A Deep Learning Approach to Automatically Classify Ice Hockey Shooting Actions Using Acceleration Signals

Area of Science:

• Sports Science
• Machine Learning
• Wearable Technology

Background:

• Objective performance evaluation in ice hockey is crucial for player development.
• Current methods often rely on subjective coach assessments.
• Wearable sensors and machine learning offer potential for automated activity detection.

Purpose of the Study:

• To evaluate a deep learning model's ability to recognize ice hockey stick actions from inertial measurement unit (IMU) sensor data.
• To compare the effectiveness of a comprehensive sensor setup versus a minimal, hands-only configuration.

Main Methods:

• A fully connected convolutional neural network (CNN) was developed to classify seven ice hockey actions.
• Data were collected from 43 elite players using two sensor configurations: 17 sensors (all-sensor) and 2 sensors (hands-only).
• Standard machine learning practices including train/test splits, cross-validation, and repeated random splits were employed.

Main Results:

• The CNN model demonstrated high classification accuracy for both sensor setups.
• The all-sensor model achieved an average F1 score of 95.0 ± 3.0%.
• The hands-only model achieved a comparable average F1 score of 93.5 ± 1.6%.

Conclusions:

• Convolutional neural networks are effective for automatic ice hockey shooting action classification.
• Minimal sensor configurations, such as integrated gloves, are feasible for practical, real-world applications.
• This technology can provide objective performance metrics to enhance coaching and player training.