Research On Methods Of Video-based Flame Detection

With the rapid development of economy and society,once the fire broke out,it will cause immeasurable loss because of the rise of high buildings,industrial expansion and excessive concentration of personnel and property.Thus more and more attention has been paid to public safety caused by fire,and the development of fire detection technology becomes a focus.In recent years,the intelligent video surveillance system is applied to the various walks of life,which provides a platform for the development of video-based fire detection technology.Contrasting to traditional fire detection technology,video-based fire detection is not only suitable for large space but also can reduce the cost.This dissertation studies on the methods for video-based flame detection.The process consists of two phases:flame candidate region extraction phase and flame features detection phase.The main work is as follows:Firstly,the extraction methods of flame candidate regions are explored.One is the methods based on color.We use RGB,YCbCr and HSV color space rules to achieve the extraction of flame pixels.The other is saliency detection method.And we apply confidence measures to further acquire accurate flame regions.We also compare and analyze the differences between color space methods and saliency detection methods.Secondly,a flame detection method based on spatiotemporal SURF feature is explored.We extend the traditional SURF descriptors to the spatiotemporal domain and combine it with the global color histogram feature to realize the flame detection.In this way,we make full use of the spatial static features and temporal dynamic features of the flame.Compared with the existing spatiotemporal feature methods,our method achieves more excellent performance.Finally,another flame detection method based on convolutional neural network is explored.We combine the saliency detection method and CNN feature learning method,and apply it to video flame detection.The experimental results show that the proposed method achieves higher detection accuracy and lower false alarm rate than the local covariance temporal feature method.