PVRED: A Position-Velocity Recurrent Encoder-Decoder for Human Motion Prediction

Human motion prediction, which aims to predict future human poses given past poses, has recently seen increased interest. Many recent approaches are based on Recurrent Neural Networks (RNN) model with exponential maps. These neglect the pose velocity as well temporal relation of different and tend converge mean or fail generate natural-looking poses. We therefore propose a novel Position-Velocity Encoder-Decoder (PVRED) for makes full use velocities positional information. A position embedding method is presented RNN (PVRNN) proposed. also emphasize benefits quaternion parameterization design trainable Quaternion Transformation (QT) layer, combined robust loss function during training. provide quantitative results both short-term prediction in 0.5 seconds long-term 1 seconds. Experiments several benchmarks show that our approach considerably outperforms state-of-the-art methods. In addition, qualitative visualizations 4 could human-like meaningful very long time horizons. Code publicly available GitHub: https://github.com/hongsong-wang/PVRNN.