Human Motion Prediction based on IMUs and MetaFormer

Human motion prediction forecasts future human poses from the histories, which is necessary for all tasks that need human-robot interactions. Currently, almost existing approaches make predictions based on visual observations, while vision-based motion capture (Mocap) systems have a significant limitation, e.g. occlusions. The vision-based Mocap systems will inevitably suffer from the occlusions. The first reason is the deep ambiguity of mapping the single-view observations to the 3D human pose; and then considering the complex environments in the wild, other objects will lead to the missing observations of the subject. Considering these factors, some researchers utilize non-visual systems as alternatives. We propose to utilize inertial measurement units (IMUs) to capture human poses and make predictions. To bump up the accuracy, we propose a novel model based on MetaFormer with spatial MLP and Temporal pooling (SMTPFormer) to learn the structural and temporal relationships. With extensive experiments on both TotalCapture and DIP-IMU, the proposed SMTPFormer has achieved superior accuracy compared with the existing baselines.