A near-real-time UAV-based computer vision framework for fire and smoke detection and localization using YOLO object detection and visible light communication

Introduction Early and reliable fire and smoke detection is essential for effective emergency monitoring, particularly in UAV-based environments characterized by variable illumination, changing object scales, and smoke-related visual disturbances. This study proposes a near-real-time UAV-based computer vision framework integrating YOLO-based object detection, geometric fire-source localization, and visible light communication (VLC). Methods A dataset of 4,200 manually annotated fire and smoke images was divided into training, validation, and test subsets using an 80:10:10 ratio. YOLOv10, YOLOv11, and YOLOv12 architectures were compared using Precision, Recall, mAP50, mAP50-95, inference time, and frames per second. Fire-source localization was modeled using camera-ray projection and ground-plane intersection based on UAV position and camera geometry. RF and VLC communication were evaluated through analytical modeling and simulation. Results YOLOv10l demonstrated a favorable balance between detection accuracy and computational efficiency, achieving Precision of 0.962, Recall of 0.968, mAP50 of 0.980, and mAP50-95 of 0.676. Under the modeled communication conditions, VLC-assisted transmission showed signal loss of 5–10%, compared with 15–25% for RF communication, with estimated latency below 40 ms. Simulated fire-source localization errors ranged from 0.5 to 1.2 m and decreased to 0.3–0.5 m under multi-view triangulation assumptions. Discussion The findings demonstrate the potential of integrating YOLO-based fire and smoke detection with geometric localization and VLC-assisted communication for near-real-time UAV monitoring. The detection experiments were conducted using workstation hardware, while communication and localization results were obtained through modeling and simulation. Further onboard UAV testing and independent validation are required to assess real-world performance and generalizability.