Human Detection in Thermal Images Using YOLOv8 for Search and Rescue Missions
Mostafa Rizk, I. Bayad
Abstract
The effectiveness of life-saving search and rescue missions relies on swiftly locating and supplying crucial information about injured or missing persons, which optimize resource allocation and operational efforts while reducing expenses. Our research work explores the advances in deep learning techniques for efficient and accurate human detection using thermal images targeting the application of search and rescue missions. This paper focuses on exploring YOLOv8, the latest version of the You Only Look Once (YOLO) object detection model, to detect humans in diverse scenarios. A novel dataset of 17,148 grayscale thermal images with 90,882 annotations of humans is constructed carefully to represent various conditions and scenarios. All available YOLOv8 variants of different sizes and architectures (from Nano to Extra Large) are trained and evaluated using this dataset. The evaluation of the trained models demonstrates the effectiveness of YOLOv8 models in human detection exhibiting a high average precision of 95%. Furthermore, the impact of model size on precision and accuracy is highlighted.