基于显著性检测和烟雾时空特征的视频火灾探测方法研究

发布时间：2019-07-01 17:02

【摘要】：火灾是严重威胁大类生产和生活安全的因素之一,通过火灾探测报警系统预防火灾对于保障人民的生命财产安全、维持社会稳定具有重要意义。目前广泛应用的常规感烟、感温火灾探测器多利用接触式感知技术,在应用场所(空间、时间等)具有一定的局限性,不适用于层高较高的建筑及开放空间中的火灾探测,具有非接触、响应迅速、探测范围大、主动可视等优点的视频火灾探测技术应运而生,被应用于地面建筑、厂区、森林等场所的火灾监测。目前的视频火灾探测技术主要集中在视频火焰探测、红外探测等方面,而在及时性方面有更大优势的视频烟雾探测技术在理论、方法和应用上还存在很多问题亟待解决,如缺乏标准测试视频库、未对疑似烟雾区域提取方法进行深入研究、在进行相似运动目标分类方面缺乏有效特征、烟雾探测误报率高等。本文旨在深入研究早期火灾烟雾视频探测理论和方法,为视频火灾探测技术的发展及应用提供理论和技术支撑。本文在中国科学技术大学火灾科学国家重点实验室的标准试验间和低压实验舱中设计搭建视频拍摄系统,获取阴燃烟雾视频图像序列,进行阴燃烟雾的视频分割、特征提取和分析、特征识别等方面的研究,进而设计完整的视频火灾探测系统,并对系统在低气压环境中应用的可行性进行实验验证。具体研究工作如下：(1)建立阴燃烟雾视频基础数据库。火灾烟雾标准视频图像数据库是研究视频火灾烟雾探测原理、发展烟雾识别方法的原始数据基础。在中科大火灾实验室的标准试验间和低压实验舱内搭建火灾阴燃烟雾视频采集平台,采用木材和棉绳作为实验材料产生阴燃烟雾,使用高清摄像机、红外摄像机、高速摄像机等设备进行烟雾视频采集,对采集的视频进行标准化处理,最终建立阴燃烟雾视频基础数据库。(2)提出并完成基于显著性检测的早期烟雾疑似区域分割方法。根据人眼视觉注意机制,阴燃烟雾可被看作是视频中湍流和灰色显著的区域,基于结合自顶向下和自底向上的视觉注意模型利用显著性检测方法进行疑似烟雾区域分割。首先使用非线性增强方法对视频的亮度图像和光流图谱进行增强,用增强后的图像计算显著性谱,然后用高斯混合模型(GMM)计算出的运动前景构造运动能量函数,对显著性谱进行估计,分割出疑似烟雾区域。对几种常用的传统烟雾区域检测方法的对比实验结果表明该方法具有更好的分割效果,并且计算速度也能够满足实时视频烟雾探测的需求。(3)发展基于烟雾时空特征和检测前跟踪方法的视频烟雾探测算法。为提高探测算法的准确率和鲁棒性,对每个候选烟雾区域进行跟踪,在一个时间窗口内综合空域和时域特征信息进行火灾识别。通过对烟雾形成的物理过程和烟雾组分的分析,提出能够描述湍流烟雾纹理的熵和对比度特征,并通过实验进行特征有效性验证。结合阴燃烟雾图像的纹理时空特征、颜色特征和速度特征,训练支持向量机分类模型,通过一个时间窗的多次分类结果的累积估计发生火灾的可能性。使用包含烟雾和非烟雾的视频对算法性能进行测试验证。实验结果表明该算法框架合理有效,而且通过修改框架中的组件,还可以快速完成火焰探测等其他功能。最后还对烟雾浓度对视频烟雾探测的影响进行实验分析,并指出烟雾浓度是火灾探测研究和工业应用中需要考虑的重要因素。(4)设计阴燃烟雾视频探测系统,并对系统在低气压环境下的适用性进行测试。根据实验结果和前期研究的视频烟雾探测方法发展出视频烟雾探测系统。用基于显著性检测的阴燃烟雾分割方法和检测前跟踪算法对不同气压下的烟雾视频进行分析,对不同气压下的阴燃烟雾区域进行分割,提取烟雾区域特征,然后对视频中烟雾区域的亮度、速度均值及方差、纹理的熵及对比度等特征进行统计分析,并使用这些特征在不同气压下使用多种分类器进行分类测试,实验研究结果表明本文提出的方法和特征在低压环境下仍然适用。
[Abstract]:Fire is one of the factors that seriously threaten the production and life safety of the major categories. The fire detection and alarm system is of great significance to protect the life and property of the people and to maintain social stability. The conventional smoke-sensing and temperature-sensing fire detector which is widely used at present has a certain limitation in the application places (space, time, etc.), and is not suitable for fire detection in high-story buildings and open spaces, has non-contact and rapid response, Video fire detection technology, which has the advantages of large detection range, active visual and other advantages, comes into being, and is applied to the fire monitoring of the ground building, the plant area, the forest and other places. The current video fire detection technology mainly focuses on the aspects of video flame detection, infrared detection and the like, and the video smoke detection technology with greater advantages in the timeliness has many problems to be solved in theory, method and application, such as the lack of standard test video library, The method for extracting the suspected smoke area is not in-depth study, and there is a lack of effective characteristics in the classification of similar moving targets, and the smoke detection is false and high. The purpose of this paper is to study the theory and method of early fire smoke detection and provide theoretical and technical support for the development and application of video fire detection technology. In this paper, a video shooting system is designed in the standard test room and the low-pressure laboratory of the fire science national key laboratory in China's science and technology university, and the video image sequence of the smoldering smoke is obtained, the video segmentation, feature extraction and analysis of the smoldering smoke are carried out, the feature recognition and the like are carried out, And then a complete video fire detection system is designed, and the feasibility of the application of the system in the low-pressure environment is experimentally verified. The specific research work is as follows: (1) Create a negative-burning smoke video base database. The fire smoke standard video image database is the raw data base of the research on the principle of video fire smoke detection and the development of the smoke identification