【信息技术】【2013.05】实时视频中车牌的自动检测跟踪与识别

本文为印度Rourkela国立技术研究所（作者：Lucky Kodwani）的硕士论文，共88页。

交通监控摄像头的视频自动分析是一个基于计算机视觉技术的新兴领域，它是公共安全、智能交通系统（ITS）和交通高效管理的关键技术。近年来，交通活动自动分析的范围不断扩大。我们将视频分析定义为基于计算机视觉的监控算法和系统，从视频中提取上下文信息。在交通场景中，可以通过应用计算机视觉和模式识别技术来支持多个目标的监控，包括检测交通违法行为（例如非法转弯和单行道）和识别道路使用者（例如车辆、摩托车和行人）。目前最可靠的方法是对车牌的识别，即自动车牌识别（ANPR或ALPR）。

介绍了一种全功能的自动车辆检测、跟踪和车牌识别系统，该系统在模式识别和机器视觉等领域有着广泛的应用，从复杂的安全系统到公共区域，从停车场准入到城市交通管控等。由于雾、雨、阴影、光照条件不均匀、遮挡、距离变化、车速、景物的帧内角度、车牌旋转、景物中车辆数量等多种影响，该系统具有复杂的特性。

本文的主要目的是展示一个解决实际场景下汽车识别问题的系统。在整个过程中，从视频采集到光学字符识别，都考虑了车牌的自动识别。视频流中运动目标的检测是信息处理的第一步，背景减法是一种非常流行的前景分割方法。本文针对不同的背景减法进行了仿真研究，以克服光照变化、阴影、背景杂波和伪装等问题。下一步是车牌提取，这是自动交通系统车牌识别的一个重要阶段。我们提出了两种提取车牌的方法，并将其与其他现有方法进行比较。利用基于区域的方法将提取出的车牌分割成单个字符，该识别方案结合了自适应迭代阈值和模板匹配算法。该方法对光照、字符大小和厚度、倾斜和小字符断开具有鲁棒性。除视频流之外，该系统的主要优点是它的实时性，不需要任何额外的传感器输入（例如来自红外传感器的输入）。该系统通过大量的车辆图像和视频进行了评估，系统计算效率高，适用于其它相关的图像识别应用。该系统在门禁、收费、边防巡逻、交通管制、查获被盗车辆等方面有着广泛的应用，而且该技术不需要在车辆上安装任何发射器或应答器。

Automatic video analysis from trafficsurveillance cameras is a fast-emerging field based on computer visiontechniques. It is a key technology to public safety, intelligent transportsystem (ITS) and for efficient management of traffic. In recent years, therehas been an increased scope for automatic analysis of traffic activity. Wedefine video analytics as computer-vision-based surveillance algorithms andsystems to extract contextual information from video. In traffic scenariosseveral monitoring objectives can be supported by the application of computervision and pattern recognition techniques, including the detection of trafficviolations (e.g., illegal turns and one-way streets) and the identification ofroad users (e.g., vehicles, motorbikes, and pedestrians). Currently mostreliable approach is through the recognition of number plates, i.e., automaticnumber plate recognition (ANPR), which is also known as automatic license platerecognition (ALPR), or radio frequency transponders. Here full-featuredautomatic system for vehicle detection, tracking and license plate recognitionis presented. This system has many applications in pattern recognition andmachine vision and they ranges from complex security systems to common areasand from parking admission to urban traffic control. This system has complexcharacteristics due to diverse effects as fog, rain, shadows, unevenillumination conditions, occlusion, variable distances, velocity of car,scene’s angle in frame, rotation of plate, number of vehicles in the scene andothers. The main objective of this work is to show a system that solves thepractical problem of car identification for real scenes. All steps of theprocess, from video acquisition to optical character recognition are consideredto achieve an automatic identification of plates. Detection of moving objectsin video streams is the first relevant step of information and backgroundsubtraction is a very popular approach for foreground segmentation. In thisthesis, different background subtraction methods are simulated to overcome theproblem of illumination variation, shadows, background clutter and camouflage.Next step is License plate extraction which is an important stage in licenseplate recognition for automated transport system. We are proposing two methodsfor extraction of license plates and comparing it with other existing methods.The Extracted license plates are segmented into individual characters by usinga region-based method. The recognition scheme combines adaptive iterativethresholding with a template matching algorithm. The method is robust toillumination, character size and thickness, skew and small character breaks.The main advantages of this system is its real-time capability and that it doesnot require any additional sensor input (e.g. from infrared sensors) except avideo stream. This system is evaluated on a large number of vehicle images andvideos. The system is also computationally very efficient and it is suitablefor others related image recognition applications. This system has wide rangeof applications such as access control, tolling, border patrol, trafficcontrol, finding stolen cars, etc. Furthermore, this technology does not needany installation on cars, such as transmitter or responder.

1. 引言
2. 车辆检测
3. 车牌提取
4. 字符提取
5. 结论与未来工作展望

更多精彩文章请关注公众号：