Wavelet-based multi-modal fire detection

Over the last decade, video fire detection is started to be explored as an alternative for traditional fire sensors. Inspection of the several flame and smoke detection algorithms that have been proposed in literature shows that most of them start from simple background subtraction in spatial domain. The influence of the background model, as such, is not yet fully explored. This paper is a first attempt in this direction and investigates the added value of two wavelet-based background subtraction methods for segmenting the input scene during video fire detection. The first of these wavelet based methods focuses on both the high energy and low-pass images of the Discrete Wavelet Transformed input video frames in spatial domain. The second wavelet-based method is a nonlinear Difference of Gaussians method, which is illumination invariant and performed in frequency domain. Experimental results show that both wavelet based methods lead to better fire detection results than non-wavelet based background subtraction methods. Especially when there are a lot of flame reflections and other fire-related illumination changes, less false alarms and missed detections occur in the wavelet-based setups.