Edgefiresmoke: a Novel Lightweight CNN Model for Real-Time Video Fire-Smoke Detection

Abstract

Forest fires can have severe impacts on both the environment and human communities. They can cause soil erosion, loss of habitat and biodiversity, as well as the release of carbon dioxide and other pollutants into the atmosphere. In addition, they can cause damage to properties, displacement of residents, and put firefighters and other responders at risk. Forest fires can also contribute to climate change by releasing stored carbon into the atmosphere and altering ecosystems. In this work, we propose a novel algorithm capable of monitoring small areas of forest reserve environment through video streaming in real-time. It will complement the existing means of forest monitoring and surveillance and provide effective solutions faced in satellite-based monitoring. The proposed algorithm is an improvement of the EdgeFireSmoke method and uses an artificial neural network together with a deep learning method. The proposed EdgeFireSmoke++ algorithm was able to detect forest fires with 95.41% accuracy, 95.49% precision, 95.38% Recall and 95.41% F1-score. The proposed algorithm recorded the best FPS rates of the HD IP camera at 33 FPS and with the USB VGA camera at 40 FPS. For its operation, the proposed algorithm proved to be quite light, being able to work on a CPU with 4 cores, 2.1GHz, with an average consumption of 540MB of RAM memory. This test was superior to the methods evaluated in the literature.