4. Conclusion
A new approach based on wavelets and a support vector
machine has been proposed for smoke detection. Characterization
of smoke was carried out by extracting wavelet features from
approximate coefficients and three levels of detailed coefficients.
The system is implemented in visual C++ on a PC and is shown to
work well using a block based approach as well as on motion
segmented images. An excellent cross validation accuracy of over
90% with sensitivity and specificity of 0.9 and 0.89 respectively is
obtained on videos taken of forest fire. This indicates that the 60
features extracted can efficiently represent smoke in the tested
scenarios. The method has the flexibility to analyze smoke every
few seconds using only a few frames rather than continuous
monitoring. It can also be used with systems which have motion
detection capabilities as discussed in Section 3. To check the
robustness of the technique, motion segmentation was carried out
on a forest fire video, a news channel video, a tunnel video and
then input to the system. A leave one out error of 8.53% is
obtained which is an indication that the recognition engine can be
plugged into any commercially available surveillance system.