LI-AGCN: A Lightweight Initialization-Enhanced Adaptive Graph Convolutional Network for Effective Skeleton-Based...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Graph Convolutional Networks (GCNs) are key in skeleton-based action recognition.
• Prior GCN models often neglect heuristic initialization's impact on spatial feature extraction, limiting performance.

Purpose of the Study:

• Introduce a lightweight initialization-enhanced adaptive graph convolutional network (LI-AGCN).
• Improve spatiotemporal feature extraction and computational efficiency in action recognition.

Main Methods:

• LI-AGCN uses three coordinate-based input branches (CIB) for dynamic graph structure adjustment.
• Incorporates a lightweight, multi-scale temporal module and a spatio-temporal-channel attention module.

Main Results:

• Achieved excellent performance on NTU RGB+D, NTU RGB+D 120, and UAV-Human datasets.
• Obtained 90.03% accuracy on NTU RGB+D cross-subject benchmark with only 0.18 million parameters.

Conclusions:

• The proposed LI-AGCN effectively captures spatiotemporal features with low complexity.
• Initialization enhancement is crucial for optimal GCN performance in action recognition.