Related Experiment Video
Updated: Sep 27, 2025
06:37
Author Spotlight: Addressing Technical and Subjective Challenges in Measuring Classroom Attention
Published on: December 15, 2023
4.2K
Wearable Sensors for Activity Recognition in Ultimate Frisbee Using Convolutional Neural Networks and Transfer
Johannes Link1, Timur Perst1, Maike Stoeve1
1Machine Learning and Data Analytics Lab, Department Artificial Intelligence in Biomedical Engineering, Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), 91052 Erlangen, Germany.
Sensors (Basel, Switzerland)
|April 12, 2022
Summary
This study explores using Convolutional Neural Networks (CNNs) and transfer learning for human activity recognition (HAR) in marginal sports like Ultimate Frisbee. Results show CNNs can classify throws effectively, with transfer learning aiding small datasets.
Area of Science:
- Sports Science
- Machine Learning
- Sensor Data Analysis
Background:
- Human Activity Recognition (HAR) typically focuses on common sports, neglecting marginal sports due to sparse and costly data acquisition.
- Existing HAR approaches are not directly applicable to marginal sports with limited datasets.
Purpose of the Study:
- Investigate the efficacy of Convolutional Neural Networks (CNNs) and transfer learning for classifying Ultimate Frisbee throws using inertial measurement unit (IMU) data.
- Address the challenge of data scarcity in marginal sports HAR.
Main Methods:
- Recorded and annotated IMU data for various Ultimate Frisbee throws.
- Developed and applied a CNN pipeline for automatic action detection and classification.
- Evaluated a transfer learning approach using a pre-existing beach volleyball dataset.
Main Results:
- The CNN pipeline achieved 66.6% accuracy for nine fine-grained throw classes and 89.9% for three basic throwing techniques.
- Transfer learning reduced training time but did not significantly improve overall classification accuracy.
- On reduced datasets without augmentation, transfer learning enhanced network generalization, improving accuracy and F1 score.
Conclusions:
- CNNs and transfer learning show promise for HAR in marginal sports, particularly with limited data.
- Transfer learning can improve model generalization and reduce training time, making HAR more feasible for underrepresented sports.
- This approach can enhance the tracking and analysis capabilities for marginal sports.
