火焰是火灾初期的物理现象，如果能够及时发现火焰、及时告警并及时采取措施将其扑灭，则可避免大面积火灾的发生，将生命财产损失降到最低。基于计算机视觉的火焰监测系统，解决了传统传感火焰检测装置不便在室外使用且探测精度低的缺点。但是现有的视觉监测系统还存在使用红外热成像设备成本较高，使用可见光设备检测算法不成熟且工程应用较少等缺点。因此，论文设计了一个基于监控图像的火焰监测系统，通过对系统结构和视觉算法的设计，实现了对可见光视频监控设备图像的采集、处理和分析，检测出火灾初期火焰，判定异常的发生并进行报警。
论文运用智能监控技术，计算机视觉技术、机器学习原理设计了火灾视觉监控系统。系统包括据数据管理服务器，图像数据处理终端，人机交互终端。数据管理服务器负责管理系统数据并协调各部分间数据通信，通信部分不是论文的主要工作，论文设计了MySql数据库实现对系统数据的管理，设计了数据库接口实现数据操作。图像数据处理终端使用面向对象和多线程技术开发，将每一个摄像头设定为一个实例，并为每个摄像头开启一路线程；使用智能监控技术，通过SDK二次开发网络摄像头后直接驱动设备获取监控图像；使用计算机视觉技术实现对数字图像的处理和分析，包括初始化、预处理和火焰判定三个阶段，初始化阶段通过ViBe模型构建初始背景模型；预处理阶段首先通过ViBe模型+Haar-like的方法过滤掉光线影响，提取到16 *16像素的运动目标块，然后使用M(品红),Cb(蓝色)等颜色分量过滤掉非火焰目标块，获得疑似目标块；火焰判定是对疑似目标块在时序上的图像序列时空块进行特征提取和随机决策森林分类后进行综合决策，提取的特征包括颜色运动特征，光流直方图(HOF)特征，M、Y、V静态颜色特征，LBP静态纹理特征，随机决策森林是由自行标记的火焰非火焰样本集根据特征训练而成。人机交互终端基于Wpf前端框架开发，通过对硬盘录像机的驱动实现实时监控预览和视频回放，根据通信协议与服务器进行数据交互，从而实现对设备和人员的信息管理、异常实时推送和历史异常追溯。经过实验室模拟测试，系统能够快速高效的完成视频监控图像的采集，分析判断，检测到火焰并进行异常处理和报警，同时实现了设备人员信息设置管理和历史异常追溯等人机交互。
论文设计的基于视频监控图像的火焰监测系统，能够在不同环境中实时、稳定、高效的完成对火焰的检测，判定异常的发生，系统具有稳定性好，实时性高，漏检率低，成本低的特点。目前系统处在实验室测试运行阶段，运行稳定，内存泄漏较低，在存在部分漏检的情况下，准确率达92%。经过理论研究、算法设计、软件设计和实验调试，达到了设计的目标，为火灾早期探测的进一步研究提供了一定的参考。
 

Flame is the physical phenomenon in the early stage of a fire. If it can be discovered and alerted in time, measures can be taken in time to extinguish the flame to avoid the occurrence of large-scale fires and minimize the loss of life and property. Computer vision-based flame monitoring system solves the shortcomings of traditional sensing fire detection that cannot be used in outdoor large spaces and has low detection accuracy. However, the existing visual detection systems also have the disadvantages of high cost of using infrared thermal imaging equipment, immature identification of visible light equipment, and poor engineering implementation. Therefore, the dissertation designs a fire monitoring system based on motion images, Therefore, the thesis designs a fire monitoring system based on motion picture. By improvement of the system structure and visual algorithm, it realizes the acquisition, processing and analysis of the moving image of the visible light camera monitoring equipment, detects the flame, determines the occurrence and the alarm of the anomaly.
The flame monitoring system designed in this thesis includes three parts: data management server, image data processing terminal and human-computer interaction terminal. Data management server manages system data and coordinates data communication between parts. Data management server deploys MySql database to manage system data, and designs database interface to implement data operation. The image data processing terminal is developed using object-oriented and multi-thread technology, setting each camera as an instance, and opening a thread for each camera; using intelligent monitoring technology, the terminal uses the SDK to redevelop the webcam and directly drives the device to obtain the moving image; The process includes initialization, pretreatment and flame determination. In initialization, system builds initial background through ViBe model. The preprocessing stage, first filters out the light effects through the ViBe model + Haar-like method, extracts the motion target block of 16 * 16 pixels, and then uses the color components such as M (Magenta), Cb(blue) to filter out the non-flame, and get suspected target block; The flame determination is based on the feature extraction of the temporal image and temporal block of the suspected target block and the random decision forest classification. The extracted features include color motion characteristics, Level insensitivity optical flow histogram (HOF) features, M, Y, and V Static color features, LBP static texture features, and random decision forests are trained on self-labeled flame and non-flame sample sets. The human-computer interaction terminal is developed based on the Wpf front-end framework, real-time monitoring preview and video playback through the drive of the NVR, and data exchange with the server according to the communication protocol, thereby realizing the information management of equipment and personnel, abnormal real-time pushing and historical anomaly tracing. After the laboratory simulation tests, the system can quickly and efficiently complete the collection, analyze and judgement of monitoring moving images, and once the fire phenomenon is detected, it immediately performs exception handling and alarming, at the same time this system realizes the information management of equipment personnel and the historical anomaly tracing and other human-computer interaction.
The flame monitoring system based on motion picture designed in this thesis can detect fire flame in real time, steadily and efficiently in different environment, and determine the occurrence of fire anomaly. The system has the characteristics of good stability, high real-time, low leakage rate and low cost. At present, the system is in the laboratory test running stage, with stable operation and low memory leakage. In the presence of partial missed detection, the accuracy rate is about 92%. After the theoretical research, the design of the algorithm and software, and the experiment debugging, the design goal is basically reached, which provides some reference for the further research of the early detection of the fire.