【作者】 刘欢；

【导师】 尹云飞；

【作者基本信息】 重庆大学 ， 计算机科学与技术， 2019， 硕士

【摘要】 传统接触式的火灾检测方式在一些狭小的封闭场景中能够取得很好的检测效果,但是在开放式的环境中难以达到有效的检测效果,而且这种检测方案对于安装的环境也有很苛刻的要求,否则难以达到检测火灾的目的。而基于视频分析的非接触式火灾检测方案,适用范围广,部署简便,能够有效地弥补接触式的检测方案的不足之处。但目前基于视频的火灾检测方案的研究还存在很多需要完善的地方,如对复杂场景的识别精度较低、受光照环境因素影响较大等。本文以中国船舶重工集团公司第七〇一研究所与重庆大学合作的舱室火灾检测项目为基点,通过分析火灾产生过程中的烟雾特性提出了一套完整的早期火灾检测方法,来解决目前检测方法存在的一些不足,为后续基于视频的火灾检测方案的发展提供一定的技术方向。首先,本文根据视频中运动的物体提出了一种基于多连通分量的像素自适应前景分离方法来提取出图像中前景区域。该方法首先对视频帧进行了去噪、增强等预处理工作,来提高图像中烟雾和背景对比度。然后利用Sobel算子进行边缘提取,根据边缘信息计算出最大的连通区域,并根据连通区域快速构建出初始背景样本。最后利用背景样本来实现对视频帧中前景区域的提取。其次,将疑似烟雾前景区域的边缘轮廓表示为Freeman链码,并根据链码从原图像中得到目标区域,再将目标区域通过双线性差值法将图像放大到256*256大小。然后将图像分割成大小相同的像素块,并通过纯烟样本和无烟样本训练好的贝叶斯分类器来对像素块进行烟雾识别,来判断每个像素块中是否为烟雾像素块。最后根据每个像素块中含有的烟雾情况进行判断原图像中烟雾的存在情况。最后,当视频帧中存在火灾烟雾时,则提取当前视频帧后的M帧图像对火灾烟雾的扩散性进行分析,来检测当前视频帧下的场景中是否存在火灾。同时为了验证算法的有效性,计算了不同算法在不同场景下的识别效果,并通过检测率、虚警率和漏警率等指标进行了算法的有效性对比。结果表明本文所提出的的检测方式能够有效的检测出各种场景下的火灾情况,并能达到一定的准确程度。更多还原

【Abstract】 The traditional contact fire detection method can achieve good detection results in some narrow closed scenes,but it is difficult to achieve effective effects in an open environment.For the non-contact fire detection scheme based on video analysis,it has a wide application scenario,which can effectively overcome the shortcomings of the existing contact detection.At present,video-based fire detection research still has many shortcomings,such as the lack of a unified standard fire detection video database,low identification accuracy of complex scenes,and great influence by external factors such as lighting and environment.This paper is based on cabin fire detection project co-operated by 701 Research Institute of China shipbuilding Industry Corporation and Chongqing University.By analyzing the characteristics of smoke in the fire generation process,a complete set of early fire detection methods is proposed to solve some shortcomings of the current detection methods,and provide a certain technical direction for the subsequent development of video-based fire detection methods.Firstly,this paper proposes a pixel-adaptive smoke foreground separation method based on multi-connected components to analyze the moving objects in the video to extract the foreground regions in the image.This method firstly carries out the pre-processing work of denoising,greying and enhancing the video frame to enhance the smoke contrast in the video frame.Then,Sobel operator is used for edge detection.According to the edge information,N video frames of different connected regions are calculated to quickly establish N background models.Finally,the background model is used to extract the foreground of the suspected smoke region in video frames.Secondly,the edge contour of the suspected smoke foreground area is expressed as Freeman chain code,and the target area is extracted from the original image according to the chain code.Then the target area is enlarged to 256*256 size by bilinear difference method.Then the image is divided into pixel blocks of the same size,and the smoke is identified by the Bayesian classifier trained by the pure smoke sample and smokeless sample to determine whether each pixel block is a smoke pixel block.At last,the smoke in the original image was judged according to the smoke in each pixel block.Finally,if there is fire smoke in the current video frame,the M frame image after the current video frame is extracted to analyze the diffusing property of early fire smoke,so as to distinguish the fire-like smoke and fire smoke and comprehensively judge whether there is fire in the scene under the current video frame.At the same time in order to verify the validity of the algorithm,to calculate the effect of different algorithms in different scenarios,and then through the detection rate and false alarm rate and missing alarm rate compared the effectiveness of the algorithm.The results show that the detection method proposed in this paper can effectively detect the fire in various scenarios,and can achieve good accuracy.更多还原