Enhancing keypoint detection in thermal images: Optimizing loss function and real-time processing with YOLOv8n-pose

Abstract

This article presents an investigation of the precision of keypoint detection in single-person thermal images using a novel dataset collected with a thermal camera. The dataset contains 1000 images, with keypoints annotated for pose detection. Different loss functions, including the default OKS loss, L1 loss, and L2 loss, were applied to train the YOLOv8n-pose model. The experimental results revealed 1.3% improvements in object keypoint similarity when using L1 loss or a combination of LOKS and L1 loss functions. The study presents the frame processing speed of the YOLOv8n-pose model on an RTX 2060 GPU, demonstrating efficient performance suitable for real-time applications at 64 fps 640×512 px input. These findings highlight the importance of optimizing loss functions for improved keypoint detection accuracy in thermal images, with implications for various applications such as surveillance and recognition of human behavior.