Abstract: A new smoke detection method based on video is researched due to the weakness of the
traditional fire detection technology. Firstly, according to the characteristic of the smoke color, the
suspect smoke regions are extracted in video sequences. Then, looking for three dynamic
characteristics of smoke in the suspected smoke area, there are three features extracted, which
respectively are the growth of the area in the smoke spread, irregular contour feature of the smoke
region and the background to blurred when smoke appeared. And those three dynamic characteristics
are fused by a BP neural network to determine smoke or not. Test results show that the multi-feature
fusion smoke detection algorithm can identify smoke in video accurately, real-time and effectively.

Key words: smoke, color feature, dynamic characteristics, BP neural network