Fusion learning-based recurrent neural network for human motion prediction

Human motion prediction is an important research frontier, which is a key supporting technology in the fields of human-robot collaboration, automatic driving, etc. As is well known, long-term motion prediction is one most challenging direction. This paper mainly focuses on how to eliminate cumulative errors to overcome the fossilization of long-term motion sequences and aims to improve the reliability of prediction results. This paper proposed an algorithm named "fusion loss learning network," which is based on gated recurrent unit, to solve the above-mentioned problem. A fusion training method was established by combining the sampling of each step of the GRU unit with true value and output value of each previous step, which helped recover from the errors in the long-term prediction sequences. This method achieved promising results on the Human 3.6 M dataset. The results show that the proposed method could significantly improve the performance of long-term human motion prediction, and the total prediction error is reduced by 7.25% on average.