Joint Angles Estimation Based on Center of Force With Artificial Neural Network and Convolutional Neural Network

The current trend of assessing joint angles could be conducted through machines or deep learning (DL) relies on various input data. The development of different types of DL models using foot data may avoid finding optimal sensor positions and direct contact with human skin and can be accessible for subjects. We aimed to investigate the performance of artificial neural network (ANN) and convolutional neural network (CNN) in predicted joint angles based on center of force (CoF). Twenty subjects were required to walk on a treadmill at three gait speed modes: comfortable, fast, and running. Foot data were collected through an F-scan system during the experiment. The kinematics of the lower limb joints were recorded using a camera system. Two different types of DL approaches were applied: ANN and CNN. The DL models were trained using the input data of CoF and combination of CoF with joint angle. The two matrices of root-mean-square error (RMSE) and correlation coefficient were used to evaluate each model's performance. The DL approaches of ANN and CNN provided the maximum RMSEs (°) of 12.56 and 14.62, while the minimum correlation coefficients were 0.9102 and 0.9354, respectively. Accordingly, DL based on ANN or CNN architecture is a promising approach that can accurately estimate joint angles even with running gait, beneficial for the sport and rehabilitation field.