Dynamic Edge Convolutional Neural Network for Skeleton-Based Human Action Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human Action Recognition (HAR) is crucial for applications like IoT, healthcare, and autonomous driving.
• Existing HAR methods vary in data modalities, feature design, and network configurations.
• There is a need for more efficient and accurate HAR models, especially those utilizing skeleton data.

Purpose of the Study:

• To develop a novel deep learning model for enhanced human action recognition from skeleton sequences.
• To integrate spatial-temporal attention with edge convolution for improved feature extraction.
• To evaluate the model's effectiveness and robustness on benchmark datasets.

Main Methods:

• A new deep learning model combining criss-cross attention and edge convolution was designed.
• Attention mechanisms were applied in spatial and temporal dimensions to capture frame relationships.
• Edge convolutional layers explored geometric relationships among human body joints.
• The model dynamically updated graph re-computation based on k-nearest joints for local and global information learning.

Main Results:

• The proposed model achieved 99.53% accuracy on the UTD-MHAD dataset.
• The model achieved 95.64% accuracy on the MSR-Action3D dataset.
• Performance surpassed existing state-of-the-art methods on both benchmark datasets.

Conclusions:

• The novel deep learning model effectively extracts discriminative features from skeleton sequences for HAR.
• The integration of criss-cross attention and edge convolution offers superior performance.
• The model demonstrates high accuracy and robustness, outperforming current HAR techniques.