Smoke and Fire Detection with YOLOv9-Based Real-Time Computer Vision

Fire and smoke have always been dangerous to people, property, and the environment, so detecting them early and accurately helps save lives and property from damage. Conventional fire detection approaches are not only slow but also inaccurate in complex surroundings, which is a major disadvantage in any real-time hazard detection scenario. In response to these obstacles, this research proposes a fire and smoke detection framework using the cutting-edge You Only Look Once version 9 (YOLOv9) deep learning model. Roboflow Fire Segmentation Part 1 dataset with 7,040 images labeled with various scenarios of fire and smoke was used for training and validation. The performance of YOLOv9 was compared using standard evaluation metrics like precision, recall, F1-score, and mean average precision (mAP). At (confidence threshold of 0.885) a precision of 1.00, recall of 0.91, $F 1$-score of 0.77, and mean average precision (mAP@0. 5) of 0.81. However, the confusion matrix showed a comparatively high number of false positives, indicative of a struggle to detect negative classes despite a strong ability for positive class identification. These findings underscore YOLOv9's potential for real-time fire detection but also stress the necessity of further tuning to improve negative class identification and overall efficacy. YOLOv9 could play a pivotal role in improving fire safety systems with its high accuracy and rapid detection speed, allowing for quicker response to emergencies and reducing potential damages in key environments that may suffer from fire hazards.