基于分层贝叶斯模型的智能视频监控中的异常检测
发布时间：2019-03-16 16:14
【摘要】：随着计算机视觉处理技术、电子技术、通信技术利智能信息处理技术的快速发展,智能视频监控系统在国防建设、交通管制以及智能安保等众多领域中得到广泛的应用。而现有的大多数视频监控系统仍依赖于监控人员的现场操作,造成了人力资源的浪费,也影响了整个工作系统的效率。因此,对智能视频监控系统中的关键技术进行研究并提高视频监控的性能具有重要的理论意义和实用价值。目前,智能视频监控方面的研究和应用都面临着很多难题,国内外的许多学者投身于该领域的研究,并取得了大量的成果。本文在这些成果的基础上,主要针对智能视频监控系统中的运动目标特征提取与异常事件检测两个步骤进行了研究,主要的工作概括如下：1、简要介绍了智能视频监控系统的主要任务、相关技术及其应用；概述了贝叶斯方法、分层贝叶斯方法及其基本算法；归纳总结了常用聚类方法及其适用范围和优缺点；列出了聚类算法的几个性能指标。2、针对视频文件的非结构化、以像素的形式存储目标对象的颜色、亮度和位置等低层信息且数据量巨大、表现内容多样性的特点。本文将视频文件进行预处理,借助于比较成熟的文本处理技术来实现视频文件的分析。3、针对智能视频监控中的运动目标特征提取问题,采用改进的金字塔Lucas-Kanada(PLK)光流法来提取运动目标的特征。传统的Horn-Schunck光流法属于稠密光流算法,对于运动不明确的像素,其计算量相当大；而Lucas-Kanada就是一种稀疏光流法,解决了计算量大的问题,然而该方法有很多限制条件,使得该光流法具有很多局限性。PLK光流法的基本思想是构造图像序列的一个金字塔,较高的层是下层平滑后的下采样形式,原始图像层数等于零。该方法提高了满足运动假设的可能性,从而实现对快速运动目标的特征提取。针对PLK光流算法中使用的最小二乘方法稳健性差的缺点,使用加权最小二乘法对PLK光流法进行改进。实验结果表明：相对于传统的光流法,改进的PLK光流法具有较好的特征提取效果。4、针对智能视频监控中的异常检测问题,提出了加权分层贝叶斯模型。该模型的核心思想是对先验分布的选取采用分层先验,其基本思路：人们可能同时掌握总体结构和部分细节的先验信息,则分阶段(层)有步骤地建立模型,当所给定先验分布中超参数难以确定时,可以对超参数再给出一个先验,第二个先验称为超先验。由先验和超先验共同决定的一个新先验,就称为分层先验。该模型将分层贝叶斯分析的理论用于模型的先验分布假设,有助于消除先验分布对估计结果的过度影响,增强估计的稳健性,使模型具有较强的适用性。实验结果表明：相对于传统的贝叶斯,该模型具有较好的异常事件检测效果。
[Abstract]:With the rapid development of computer vision processing technology, electronic technology, communication technology and intelligent information processing technology, intelligent video surveillance system has been widely used in many fields such as national defense construction, traffic control and intelligent security. However, most of the existing video surveillance systems still rely on the on-the-spot operation of surveillance personnel, resulting in a waste of human resources and affecting the efficiency of the entire work system. Therefore, it is of great theoretical significance and practical value to study the key technologies of the intelligent video surveillance system and to improve the performance of the video surveillance system. At present, the research and application of intelligent video surveillance are faced with many difficult problems. Many scholars at home and abroad have devoted themselves to the research in this field, and have made a lot of achievements. On the basis of these achievements, this paper mainly focuses on the two steps of feature extraction and abnormal event detection in intelligent video surveillance system. The main work is summarized as follows: 1, This paper briefly introduces the main tasks, related technologies and applications of the intelligent video surveillance system. This paper summarizes the Bayesian method, hierarchical Bayesian method and its basic algorithm, summarizes the common clustering methods, their scope of application, advantages and disadvantages; This paper lists several performance indexes of clustering algorithm. 2, aiming at the unstructured video file, the low-level information such as color, brightness and position of the target object is stored in the form of pixels, and the amount of data is huge, which shows the diversity of content. In this paper, the video file is pre-processed, and the analysis of the video file is realized with the help of mature text processing technology. 3, aiming at the problem of feature extraction of moving object in intelligent video surveillance, The improved pyramid Lucas-Kanada (PLK) optical flow method is used to extract the feature of moving target. The traditional Horn-Schunck optical flow method belongs to the dense optical flow algorithm. For the pixels whose motion is not clear, the computation is very large. Lucas-Kanada is a sparse optical flow method, which solves the problem of large amount of computation. However, this method has many limitations, which makes the optical flow method have many limitations. The basic idea of the PLK optical flow method is to construct a pyramid of image sequences, and the basic idea of PLK optical flow method is to construct a pyramid of image sequences. The higher layer is the downsampling form after the lower level smoothing, and the number of original image layers is equal to zero. This method improves the possibility of satisfying the motion hypothesis and realizes the feature extraction of the fast moving object. In order to improve the robustness of the least square method used in the PLK optical flow algorithm, the weighted least square method is used to improve the optical flow method. The experimental results show that compared with the traditional optical flow method, the improved PLK optical flow method has better feature extraction effect. 4. Aiming at the anomaly detection problem in intelligent video surveillance, a weighted hierarchical Bayesian model is proposed. The core idea of the model is to use hierarchical priori to select the prior distribution. The basic idea is that people may master the prior information of the overall structure and some details at the same time, then build the model step by step in stages (layers). When the super-parameter in the given prior distribution is difficult to determine, we can give another priori for the super-parameter, and the second priori is called the super-priori. A new priori, which is determined by both priori and hyperpriori, is called layered priori. In this model, the hierarchical Bayesian analysis theory is applied to the prior distribution hypothesis of the model, which is helpful to eliminate the excessive influence of the prior distribution on the estimation results, enhance the robustness of the estimation, and make the model more applicable. The experimental results show that compared with the traditional Bayesian model, the model has better abnormal event detection effect.
