Reconstructing Global Multi-Person 3D Motions from a Single Moving Camera

Currently, the reconstruction of global human trajectories from in-the-wild videos represents a highly promising research area, with applications derived from this technology becoming more practical and widespread. To address the challenges of depth ambiguity and occlusions in monocular videos, we propose an optimization-based framework combining SLAM techniques and human motion priors to recover global multi-person motions from a single moving camera. Our approach employs SLAM techniques to estimate camera parameters and utilizes an advanced tracking algorithm to alleviate occlusions. We also incorporate a learned motion smoothness prior into the optimization framework and optimize SMPL models and camera parameters to fit detected 2D poses. On the 3DPW dataset, our method can achieve superior performance over previous work.