GRU-Based Multi-Modal Human Activity Recognition
Adarsh Muralidharan, Sazia Mahfuz · Procedia Computer Science · 2025
Human Activity Recognition (HAR) is an important challenge faced in the healthcare industry, particularly in nursing homes for the elderly and vulnerable patients. This study explores a deep learning-based approach for HAR using multi-modal data, specifically skeletal and inertial data. We employ a Gated Recurrent Unit (GRU)-based architecture to classify human activities, using data preprocessing techniques and feature engineering. The proposed model integrates features from both skeletal and inertial sensors to predict activity labels. The final model achieves an accuracy score of 96.30% outperforming previous random forest model which has an accuracy of 90%. The findings highlight the potential of using GRU networks for real-time HAR systems and provide insights into improving classification in multi-modal activity recognition.