A Deep-Learning-Based Method for Gaining More Biomechanical Parameters with Fewer Sensors in Fast and Complex Movements

Abstract

Real-time biomechanical feedback can provide direct and objective quantified information for any practitioners, such as the athletes/coaches, to assist in accelerating their motor skills’ learning and training process. However, it is usually difficult to monitor the human motion in full and acquire the key biomechanical parameters (i.e., the kinematic and kinetic data, the EMG, etc.) just with few sensors in some elite sports involving fast movements and complex motor skills. Using too many sensors in the field tests may limit the athletes’ motor ability and affect the collected data’s validity and reliability. In this paper, we employ a deep learning method to immensely reduce the number of sensors required for providing the real-time biomechanical feedback in field, according to the hammer-throw local motion features found from our pilot study. Based on the Keras API imported from the TensorFlow open-source platform, two Sequential Neural Network models are implemented and compared. One model has two inputs (i.e., vertical displacements and velocities on waist) and six outputs (i.e., vital joint angles on lower limbs). The other one has four inputs (i.e., vertical displacements and velocities on wrist and waist) and thirteen outputs (i.e., vital joint angles on both upper and lower limbs). The experimental results demonstrate that the vital joint angles on the upper and lower limbs have strong correlation with the vertical wrist and waist/hip displacements respectively. This study indicates that fewer wearable sensors can be applied in fast and complex movements to obtain the most significant kinematic data, whereas more biomechanical parameters can be further gained by prediction.