AI-Driven Action Recognition and Pose Estimation for Traditional Chinese Performing Arts in Digital Cultural Communication

The proposed AI framework for Traditional Chinese dance pose recognition uses multi-view alignment and attention-driven temporal modeling, capturing expressive motion semantics. It preprocesses, extracts features, and classifies poses to preserve cultural heritage, outperforming existing approaches in accuracy. However, existing dance recognition systems often lack robust cross-view adaptability and effective long-range temporal modeling, limiting their ability to capture expressive motion dynamics in traditional dance. This reveals a research gap in developing a culturally adaptive and temporally attentive recognition framework. Skeletal pose sequences are normalized and segmented, with ResNet extracting discriminative spatial features. These features are modeled using BiLSTM with self-attention to capture long-range past and future temporal dependencies, enabling robust recognition of culturally expressive dance motions. Generative adversarial training using the Archive of Motion Capture as Surface Shapes (AMASS) dataset and spatial feature extraction through ResNet enhance motion realism and generalization. Evaluated across multiple dance categories, the model achieves 96% accuracy, 94.90% precision, 96.17% recall, and 95.53% F1-score, demonstrating robust classification performance. The framework supports digital preservation of Traditional Chinese dance and enables applications in interactive performances, cultural heritage initiatives, and AI-driven dance research.