Enriching Intention of Human Motion Prediction
Junyi Wang, Xinyu Su · 2020
Human motion prediction involves visualization, spatial-temporal data mining, which is a classical problem in artificial intelligence. Recent works focus on recurrent neural network (RNN). However, we find that previous methods could only analyse the inertia motion of body joint nodes. Inertia motion means that joints would maintain the motion state while satisfying environmental factors, such as body structure and gravity. Short-term human motion is similar to inertia motion, but long-term human motion is activity with a clear intention. The intention can be distinguished by action types, such as walking and eating. We propose a new sequence-to-sequence RNN model and gate auto encoder (GAE), to analyze the intention information for different actions. The experimental result shows that our model achieves state-of-the-art performance in long-term motion prediction. Furthermore, our model can generate composite motion by injecting subjective intention.