<p>Mining temporal and spatial correlations from observed data is a crucial challenge in human motion prediction. Traditional methods mainly rely on artificially designed features, which cannot accurately capture complex motion patterns and often ignore the reliability of motion. In recent years, deep learning methods have made significant progress in spatiotemporal representation, especially when dealing with human motion prediction tasks. Although temporal and spatial features have different characteristics, existing methods often fail to fully consider the temporal and spatial features. Correlation leads to inaccurate modeling of spatiotemporal information. In this paper, we propose a novel branch network based on the Transformer and augmented graph convolutional network (TAGCN). This method innovatively combines the global dependency modeling capability of the Transformer model with the local feature extraction capability of the graph convolution network. The Transformer model effectively captures long-distance dependencies in time series through a self-attention mechanism and is able to identify complex temporal patterns. At the same time, GCN models the spatial dependence between joints through a graph structure and can effectively extract local spatial features. Our method significantly improves the expressive ability of motion information by learning temporal and spatial features simultaneously, making it richer and more representative. In addition, we design an improved loss function to further improve the accuracy of spatiotemporal prediction. Experimental results show that our method achieves state-of-the-art prediction performance on Human3.6M, CMU Mocap and 3DPW datasets, significantly outperforming existing mainstream methods.</p>