Angle information assisting skeleton-based actions recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Human Action Recognition

Background:

• Graph Convolutional Networks (GCNs) excel in skeleton-based action recognition but falter with complex actions due to lack of invariance and reliance on variable motion data.
• Existing methods struggle with perspective distortions and individual motion variability, hindering performance in tasks like figure skating recognition.

Purpose of the Study:

• To improve the robustness and accuracy of skeleton-based action recognition for complex human activities.
• To address the limitations of Graph Convolutional Networks (GCNs) in handling scale, rotation, and motion variability.
• To introduce novel methods for spatial and temporal feature extraction without modifying the GCN backbone.

Main Methods:

• A novel cosine stream was developed to enhance the robustness of spatial features against distortions.
• A Keyframe Sampling algorithm was introduced for efficient and effective temporal feature extraction, bypassing motion information.
• The proposed methods were integrated without requiring modifications to the existing GCN backbone architecture.

Main Results:

• Achieved a 2.6% improvement in Top-1 accuracy on the FSD-10 figure skating dataset compared to state-of-the-art methods.
• Demonstrated superior performance on FineGYM and NTU RGB+D datasets, validating the generalizability of the approach.
• The Keyframe Sampling algorithm effectively extracts temporal features, reducing reliance on potentially misleading motion data.

Conclusions:

• The proposed cosine stream and Keyframe Sampling significantly enhance skeleton-based action recognition for complex actions.
• The method offers improved accuracy and robustness, particularly for activities with high motion variability like figure skating.
• This approach provides a valuable advancement for human action recognition, offering better performance without architectural changes.