【作者】 钟取发；

【导师】 周平；

【作者基本信息】 浙江理工大学 ， 计算机应用技术， 2010， 硕士

【摘要】 视频火灾探测是计算机视觉中一项理论意义与实际价值兼备的重要课题,对烟火事故的消防安全具有重要的实际意义。但由于火灾衍生物的多变性和火灾场景的复杂性,使得火灾的视频探测研究成为一个极具挑战性的工作,目前尚未形成具有普适性的理论和算法。本文主要是进行视频火灾探测方法的研究,旨在提高火灾预警系统的灵敏度,降低误报率,从而更好地对火灾的发生进行早期预警。全文研究内容主要分成四个部分:背景重建与目标的提取分割,静态特征的提取,动态特征的提取和基于BP神经网络的火灾分级检测。在对传统的运动目标检测算法进行深入研究的基础上,分析了高斯背景模型等几类算法的基本原理,然后利用适合本文场景的二级背景模型和背景差分方法,结合数学形态学提取出初始目标,从而去除一些静态因素的干扰。在此基础上,通过对大量火灾的烟雾、火焰图像的调查研究,找出烟雾、火焰在特定颜色空间中的分布,建立了相应的颜色模型,分割出类似火焰、烟雾的区域。本文在对比烟雾、火焰及干扰物的动态特征的基础上,分析了所得分割区域的运动累积量、闪烁规律、运动方向、运动一致性、运动程度等火灾判据,并给出了各种判据的分析和计算方法。讨论了基于BP神经网络的动态特征判据融合方案,首先简单介绍了人工神经网络的内容,随后给出了本文神经网络的特征定义、输入输出单元及设计方案,并利用设计的神经网络实现了动态环境下早期火灾的分级检测。本文采用25个不同条件下的视频作为样本进行训练,并利用训练好的神经网络测试另外的35个视频,其中,有两个视频出现漏判,一个视频出现误判。结果表明,本文方法能在200帧内有效地识别出早期烟雾、火焰,并可以抵抗常见干扰对系统的影响,较好地实现了识别系统的鲁棒性与敏锐性的统一。更多还原

【Abstract】 Video fire detection is one of the most active research topics being valuable for both theoretical and practical research in computer vision especially has a wide spectrum of promising applications in video surveillance for early fire alarms in public security. However, because of the polytropy of the fire derivatives and complexity of scene, video fire detection becomes a difficult problem with large challenges, yet there are no general theories or algorithms have formed so far. In this paper, it is mainly research the methodology of video fire detection, in order to improve the sensitivity of fire alert system and reduce false alarm, so as to promote the performance of video fire detection system.The research contents of this paper are mainly composed of 4 parts: background-rebuild and moving object extraction, static features extraction, dynamic features extraction, fire grading-detection base on BP neural network.The principle of Gaussian and some other background models are analyzed on the basis of further study on traditional moving detection algorithms and, then the two step background model which is suitable for this paper is selected to extract the initial object, combine with background subtraction and mathematical morphology, and thus remove the static interferences. After extracting the moving region the distribution in specific color space of flame and smoke are found by investigation on flame and smoke images, and the corresponding color models are build to segment the flame and smoke like regions.The motion accumulation, flicker frequency, motion orientation, motion consistency, motion degree, etc, those dynamic criterions of the segment regions are analyzed based on comparisons between flame, smoke and other disturbing objects, more over, the analysis and computation methods of the criterions are brought forward. The fire criterions fusion scheme based on BP neural network is discussed, firstly, the basic contents of artificial neural network is introduced, and then the definition of characters, the input/output unit and design scheme of BP neural network are presented, at last, the designed neural network is used for grading-detection of early smoke and flame images.25 video clips under different condition are used for training neural network and, the neural network is used for recognizing other 35 video clips, among these, only 2 video clips are missed and 1 video clips is error recognized. The results show that the system can recognize the flame and smoke in 200 frames and has good anti-interference ability.更多还原