【摘要】：社會(huì)在不斷的發(fā)展，經(jīng)濟(jì)與科技的進(jìn)步也越來(lái)越大，人們對(duì)工作和生活中的安全問(wèn)題的重視程度越來(lái)越高，視頻監(jiān)控設(shè)備也隨之大大普及，現(xiàn)在，無(wú)論你在任何公共的場(chǎng)合，基本上都在視頻監(jiān)控范圍內(nèi)。視頻監(jiān)控的普及帶來(lái)的數(shù)據(jù)量是巨大的，像過(guò)去那樣找專(zhuān)門(mén)的安保工作人員來(lái)24小時(shí)監(jiān)視監(jiān)控視頻數(shù)據(jù)變得不可能，這使得現(xiàn)在的監(jiān)控視頻更多只是作為事故發(fā)生后用于偵察破案的依據(jù)，而不能作為預(yù)防犯罪事故發(fā)生的主要手段。因此，智能視頻監(jiān)控技術(shù)得到了極大的重視，讓計(jì)算機(jī)代替人去監(jiān)視監(jiān)控設(shè)備的需求越來(lái)越大。在智能視頻監(jiān)控技術(shù)中，對(duì)運(yùn)動(dòng)目標(biāo)的分類(lèi)技術(shù)先得尤為重要。智能視頻監(jiān)控系統(tǒng)的最終目標(biāo)就是理解視頻中特定目標(biāo)的行為，從而發(fā)現(xiàn)潛在的不安全因素，向外界發(fā)出警報(bào)。然而，在監(jiān)控視頻中存在著大量的運(yùn)動(dòng)目標(biāo)，如行人、車(chē)輛、樹(shù)葉、噪聲等，要理解我們感興趣的運(yùn)動(dòng)目標(biāo)的行為，首先要做的就是確定這些運(yùn)動(dòng)的目標(biāo)的類(lèi)型。本文主要針對(duì)拍攝距離較遠(yuǎn)運(yùn)動(dòng)目標(biāo)形狀相對(duì)較小的目標(biāo)分類(lèi)問(wèn)題展開(kāi)研究。首先，介紹了智能視頻監(jiān)控中運(yùn)動(dòng)目標(biāo)分類(lèi)相關(guān)的一些技術(shù)，綜合使用這些技術(shù)，提出了一種針對(duì)運(yùn)動(dòng)目標(biāo)相對(duì)較小、清晰度不高的運(yùn)動(dòng)目標(biāo)分類(lèi)方法，并使用這種方法，對(duì)真實(shí)的場(chǎng)景進(jìn)行應(yīng)用，把運(yùn)動(dòng)目標(biāo)分為人、人群、車(chē)輛和其他四個(gè)類(lèi)別。本文的工作可歸納如下： 1.針對(duì)本文所提出的分類(lèi)算法所使用相關(guān)算法以及智能視頻監(jiān)控技術(shù)所涉及的常用算法進(jìn)行介紹，，重點(diǎn)介紹了運(yùn)動(dòng)目標(biāo)分類(lèi)技術(shù)，以基于運(yùn)動(dòng)特征和基于靜態(tài)特征兩個(gè)方面對(duì)運(yùn)動(dòng)目標(biāo)分類(lèi)技術(shù)進(jìn)行闡述，也重點(diǎn)介紹了支持向量機(jī)分類(lèi)器。 2.針對(duì)中遠(yuǎn)距離拍攝情況下，對(duì)外形較小、清晰度不高的運(yùn)動(dòng)目標(biāo)提出了一種基于輪廓HOG的特征包分類(lèi)方法，并描述了如何在輪廓圖像上對(duì)HOG特征進(jìn)行快速運(yùn)算，使用特征包技術(shù)進(jìn)行降維形成最最終的特征向量。 3.對(duì)本文提出的分類(lèi)方法進(jìn)行實(shí)驗(yàn)分析，得到最優(yōu)的參數(shù)組合，其他文獻(xiàn)中的分類(lèi)算法進(jìn)行對(duì)比，證明本分類(lèi)方法的可行性。
[Abstract]:With the development of society, the progress of economy and science and technology, people pay more and more attention to the security problems in their work and life, and video surveillance equipment becomes more and more popular. Now, no matter you are in any public place, Basically in the video surveillance range. The amount of data generated by the popularity of video surveillance is so large that it is impossible to find specialized security personnel to monitor video data 24 hours a day, as in the past. This makes the current surveillance video more as the basis of detection and detection after the accident, but not as the main means to prevent crime accidents. Therefore, intelligent video surveillance technology has received great attention. In the intelligent video surveillance technology, the classification of moving targets is very important. The ultimate goal of an intelligent video surveillance system is to understand the behavior of a specific target in a video, thereby detecting potential unsafe factors and alerting the outside world. However, there are a large number of moving targets in surveillance video, such as pedestrians, vehicles, leaves, noise and so on. In order to understand the behavior of moving targets of interest to us, the first thing to do is to determine the types of moving targets. This paper focuses on the classification of objects with relatively small shape of moving objects. Firstly, this paper introduces some technologies related to moving object classification in intelligent video surveillance. By using these technologies, a new method of moving target classification is proposed, which is aimed at the relatively small moving target and low definition, and uses this method. The real scene is used to classify the moving targets into people, crowds, vehicles and other four categories. The work of this paper can be summarized as follows: 1. The related algorithms used in the classification algorithm and the common algorithms involved in the intelligent video surveillance technology are introduced in this paper, and the moving target classification technology is emphasized. Based on motion feature and static feature, the technology of moving target classification is discussed, and the support vector machine classifier is also introduced. 2. In this paper, a feature packet classification method based on contour HOG is proposed for moving objects with small shape and low definition in the case of medium and long range shooting, and how to quickly calculate HOG features on contour images is described. The feature packet technique is used to reduce the dimension to form the final feature vector. 3. Through the experimental analysis of the classification method proposed in this paper, the optimal parameter combination is obtained, and the classification algorithms in other literatures are compared to prove the feasibility of this classification method.
【學(xué)位授予單位】：華南理工大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2014
【分類(lèi)號(hào)】：TN948.6

【參考文獻(xiàn)】

相關(guān)期刊論文 前5條

1 於建生;汪宏;顧晶;;一種基于視頻的新型行人識(shí)別方法[J];信息化研究;2010年03期

2 毛建昌,王成道,萬(wàn)嘉若;多分辨率自回歸紋理模型[J];電子學(xué)報(bào);1988年06期

3 周開(kāi)軍;陳三寶;徐江陵;;復(fù)雜背景下的車(chē)牌定位和字符分割研究[J];計(jì)算機(jī)工程;2007年04期

4 王亮,胡衛(wèi)明,譚鐵牛;人運(yùn)動(dòng)的視覺(jué)分析綜述[J];計(jì)算機(jī)學(xué)報(bào);2002年03期

5 溫靜;高新波;;一種基于膚色模型的貝葉斯人臉檢測(cè)算法[J];西安電子科技大學(xué)學(xué)報(bào);2006年05期

本文編號(hào)：2417510