Deep Learning for Human Detection from Aerial Images in Search and Rescue Missions

This paper addresses the critical need for effective Search and Rescue (SAR) operations in the face of increasingly frequent natural disasters. Existing datasets and models often fail to account for the diverse lighting scenarios encountered in real-world SAR operations, leading to suboptimal performance. Utilizing SARD dataset, we augmented the dataset to include varied lighting conditions, enhancing the model's robustness. We fine-tuned and compared six YOLO model variations (v8n, v8s, v8m, v10n, v10s, v10m) to identify the most effective for human detection in aerial images. Our results indicate that the YOLOv10m model offers the best balance of accuracy, precision, recall, F1 score, and mAP, making it highly suitable for SAR missions. The model's adaptability to diverse lighting conditions underscores its practical value in real-world scenarios, ensuring efficient and reliable human identification and resource deployment during SAR operations.