Machine Learning Method for Image Recognition-based Fire Detection System

Abstract

Ideal smart fire detection systems should be able to detect the fire and trigger the automatic alarm at an early stage. Meanwhile, it should also trigger the automatic fire extinguishing system and broadcast the fire alarm under different fire conditions in order to avoid further damage and losses. The most critical problem for a smart fire detection system is to determine the existence/condition of the fire based on different kinds of information. The smart fire detection system should have fast reaction and be very reliable in the action process. Most of the modern smart fire detection systems are based on multi-sensor systems, or image/video surveillance system. In this work, we propose to combine the multi-sensor detection system with image recognition. When the decision from the multi-sensor system is uncertain or the data is not available/faulty, images are used to assist the fire detection process, which could make the whole system more robust and reliable. We are aiming at extracting important features from the images by using machine learning methods. Then, different classification methods will be applied to detect the fire conditions. We make use of the existing images collected from real environments to evaluate the proposed approach. In addition, we investigate and discuss the detection results using different classification methods, which verifies that the image-based fire detection scheme combined with multi-sensor system can achieve better efficiency and accuracy.