Thermal vision-based victim detection and rescue prioritization in collapsed buildings using posture and temperature information

초록

Post-disaster Search And Rescue (SAR) operations require rapid and condition-aware decision making; however, most existing artificial intelligence systems focus primarily on victim detection and provide limited support for rescue prioritization in low-visibility and debris-rich environments. This study presents an intelligent decisionsupport framework that integrates thermal perception with rule-based reasoning to detect entrapped victims and assign rescue priorities in collapsed-building scenarios. A custom-trained You Only Look Once version 12 nano (YOLOv12n) model performs single-pass victim detection and posture recognition (standing, sitting, and lying), while also identifying partially visible cases such as body parts. For each detection, calibrated surface temperature derived from radiometric thermal imagery is combined with posture information through transparent decision logic to determine rescue priority levels (P1, P2, and P3). The framework was evaluated using a custom radiometric dataset collected under simulated disaster conditions. Experimental results demonstrate reliable performance, achieving a mean detection accuracy of mAP@0.5 = 0.898, posture classification accuracy exceeding 91%, and strong agreement between calibrated thermal measurements and thermometer references (R2 = 0.90; mean absolute error = 0.12 degrees C). The end-to-end prioritization process achieved 96.3% accuracy compared with manual expert assessment. This accuracy agreement with rule-derived reference labels based on annotated posture and reference temperature information. These results indicate that integrating perception outputs with interpretable reasoning enables consistent and rapid victim prioritization beyond simple localization. Although the experiments were conducted in a controlled indoor environment, the proposed framework provides a foundation for explainable decision-support systems for disaster response and other low-visibility emergency scenarios.

키워드