Towards Human Activity Recognition in Smart Environments through Thermal Imaging

Abstract

This paper presents the results of keypoint detection achieved through training YOLOv8 pose models on a thermal person dataset, aiming towards human activity recognition in smart environments. The YOLOv8 pose models, ranging from Nano to Extra Large configurations, were evaluated based on their parameters, Object Keypoint Similarity (OKS) percentage, and processing time per frame. Results indicate that even the smallest model, Nano, achieved a satisfactory OKS of 95.6% with a processing time of 10.4 milliseconds per frame. As model complexity increased, OKS remained high, with a marginal decrease in performance. The findings suggest promising prospects for utilizing thermal imaging in human activity recognition systems, with the potential for real-time deployment in smart environments.