Refined Feature-Enhanced Spatial-Temporal Graph Convolutional Network for Labanotation-Based Action Recognition

Labanotation is a robust system for recording and preserving dance movements. Despite advancements in deep learning techniques for recognizing and classifying Laban symbols for single Laban elemental actions, current methodologies often fail to capture the spatial-temporal dynamics of skeletal movements, leading to decreased accuracy with ambiguous actions. To address this challenge, we propose a Refined Feature-Enhanced Spatial-Temporal Graph Convolutional Network (RF-STGCN) for Labanotation-based action recognition. Our network utilizes a feature refinement module to precisely extract spatial-temporal information from hidden features, effectively identifying and reclassifying ambiguous actions. Additionally, an improved channel-attention mechanism isolates and emphasizes critical skeletal features, enhancing the model's focus and improving action recognition accuracy. Extensive experiments on the Laban-16 and Laban-48 datasets demonstrate that our method significantly outperforms existing algorithms, achieving improvements in Laban elemental action recognition accuracy.