【学位授予单位】：西南大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN948.6
本文编号：2441713
[Abstract]:With the rapid development of computer vision processing technology, electronic technology, communication technology and intelligent information processing technology, intelligent video surveillance system has been widely used in many fields such as national defense construction, traffic control and intelligent security. However, most of the existing video surveillance systems still rely on the on-the-spot operation of surveillance personnel, resulting in a waste of human resources and affecting the efficiency of the entire work system. Therefore, it is of great theoretical significance and practical value to study the key technologies of the intelligent video surveillance system and to improve the performance of the video surveillance system. At present, the research and application of intelligent video surveillance are faced with many difficult problems. Many scholars at home and abroad have devoted themselves to the research in this field, and have made a lot of achievements. On the basis of these achievements, this paper mainly focuses on the two steps of feature extraction and abnormal event detection in intelligent video surveillance system. The main work is summarized as follows: 1, This paper briefly introduces the main tasks, related technologies and applications of the intelligent video surveillance system. This paper summarizes the Bayesian method, hierarchical Bayesian method and its basic algorithm, summarizes the common clustering methods, their scope of application, advantages and disadvantages; This paper lists several performance indexes of clustering algorithm. 2, aiming at the unstructured video file, the low-level information such as color, brightness and position of the target object is stored in the form of pixels, and the amount of data is huge, which shows the diversity of content. In this paper, the video file is pre-processed, and the analysis of the video file is realized with the help of mature text processing technology. 3, aiming at the problem of feature extraction of moving object in intelligent video surveillance, The improved pyramid Lucas-Kanada (PLK) optical flow method is used to extract the feature of moving target. The traditional Horn-Schunck optical flow method belongs to the dense optical flow algorithm. For the pixels whose motion is not clear, the computation is very large. Lucas-Kanada is a sparse optical flow method, which solves the problem of large amount of computation. However, this method has many limitations, which makes the optical flow method have many limitations. The basic idea of the PLK optical flow method is to construct a pyramid of image sequences, and the basic idea of PLK optical flow method is to construct a pyramid of image sequences. The higher layer is the downsampling form after the lower level smoothing, and the number of original image layers is equal to zero. This method improves the possibility of satisfying the motion hypothesis and realizes the feature extraction of the fast moving object. In order to improve the robustness of the least square method used in the PLK optical flow algorithm, the weighted least square method is used to improve the optical flow method. The experimental results show that compared with the traditional optical flow method, the improved PLK optical flow method has better feature extraction effect. 4. Aiming at the anomaly detection problem in intelligent video surveillance, a weighted hierarchical Bayesian model is proposed. The core idea of the model is to use hierarchical priori to select the prior distribution. The basic idea is that people may master the prior information of the overall structure and some details at the same time, then build the model step by step in stages (layers). When the super-parameter in the given prior distribution is difficult to determine, we can give another priori for the super-parameter, and the second priori is called the super-priori. A new priori, which is determined by both priori and hyperpriori, is called layered priori. In this model, the hierarchical Bayesian analysis theory is applied to the prior distribution hypothesis of the model, which is helpful to eliminate the excessive influence of the prior distribution on the estimation results, enhance the robustness of the estimation, and make the model more applicable. The experimental results show that compared with the traditional Bayesian model, the model has better abnormal event detection effect.
【学位授予单位】：西南大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN948.6
【参考文献】
相关期刊论文 前2条
1 张立成;蒋楠;杭亦文;张阳;骆蔓;;基于视频控制的智能交通灯[J];电子测量技术;2011年06期
2 李双虎;张风海;;一个新的聚类有效性分析指标[J];计算机工程与设计;2007年08期
相关硕士学位论文 前3条
1 尹波;聚类分析及其在移动通信企业数据挖掘分析中的应用研究[D];湖南大学;2008年
2 周帅伟;基于MCMC抽样算法的贝叶斯金融面板数据模型研究[D];湖南大学;2011年
3 郭美君;湖南省种植业保险分类定价研究[D];湖南大学;2014年
，本文编号：2441713
本文链接：https://www.wllwen.com/kejilunwen/wltx/2441713.html
