【作者】 韩亚伟；

【导师】 张有志；

【作者基本信息】 山东大学 ， 通信与信息系统， 2010， 硕士

【摘要】 当前社会是一个高速发展的社会。经济在不断的发展,硬件水平也在不断提高。随着人们生活现代化程度的提高,视频监控系统的应用也变得无处不在。利用计算机图像处理技术提高视频监控系统的自动化程度,尽量减少人工作业,将是视频监控系统未来的发展方向。特别是在各种公共场所中,人群的流动越来越频繁。如何对公共场所中的人群有效的管理和控制,是当前需要解决的一个关键问题。由此,动态场景下的人数统计方法就应运而生。本文介绍了目标检测与跟踪技术的现状,研究了主要检测、跟踪算法的基本理论,分析了当前主要算法的优缺点。在此基础上构建了一个人物分割与人群跟踪相结合的人物计数系统,在计算速度与稳定性方面较已存在系统有了较大提高。系统主要包括四个部分：目标检测、粗略分割、单人计数、人群计数。首先对监控区域运动目标检测提取并滤除噪声,将背景区域中被认定为人群的区域分割并标注,然后分别对人群和单人跟踪计数,最后将数据相加得到具体人数。在目标检测中,针对提取效果和实时性方面的要求,本文使用了背景差分与帧间差分相结合的检测算法,使用帧间差分完成背景更新,使用背景差分提取前景。在分割过程中,为了避免遮挡物的影响,采用重叠率代替边框中心点之间距离来比较相似度。在人群跟踪步骤中,使用面积统计法,对人物的数量进行粗略统计,并利用融合与分裂检测,补偿第一次分割时产生遮挡所造成的误差。在人群统计过程中,依据相关性进一步补偿了由于人物分割导致的误判。实验证明,本文构建的视频监控系统,可以对进入监控范围内的人物进行数量统计,并在提高实时性的同时也兼顾了统计的准确率。更多还原

【Abstract】 At present society is a rapidly developing society. Along with economical development, progress of the hardware is also constantly. With the modernization of people’s lives, the application of video surveillance systems has become ubiquitous. The future of video surveillance system development focus on enhance the degree of automation and minimizing manual operations, so make full use of computer image processing technology is one of the most important way. In particular, in various public places, the crowd movements become more frequent. How the crowd thoroughly enjoyed themselves in public places, effective management and control, is currently a key issue to be resolved. As a result, the counting of the people in dynamic scenes under the statistical methods came into being.This article describes the development status of target detection and tracking technology. Study the basic theory of tracking algorithm and main detection, analysis of the advantages and disadvantages of current algorithm. On this basis, combine character segmentation and crowned tracking to build people counting system. Our system has been greatly enhanced at calculation of speed and stability. System mainly includes four components:target detection, a rough segmentation, single count, the crowd count. Firstly, the regional campaign on the monitoring target detection extract and filter out noise, the background regions have been identified as the crowd of region segmentation and annotation, and then separately track the population and single count, the final data to be added to the exact number.Completion of background updating using the frame difference, foreground extracted using background difference. In the segmentation process, in order to avoid the impact of occlusion, we consider similarity between tracker and measure not only by comparing central positions but also height and width of the bounding box:using overlapping area. The overlapping rate matrix functions as similarity and can reduce computing complexity effectively. Then we used an area of statistics on the number of people to get a rough statistics, and using merging and splitting to compensate weakness of multiple-human segmentation from handle occlusion.From experiments we can see, this paper builds a real-time, effective monitoring system that can monitor the right to enter characters within the scope of quantitative statistics. Using the Coarse-to-fine segmentation, instead of overlapping frame rate to compare the distance between the centers of similarity, reducing the computational complexity; in the population statistics of the time, according to the relevance of compensation due to character segmentation result of misjudgment. Experiments show that the system in improving the real-time, while also taking into accounts its accuracy.更多还原