【作者】 郑河荣；

【导师】 卢炎麟；

【作者基本信息】 浙江工业大学 ， 化工过程机械， 2009， 博士

【摘要】 近年来,工业园区由于安防措施不到位,导致安全事故频发,给社会造成了巨大的生命和财产损失。为了有效提高安防能力,各级政府和工业园区纷纷投巨资安装监测、监控系统,构建应急救援体系。但已有监控系统一般只是把现场画面传回控制中心,由管理员肉眼观测判断现场是否安全。这种方式无法使计算机自动分析现场状态并启动相应的应急预案,还存在重大的安全防范漏洞。因此,如何采用视频分割技术实现区域状态自动分析和判断,成为当前研究热点,也是构建工业园区应急救援体系的关键技术之一。论文主要研究如何从各种工业环境监控视频中自动提取出运动目标,这正是视频分割技术的难点。论文主要贡献包括:1)针对非特定目标、视频背景相对稳定的分割问题,提出背景阈值自适应的分割方法。在已有背景建模和条件随机场理论基础上,提出采用Isillg模型自适应地计算视频邻域关系的特征强度函数,避免已有算法需要预先定义经验值而无法适用不同视频。实验结果表明,背景阈值自适应算法能够很好地逼近采用最优经验参数得到的分割效果。2)针对特定目标、视频背景不固定的分割问题,采用基于识别的分割方法,提出把随机蕨丛理论引入到视频分割中,通过该理论所特有的蕨丛联合共享规则构造高精度分类器,建立双层识别分割模型。该算法对于背景变化明显的视频,也能得到稳定的分割效果。视频分割实验表明,和已有的识别分割算法相比较,随机蕨丛理论能够显著提高分割准确率。3)针对非特定目标、视频背景也不固定的分割问题,提出模拟人类视觉感知机制建立数学模型,利用人类视觉对运动目标的敏感性提高分割质量。算法模拟人类视觉感知机理,对视频时域和空域的显著信号进行提取,并采用时域和空域显著信号的动态混合模型来确定视频目标区域。在此基础上,通过设计特征函数,构造分层随机场模型对特征函数进行约束求解,获得最终的分割结果。实验结果表明,建立合理的视觉激励信号,能够有效地解决非特定目标在不固定背景中的视频分割问题。4)针对工业园区的不同应用场合,开发工业园区视频监控原型系统。原型系统包括两层次的应用,第一层次是视频分割的简单应用,主要是检测当前监控区域中是否有运动目标出现。第二层次应用是在视频分割基础上,对分离出来的运动目标进行跟踪或识别,实现运动目标计数或异常目标报警。更多还原

【Abstract】 Recently, due to lack of powerful defending system, plenty of accidents occurred in Industry Park, which caused huge damage to life and property. To improve the capability of handling accidents, governments and Industry Park administration have built emergency response systems by fixing monitor and video surveillance. Existing video surveillance mainly transfer to control center the image data which are observed and judged by the system administrator. However, this kind of video surveillance still has serious problem because it can not analyze status and start emergency strategy automatically. In consequence, how to develop intelligent video surveillance has become a hot research topic. It is also the key technology in modern emergency response system.This paper focuses on video segmentation namely how to extract moving objects reliably. It is the most important part for intelligent video surveillance. The main contributions of this paper include:1) For the video with unspecific objects and minor changes in background, this paper proposed an improved background modeling method, which can compute threshold adaptively in the extraction of any moving objects. Background modeling method is relative simple but has serious misclassification problem. Some methods have been presented to overcome this problem by inocrpoating neighboring relationships of video. However, existing methods have to define empirical values manually, which is not robust for different kinds of videos. This paper uses Ising model to compute energies of neighboring relationships adaptively instead of empirical values. As a result, it makes the improved method is suitable for different kinds of videos. Experiments show the improved method can get expected segmentation.2) For the video with specific objects and obvious changes in background, this paper proposed a novel segmentation method based on recognition for the extraction of specific moving objects. The core idea of the proposed method is to use random fern theory to construct robust classifier according to the features of specified objects. Due to sharing voting rules of ferns, random fern theory can combine with different feature functions to strengthen segmentation result. Experiments show random fern theory can greatly improve segmentation precision greatly over the similar algorithms.3) For the video with unspecific objects and obvious changes in background, this paper proposed a segmentation model by simulating human vision perception. It is suitable for extracting any moving objects. The idea of the proposed algorithm is that human vision can extract any moving objects even with more complex backgrounds. The proposed method first extracts salient signals both in temporal and spatial domain for the video. Then dynamic model combine two kinds of signals to get the region of moving object. Finally, hierarchical conditional random field is used to obtain the final segmentation result. Experiments show it can extract any moving objects in complex backgrounds with the reasonable vision excite signal.4) This paper develops a prototype for the different applying environments in Industry Park. It mainly consists of two levels. The fist application is to detect whether or not moving object occurs in surveillance region. This problem can be resolved directly by video segmentation. The more deep applications are the counting of moving objects and alteration of abnormal objects. Other technologies, such as tracking and recognition, are required for these applications. Finally, some experiment results can prove these applications.更多还原