An Improved YOLO-Based Framework for Real-Time Fire and Smoke Detection

Abstract

Fire and smoke detection is highly relevant in terms of protecting people, buildings, and the environment as it helps people to be aware of dangerous events in time. Conventional vision-based detection algorithms rely heavily on hand-crafted features, which do not always work well in challenging environments with variable light conditions, cluttered backgrounds and high smoke content. To break these limitations, a deep learning-based fire and smoke detector system is developed based on advanced YOLO models. An image gallery consisting of images of fire and smoke scenarios is used, which is bound by annotation in the form of an object localization (YOLO). Image normalization, organizing the data, and establishing a single data file to be used with the classification as well as the detection process are all aspects of preprocessing. We train and test various different versions of YOLO, including YOLOv5, YOLOv8, YOLOv9, YOLOv10, YOLOv11s, and a better version of YOLOv11-DH3. The latest YOLOv11x model is also provided to determine whether the performance has improved. Models are measured by using standard metrics such as precision, recall, and mAP. The experimental results indicate that YOLOv11x is the most effective, and its precision is 0.930, recall is 0.981 and mAP is 0.967. It is also linked to a Flask-based web interface in such a way that it is capable of detecting smoke and fire in real time. There is also an automated system of email alert. It provides real time email messages with the count of detections, confidence, and the geographical location of fire and smoke events. This allows the need to respond swiftly and implement the system in a more efficient manner.