Real-time Non-Rigid 3D Reconstruction System Specialized for Human Body Reconstruction

Abstract

This thesis proposes a real-time non-rigid 3D human body reconstruction framework from depth images taken by a commercial color-depth camera. We take a consecutive color-depth image stream as the input and incrementally reconstruct a surface of the model in a 3D canonical TSDF volume space based on the estimated non-rigid warp field from the canonical space to the current input frame. Since the target domain is a human body, we exploit information of a skeletal structure for estimating the warp field of each bone in a hierarchical way and accelerate the computation time compared to previous works. Per-voxel non-rigid warpings are calculated via GPU parallel processing for real-time performance.