RNN-based Human Motion Prediction via Differential Sequence Representation

Human motion prediction-to predict the future motion sequences given the past sequences using 3D skeleton data-is a classical research field in computer vision. With the deep learning method has been successful in several computer vision tasks, recent attempts on this field suggested to use RNN-based methods to model the movement of human in the temporal domain. In this kind of task, as the error will accumulate with time steps, the stability of human body movement data will affect the accuracy of prediction. In this paper, aiming at the stability of human body data, we focus on the motion of human body can be represented by velocity and acceleration instead of posture to obtain more stable sequence to be predicted. We propose a framework based on the velocity and acceleration and obtain better results.