Real-Time Smoke Detection in Video Sequences: Combined Approach

This paper presents a combined approach for rapid smoke detection from video sequences using pre-improvement methods. Smoke is characterized by different properties like a color, irregularities in motion, smoothing the edges, etc. There are hard to describe it using basic image features. Usually smoke detection algorithms use some color and turbulence features for the smoke description. But our experiments shown the high false alarm rate of such algorithms. In our work we propose the additional smoke description features based on smoke transparency. Before the recognition we use the regions matching method to increase the flexibility of our system. As a first step in processing we extract background. Moving objects are candidates for smoke. The Gray World algorithm is used here. Compare the results with the original frames in order to get image features within some particular gray scale interval. After that we use the rate of color changing to checking of the transparency of current area. Last we calculate complexity of turbulent phenomena of the smoke shape and apply it to the incoming video stream. As a result we have just smoke regions on the video stream. There are different objects, shadows or illumination changes will not be mistaken for smoke by the algorithm. This method gives an early recognition of smoke in the observed scene.