A Survey on 3D Skeleton-Based Action Recognition Using Learning Method.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Three-dimensional skeleton-based action recognition (3D SAR) leverages skeleton data advantages.
• Existing surveys predominantly focus on RGB data, with limited reviews on skeleton data.
• Deep learning methods are widely applied, yet a comprehensive review from an architectural perspective is lacking.

Purpose of the Study:

• To underscore the importance of action recognition and 3D skeleton data.
• To provide a comprehensive review of 3D SAR techniques using deep learning architectures.
• To address the gap in literature regarding deep learning-based 3D SAR reviews.

Main Methods:

• Review of mainstream 3D SAR techniques based on four fundamental deep architectures: Recurrent Neural Networks (RNNs), Convolutional Neural Networks (CNNs), Graph Convolutional Networks (GCNs), and Transformers.
• Data-driven presentation and detailed discussion of methods within each architecture.
• Analysis of leading 3D skeleton datasets, including NTU-RGB+D and NTU-RGB+D 120.

Main Results:

• Identification and categorization of 3D SAR methods based on core deep learning architectures.
• Discussion of the strengths and applications of different deep learning models for skeleton-based action recognition.
• Overview of state-of-the-art performance on benchmark datasets like NTU-RGB+D and NTU-RGB+D 120.

Conclusions:

• This survey is the first to comprehensively discuss deep learning-based action recognition using 3D skeleton data.
• It provides a foundational understanding of deep learning architectures for 3D SAR.
• Highlights key datasets and algorithms, offering insights for future research.