SAR-YOLO11: A High-Resolution YOLO11 Framework for UAV Thermal Infrared Person Detection in Search-and-Rescue Scenarios

Abstract

For search-and-rescue (SAR) teams, unmanned aerial vehicle (UAV) thermal infrared imagery offers a way to look across unsafe, dark, or inaccessible terrain before rescuers enter the scene. Reliable person detection in this setting is still difficult. Victims seen from altitude often cover only a few pixels; thermal images carry limited texture; clutter, partial occlusion, and changing viewpoints can all weaken localization. This paper addresses these constraints through SAR-YOLO11, a high-resolution YOLO11-based framework designed for thermal person detection in SAR-like UAV scenes. We build a strict person-only protocol from the public HIT-UAV dataset, retaining 2,898 labeled thermal images and splitting them into 2,029 training images, 290 validation images, and 579 testing images. The framework uses a 960 x 960 input setting so that small thermal cues are less likely to be lost during resizing, and it narrows the recognition task to the person category to reduce class ambiguity and focus model capacity on rescue-relevant targets. Under the same evaluation protocol, YOLO11n, YOLO11s, YOLO11s-P2, and SAR-YOLO11 are compared. SAR-YOLO11 reaches 0.894 precision, 0.880 recall, 0.923 mAP@0.5, and 0.519 mAP@0.5:0.95, giving the strongest strict-localization result among the tested variants. The results suggest that high-resolution, person-only optimization is a practical route for improving small thermal target localization in UAV-assisted SAR perception.