Spatio-temporal smoke clustering in outdoor scenes based on boosted random forests

Nowadays, vision-based techniques for automatic early smoke detection in the outdoor scenes are in a hot topic of computer vision. The basic set of features includes the traditional features describing the spatial ones, such as color, shape, transparency, energy, and fractal property, and the temporal ones, such as frame difference estimator, motion estimator, and flicker on boundaries. The main problem of the early smoke detection is to obtain the low values of the clustering errors. Our contribution deals with a reasonable clustering of the smoke/non-smoke regions based on the Boosted Random Forests (BRFs). The BRFs provide better clustering results in comparison with the traditional clustering techniques, as well as the ordinary random forests. Forty test video sequences with and without smoke were analyzed during experiments. The true recognition results of a smoke detection achieved 97.8% that is better on 3-4% of the results obtaining by the Support Vector Machine (SVM) application. False reject rate and false acceptance rate values were significantly decreased till 3.68% and 3.24% in average, respectively. (C) 2016 The Authors. Published by Elsevier B.V.