Abstract
Lightweight global trajectory reconstruction in 3 dimensions is a yet to be solved problem for animation data from IMU based motion capture systems. In this work we use a lean U-Net neural network architecture to estimate global displacements in a stable and accurate manner. Our network takes as an input a combination of local pose information and acceleration signals from IMU sensors and estimates short, character centred trajectories of 8 frames. We use a weighted average of the predictions to reduce the effect of estimation bias and noise during integration of the displacement and empirically show its advantage over other methods. We test our trained model on a dataset of unseen data, from actors that were not included in the training set. Our method is capable of accurately reconstructing the ground truth trajectory, without significant drift effects, for both horizontal planar motion as well as motion in the vertical direction. We further show how with declining pose data quality, estimation accuracy deteriorates and how acceleration signals are pivotal to maintain high quality trajectory reconstruction.