Research On Early Fire Detection Algorithm Based On Machine Learning

Fire is the most important work of disaster prevention and mitigation in today's society,which has the characteristics of frequent occurrence,strong destructive power and rapid development.According to statistics,fire accidents cause hundreds of deaths and property losses of billions of RMB every year,so timely and accurate detection of fire is of great significance to safeguarding people's life and property.At present,smoke,temperature and other sensors are often used in practical application to implement fire detection,which has the disadvantages of long response time,high equipment cost and single alarm information.With the rapid development of machine vision,image-based fire detection technology has been proposed,which has the advantages of short response time,wide detection range,low cost and abundant alarm information.This paper proposes a new fire detection algorithm aiming at solving the problems of the existing algorithm for video-image fire detection,including serious loss of foreground information,high false alarm rate and weak generalization ability.The algorithm consists of two modules: foreground extraction and classification decision.The foreground extraction module first uses the improved ViBe algorithm to extract the motion area.Then the motion area is divided into yellow,white and other color region by random forest model.Finally,the Support Vector Machine algorithm is used to classify the yellow and white pixels of foreground regions in order to obtain accurate foreground regions.The classification decision module combines spatial feature and temporal feature to implement the classification of foreground region.Considering the lack of representativeness of the existing early flame features,two new early flame time domain features are proposed to describe the overlap rate of the flame region between frames and the ratio of motion intensity of different flame region parts,and then the decision classifier is trained with the Hu moment feature.According to experiment testing,it is found that there are a few false positives for outdoor scenes,which is resulted from the awfully simple feature through detailed analyzing.Therefore,convolutional neural network is adopted in this paper to classify the foreground area.Experimental results show that the convolution neural network can classify the above false positives correctly.In view that there is a great shortage of fire videos and there exists no public fire video datasets,this paper collects some actual and experimental fire videos to evaluate the above algorithm.The experiment shows that the proposed algorithm has good detection results,strong generalization ability and robustness,and is able to meet the demand for real-time detection.Besides,it can still obtain a low false alarm rate and a high detection rate even in some complex scenes.Therefore,the proposed algorithm can be applied into different kinds of actual environment.