| Title |
Enhancing Human Action Recognition with Demographic Attribute Information in Metaverse Environments |
| Authors |
한석호(Seok-Ho Han) ; 김종성(Jong-Sung Kim) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.3.63 |
| Keywords |
Human action recognition; Demographic attribute information; Transformer; LSTM; GRU; 1-D CNN |
| Abstract |
This study proposes a method to enhance the recognition performance of Human Action Recognition (HAR) models by incorporating Demographic Attribute Information such as gender and age group in a metaverse environment. To achieve this, participants were grouped by gender and age, and action data of golf swings and bowling throws were collected using full-body XR devices. Each action sequence was labeled into detailed sub-actions for fine-grained analysis. Based on the collected data, LSTM, GRU, 1D-CNN and Transformer models were trained and compared according to whether Demographic Attribute Information was included in the input. Experimental results show that the models incorporating Demographic Attribute Information consistently outperformed those without it, achieving an average improvement of approximately 3.0% for golf swings and 1.5% for bowling throws. In particular, the Transformer model demonstrated the highest performance, showing improvements of 3.8% and 1.8% for each action, respectively. Furthermore, group-wise evaluation revealed performance improvements across all gender and age groups, while t-SNE visualization demonstrated clearer separability among feature clusters when Demographic Attribute Information was utilized. These findings confirm that integrating Demographic Attribute Information can improve both the overall recognition performance and the balance across user groups in HAR models, contributing to the development of more inclusive and user-friendly metaverse environments. |