Deep Learning Models for Fire Detection using Surveillance Cameras in Public Places

Abstract

In a world led by automation and robotics, an exposed wiring, an overloaded outlet, or a single spark of electricity is capable of inducing a fire. Fire threats are one of the most recurring accidents in the fast-paced world. The magnitude of damage left behind by these incidents depends on how fast help arrives or the working efficiency of fire extinguishing systems. The field of deep learning using computer vision puts forth a fast and efficient technique for fire detection. This, when coupled with necessary additions, can result in an early dousing of fire. This paper puts forth five deep training models for fire detection - RegNet, ResNet, Mobile Net, GoogLeNet, and a customized Feed Forward Neural Network model. The project aims to build a fire discovery method that can discover fire utilizing surveillance feed. An analytic comparison of the performance of all the models suggests that ResNet and Mobile Net top the chart with an accuracy of 83.06%, followed by GoogLeNet, RegNet, and Feed Forward neural network.