Reconstructing Global Multi-Person 3D Motions From A Single Moving Camera

Abstract

The reconstruction of 3D human models from in-the-wild videos has emerged as a highly promising research direction, with increasingly practical and widespread applications. To address the challenges of depth ambiguity and human occlusions in monocular videos, an optimization-based framework is proposed that combines SLAM techniques with human motion priors to recover global multi-person motion from a single moving camera. Camera parameters are estimated using SLAM, while a robust tracking algorithm is adopted to mitigate occlusion effects. A learned motion smoothness prior is further integrated into the optimization process. The SMPL model and camera parameters are jointly optimized to align the reconstructed 3D poses with 2D detections. Experimental results on the 3DPW dataset demonstrate that the proposed method outperforms existing approaches in both accuracy and robustness.