Precise object pose estimation in range image considering measurement error

In this paper, we describe a method for estimating the position and pose of an object from its range image when its shape model is given. The measurement errors of the range image in the viewing direction are much larger than those in the directions parallel to the image plane and their magnitudes differ at every measurement point. We estimate the precise position and pose by considering these characteristics. The observed shape under consideration obtained discretely as a range image is regarded as a plane locally and expressed by a set of triangular patches. On the other hand, the model is expressed by a set of points. The respective points of the model are regarded originally on their corresponding triangular patch plane. The gaps between the model points and their corresponding triangular patches are considered as caused by the measurement errors. Its possibility of occurrence is called the matching value and the goodness of the registration of the set of triangular patches to the set of model points is expressed by the sum of the square of these matching values at the respective model points. The distribution of these matching values in the space is called the matching map. The coordinate transformation in the space the best match is determined by moving the model points in the space. For this purpose, to evaluate the goodness of the registration by the coordinate transformation, the gradient of the matching map is directly calculated.