An Analysis of Synthetic Data for Improving Performance of Skeleton-Based Fall Down Detection Models

Skeleton-based human action recognition technology, based on a skeleton framework, is increasingly adopted in visual safety monitoring systems as it does not require exposure of personal identity information. Among various visual-based safety monitoring tasks, fall incidents can sometimes be fatal, emphasizing the need for accurately classifying human body activities and providing prompt assistance. While artificial intelligence has been applied to visual-based solutions for action recognition, accurately classifying actions remains challenging due to the lack of training data. Research has attempted to improve model performance using synthetic data, yet discussions on the relationship between the quality of skeleton data obtained from synthetic data and model performance have been limited. In this proposed study, we demonstrate how the quality of skeleton data used in fall detection model training affects the performance of fall detection. Therefore, it is expected that the results of this study will serve as valuable foundational material for improving the performance of skeleton-based fall detection models.