Dynamic Texture Modeling Applied on Computer Vision Based Fire Recognition

In computer vision based fire detection systems, the fire sensing algorithm usually consists of two main parts: fire pixel classification, and analysis of the candidate regions. Purpose of the first step is to find a set of pixels which may be fire, and the second step is used to make decisions whether fire exists in the region. The algorithm proposed in this paper works on the second step. Different with the traditional methods using color or shape information, candidate fire regions in video are taken as dynamic textures in our approach, and dynamic texture modeling is adopted for fire testing. That is, for a candidate region, we build dynamic texture models, e.g. LDS, NLDS, and our proposed MRALDS, then parameters of the models are employed to determine the presence or absence of fire using our proposed recognition framework. Performance of our method has been tested with 68 videos including 32 fire videos and 36 non-fire videos. The experimental results are quite encouraging in terms of correctly classifying the videos.