Local 3D pose estimation of feature points based on RGB-D information for object based augmented reality

Abstract

We here describe a novel approach for locally obtaining pose estimation of match feature points, observed using RGB-D cameras, in order to apply to locally planar object pose estimation with RANSAC method for augmented reality systems. Conventionally, object pose estimation based on RGB-D cameras are achieved by the correlation between observed 3D points captured by feature point matching and known 3D points of the object. However, in such methods, features are simplified as single 3D points, losing information of the feature and its neighborhood surface. This approach based on local 3D pose estimation of locally planar feature points, brings richer information for 3D pose estimation of planar, 3D rigid or deformable objects. This information enables more stable pose estimation across RANSAC settings than conventional three-points RANSAC methods.