An End-to-End UAV–Cloud–AI Pipeline for Infrared-Based Real-Time Person Detection in SAR Missions

To effectively respond to the growing number of search and rescue (SAR) incidents in mountainous areas, a real-time and automated detection system is essential. Traditional SAR operations still rely heavily on manual visual detection, which becomes signif icantly limited in low-light or night-time environments. The proposed system is an integrated UAV–Cloud–AI automated architecture designed to enable real-time detection even in low-visibility environments, with the goal of improving rescue efficiency. The entire pipeline-ranging from Uncrewed Aerial Vehicles (UAV) operation and IR data acquisition to MinIO-based storage, event-driven file conversion, SFTP-based transmission, and AI-based inference is fully automated without manual intervention. To evaluate system latency, the entire process was divided into four key stages image acquisition, cloud upload and conversion, server transmission, and object detection. A total of 231 Infrared (IR) images were collected across five sorties, with an average processing time of 12.4 seconds per image. The upload and conversion stage showed the longest delay at 10.474 seconds, while file transfer and model inference recorded stable performances of 0.532 seconds and 1.402 seconds, respectively. In addition, detection experiments u sing YOLOv12 demonstrated that the model consistently identified human targets in thermal imagery, even under complex backgrounds and low thermal contrast. This study experimentally validated the feasibility of a UAV-based SAR system capable of real-time detection and response. Its scalability and field applicability are expected to be further enhanced through the future integration of lightweight detection models and collaborative multi-drone architectures.