Optimizing Survivor Detection Models on Thermal Imagery for UAV-Assisted Search and Rescue
2026 8th International Congress on Human-Computer Interaction, Optimization and Robotic Applications (ICHORA), Ankara, Türkiye, 21 - 23 Mayıs 2026, ss.1-10, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/ichora69329.2026.11537125
- Basıldığı Şehir: Ankara
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.1-10
- Hacettepe Üniversitesi Adresli: Evet
Özet
This study presents a systematic hardware-aware optimization and deployment analysis of deep neural networks for human detection in disaster areas using thermal imaging in Search and Rescue (SAR) operations. The study specifically targets real-time inference on AI-mounted UAVs, particularly the NVIDIA Jetson Orin Nano. Various state-of-the-art object detection architectures are evaluated, including YOLO models (YOLO v5-v11, v26) and transformer-based detectors (RT-DETR, RFDETR). In 278 experimental configurations, combinations of post-training quantization (FP16/INT8 via TensorRT), unstructured pruning (30−85% sparsity), power mode scaling (7 W−25 W) and input resolution reduction (160-512 px) were applied. Models are trained and validated on the WiSARD wilderness thermal UAV dataset to reflect realistic SAR deployment conditions. The contributions of the project are overall comparative analysis of detection methods and a scalable/optimized deep learning framework capable of real-time survivor detection from thermal images on edge AI devices.