基于深度学习的车辆检测和车牌定位

发布时间：2018-03-18 05:23

本文选题：深度学习　切入点：卷积神经网络　出处：《江西理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着交通管理技术的逐步发展和完善,车牌识别技术如今已普遍使用在道路监控和道路指挥系统中,如高速路收费站、路口车流行人监控、小区以及停车场自动收费放行系统等。同时它也是智能交通系统的重要组成部分,为解决道路拥堵的状况提供了新的方案,能够帮助决策者快速高效的制定执行计划,节约了劳动成本。车辆牌照是汽车独有认证标志,因而对车牌相关技术的学习和探讨能带来重大社会价值。本文介绍了车牌识别的历史和背景,并深入了解了车牌识别技术在国内外的发展现状,介绍近几年一直很流行的机器学习算法及其在图像检测识别等方面的应用。克服现有算法的一些局限性,并结合机器学习的相关算法,提出一种机器学习和图像处理技术相结合的车牌识别系统,利用深度学习和图像处理技术来实现。本文车牌识别着重从车辆位置的检测、车牌位置的确定这两方面来介绍。文中对各个部分的常见算法进行了总结,并对相关算法进行了改进,利用深度学习和图像处理的知识进行优化。主要工作如下:⑴介绍数字图像处理相关技术,对图像中三种颜色空间以及它们之间相互转换进行了简要的描述和分析,介绍了数学形态学原理及其在图像滤波去噪方面的作用,针对文中要使用的卷积神经网络和角点密度聚类,描述了这两种算法的基本概念,实现方式。⑵为了解决传统车辆检测存在的问题,提高车辆检测的准确率,提出将区域卷积神经网络算法应用到车辆检测中。该方案依照图像颜色层次相关特征,产生潜在车辆待选区域。建立相应卷积神经网络模型,提取每个候选区域局部特征。对卷积神经网络模型做出改良,修改原输入图像大小,其网络参数也做出相应调整。选定正负样本进行SVM分类器训练,采取SVM分类器进行车辆候选区域分类,最后判断车辆信息。通过实验数据论证,本文改进的卷积模型在车辆检测测试中获得较优异的效果。⑶为了解决传统车牌定位算法性能不够理想的情况,提出一种角点密度统计方法对车牌进行定位。第一步,依据车牌自身的颜色特性,将整幅图像从RGB彩色空间变换为HSL彩色空间,对获取的HSL图像进行阈值化处理,然后采用一系列形态学方法完成图像滤波,剔除无用信息。接着,对滤波后图像使用角点检测算法,获取角点数量、坐标信息。最后采取DBSCAN角点密度判定准则确定车牌位置。实验结果表明,此算法定位精度也较高,定位时间较快,能满足实时性需求。
[Abstract]:With the development and improvement of traffic management technology, license plate recognition technology has been widely used in road monitoring and road command system, such as highway toll stations, traffic and pedestrian traffic monitoring, It is also an important part of the Intelligent Transportation system, which provides a new solution to the congestion of roads, and can help decision makers to make implementation plans quickly and efficiently. This paper introduces the history and background of license plate recognition, which can bring great social value to the study and discussion of license plate related technology. And deeply understand the development of license plate recognition technology at home and abroad, introduce the machine learning algorithm and its application in image detection and recognition, which has been very popular in recent years, overcome some limitations of existing algorithms. Combined with the related algorithms of machine learning, a license plate recognition system combining machine learning and image processing technology is proposed, which uses depth learning and image processing technology to realize the license plate recognition. In this paper, the common algorithms of each part are summarized, and the related algorithms are improved. Using the knowledge of depth learning and image processing to optimize. The main work is as follows: 1 introduces the digital image processing technology, describes and analyzes the three color spaces in the image and the conversion between them. This paper introduces the principle of mathematical morphology and its role in image filtering and denoising, and describes the basic concepts of these two algorithms for the convolution neural network and corner density clustering used in this paper. In order to solve the problems existing in traditional vehicle detection and improve the accuracy of vehicle detection, a regional convolution neural network algorithm is proposed for vehicle detection. The corresponding convolution neural network model is established to extract the local features of each candidate region. The convolution neural network model is modified to modify the original input image size. The network parameters are adjusted accordingly. The positive and negative samples are selected for SVM classifier training, and the SVM classifier is used to classify vehicle candidate regions. Finally, the vehicle information is judged. In order to solve the problem that the performance of the traditional license plate location algorithm is not ideal, a corner density statistical method is proposed to locate the vehicle license plate. According to the color characteristics of license plate, the whole image is transformed from RGB color space to HSL color space, and the obtained HSL image is thresholded. Then, a series of morphological methods are used to filter the image and eliminate the useless information. Corner detection algorithm is used for filtered images to obtain corner number and coordinate information. Finally, DBSCAN corner density criterion is adopted to determine the location of license plate. The experimental results show that the algorithm has higher accuracy and faster localization time. It can meet the real-time requirement.
【学位授予单位】：江西理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：U495;TP391.41

【参考文献】