This paper presents a vision-based framework for detecting humans and estimating head surface temperature from aerial thermal imagery acquired by Unmanned Aerial Systems (UAS). A comparative evaluation of recent object detection architectures was conducted to identify the most stable and reliable model for thermal human detection under varying flight altitudes. The selected framework integrates two head localization strategies, namely, segmentation-based mask slicing and pose-assisted keypoint localization, to extract head regions and compute per-pixel temperature values from radiometric metadata. The results show that cross-domain inference using pre-trained YOLOv11 models achieves reliable human detection across controlled outdoor environments. Between the two pipelines, the pose-assisted method produced temperature estimates closer to the expected human physiological range (36–38 ∘C), whereas the segmentation-based approach exhibited higher values attributable to mask boundary contamination and solar surface heating. In the absence of ground-truth validation from medical-grade sensors, these findings are characterized as relative comparisons rather than absolute accuracy claims. This study establishes a methodological foundation for future UAS-based thermal assessment systems and identifies critical calibration and validation requirements for field deployment.
Ahmed et al. (2026) studied this question.