April-GCN: Adjacency Position-velocity Relationship Interaction Learning GCN for Human motion prediction

Abstract

The Human Motion Prediction (HMP) task attempts to model human kinematics, which requires considering both the physical connections between joints and the continuity of the joints’ trajectories. To handle this complex task, some recent works have employed Graph Convolutional Networks (GCN) in learning dynamic relations among joints. However, HMP task is essentially a time series task, to model the temporal and spatial features of motion, some works extended the common adjacency matrix formulation to create a large-scale mixed adjacency matrix which introduces redundant connection relationships in single-path GCN. Other works proposed dual-path GCNs and multi-path GCNs to make feature representations more informative, while ignoring the significant differences of features. To tackle these issues, we propose a novel adjacency matrix interaction learning framework called Adjacency Position-velocity Relationship Interaction Learning GCN (April-GCN) which alleviates the differences between dual-path features and improves the feature fusion results. The proposed network utilizes a dual-path GCN structure to exploit the adjacency relationship between physical connections and movement amplitude by interactively learning the GCN adjacency matrices of the two paths. The network excavated closely aligned position-velocity correlations to obtain a more comprehensive and richer representation of joint connections. On three datasets, H3.6M, CMU, and 3DPW, it achieves state-of-the-art performance for both short-term and long-term predictions.