An explainable real time sensor graph transformer for dance recognition.

Area of Science:

• Computer Science
• Robotics
• Human-Computer Interaction

Background:

• Automatic dance motion interpretation from sensor data is challenging due to temporal fidelity, orientation invariance, and real-time constraints.
• Existing methods struggle with diverse choreography, sensor placement, and performer kinematics, offering limited explainability.

Purpose of the Study:

• To introduce an explainable, sensor-driven dance recognition framework for real-time, edge-class deployment.
• To enhance accuracy and provide verifiable reasoning for dance motion analysis.

Main Methods:

• Developed an Adaptive Sensor Normalisation module with quaternion-based orientation correction and Kalman refinement.
• Implemented a Multi-Scale Motion Feature Extractor using tempo-conditioned dilation for rhythmic structure analysis.
• Utilized a Spatio-Temporal Graph Attention Core with graph convolutions and dual attention for sensor saliency and temporal concentration.

Main Results:

• Achieved up to 94.2% classification accuracy and 92.8% movement-quality estimation on inertial datasets.
• Demonstrated sub-8.5 millisecond per-frame latency, confirming stability under tempo variation, sensor drift, and channel loss.
• Optimized for edge-class execution using kernel-level compression and causal attention windows.

Conclusions:

• The proposed framework offers robust and explainable dance recognition, overcoming limitations of existing approaches.
• The system's real-time performance and accuracy make it suitable for constrained hardware and dynamic environments.
• Explainable decision-making provides insights into class-specific motion determinants, enhancing model transparency.