Person Detection in Thermal Images: A Comparative Analysis of YOLOv8 and YOLOv9 Models
Valentinas Breivė, Tomyslav Sledević
Abstract
Person detection in thermal imagery is crucial for surveillance and monitoring applications. We assess the performance of YOLOv8 and YOLOv9 models using a new thermal image dataset. Our study reveals that both models achieve high precision, with YOLOv8 showing a superior training time and inference speed compared to YOLOv9. Specifically, YOLOv8 models achieve precision rates of 89% to 90 %, outperforming YOLOv9 with precision ranging from 87% to 88%. Moreover, YOLOv8 models demonstrate faster training times and lower in-ference processing times, making them more suitable for real-time applications. Despite challenges such as false positives and false negatives, our findings provide valuable insights into improving the accuracy and efficiency of thermal-based person detection, thereby enhancing surveillance and monitoring systems.