Adaboost机器学习算法的车牌检测算法。在该算法中,定义了一种表征局部灰度与全局灰度比例的矩形特征,使用CS-Adaboost算法训练了一个由85个特征组成的级联分类器,并由此实现了一种车牌检测算法。以上两种算法都在PC平台上编程实现,并且在一个583张图像组成的测试集上进行了实验,给出了结果和性能分析。此外,图像中的倾斜车牌不利于后续的字符分割等操作。针对该问题,本文实现了一种基于“跨栏模型”和“窄孔透射模型”的倾斜车牌矫正算法,给出了实验结果。最后,本文将启发式车牌检测算法移植到了DSP平台上,构建了一个车牌检测系统,可以在CIF格式的图像上达到10f/s的检测速度。文中介绍了算法在DSP平台上的移植和优化过程,并给出了实验结果。更多还原

【Abstract】 License Plate Recognition (LPR) system is an important part of the Intelligent Transportation System (ITS), and plays an important role in modern traffic management and social security. License Plate Detection is one of the key technologies in LPR systems. With the extensive application of LPR systems, the algorithms used to work under controlled conditions and simple scenes is unable to fulfill the requirement. Therefore, license plate detection under complex scenes and imaging conditions becoming a research focus.In this thesis, license plate detection under complex scenes and imaging conditions is investigated, contributions are listed below.First, an improved heuristic license plate detection algorithm has been realized. In this algorithm, edge extraction is performed first according to the characteristics of regions contain license plate that they have a high density of edge information. By connecting the edges, several candidates are generated. The license plate was obtained by filtering these candidates with geometry features, edge distribution and color.Second, an algorithm based on Adaboost has been implemented. A kind of rectangle feature is selected, which indicates the percentage of the sum of a local area in the whole region. A cascade classifier with 85 features was trained by CS-Adaboost algorithm, and then a license plate detection algorithm was implemented. An experiment on a test set contains 583 images is carried out, the result and performance analysis was presented.Since the slant of the license plate would affect the character segmentation and other subsequent operation, a slant correction algorithm was implemented based on the hurdle model and tubiform interspace projection model. The result of experiment was presented.Finally, a license plate detection system was implemented by transplant the heuristic algorithm to DSP platform. This system can perform license plate detection on CIF images at the speed of 10 f/s. The result of experiment was reported.更多还原