Improved Fire Detection by YOLOv8 and YOLOv5 to Enhance Fire Safety

Fire hazard has always been a major concern for mankind. To safeguard our lives and assets from this threat, its precise detection is a prerequisite to initiate protective measures. In this paper, we propose a fire detection framework based on the You Only Look Once (YOLO) algorithm. The YOLO is a Convolutional Neural Network (CNN) backboned object detection model. We have evaluated the most popular YOLO version i.e., the YOLOv5 and also the latest YOLOv8 model. Three separate submodels i.e., the Nano, Medium and Large submodels of each of YOLOv5 and YOLOv8, are trained on a benchmark fire dataset and our custom dataset. Thereby, a total of six YOLO models are analyzed to examine their efficacy in fire detection and compared with existing CNN-based fire detection systems. Both the YOLOv5 and YOLOv8 models have outperformed other contemporary methods with the YOLOv8 resulting in the highest precision. A precision rate of 99.1% on the benchmark dataset and 94.5% on our custom dataset is recorded for the YOLOv8 in fire detection. This accuracy rate opens a new door for incorporating our model in modern fire alarm systems.