Abstract:

Fire detection based on videos is an effective method to prevent fire in large spaces. The texture of flame is special. Multiscale texture features were extracted to improve the flame detection performance due to its much more discrimination information. The flame candidates were located by character of flame brightness at first. Then different patterns of LBPfeature with different scales were extracted from these candidate areas. Finally, these features were put into SVM classification to recognize whether it was a flame or not. Experimental results showed that the method had a simple computation and could accurately recognize flame in video sequences and the false positive was low.

Key words: multiscales, uniform pattern LBP, rotationinvariant pattern LBP, rotationinvariantuniform pattern LBP, SVM