3D Human Motion Prediction via Activity-driven Attention-MLP Association

Human motion prediction aims to forecast the future motion based on historical sequences of human poses. However, due to incomplete aggregation of the spatio-temporal information, the predicted poses based on existing methods often suffer from discontinuity and error accumulation. To tackle these problems, we design an activity-driven attention-MLP association method, which adopts discrete cosine transform(DCT) to encode temporal information and extracts spatial features by attention-MLP association layers. For better enhancing spatial feature representation capability of the global and local united graph attention, a channel excitation module is inserted into each layer. Meanwhile, the spatio-temporal MLP is associated with attention to further aggregate the features both in temporal and spatial dimension. Moreover, we introduce a novel activity-driven supervision of the optimization. Empirically, the proposed method outperforms other state-of-the-art algorithms on datasets human3.6m and CMU Mocap. Code is available at https://github.com/alanzhangv123/HMP\_am.