Virtual feature point extraction from polyhedral structure

Recently, recognition/modeling of 3D objects in sequential point clouds is a major issue. In this research, estimation of the transformation matrix between point clouds is an important subject to eliminate the influence of rotation and translation by motion of the sensor and objects. Generally, ICP is used to calculate the transformation matrix with huge computational costs. On the other hand, feature point based method is lower costs than ICP, but its accuracy depends on sensor resolution and environmental distortion strongly. In this report, we propose novel feature point extraction methods to obtain high accuracy fusion of sequential point clouds, focus on the relatively stable locations of polyhedral structure. These feature points are determined virtual locations instead of actual points, their properties are described aligned normal vectors of planar surfaces. In experimental results, we confirmed that proposal methods are able to obtain high accuracy correspondences and the translation vector with low computational costs.