Related Experiment Video
Updated: Jun 27, 2026
09:41
Estimation of Contact Regions Between Hands and Objects During Human Multi-Digit Grasping
Published on: April 21, 2023
DHA-eGCN: Differential Hyperedge Attention-Enhanced Graph Convolution Network for Skeleton-Based Human Action
Oskar Ika Adi Nugroho1, Wen-Nung Lie1
1Department of Electrical Engineering and Advanced Institute of Manufacturing with High-Tech Innovations (AIM-HI), National Chung Cheng University, Chia-Yi 621, Taiwan.
Sensors (Basel, Switzerland)
|June 26, 2026
Summary
This study introduces a new model for human action recognition using skeleton data. The Differential Hyperedge Attention-enhanced GCN (DHA-eGCN) improves accuracy by better understanding body structure and motion, outperforming existing methods.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Machine Learning
Background:
- Skeleton-based human action recognition (HAR) faces challenges in preserving body structure and capturing long-range dependencies, especially with noisy data.
- Traditional Graph Convolutional Networks (GCNs) excel at local information but struggle with global interactions.
- Attention mechanisms capture global patterns but can be misled by spurious correlations without strong skeletal constraints.
Purpose of the Study:
- To develop a hybrid architecture that effectively integrates local structural information with global spatiotemporal dependencies for HAR.
- To enhance the robustness and accuracy of skeleton-based HAR models, particularly under challenging conditions like incomplete or noisy joint data.
Main Methods:
- Proposes Differential Hyperedge Attention-enhanced GCN (DHA-eGCN), a hybrid model combining differential hyperedge attention with multi-scale temporal convolution.
- DHA injects skeletal structure using hop-distance relative positional encoding and hyperedge context tokens, employing differential attention to reduce noise.
- Incorporates an explicit GCN branch for spatial grounding and an ensemble strategy for improved prediction robustness.
Main Results:
- DHA-eGCN achieves state-of-the-art or competitive performance across multiple benchmark datasets (NTU RGB+D 60/120, Northwestern-UCLA).
- Achieved high accuracy rates: 93.7% (NTU RGB+D 60 X-Sub), 97.0% (NTU RGB+D 60 X-View), 90.9% (NTU RGB+D 120 X-Sub), 91.9% (NTU RGB+D 120 X-Set), and 97.6% (Northwestern-UCLA).
- Demonstrates consistent outperformance compared to strong graph-based, transformer-based, and hybrid methods.
Conclusions:
- DHA-eGCN offers a superior approach to skeleton-based HAR by effectively fusing structural information and spatiotemporal dependencies.
- The hybrid architecture and proposed attention mechanism significantly enhance model accuracy and robustness.
- The model represents a significant advancement in recognizing human actions from skeletal sequences.
