Indoor Smoking Detection Method based on Dual Spectral Fusion Image and YOLO Framework

Abstract

Indoor fires are a major problem for public safety, with smoking being the most hidden threat. Traditional fire detection systems, such as smoke detectors, are only useful in the early stages and face challenges due to low light and limited visibility. This article describes an indoor smoking detection system that combines visible and infrared image fusion with the YOLO (You Only Look Once) detection framework. This technique improves indoor smoking detection performance by combining infrared thermal data with deep learning concepts. The YOLOv9 system detects indoor smoking behavior using a deep neural network for feature presentation and inference. The approach is optimized at the data, feature extraction, and model training levels to improve scene adaptability. The experimental results showed that on the custom indoor smoking dual spectral fusion image dataset, the average accuracy mAP (@ 0.5) of the Modified YOLOv9c detection model reached 95.8%, which was much better than the baseline models YOLOv5s (81.4%), YOLOv7 (89.7%), YOLOv8 (90.8%), and YOLOv9c (89.9%) mAP, respectively with significant performance improvements. Strategies like dual spectrum fusion, data augmentation, attention mechanism, and loss function were implemented to improve model detection performance. This paper presents a practical solution for indoor smoking detection tasks, demonstrating the approach's superiority in detection performance and providing a viable toolset for public safety against indoor fire hazards.