method. in that standard test room and the low-pressure experiment cabin of the middle and large fire laboratory, a fire-smoldering smoke video acquisition platform is built, wood and cotton ropes are used as experimental materials to generate smoldering smoke, and the smoke video acquisition is carried out by using a high-definition camera, an infrared camera, a high-speed camera and the like, The collected video is standardized, and a negative-burning smoke video base database is finally established. (2) An early smoke suspected area segmentation method based on saliency detection was proposed and completed. In light of that visual attention mechanism of the human eye, the smoldering smoke can be regarded as a region of turbulence and gray in the video, and the suspected smoke region is divide by a saliency detection method based on a visual attention model combined with the top-down and the bottom-up. firstly, the brightness image and the optical flow pattern of the video are enhanced by using a non-linear enhancement method, the saliency spectrum is calculated by the enhanced image, and then the energy function of the motion foreground tectonic movement calculated by the Gaussian mixture model (GMM) is used to estimate the significance spectrum, The suspected smoke area is divided. The experimental results show that the method has better segmentation effect and can meet the demand of real-time video smoke detection. And (3) developing a video smoke detection algorithm based on the smoke-time characteristic of the smoke and the tracking method before the detection. In order to improve the accuracy and the robustness of the detection algorithm, each candidate smoke area is tracked, and the air space and time domain characteristic information are integrated in a time window for fire identification. The entropy and contrast characteristics of the turbulence smoke texture can be described by the analysis of the physical process and the smoke component of the smoke formation, and the characteristic validity verification is carried out by the experiment. The probability of fire is estimated by the cumulative estimation of multiple classification results of a time window in combination with the texture space-time characteristics, the color characteristics and the speed characteristics of the smoldering smoke image, the training support vector machine classification model. The algorithm performance is tested and verified using video containing smoke and non-smoke. The experimental results show that the framework is reasonable and effective, and other functions such as flame detection can be accomplished quickly by modifying the components in the frame. Finally, the influence of the smoke concentration on the detection of the video smoke is analyzed, and the smoke concentration is an important factor that needs to be considered in the research of fire detection and industrial application. And (4) designing a smoldering smoke video detection system and testing the applicability of the system under the low pressure environment. And the video smoke detection system is developed according to the experimental results and the video smoke detection method of the earlier research. the smoke video under different air pressure is analyzed by using a negative-burning smoke segmentation method based on the significance detection and a detection front tracking algorithm, the smoke area of the smoldering smoke under different air pressure is divided, the smoke area characteristic is extracted, and then the brightness, the speed mean value and the variance of the smoke area in the video are extracted, The characteristics of the entropy and contrast of the texture are analyzed and tested by using a variety of classifiers under different air pressure. The experimental results show that the proposed method and features are still applicable in the low-pressure environment.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：X924.3;X932

【相似文献】