Abstract: Using the random forest algorithm, a video smoke detection method with features selection was proposed. The method first extracted four original smoke image features including color features in RGB space, wavelet high frequency sub-images, multi-scale local max saturation, and multi-scale dark channel to input the random forest(RF). Then it utilized haze image formation model to make the synthetic smoke images from non-smoke images and partitions these images into blocks as the samples for RF. Thirdly, it trained RF to get the selected features from the original features and used support vector machine(SVM) to get a classifier which recognizes the smoke blocks and the non-smoke blocks. And then the smoke region candidate can be extracted from video images by the classifier. Finally, the method analyzed the detected smoke region with the features of the growth rate and the perimeter to area ratio to make the final decision on video smoke detection. The experimental results show that the proposed method can detect the smoke timely and give a fire alarm with a lower false-alarm rate.

Key words: smoke detection, random forest, SVM, features selection, wavelet transform, smoke growth, dark channel