EAN: Error Attenuation Network for Long-term Human Motion Prediction

Human motion prediction is an important problem in human-robot interaction, computer graphics, and autonomous driving. Currently, there are still some difficulties. Firstly, the error of reasoning about temporal relations accumulates over time. Secondly, the local optimal caused by unbalanced data. For example, leg walking is more common than knee bending or standing. Thirdly, the problem of mean pose is easy to occur in long sequence prediction. In this work, we propose a novel prediction method named Error Attenuation Network (EAN) by taking the Recursive Attenuation Mechanism into consideration with attention model. Firstly, an error attenuation wrapper for optimal function is introduced to alleviate the effect of error accumulation and mean pose. Secondly, the attention model is introduced to restrain incidental actions to balance motions, which aims to mitigate the impact of unbalanced data and mean pose. Experimental results demonstrate that our method predicts the future human motion more accurately, which outperforms the related state-of-the-art approaches on long-term prediction in most cases while having a comparable performance on short-term prediction.