Automated detection of quiet eye durations in archery using electrooculography and comparative deep learning models

Area of Science:

• Sports Science
• Biomedical Engineering
• Machine Learning

Background:

• Quiet Eye (QE) is crucial for precision in sports, but traditional detection is subjective and time-consuming.
• Electrooculography (EOG) offers a potential objective measure for gaze tracking.
• Automating QE detection can enhance sports performance analysis and training.

Purpose of the Study:

• To develop and evaluate deep learning models for automated Quiet Eye detection using EOG signals.
• To compare the performance of various deep learning architectures against a traditional model.
• To establish an objective and scalable method for analyzing QE in archery.

Main Methods:

• Collected EOG data from 10 archers during shooting sessions.
• Preprocessed EOG signals using wavelet transform and Butterworth bandpass filtering.
• Implemented and compared Support Vector Machine (SVM), CNN+LSTM, CNN+GRU, Transformer, UNet, and 1D CNN models for QE detection.

Main Results:

• The CNN+LSTM model achieved the highest accuracy (95%), followed by CNN+GRU (93%).
• Deep learning models significantly outperformed the traditional SVM model.
• CNN+LSTM and CNN+GRU demonstrated superior ability in capturing spatio-temporal EOG signal dependencies.

Conclusions:

• Deep learning provides an effective, scalable, and objective solution for Quiet Eye analysis.
• Automated QE detection reduces reliance on subjective expert annotations.
• This approach can enhance sports training with real-time, data-driven feedback and has potential for broader skill assessment applications.