基于视频的运动车辆检测算法研究与系统实现

发布时间：2018-10-18 09:32

【摘要】：车辆检测和车流统计是智能交通系统（Intelligent Transportation System，ITS）中最重要的研究内容之一。在研究分析了现有的车辆检测技术基础上，本文研究了数字图像运动目标检测以及基于视频的车辆检测的相关算法，并设计实现了一个基于视频的车辆检测系统。在本文中，针对目前采用单一方法进行车流检测的缺陷与不足，提出了自动判断道路拥堵状态的算法，以及适合于不同拥堵状态下的基于检测线与基于虚拟线圈的车流检测算法，大大提高了系统的检测准确率。本文的研究还涉及到了运动车辆检测、车辆阴影去除、夜晚车灯去除、拥堵检测和不同检测方式之间的切换等。主要研究内容如下：（1）关于运动车辆检测的研究在分析了多种车辆检测算法之后，本文采用了基于背景差分的车辆检测方法。为了满足实时性的需要，本文提出了一种基于采样点多帧累计帧差的分区块的自动背景更新方法，相比传统的方法，此方法计算量小，大大提高了计算速度与背景更新速率，保障了系统的实时性。（2）关于夜晚车灯去除的研究在夜晚，车灯灯光会在运动车辆的前方路面照射出高亮度的光斑，如果不加以去除会对车辆检测识别造成极大的干扰，本文在实验的基础上，，提出了一种基于梯度差异的灯光滤除算法，取得了良好的效果。（3）关于拥堵检测和计数方式切换的研究本文共提出了两种车辆检测、计数的方法，分别适用于拥堵的交通状况和非拥堵的交通状况，这样不管交通状况如何，均能取得良好的效果。非拥堵状况下，本文采用背景差分检测出车辆目标，采用基于检测线的方式进行车流计数，即使车辆出现变道的情况也不发生检测计数错误，且背景的实时更新使得运动目标分割更加准确。在拥堵状况下，停止背景更新而采用备用背景，避免了拥堵状况下前景融合到背景中去，继而采用基于虚拟线圈的计数方式进行检测、计数，避免了基于检测线的检测方式在交通拥堵时由于运动目标黏连而导致的错误分割。根据检测计数方式的不同，本文相应解决了如何在两种检测计数方式间实现自动切换的问题。为了保证系统的准确运行，如何准确实时地检测交通状况是否拥堵是极其必要的，本文提出了一种基于多帧帧差均值变化的拥堵检测方法。在上述运动目标检测的理论研究基础上，本文设计并实现了视频车辆检测及车流计数系统。通过在重要路口架设摄像头，运用图像处理技术进行运动车辆目标的检测。该系统作为武汉市智能交通系统的重要组成部分，车辆检测计数准确率达到99%以上，能以自动化的方式检测主要路口桥梁的车流量以及拥堵状况，实现了交通的智能管理，提高武汉市交通系统的效率。
[Abstract]:Vehicle detection and traffic statistics is one of the most important research contents in Intelligent Transportation system (Intelligent Transportation System,ITS). Based on the research and analysis of the existing vehicle detection technology, this paper studies the digital image moving target detection and the video based vehicle detection algorithm, and designs and implements a video-based vehicle detection system. In this paper, an algorithm to automatically judge the traffic congestion state is proposed to solve the defects and shortcomings of the single method used to detect the traffic flow. The detection accuracy of the system is greatly improved, and the detection algorithm based on detection line and virtual coil is suitable for different congested conditions. The research in this paper also involves moving vehicle detection, vehicle shadow removal, night lights removal, congestion detection and switching between different detection methods. The main research contents are as follows: (1) after analyzing various vehicle detection algorithms, this paper adopts the vehicle detection method based on background difference. In order to meet the need of real-time, this paper presents an automatic background updating method based on multi-frame accumulative frame difference of sampling points. Compared with the traditional method, this method has less computation, and greatly improves the speed of calculation and the rate of background updating. (2) Research on the removal of vehicle lights at night. At night, the lights of the vehicle will shine bright spots on the road surface in front of the moving vehicle. If it is not removed, it will cause great interference to vehicle detection and recognition. On the basis of experiments, a light filtering algorithm based on gradient difference is proposed in this paper. (3) the research on congestion detection and counting mode switching. In this paper, two methods of vehicle detection and counting are proposed, which are suitable for congested traffic conditions and non-congested traffic conditions, respectively. In this way, regardless of the traffic conditions, can achieve good results. In the condition of non-congestion, the background differential detection method is used to detect the vehicle target, and the traffic flow count is carried out based on the detection line. Even if the vehicle changes the track, there is no detection error. And the real-time updating of background makes moving object segmentation more accurate. In the congested condition, the background update is stopped and the backup background is adopted, which avoids the foreground fusion into the background in the congested condition, and then uses the counting method based on virtual coil to detect and count. The detection method based on detection line can avoid the wrong segmentation caused by the adhesion of moving objects in traffic jam. According to the different ways of detecting and counting, this paper solves the problem of how to realize automatic switching between two kinds of detection and counting methods. In order to ensure the accurate operation of the system, it is extremely necessary to detect traffic congestion accurately and in real time. In this paper, a congestion detection method based on the change of the mean value of frame difference is proposed. Based on the theoretical research of moving target detection, a video vehicle detection and vehicle flow counting system is designed and implemented in this paper. By setting up cameras at important junctions and using image processing technology to detect moving vehicle targets. As an important part of Wuhan Intelligent Transportation system, the accuracy of vehicle count is over 99%. The system can automatically detect the traffic flow and congestion of the bridge at major junctions, and realize the intelligent traffic management. Improve the efficiency of Wuhan traffic system.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：U495;TP391.41

【相似文献】