3D Non-rigid Registration of Deformable Object Using GPU

Abstract

We propose a method to quickly process non-rigid registration between 3D deformable data using a GPU-based non-rigid ICP(Iterative Closest Point) algorithm. In this paper, we sequentially acquire whole-body model data of a person moving dynamically from a single RGBD camera. We use Dense Optical Flow algorithm to find the corresponding pixels between two consecutive frames. Next, we select nodes to estimate 3D deformation matrices by uniform sampling algorithm. We use a GPU-based non-rigid ICP algorithm to estimate the 3D transformation matrices of each node at high speed. We use non-linear optimization algorithm methods in the non-rigid ICP algorithm. We define energy functions for an estimate the exact 3D transformation matrices. We use a proposed GPU-based method because it takes a lot of computation time to calculate the 3D transformation matrices of all nodes. We apply a 3D transformation to all points with a weight-based affine transformation algorithm. We demonstrate the high accuracy of non-rigid registration and the fast runtime of the non-rigid ICP algorithm through experimental results.