Artificial intelligence based fire detection system

Abstract

Fires cause great damage when they burst, and often have great destructive effects on environment and surroundings. The most effective way to limit the damage is the early detection of fire before it spreads. This work investigates the ability of Deep Learning to identify and distinguish fire, as well as reduce detection time, by applying object detection on a video or image stream. Over the previous years, object detection has advanced gradually in terms of speed and accuracy. In this work, we proposed a solution based on Deep Learning to deal with such phenomena from collecting various datasets to training them using the one-stage detector YOLOv8 and YOLOv5 and the two-stage detector Faster RCNN with VGG16/19, Xception, Inceptionv3, MobileNet, and our proposed hybrid model Xception-VGG19 which is a concatenation of both VGG19 and Xception as backbones. We gathered 6 different datasets of fire and smoke. The obtained results were satisfying using YOLOv8, with D4 which contains smoke-only images with 99% mAP@0.5 followed by D6 with 93%, D3 with 92%, and D5 with 70%. With YOLOv5 D2 43% and D1 34%. Moving forward to the two-stage detection the best outcomes were obtained by Xception-VGG19 with 44% followed by Inceptionv3 with 43%, VGG16 with 40%, VGG19 with 35%, MobileNet with 33%, and Xception with 23%. We also proposed a simulation scenario using UAVS to take down fire once detected by cameras as an extension of our work.