An Enhanced Approach for Graph Neural Network Human Activity Recognition Using Deep Learning Technique

Abstract

Human activity recognition (HAR) has emerged as an essential area of study in video analysis. It has evolved substantial attention as a result of its applications in a diverse range of disciplines, such as healthcare, surveillance, and human-computer interaction. But achieving real-time robustness remains a challenge due to environmental factors such as occlusions and shifts in motion dynamics. This study proposes the Temporal Graph Neural Network (TGNN) for the recognition of human activity in order to integrate multimodal feature extraction and temporal graph adaptive fusion techniques. For the purpose of structured representation learning, the temporal graph neural network is implemented to encapsulate the spatial and temporal dependencies of human activities, Contrastive learning is used to which refines feature discrimination, thereby improving generalization across a variety of conditions. An EfficientNetB0-based whale optimization algorithm is further integrated for optimal hyperparameter tuning, which reduces the computational overhead Extensive experiments were carried out on HMDB51 datasets with an accuracy of 90.06%. This method is superior to existing HAR baseline models. It offers a real-time and scalable HAR solution that is appropriate for deployment in various applications.