Video detection for dynamic fire texture by using motion pattern recognition

Abstract

Significant motion features which are able to be used for fire video detection in regard to the dynamic fire texture are proposed in this article. We are now interested in motion characteristics rather than color schemes. Since colors of fire textures observed on video medium nowadays are possibly illustrated with whimsical colors. It is not caused by only nature chemical phenomena but also by special effect application technologies in video industry. We propose four data series of motion features gained from motion vector field or optical flow estimation, namely, the series of average radius, the series of motion coherence index, the covariance stationary series of average radius, and the covariance stationary series of motion coherence index, respectively. The extracted data is used by machine learning part to form training set and test set for video classification using support vector machine method. Our four proposed data series are able to leverage fire video detection. Our experimental results demonstrate that the accuracy of video detection in regard to fire texture is significantly high and its time elapsed only few seconds of gaining data.