Human Motion prediction based on attention mechanism

Abstract

Human motion prediction, although in the field of human-computer interaction, personnel tracking, automatic driving and other fields have very important significance. However, human motion prediction is affected by uncertainties such as motion speed and amplitude, which results in the predicted first frame is discontinuous and the time for accurate prediction is short. This paper proposes a method that combines sequence-to-sequence (seq2seq) structure and Attention mechanisms to improve the problems of current methods. We refer to the proposed structure as the At-seq2seq model, which is a sequence-to-sequence model based on GRU (Gated Recurrent Unit). We added an attention mechanism in the decoder part of the seq2seq model to further encode the output of the encoder into a vector sequence containing multiple subsets so that the decoder selects the most relevant part of the sequence for decoding prediction. The At-seq2seq model has been validated on the human3.6 m dataset. The experimental results show that the proposed model can not only improve the error of short-term motion prediction but also significantly increase the time of accurate prediction.