Sensor Data Required for Automatic Recognition of Athletic Tasks Using Deep Neural Networks.

Area of Science:

• Biomechanics
• Sports Science
• Machine Learning

Background:

• Movement screens assess athlete movement quality via subjective visual observation.
• Current data-driven methods require manual data pre-processing and optical motion capture.
• Objective, automated scoring of athletic movements is needed for practical application.

Purpose of the Study:

• To apply deep learning for automatic identification of athletic movements in screens.
• To assess the feasibility of classifying movements using wearable sensor data.
• To determine the minimum sensor configuration for accurate movement classification.

Main Methods:

• Trained a deep neural network (CNN-RNN) on optical motion capture data from 417 athletes performing 13 movements.
• Generated simulated inertial measurement unit (IMU) data from optical motion capture.
• Tested network performance with varying numbers and placements of simulated IMU sensors.

Main Results:

• Classification accuracy reached 90.1% with optical data and 90.2% with full-body simulated IMU data.
• Achieved 85.9% accuracy using only three simulated sensors (torso and shanks).
• Fewer than three sensors or alternative placements (torso/upper arms) yielded poor accuracy.

Conclusions:

• Deep learning can accurately classify athletic movements using simulated wearable sensor data.
• A minimal sensor configuration (e.g., torso and shanks) is feasible for objective movement quality assessment.
• This approach facilitates the field implementation of automated, data-driven movement analysis.