Addressing false alarms from high-voltage structures in subpixel fire detection

Abstract

False alarms in subpixel fire detection often arise when high-voltage structures, such as powerlines or towers near thermographic cameras, emit intense infrared radiation that mimics early fire signals at long distances. This paper proposes the study and statistical analysis of You Only Look Once version 8 (YOLOv8) to detect, segment, and isolate these sources of false alarms. YOLOv8 is trained on the Addressing False Alarm Situations (AFAS) dataset, which includes a variety of Long-Wave Infrared (LWIR) and Near-Infrared (NIR) imagery from both aerial and ground-level perspectives. The model achieves a mean Average Precision (mAP) of 0.784 at an Intersection over Union (IoU) threshold of 0.5. The contribution of this work lies in a detailed statistical analysis of YOLOv8 outputs, introducing, among others, the Empirical Cumulative Distribution Function (ECDF) as a metric to assess the relationship between mask overlap and detection confidence. To evaluate the model's robustness under thermal disturbances, synthetic fires are introduced to simulate changes in the scene. The two-sample Kolmogorov-Smirnov (KS) test compares prediction distributions with and without these anomalies, important to ensure that the model performs reliably over a wide range of scenarios so that the presence of these structures can always be determined and isolated. Finally, an energy retention metric is introduced to quantify the probability that the model's predicted masks obscure at least half of an early fire's energy. In critical cases where the fire appears at 2, 3, and 4 pixels from the segmented structures, these probabilities are approximately 7%, 4%, and 3%, respectively.