Wearable-based Human Activity Recognition with spatial-temporal Graph Convolutional Transformer Network
Lu Ma, Xiaodong Yang · ZKG International · 2025
Human Activity Recognition(HAR) is an important topic in the field of wearable computing as it involves the study of Spatio-Temporal (ST) interactions.Nowadays, algorithms based on Graph Convolutional Networks (GCN) are widely used.From a graphical point of view, this approach is effective in capturing a large amount of spatial information, however, it ignores the connection between different sensor information at different times with the global information.To overcome this limitation, we propose a novel network, Graph Convolutional Transformer Network (GCTNet), combining Transformer and GCN blocks, and using a fully-connected (FC) graph to optimize the graph structure constructed by the GCN model.To verify the superiority of the model, we are using the UCI-HAR dataset to validate our model.Experimental results demonstrate that our network achieves a 4% higher accuracy com-pared to the transformer model and a 0.5% higher accuracy compared to the latest FC-STGNN model on the UCI-HAR dataset.Compared to soat methods such as DeepConvLSTM, our model improves by 1.5%.