Intra-frame Skeleton Constraints Modeling and Grouping Strategy Based Multi-Scale Graph Convolution Network for 3D Human Motion Prediction

Attention-based feed-forward networks and graph convolution networks have recently shown great promise in 3D skeleton-based human motion prediction for their good performance in learning temporal and spatial relations. However, previous methods have two critical issues: first, spatial dependencies for distal joints in each independent frame are hard to learn; second, the basic architecture of graph convolution network ignores hierarchical structure and diverse motion patterns of different body parts. To address these issues, this paper proposes an intra-frame skeleton constraints modeling method and a Grouping based Multi-Scale Graph Convolution Network (GMS-GCN) model. The intra-frame skeleton constraints modeling method leverages self-attention mechanism and a designed adjacency matrix to model the skeleton constraints of distal joints in each independent frame. The GMS-GCN utilizes a grouping strategy to learn the dynamics of various body parts separately. Instead of mapping features in the same feature space, GMS-GCN extracts human body features in different dimensions by up-sample and down-sample GCN layers. Experiment results demonstrate that our method achieves an average MPJPE of 34.7mm for short-term prediction and 93.2mm for long-term prediction and both outperform the state-of-the-art approaches.