AN EMPIRICAL EVALUATION OF REAL TIME FIRE AND SMOKE DETECTION IN COMPLEX ENVIRONMENTS USING THE YOLOV8 ARCHITECTURE

Abstract

Automated real time fire and smoke detection is critical for modern disaster mitigation and smart city surveillance infrastructure. However, standard single stage deep learning object detection models frequently suffer from high false positive rates due to the amorphous, dynamic nature of fire and smoke, often misclassifying environmental artifacts such as sun glare, clouds, fog, and artificial reflections. This study presents a rigorous empirical evaluation of the baseline YOLOv8 architecture deployed for vision based hazard detection under complex environmental constraints. Utilizing a comprehensive dataset of over 13,000 images characterized by a heavy distribution of small scale targets, advanced preprocessing and augmentation strategies including Mosaic augmentation, Letterboxing, and HSV color jittering were deployed to optimize model robustness. The baseline model was trained and evaluated over 50 epochs, achieving an overall mean Average Precision (mAP@0.5) of 53.9%, with individual class performances reaching 62.3% for fire and 45.5% for smoke. Detailed error analysis using a normalized confusion matrix reveals a critical challenge in separating semi transparent smoke from complex background noise, yielding a 58% background confusion rate. These findings establish a baseline performance benchmark for edge ready disaster management systems and outline the exact architectural boundaries where standard single stage detectors require future spatio-temporal or structural modifications.