Development of a Human Activity Recognition System for Ballet Tasks.

Area of Science:

• Sports Science
• Biomechanical Engineering
• Machine Learning Applications

Background:

• Quantifying dancer training volume is crucial for understanding the relationship between pain and physical exertion.
• Existing methods lack the specificity to measure training volume based on distinct movement activities.
• Machine learning with wearable sensors has proven effective for human activity recognition in other sports.

Purpose of the Study:

• To develop a human activity recognition system utilizing wearable sensor data to precisely identify key ballet movements.
• To assess the efficacy of machine learning in accurately recognizing specific ballet movements during training.
• To investigate how sensor location and quantity impact the accuracy of movement identification.

Main Methods:

• Convolutional neural networks were employed to create classification models.
• Models were developed for various sensor combinations (six down to one) and with/without transition movements.
• The system used data from wearable sensors placed on dancers.

Main Results:

• A model utilizing all six sensors without transition data achieved 97.8% accuracy in identifying ballet movements.
• Classification accuracy decreased with the inclusion of transition movements, fewer sensors, and different sensor combinations (83.0% and 75.1% for subsequent levels).
• The developed models demonstrated robustness in recognizing jumping and leg-lifting actions in real-world dance scenarios.

Conclusions:

• The study successfully developed a robust system for identifying specific ballet movements using wearable sensor data.
• This novel system enables accurate quantification of dancer training volume, facilitating research on pain-training volume relationships.
• The findings offer a proof of concept applicable to monitoring athlete training in other lower-limb dominant sports.