Human motion prediction with gated recurrent unit model of multi-dimensional input

The issue of human motion prediction aimed to predict sequences of joint positions or joint rotations of human skeleton has recently grown in importance. The Recurrent Neural Network is widely applied on the sequence prediction problems which has been proved effective. However it is difficult to train the model with human skeleton data of multi-dimensional as input, which would do naive forecasting to produce motionless sequence. To address the problem, it is a consensus that additional information will help to improve the accuracy, thus the angular velocities are extracted from the joint rotations as the input to enhance the prediction. Further more, this work adopts proper strategies on the basis of a stacked Gated Recurrent Unit network and verify them on the human motion prediction task. The experimental results show that our network outperforms the state-of-art on the short-term prediction task, and generates plausible action sequences in a relatively long period of time.