TST-Attention: A Diffusion-Based Approach for Human Motion Prediction

Human motion prediction(HMP), a crucial problem in human-computer interaction, will significantly impact our lives in the future. Current research still fails to provide precisely enough and detailed predictions of future motion based on observed sequences. To address the issue, we propose an innovative method that sequentially adds spatial and time-step attention modules to the posture prediction task, which usually focused only on temporal dependencies. The model combines spatial-temporal dependencies and time-step information with posture sequences, aiming to enhance the accuracy and fidelity of stochastic predictions. Our model utilizes a diffusion process with a transformer backbone and incorporates a squeeze-and-excitation module to calibrate the influence of each time step in the posture sequence, enhancing prediction performance, Improving the model's adaptability to varying levels of channel feature importance. We evaluate our approach on the two benchmark datasets, demonstrating state-of-the-art performance. Notably, our model is lightweight, with a reduced number of parameters.