Attention Mechanism for Human Motion Prediction

Human motion prediction aims to forecast the most likely future frames of motion conditioned on a given sequence of frames. Because of its importance to many applications especially robotics, it has received a lot of interest and has become an active area of research. Since humans are very flexible in nature, human motion prediction is very challenging Recently, deep learning methods have been dominant in many tasks due to their successful results. Particularly, Recurrent Neural Networks (RNNs) have shown excellent performance on human motion prediction task and other tasks that depend on sequential data, where preserving the order of the sequence items is crucial. The well-known Sequence-to-Sequence (Seq2Seq) architectures have been used for sequence learning where two RNNs namely the encoder and the decoder work cooperatively to transform one sequence to another. In the context of neural machine translation, the use of attention decoders yields state-of-the-art results. This work employs a simple but efficient Seq2Seq model with attention decoder. Both encoder and decoder are trained jointly to predict 15 different categories of human motion. Our experiments have shown that the attention decoder clearly outperforms earlier methods and achieves state-of-the-art results in the short-term (< 500ms) motion prediction task. Contrary to earlier methods that show progressive deterioration as the time of prediction increases, our model shows high quality long-term (> 500ms) motion prediction which stays as high even after 1000ms of prediction.