Golf Swing Segmentation from a Single IMU Using Machine Learning.

Area of Science:

• Sports Science
• Biomechanics
• Wearable Technology

Background:

• Golf swing analysis relies on accurate segmentation of movement phases.
• Inertial Measurement Units (IMUs) are key for wearable-based motion capture.
• Machine learning has not been previously applied to segment golf swing phases.

Purpose of the Study:

• To propose and verify machine learning models for segmenting golf swings.
• To divide a full golf swing into five distinct phases using IMU data.
• To assess the accuracy and efficiency of ML-based segmentation.

Main Methods:

• Developed bidirectional long short-term memory (LSTM) and convolutional neural network (CNN) models.
• Utilized data from IMUs attached to the head, wrist, and waist of golfers.
• Trained and verified models using data from 9 professional and 11 skilled male golfers.
• Employed leave-one-out cross-validation for robust accuracy assessment.

Main Results:

• Achieved average segmentation errors of 5-92 ms across different IMU placements.
• Demonstrated accuracy comparable to heuristic methods for golf swing phase segmentation.
• Successfully segmented all swing phases using only acceleration data, reducing power consumption.
• Showcased the potential for flexible IMU placement in motion analysis.

Conclusions:

• Machine learning models, specifically bidirectional LSTMs and CNNs, provide accurate golf swing segmentation.
• The proposed methods are efficient, requiring only acceleration data and offering low power consumption.
• These findings enable broader applications of wearable motion analysis in various environments.