Human Motion Modeling Based on Single Action Context

Motion modeling is applied in pattern recognition, image and video processing, and human-computer interaction, including motion analysis, prediction and synthesis. With the development of deep learning method, recent works focus on recurrent neural network (RNN). According to experimental results of recent works, RNN is effective for short-term motion analysis. However, RNN has distortion problem in long-term motion analysis and prediction. We find that long-term motion is always based on a explicit action context, such as walking and sitting. RNN could hardly analyze action context through short-term historical sequence, while a longer historical sequence difficulties in network fitting. We propose a encoder-decoder model with a context unit, to keep the context information of each actions. Our model obtains state-of-the-art performance on long-term human motion prediction. Furthermore, our model can synthesis composite motion by changing action context.