基于分層貝葉斯模型的智能視頻監(jiān)控中的異常檢測
發(fā)布時(shí)間：2019-03-16 16:14
【摘要】：隨著計(jì)算機(jī)視覺處理技術(shù)、電子技術(shù)、通信技術(shù)利智能信息處理技術(shù)的快速發(fā)展,智能視頻監(jiān)控系統(tǒng)在國防建設(shè)、交通管制以及智能安保等眾多領(lǐng)域中得到廣泛的應(yīng)用。而現(xiàn)有的大多數(shù)視頻監(jiān)控系統(tǒng)仍依賴于監(jiān)控人員的現(xiàn)場操作,造成了人力資源的浪費(fèi),也影響了整個(gè)工作系統(tǒng)的效率。因此,對智能視頻監(jiān)控系統(tǒng)中的關(guān)鍵技術(shù)進(jìn)行研究并提高視頻監(jiān)控的性能具有重要的理論意義和實(shí)用價(jià)值。目前,智能視頻監(jiān)控方面的研究和應(yīng)用都面臨著很多難題,國內(nèi)外的許多學(xué)者投身于該領(lǐng)域的研究,并取得了大量的成果。本文在這些成果的基礎(chǔ)上,主要針對智能視頻監(jiān)控系統(tǒng)中的運(yùn)動目標(biāo)特征提取與異常事件檢測兩個(gè)步驟進(jìn)行了研究,主要的工作概括如下：1、簡要介紹了智能視頻監(jiān)控系統(tǒng)的主要任務(wù)、相關(guān)技術(shù)及其應(yīng)用；概述了貝葉斯方法、分層貝葉斯方法及其基本算法；歸納總結(jié)了常用聚類方法及其適用范圍和優(yōu)缺點(diǎn)；列出了聚類算法的幾個(gè)性能指標(biāo)。2、針對視頻文件的非結(jié)構(gòu)化、以像素的形式存儲目標(biāo)對象的顏色、亮度和位置等低層信息且數(shù)據(jù)量巨大、表現(xiàn)內(nèi)容多樣性的特點(diǎn)。本文將視頻文件進(jìn)行預(yù)處理,借助于比較成熟的文本處理技術(shù)來實(shí)現(xiàn)視頻文件的分析。3、針對智能視頻監(jiān)控中的運(yùn)動目標(biāo)特征提取問題,采用改進(jìn)的金字塔Lucas-Kanada(PLK)光流法來提取運(yùn)動目標(biāo)的特征。傳統(tǒng)的Horn-Schunck光流法屬于稠密光流算法,對于運(yùn)動不明確的像素,其計(jì)算量相當(dāng)大；而Lucas-Kanada就是一種稀疏光流法,解決了計(jì)算量大的問題,然而該方法有很多限制條件,使得該光流法具有很多局限性。PLK光流法的基本思想是構(gòu)造圖像序列的一個(gè)金字塔,較高的層是下層平滑后的下采樣形式,原始圖像層數(shù)等于零。該方法提高了滿足運(yùn)動假設(shè)的可能性,從而實(shí)現(xiàn)對快速運(yùn)動目標(biāo)的特征提取。針對PLK光流算法中使用的最小二乘方法穩(wěn)健性差的缺點(diǎn),使用加權(quán)最小二乘法對PLK光流法進(jìn)行改進(jìn)。實(shí)驗(yàn)結(jié)果表明：相對于傳統(tǒng)的光流法,改進(jìn)的PLK光流法具有較好的特征提取效果。4、針對智能視頻監(jiān)控中的異常檢測問題,提出了加權(quán)分層貝葉斯模型。該模型的核心思想是對先驗(yàn)分布的選取采用分層先驗(yàn),其基本思路：人們可能同時(shí)掌握總體結(jié)構(gòu)和部分細(xì)節(jié)的先驗(yàn)信息,則分階段(層)有步驟地建立模型,當(dāng)所給定先驗(yàn)分布中超參數(shù)難以確定時(shí),可以對超參數(shù)再給出一個(gè)先驗(yàn),第二個(gè)先驗(yàn)稱為超先驗(yàn)。由先驗(yàn)和超先驗(yàn)共同決定的一個(gè)新先驗(yàn),就稱為分層先驗(yàn)。該模型將分層貝葉斯分析的理論用于模型的先驗(yàn)分布假設(shè),有助于消除先驗(yàn)分布對估計(jì)結(jié)果的過度影響,增強(qiáng)估計(jì)的穩(wěn)健性,使模型具有較強(qiáng)的適用性。實(shí)驗(yàn)結(jié)果表明：相對于傳統(tǒng)的貝葉斯,該模型具有較好的異常事件檢測效果。
[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.
【學(xué)位授予單位】：西南大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2015
【分類號】：TN948.6
本文編號：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.
【學(xué)位授予單位】：西南大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2015
【分類號】：TN948.6
【參考文獻(xiàn)】
相關(guān)期刊論文 前2條
1 張立成;蔣楠;杭亦文;張陽;駱蔓;;基于視頻控制的智能交通燈[J];電子測量技術(shù);2011年06期
2 李雙虎;張風(fēng)海;;一個(gè)新的聚類有效性分析指標(biāo)[J];計(jì)算機(jī)工程與設(shè)計(jì);2007年08期
相關(guān)碩士學(xué)位論文 前3條
1 尹波;聚類分析及其在移動通信企業(yè)數(shù)據(jù)挖掘分析中的應(yīng)用研究[D];湖南大學(xué);2008年
2 周帥偉;基于MCMC抽樣算法的貝葉斯金融面板數(shù)據(jù)模型研究[D];湖南大學(xué);2011年
3 郭美君;湖南省種植業(yè)保險(xiǎn)分類定價(jià)研究[D];湖南大學(xué);2014年
，本文編號：2441713
本文鏈接：http://sikaile.net/kejilunwen/wltx/2441713.html
最近更新
教材專著
