Machine vision based fire flame detection using multi-features

Video fire detection has many advantages over traditional methods, such as fast response, non-contact. But most of current methods for video fire detection have high rates of false alarms. In point of general fires, the flames usually display reddish colors. And as an important physical feature of fire, the flame turbulent has a chaotic nature with abundant size and shape variation. If we consider the flame is made up of lots of spots, as a result of the turbulent movement, the spots´ velocity vector will be different from each other. A novel video fire flame detection method based on color and dynamic features is presented. The method is proposed as followed, first, candidate fire regions are determined by frame differential method and a flame color model. Then, a pyramidal Lucas Kanade feature tracker is used to calculate the velocity vectors of the feature points of the fire candidate regions. Finally, examples consisting of features extracted from sequences of off-line videos are collected for the training of a discriminating model which is used to differentiate fire from some other moving objects. Experiments show that the algorithm has fast response and encouraging false alarm rate for fire flame detection.