Beyond the Motion-Lab: Low-Cost Fatigue Monitoring in Cycling via Deep Learning

Abstract

Fatigue detection during exercise currently relies on costly and specialized technologies, limiting its accessibility outside of controlled laboratory environments. This work explores a marker-less, low-cost alternative for detecting fatigue-related changes in cycling biomechanics using OpenPose, a deep convolutional neural network for human pose estimation, together with a custom supervised classification framework. We recorded sagittal-plane video of five participants performing a cycling ramp test under two conditions, Initial and Fatigue, yielding ten sessions at 29.97 fps. From the left-side keypoints we extracted four joint angle signals (knee, hip, trunk, and ankle), filtered with a 4th-order zero-phase Butterworth low-pass filter at 6 Hz. Three classifiers were compared under leave-one-subject-out cross-validation: an LSTM with attention, a plain LSTM, and XGBoost. Averaged across five random seeds, the plain LSTM achieved 65.4% mean accuracy, slightly outperforming the attention-based LSTM (63.4%), while XGBoost reached 52.3%. The attention model achieved the highest single-seed accuracy (78.7%), but at the cost of greater variance across initializations. The most consistent biomechanical finding was an increase in ankle range of motion under fatigue, consistent with a proximal-to-distal compensation strategy. These results constitute preliminary evidence that a marker-less pipeline based on OpenPose is a viable foundation for an accessible fatigue monitoring system in cycling.