Predicting Hip Kinematics during Cycling Task with CNN

Abstract

To understand the biomechanics of cycling, it is essential to precisely measure and monitor joint kinematics and kinetics, which aids in optimizing performance and preventing injuries. The traditional method of analyzing hip joint kinematics in cycling using motion capture system is constrained by high costs, limited accessibility, and lack of real-time applicability in clinical and training environments. To address this gap, this study leveraged a Convolutional Neural Networks (CNNs) model to predict the hip joint kinematics during pedaling task on a stationary cycling ergometer. The present study involved 10 participants, comprising eight males and two females, devoid of musculoskeletal disorders, to ensure the purity and relevance of the data. Using the Vicon motion capture system, a comprehensive dataset of hip joint kinematics parameters were collected. The CNN model, designed with five hidden layers, was meticulously trained on this dataset.

Our results showed a notable accuracy in predicting hip joint kinematics parameters using CNN, with a Root Mean Square Error (RMSE) of 2.98±0.61° for hip flexion, 1.51±0.46° for hip adduction, and 1.73±0.49° for hip rotation. These values, within acceptable limits, demonstrate the model's robustness in hip measurements. This research tried to underscore the efficacy of deep learning in biomechanical applications, particularly for healthcare and sports applications.

This study contributes significantly to the biomechanical study of the hip joint, offering a potential integration of predictive models and real-time monitoring of hip joint kinematics during cycling exercise with stationary ergometers. This CNN-based technology has the potential to be employed in studies where a motion capture lab is not available.