Versatile Graph Neural Networks Toward Intuitive Human Activity Understanding.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Human Activity Recognition

Background:

• Human activity classification and motion prediction are often studied separately.
• Existing skeleton-based methods struggle with view variations.
• Current graph operators have limitations in exploring multiscale relationships.

Purpose of the Study:

• To propose a versatile graph-based model (Vers-GNN) for simultaneous human activity classification and motion prediction.
• To integrate view adaptation into graph-based human activity analysis.
• To develop novel graph operators for enhanced relationship modeling.

Main Methods:

• A versatile graph-based model (Vers-GNN) is proposed.
• A skeleton representation self-regulated scheme with view adaptation is introduced.
• Novel graph operators are developed to model joint relationships and dynamics.
• A multitask and multiobjective self-supervised learning framework is implemented.

Main Results:

• Vers-GNN outperforms state-of-the-art methods on multiple datasets.
• Highest recognition accuracies achieved: NTU RGB + D (97.2%), UWA3D (88.7%), CMU (1000 ms: 1.13).
• The model effectively handles view variations and multiscale relationships.

Conclusions:

• Vers-GNN offers a unified approach to human activity analysis.
• The proposed methods significantly advance skeleton-based human action understanding.
• The model demonstrates superior performance and robustness in complex scenarios.