Human Skeleton Motion Prediction Using Graph Convolution Optimized GRU Network

Analysis on the human motion can reveal patterns proven very useful in human-machine interactions, medical applications and ambient assisted living. One such analysis is human motion prediction consisting of predicting human pose in a set of time instances contained in constrained time window of up to 1 to 2 seconds. This prediction is done by analyzing previous motion, i.e. set of previous poses, within a selected time window. In this paper we propose to predict human motion using Gated Recurrent Unit (GRU) network, a variant of Recurrent Neural Network. The prediction is based on human skeleton model and joints position change in time. We further optimize the GRU by substituting the weighting of inputs and recurrent outputs with convolution utilizing the graph structure of the human skeleton. We validate our proposed network by testing it on publically available dataset and providing state of the art results in comparison with other popular methods.