Action-Attending Graphic Neural Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human action recognition is vital for computer vision applications.
• Skeleton-based analysis offers a robust method for understanding human motion.
• Existing methods may not optimally utilize all skeletal joint information.

Purpose of the Study:

• To propose a fully end-to-end action-attending graphic neural network (A2GNN) for skeleton-based action recognition.
• To develop a method that adaptively identifies and weights salient skeletal joints for improved action recognition.
• To enhance the intelligibility and robustness of human action recognition models.

Main Methods:

• Representing skeletons as undirected attribute graphs.
• Applying local spectral graph filtering for feature extraction, analogous to image convolution.
• Introducing an action-attending layer to dynamically weight skeletal joints based on action relevance.
• Encoding temporal motion variations using a recurrent gated network integrated with graph features.

Main Results:

• The A2GNN achieved state-of-the-art performance on multiple benchmark datasets, including the large-scale NTU RGB+D dataset.
• The action-attending mechanism effectively identified and utilized informative skeletal joints.
• The integrated approach demonstrated robust and intelligible human action recognition.

Conclusions:

• The proposed A2GNN offers a powerful and effective framework for skeleton-based action recognition.
• Adaptive joint weighting and spectral graph filtering are key components for high performance.
• The method advances the field by jointly training spectral filtering, attention mechanisms, and temporal encoding.