Hypergraph Neural Network for Skeleton-Based Action Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Human Action Recognition

Background:

• Skeleton-based human action recognition is a key area in computer vision.
• Graph Convolutional Networks (GCNs) are effective but focus only on local joint connections.
• Existing methods overlook non-physical, high-order dependencies among human body joints.

Purpose of the Study:

• To propose a novel Hypergraph Neural Network (Hyper-GNN) for skeleton-based action recognition.
• To capture both spatial-temporal information and high-order dependencies among joints.
• To enhance action recognition accuracy by addressing limitations of existing GCN methods.

Main Methods:

• Developed a Hyper-GNN to model skeletons as hypergraphs, capturing non-physical joint relationships.
• Designed hyperedge constructions for extracting local and global structure information, mitigating noise.
• Incorporated a hypergraph attention mechanism and an improved residual module for discriminative features.
• Utilized a three-stream Hyper-GNN fusion architecture for comprehensive action recognition.

Main Results:

• The proposed Hyper-GNN method demonstrated superior performance on two benchmark datasets.
• Achieved state-of-the-art results compared to existing skeleton-based action recognition techniques.
• Successfully captured complex, high-order dependencies beyond simple physical connections.

Conclusions:

• The Hyper-GNN framework offers a powerful approach for skeleton-based human action recognition.
• Modeling non-physical dependencies via hypergraphs significantly improves recognition accuracy.
• The method provides a robust solution for complex human action understanding in computer vision.