Wearable sensing for badminton stroke recognition with one-dimensional convolutional neural network.

Area of Science:

• Sports Science and Biomechanics
• Wearable Technology
• Machine Learning in Sports

Background:

• Traditional badminton motion analysis methods (video, optical) are limited by fixed setups and extensive postprocessing.
• Inertial Measurement Units (IMUs) offer a portable and lightweight solution for real-time motion monitoring.
• Improving badminton player performance necessitates accurate and accessible technique analysis tools.

Purpose of the Study:

• To develop a wearable sensing network using IMUs for capturing badminton player wrist motions.
• To implement a 1D-CNN model for classifying badminton stroke actions and movement trajectories.
• To evaluate the accuracy and effectiveness of the proposed system compared to traditional methods.

Main Methods:

• A network of two IMUs was utilized to collect wrist motion data during badminton strokes.
• A 1D-CNN model, featuring convolutional layers, batch normalization, ReLU activation, max-pooling, and a Softmax classifier, was designed.
• Data from six national-level athletes were used for training and validation via stratified 5-fold cross-validation.

Main Results:

• The 1D-CNN model achieved high classification accuracies: 97.16% for six stroke actions and 86.07% for fifteen movement trajectories.
• The proposed IMU-based system significantly outperformed traditional machine learning algorithms like KNN, SVM, and DTA.
• Feature differences were visualized using heatmaps and t-SNE, providing intuitive insights into motion patterns.

Conclusions:

• A lightweight system combining two IMUs and a 1D-CNN provides accurate badminton motion monitoring.
• This approach offers a practical and effective solution for sports training and skill enhancement in badminton.
• The developed system contributes to advancing sports analytics through wearable sensor technology and deep learning.