Enhanced fire detection: a weighted bounding box fusion approach based on Dempster–Shafer theory

In this study, we present an advanced fire detection approach using an improved weighted bounding box fusion technique based on Dempster–Shafer theory. A key challenge in fire detection is the inherent uncertainty in distinguishing true fire events from false positives ( e.g ., fire-like objects or background noise). Our approach addresses this by fusing overlapping bounding boxes from a You Only Look Once (YOLO)-based detector, explicitly modeling conflicts between fire and non-fire classifications. Dempster–Shafer theory plays a critical role in quantifying uncertainty and reconciling contradictory evidence, enabling more robust decision-making. We further introduce a confidence-aware weighting scheme to prioritize highly reliable detections during fusion. On a diverse fire image dataset, YOLOv5 combined with our Dempster–Shafer Weighted Boxes Fusion (DS-WBF) method achieved a precision of 0.814, recall of 0.863, AP 0.5 of 0.875, and AP 0.5–0.95 of 0.542, representing gains of +1.24%, +1.17%, +2.46%, and +4.63% over traditional Non-Maximum Suppression (NMS). These improvements reduce false alarms and enhance true positive rates, enabling more reliable early fire detection for real-time monitoring systems.