Integrating Perspectives: A Unified Deep Learning Framework for Human Activity Recognition

Human Activity Recognition (HAR) is vital in healthcare, smart environments, and human-computer interaction. Traditional methods struggle with real-world data complexity. This research presents a new framework combining Graph Convolutional Networks (GCNs) and Recurrent Neural Networks (RNNs) with a dual attention mechanism to boost HAR performance. GCNs model spatial relationships in graph-structured data, capturing dependencies between body joints or sensors. Our approach constructs a graph with nodes for sensors or joints and edges for their spatial connections. GCNs extract spatial features, which are then input into an RNN to model the temporal dynamics of activities. A dual attention mechanism is introduced to enhance the model's focus on relevant information. Spatial attention highlights important nodes, while temporal attention focuses on significant time steps. Experiments on benchmark HAR datasets show that our GCN-RNN dual attention framework outperforms existing methods, improving accuracy and robustness, especially for complex activities. This research demonstrates the potential of combining graph-based spatial modelling with temporal sequence learning and dual attention mechanisms for advanced HAR applications.