A study of video fire detection and its application

Abstract

In order to save life and property from fire, early and accurate fire detection is one of the important aspects in fire safety design. For traditional spot type and line type fire detectors, information concerning the fire parameters such as flame height, fire growth rate and fire location are difficult to obtain. Such information is very useful especially to enable effective fire evacuation and firefighting. In addition, traditional fire detection technology has 8 - 10% false alarm rates. With the fast development of computer technology and image processing techniques, it is now possible to use video images to detect fire in different locations and obtain those fire parameters. It is able to supplement traditional fire detection technology so the development of video fire detection is becoming important. Besides, it is also useful in fire safety design for outdoor environments. This thesis presents the study of video fire detection and its application. Based on the literature review, video fire detection systems have already been developed to detect forest fires. Four kinds of video fire detection analysis methods including the digital image processing method, statistical colour model method, artificial neural network method and combined different approaches are used. The primary objective for the development of different analysis methods is to enhance the accuracy of video fire detection. In addition, video fire detection system is not only connected to surveillance system but it can also be combined with the fire suppression system for better fire control and mitigate unnecessary water damages. In this study, computer program analysis of the fire images is conducted using MATLAB toolbox, Visual C++, C, C++. Open source computer vision library (OpenCV) and software is developed to capture the fire image and conduct analysis. In addition, image processing techniques using OpenCV are able to processing the real time images for fire detection. Experimental study was conducted using thermal still images and normal still images to develop flame region segmentation codes using the Otsu's method. The traditional Otsu's method is able to segment the flame region and background but unable to segment the flame region completely with complicated background such as the shading and the objects reflection. Therefore, modified Otsu's multi - threshold analysis method is developed to segment the flame region from complicated background video images. To recognize the fire images, parameters such as colour, flame size, motion characteristics of fire are used. Flickering frequency and direction of flame spread are two important fire images characteristics used. Logistic discrimination rules are then set up to determine the probability of the true fire conditions. Once the fire condition is determined, the analyses of real time flame motion images are conducted. Optical flow analysis is then used to track the flame spread direction and flame height can then be estimated. With the information of the flame height, the fire size is estimated using the flame height empirical equation developed by other fire researchers.