Development of a Deep Learning–Based Object Detection Model Optimized for Edge Devices Using Thermal Imaging Data

Thermal imaging has become an effective alternative to RGB vision for human detection in low-light, night-time, and visually challenging environments. Despite its robustness to illumination changes, deploying deep learning–based object detection models on resource-constrained edge devices remains challenging due to high computational and memory requirements. In this study, we investigate the performance and deployability of YOLO-based object detection models for thermal human detection under edge-oriented constraints.YOLOv5 and YOLOv8 architectures are trained and evaluated using thermal infrared images from the LLVIP dataset, with a focus on detection performance, inference speed, and model size. Experimental results show that YOLOv5 achieves slightly higher mAP@0.5 than YOLOv8 (0.983 vs 0.971) while offering significantly better computational efficiency, making it more suitable for edge deployment. Based on this observation, the YOLOv5 model trained on thermal data is selected for further optimization.The selected model is exported to the ONNX format and optimized using NVIDIA TensorRT with FP16 and INT8 precision modes. Performance evaluations demonstrate that FP16 optimization significantly reduces inference latency without degrading detection performance, enabling real-time processing at approximately 170 frames per second. In contrast, INT8 quantization does not yield the expected speed or memory benefits under the tested conditions, mainly due to mixed-precision execution and batch size limitations.The results indicate that FP16-optimized YOLOv5 offers a reliable and efficient solution for real-time thermal human detection on edge-oriented hardware, providing practical design insights for deploying thermal vision systems in security and surveillance applications.