IMU Airtime Detection in Snowboard Halfpipe: U-Net Deep Learning Approach Outperforms Traditional Threshold...

Area of Science:

• Sports Science
• Biomechanical Engineering
• Machine Learning Applications

Background:

• Airtime is critical for snowboard halfpipe trick difficulty and scoring.
• Manual video analysis for detecting take-off and landing is time-consuming.
• Accurate event detection is needed for real-time feedback and biomechanical analysis.

Purpose of the Study:

• To enhance the detection of take-off and landing events in snowboard halfpipe using inertial measurement unit (IMU) data and machine learning.
• To establish new performance benchmarks for event detection in this sport.
• To facilitate real-time feedback and biomechanical analysis for athletes and coaches.

Main Methods:

• Collected synchronized IMU (lateral lower legs) and video data from elite snowboarders.
• Manually labeled IMU data for training and validation of machine learning models.
• Utilized a 1D U-Net convolutional neural network (CNN) for binary segmentation of events.

Main Results:

• The 1D U-Net CNN significantly outperformed traditional threshold approaches in detecting take-off and landing events.
• Achieved substantial reductions in mean Hausdorff distance for unseen runs (80.34% with single IMU, 83.37% with dual IMUs).
• Demonstrated strong performance in Zero-Shot (67.58% improvement) and Few-Shot (78.68% improvement) segmentation scenarios, with highly accurate time deviations for takeoffs, landings, and airtime.

Conclusions:

• The developed U-Net model provides a highly accurate and efficient method for detecting critical events in snowboard halfpipe.
• This advancement enables precise real-time feedback and detailed biomechanical analysis, crucial for improving athlete performance.
• The study establishes new benchmarks for IMU-based event detection in high-performance sports.