Human Activity Recognition Using Graph Structures and Deep Neural Networks

Human activity recognition (HAR) systems are essential in healthcare, surveillance, and sports analytics, enabling automated movement analysis. This research presents a novel HAR system combining graph structures with deep neural networks to capture both spatial and temporal patterns in activities. While CNN-based models excel at spatial feature extraction, they struggle with temporal dynamics, limiting their ability to classify complex actions. To address this, we applied the Firefly Optimization Algorithm to fine-tune the hyperparameters of both the graph-based model and a CNN baseline for comparison. The optimized graph-based system, evaluated on the UCF101 and Kinetics-400 datasets, achieved 88.9% accuracy with balanced precision, recall, and F1-scores, outperforming the baseline. It demonstrated robustness across diverse activities, including sports, household routines, and musical performances. This study highlights the potential of graph-based HAR systems for real-world applications, with future work focused on multi-modal data integration and improved handling of occlusions to enhance adaptability and performance.