New Paper: CNN-based Estimation of Sagittal Plane Walking and Running Biomechanics from Measured and Simulated Inertial Sensor Data

We would like to directly estimate biomechanical data from inertial measurement unit (IMU) data. Biomechanical models are time consuming, but at the same time, machine learning is limited because it is difficult to create large enough experimental datasets to perform training. In this paper, we combine the advantages of physical modeling with machine learning methods to create fast estimations. We augmented an experimental dataset with gait simulations. The simulations were created by randomly generated motion patterns, which were tracked in a simulation to create a physically feasible gaits. Simulated IMU data was calculated from these gaits. Convolutional neural networks (CNNs) were trained with the original experimental dataset, as well as the augmented dataset to estimate joint angles, joint moments, and ground reaction forces. Estimation accuracy improved for the joint angles, while there was no improvement in the joint moments and ground reaction forces due to a reality gap between the simulation and the experimental data.

The paper can be found: