【作者】 周金旺；

【导师】 章权兵；

【作者基本信息】 安徽大学 ， 计算机应用技术， 2012， 硕士

【摘要】 随着社会经济的发展,人们对安全防范的需求也越来越大。例如,机场、地铁、候车厅等人多密集的重要场所,容易被恐怖分子等利用,通常具有很大的安全隐患,因此需要进行实时监测。而传统的监控系统基本只能用于事后证据呈现,却不能够实时检测异常事件发生。所以人们期待一种智能化的技术用于安全防范,为此出现了智能视频监控系统。本文主要研究监控场景中包裹遗留事件的检测。关于遗留事件的判别,目前主要有基于目标检测的方法和基于目标跟踪的方法,但现有算法多数并不鲁棒。在已有的基于目标检测的方法中,由于初始背景在慢背景更新模型中具有较大的权值,若初始背景存在运动目标,该目标区域不能及时更新背景,则该区域易被判别为遗留目标区域；当监控场景中有人体停留时间过长,并且保持不动,也会误将人体判别为遗留物体。在背景更新阶段,前景物体有可能很快融入到背景中,如果更新速率选择不好,有可能会出现“鬼影”现象等。基于目标跟踪的方法主要是对运动目标的位置进行预测估计,这种方法容易受光线等因素的影响。人体等非刚性物体在跟踪时很难用模板的形式完成检测和跟踪,从而导致预测不准确,以致丢失检测目标。本文对相关算法进行了综述,分析了遗留物检测原理和重难点,在已有成果的基础上做出了一些改进和创新。本文的工作内容如下：第一,针对现有的遗留物检测算法,分析了遗留物检测原理、遗留物检测的重点和难点以及实现步骤。第二,本文主要采用基于目标检测的方法,首先通过双背景模型检测暂时静止目标,双背景模型分别采用慢背景更新的滑动平均背景模型和快背景更新的混合高斯背景模型,然后利用Blob跟踪的方法对每一个暂时静止目标进行跟踪,分析目标边缘和中心—外围直方图信息,判别目标是遗留目标还是遗失目标。第三,算法中加入了基于HOG的人体检测,对目标感兴趣区域进行人体检测,从而消除因行人驻留产生的误判。最后,编程实现了遗留物检测系统,并对下载的标准视频库和采集的场景模拟视频进行测试,都取得了比较好的效果。更多还原

【Abstract】 With the development of society and economy, people’s security needs is increasing. Terrorists usually commit attacks in followed places, Such as airport, subway, and waiting hall and so on. So a real time monitored system is needed. But traditional monitoring system can only be used for the basic evidence and not capable of real-time detecting abnormal events, so people look forward to a kind of intelligent technology for security. So the intelligent video surveillance system is emerging.The main study in this paper is detecting abandoned object in surveillance video. So far the abandoned event detection is based on target detection and target tracking, but they are not robust sometimes. For example, the method which based on target detection can do false alarm, due to the initial background scene has the bigger weight in slow update rate background model and if has some move target, the area of the move target can’t update in time, so the area easily be judged as abandoned object area. And when somebody enter in scene and keep still during a period time, the algorithm will judge the person as abandoned object. In background update stage, foreground object sometimes will quickly melt into the background; if the background learning rate is not good,"ghost" may be happening. Abandoned object detection which based on tracking method, mainly focus on the moving target position prediction and identification, the method easily affected by light and other factors, human body as non-rigid objects in tracking is difficult to use a template to detect and track, often leads to inaccurate prediction, so that target will lose sometimes.In this paper, first of all, I do amount of research and reviewed on the basis of related algorithms, and then analysis the principle and difficulty, made some improvements and creation based on the existing achievement. The works and creation of this paper are as follows:Firstly, review the existing detection algorithm, do analysis on detecting principle, key points, difficulty and the step of implementation. Secondly, the main idea of this paper is based on detection method. At the beginning detect temporarily stationary target using the dual-background model, they are slow update moving average background model and fast update Mixture Gauss background model. And then, using Blob tracks every target which detected at beginning. Do analysis on target edge and center-peripheral region histogram, discriminate the target whether abandoned or not.Thirdly, human detection based on HOG (Histogram of Oriented Gradient) joint in proposed algorithm can detect human body and eliminate misjudgment caused by human keep still in scene.Finally, implement the detection system. Through testing the video which are from the online standard video lib and captured by ourselves, the system and the algorithm have achieved good results.更多还原