Vision Based Fire Detection Using Mixture Gaussian Model

Abstract

Vision based fire detection has many advantages over traditional methods. In vision based fire detection approaches, it is required that systems must have enough robustness and be insensitive to environment. We mainly take advantage of mixture Gaussian model and frame difference techniques to adaptively extract a background image from image sequences captured by ordinary color cameras. These techniques are able to mostly eliminate influences of artificial lights, wind and moving objects disturbance. By subtracting the background image from the incoming frame, foreground objects which are possible fire pixels are thus obtained. After analyzing behavior and spectroscopy of fire, color, shape fluctuation and growth rate are used to determine if a possible pixel is an actual fire pixel. Experiments show that our algorithm is robust for a stationary camera.