This study presents a deep learning-based approach for the automated detection of Quiet Eye (QE) durations from
electrooculography (EOG) signals in archery. QE—the final fixation or tracking of the gaze before executing a motor
action—is a critical factor in precision sports. Traditional detection methods, which rely on expert evaluations, are
inherently subjective, time-consuming, and inconsistent. To overcome these limitations, EOG data were collected
from 10 licensed archers during controlled shooting sessions and preprocessed using a wavelet transform and a
Butterworth bandpass filter for noise reduction. We implemented and compared a traditional model (SVM) and
five deep learning models—CNN+LSTM, CNN+GRU, Transformer, UNet, and 1D CNN—for QE detection. The
CNN+LSTM model achieved the highest accuracy (95%), followed closely by CNN+GRU (93%), demonstrating
superior performance in capturing both spatial and temporal dependencies in the EOG signals. Although
Transformer-based and UNet models performed competitively, they exhibited lower precision in distinguishing QE
periods. The performance of the traditional model was inferior to deep learning approaches. These results indicate
that deep learning provides an effective and scalable solution for objective QE analysis, substantially reducing
the dependence on expert annotations. This automated approach can enhance sports training by offering realtime, data-driven feedback to athletes and coaches. Furthermore, the methodology holds promise for broader
applications in cognitive and motor skill assessments across various domains. Future work will focus on expanding
the dataset, enabling real-time deployment, and evaluating model generalizability across different skill levels and
sports disciplines.
Keywords Quiet eye, Electrooculography, Wavelet transform, Convolutional neural networks, Long-short term
memory, Transformer, UNet, GRU