Characterizing Confidence Compression in Cross-Domain Thermal Human Detection for Search and Rescue

This paper addresses the "thermal crossover" phenomenon in autonomous Search and Rescue (SAR) drones, in which solar-heated objects mimic human body heat and drive high false-alarm rates in single-stage detectors. We propose a two-stage cascade pairing a permissive YOLOv8-Nano scanner with a fine-tuned EfficientNet-B2 refiner adapted to the HIT-UAV aerial thermal benchmark. On a held-out evaluation, the proposed model attains a balanced classification accuracy of about 95% at its operating threshold, substantially reducing false alarms over a single-stage baseline while preserving human recall. Separately, we characterize a Confidence Compression phenomenon: using a ground-level thermal dataset as a reference, we find that under domain mismatch the verifier’s output confidence scores are pinned below roughly 0.64 on the 0–1 probability scale. This is a calibration effect—it concerns the confidence values the verifier assigns to each detection, not how often it is correct—and domain-matched adaptation relieves it. An analytical model further projects real-time feasibility on the NVIDIA Jetson Orin Nano.