FLAME AND SMOKE ESTIMATION FOR FIRE DETECTION IN VIDEOS BASED ON OPTICAL FLOW AND NEURAL NETWORKS

Abstract

Detecting break out of fire at the initial stage itself is vital for the prevention of material as well as human loss. Traditional point based sensors for heat or smoke detection will detect the presence of fire only if the particles produced as a result of combustion reach the sensors. Video based fire detection approach is implemented in scenarios where point sensors may fail. Video based fire detection systems has got a wide variety of applications in large scale industries, naval vessels, forest fire detection etc. This is a new video based fire detection system that is making use of both flame and smoke features calculated from optical flow vectors created by different optical flow methods for fire detection. The technique used in this system is the optical flow vector creation that is used to represent the magnitude and direction of motion undergone by an object while moving from one frame to another one. For flame flow vector creation OMT and NSD methods are used which are used for modeling flame with dynamic texture and saturated fire blobs respectively. For smoke flow vector creation pyramidal Lucas-Kanade optical flow method is being used. The flow vectors created are further analyzed to create feature vector for flame and smoke respectively. Then two feed forward neural networks are used for flame and smoke feature vector classification. The outputs from neural networks are analyzed to find the presence of flame and smoke in the frame. The output from both networks in combination is used to make a final decision.