A behavior-based flame detection method for a real-time video surveillance system

Some characteristics of flames, including vague shapes and fire-like colors, are irregular, so it is difficult to use a video camera for real-time flame detection. To achieve fully automatic surveillance of fires, a real-time incipient flame detection approach that uses three stages of video processing is proposed in this study. Firstly, a thumbnail method is applied to the flame image for the various resolutions for real-time processing. A median filter and a Gaussian filter are used to remove the image noise. The flame edge information in the image is then enhanced, using a sharpening filter. Secondly, quasi-periodic behavior in flame boundaries is detected using motion detection and a background edge model, so some candidates for real flames can be selected. Thirdly, an evaluation method that uses four criteria, including compactness, fill rate, corner flicker rate and growth rate is used to identify the correct flame. The results are verified using videos of complex scenes at different resolutions. A comparison with other systems is made, to demonstrate that the proposed method is more accurate.