Automated Victim Detection from UAV Thermal Infrared Imagery for Nighttime Search and Rescue Using Multi-Pose Ground Camera Data

The survival probability of persons requiring rescue after large-scale earthquakes or landslides decreases rapidly with time, and nighttime search operations are constrained by limited visibility. Satellite remote sensing enables wide-area observation at night, but its limited spatial resolution restricts the identification of individual persons. This study develops an automated method for detecting persons in thermal infrared imagery captured by Unmanned Aerial Vehicles (UAVs) using the deep learning model Grounding DINO. To address the limited availability of Unmanned Aerial Vehicle (UAV)-based training data, thermal infrared imagery captured by fixed-point ground cameras was used for training. Spatial resizing and padding were applied to emulate UAV viewpoints and mitigate the ground-aerial domain gap. The proposed model outperformed the baseline on real-world datasets, with substantial Recall improvements in environments with low thermal contrast. An ablation study confirmed that spatial resizing, padding, and position-based augmentation each contribute progressively to detection performance. A comparison with the lightweight YOLOv8n and DETR-based RT-DETR detectors indicated that the Transformer architecture alone does not account for reliable detection in scenes with complex thermal noise. This robustness instead derives from the language-grounded semantic priors of Grounding DINO. An extended evaluation on nighttime imagery indicated that the proposed approach generalizes to nighttime acquisition.