Video Based Fire Smoke Recognition

This paper mainly studies an approach for video fire smoke detection, which can be effective in detecting the early fire smoke and monitoring a broad area. In order to overcome the insufficiency of traditional video fire smoke detection methods, this paper firstly proposes a suspcious smoke regions location algorithm that combines the Gaussian mixture model and color smoothness function. This algorithm utilizes the motion and color smoothness property of smoke to locate the most suspicious fire smoke regions with few non-fire regions, and hence decrease the computation cost in feature producing. Then Gabor wavelet is applied to model the static textures of the suspicious regions, and the energy variation model and the model of texture varying orientation are then presented to analyze the dynamic characteristics of the locating regions. The static texture features and dynamic texture features of suspicious regions are represented with histograms, so as to improve the statistic sense of the features. Finally, GentleBoosting classifier is trained to recognize the fire smoke according to the inputed joint features. The experiment proves that the proposed method can effectively detect the fire smoke in a windy environment.