Real Time Fire and Smoke Detection Techniques Using Improved Transfer Learning Algorithms

Classical fire and smoke detection systems cannot be used outdoors and do not guarantee the necessary availability for real-time applications. The objective of this research paper is to use a deep neural network to detect fire and smoke in outdoor environments using an embedded system with a camera. Recently, after significant progress in programming structures and, above all, in the field of machine learning, embedded systems are now able to perform real-time edge computing. This work proposes improved transfer learning approaches to perform computer vision fire and smoke detection using a convolutional neural network (CNN). Three transfer learning models (VGG-16, VGG-19, and YOLOv5) have been enhanced and used in order to detect once fire or smoke appears in any camera, then the system can detect them and immediately send an alert to the corresponding supervisors. The performances of the proposed algorithms are verified on a common database; then on real scenarios.