Smoke Detection Using GMM and Deep Belief Network

摘要

The objective of this work is to develop a deep learning model for classification of smoke and no smoke regions in aerial recorded videos. For that, a deep belief network model was selected and implemented. First, frames were extracted from the provided videos. The Gaussian Mixture Model (GMM) was applied as background estimation algorithm. Then, the Deep Belief Network algorithm was applied to detect the smoke for the candidate region. Deep Belief Network was implemented and tested on different datasets. Overall, the obtained results reveal that our implemented model was able to accurately classify smoke and no smoke regions. Through the experiments with input videos obtained from various weather conditions, the proposed algorithms were useful to detect smoke in forests to minimize the damage caused by forest fires onto vegetation, animals and humans.