HUMAN POSE ESTIMATION FROM MONOCULAR VIDEOS

Abstract

In this thesis, our primary goal is to develop algorithms to robustly estimate 3D human poses for single and multiple persons under severe conditions, including occlusion, large-scale variation, and rare poses, from monocular videos. In single-person scenario, we propose to use a “Cylinder Man Model” which is a simplified human body model to estimate the self-occlusion in existing 3D pose datasets, and a more flexible occlusion augmentation to simulate different occlusion cases. The TCN (Temporal Convolutional Network) is used to smoothen the human motion and a discriminator is used to regularize the kinematics and motion. In multi-person scenario, we propose to use GCN (Graph Convolutional Network) with directed affinity matrix to refine the poses under occlusion, and use different TCNs to estimate poses, velocity, and root coordinates. Furthermore, we propose a pipeline that incorporates both top-down and bottom-up approaches to be better aware of the global position of each person as well as improve the pose accuracy.