Enhancing fire detection capabilities: Leveraging you only look once for swift and accurate prediction

Abstract

Detecting fires is crucial to prevent potentially catastrophic outcomes. Traditional fire detection methods, relying on electronic, chemical, or mechanical sensors, often suffer from time delays in activation due to threshold parameters. An emerging alternative utilizes artificial intelligence, particularly image-based fire detection, using convolutional neural networks (CNNs). You only look once (YOLO) is a state-of-the-art object detection framework prized for speed and real-time capabilities. In our research, we conducted multiple training experiments employing various deep neural network (DNN) architectures as feature extractors for object detection within the YOLOv5 framework. These architectures included MobileNetV3, ResNet, and CSP-Darknet53. Among these configurations, YOLOv5 with CSP-Darknet53 (scale s) emerged as the most accurate, boasting mAP@50 of 0.88 and an impressive FPS of 73, with training model size of 14.50 MB. Furthermore, we integrated the selected model with the streamlit package to create a user-friendly web application interface for fire detection testing. The resulting model demonstrates remarkable efficiency, detecting fires within 0.01 seconds. This research represents a significant advancement in fire detection technology, offering both rapid detection and enhanced accuracy, with potential applications in various settings, from indoor facilities to outdoor environments.