【作者】 沈涛；

【导师】 邵春福；

【作者基本信息】 北京交通大学 ， 交通运输规划与管理， 2011， 硕士

【摘要】 摘要：目前交通监控的主要方式是两种：人工观察；摄像机记录,人工统计这两种方式都需要投入大量的人力和物力资源,并且人工统计的准确性有限,可能会出现疏漏,对异常情况不能及时做出反应。为此需要建立一种不需人工干预、或者只需要很少人工操作的智能交通管理系统,通过安装在固定位置的摄像机拍摄视频,实现对目标的定位、识别和跟踪及交通参数的提取和分析,并在此基础上进一步实现目标(例如行人、车辆等)行为的分析与判断。目前,针对车辆的目标检测、跟踪、识别研究已经比较成熟,但针对行人的目标检测、跟踪、识别研究则相对滞后,而且具体情况异常复杂,例如行人为非刚体,在视频中的面积小,行动较为随机,速度并不连贯,而且行人之间的间距有可能较小,这些都给相关技术造成了很大的挑战。本文对上述难点进行了研究,通过提取交叉口的行人流量、速度、步行方向等参数,给行人的交通管理与控制以及仿真系统模拟现实提供基础数据。本文的主要研究内容包括：(1)分析了智能交通监控视频中的降质因素,并研究了各种噪声的数学模型,针对交通视频中最常见的椒盐噪声和高斯噪声实现了经典滤波算法的去噪,并设计了一种改进算法,将加入图像像素点是否为噪声的判决预处理,只针对可能为噪声点的区域进行去噪,而对于非噪声点不进行去噪,更好地保持了图像的细节信息。(2)对于行人检测和识别问题,针对各种具体难点设计了具体的解决方案和步骤。比较了常用目标检测算法,并根据交叉口智能交通监控视频的特点选用了合适的行人检测算法。由于检测算法可能造成区域的不完整和空洞,采用形态学操作对行人目标区域的断裂区域实现连接,并对目标区域内的空洞进行填充。对可能存在噪声的区域采用区域面积统计的方法计算面积,并去除小面积的噪声斑块。对阴影采用几何形状的方法予以去除。对交叉路口复杂环境中的各种运动目标进行建模,并总结了行人目标的特点,实现了基于视频检测的行人目标识别。(3)实现了交通视频中交叉口行人目标的跟踪。通过对摄像机视场的标定,完成了摄像机空间拍摄的视频数字图像中的坐标到实际物理世界空间坐标系的映射换算,提取了包括行人目标的质心、位移、速度、加速度和流量等交通参数,为仿真系统模拟现实提供实际状况提供了翔实的基础数据。综上所述,本文的研究成果能够为仿真系统模拟现实提供实际状况,并分析所在路口的交通状况,为交通管理与控制提供依据。更多还原

【Abstract】 ABSTRACT:Currently, most monitoring of traffic scene is mainly achieved through manual supervisory control or artificial observation of video after the event. These methods need a lot of manpower, material resources; furthermore, omissions may occur abnormal situations can’t be responded due to the limitation of human’s energy and attention. Thus there is need to establish intelligent traffic management system that is without human intervention or requires very little manual operation. In the system, installing cameras are fixed to capture video to achieve targets location, identification and tracking. Behavior of goals (such as pedestrians, vehicles, etc.) are analyzed and judged based on the traffic parameters that are analyzed and extracted.The methods of vehicle detection, tracking and recognition have been proposed and achieved some fruits and contributions, however, pedestrian detection, tracking and recognition is relatively lagging behind. Besides, the specific situation of pedestrian is much more complex. The pedestrian area in the video is small; the action of pedestrian is random; the speed of pedestrian is not consistent; and the spacing between the pedestrians may be very small. The above problems pose a great threat on technological realization. Therefore, it is required to search and solve the difficulties of parameters extracting, including the amount of pedestrian on intersection of traffic, speed, direction and other parameters important traffic data which are important to the simulation system that can simulate the actual state of reality.The main contents of this paper include:(1) The lower quality factors of the intelligent traffic surveillance video were analyzed and various noise models are summarizes. The most common noise in intelligent traffic video, salt & pepper noise and Gaussian noise, were reduced by classical filtering algorithms. An improved algorithm was proposed in which the image pixels will be pre-judged as noise or non-noise. Only possible pixels in the region of noise were filtered while non-noise pixels were not filtered, thus the image details were remained better.(2) For pedestrian detection and identification, specific solutions and specific steps were developed to solve the difficulties. The commonly algorithms used for target detection were compared and suitable pedestrian detection algorithm was selected according to the characteristics intelligent traffic intersection surveillance video. The detected region of pedestrians is often incomplete and has empty regions, morphological operations were employed on the pedestrian region to connect the fracture of the target area and fill the holes in the target region. Small noise patches that may exist in region of pedestrians were removed by area calculation. The shadow of pedestrian was removed by geometric-based method. Various moving object models in the complex environment of intersection were established and the characteristics of pedestrian goals were summarized to achieve video-based pedestrian detection and recognition.(3) Pedestrian tracking algorithm in intelligent traffic video intersections was realized. Through calibration of the camera field, the camera space of digital image coordinates in video was mapped to the actual physical world space coordinate. A lot of important traffic parameters were extracted, including center of pedestrian target, displacement, walking speed, walking acceleration, flow rate of pedestrian traffic and so on. The valuable traffic information and data are important for the simulation system to provide realistic simulation of the actual situation.In conclusion, the results and conclusions of this research project can provide practical simulation system with the real traffic situation which analyzes the traffic situation in the junction of the transportation system. Besides, it provides the basis for comprehensive management and maintenance. Thus, this paper is not only of great theoretical significance, but also has extensive application value.更多还原