Determination of 3D object pose in point cloud with CAD model

Abstract

This paper introduces improvements to estimate 3D object pose from point clouds. We use point-pair feature for matching instead of traditional approaches using local feature descriptors. In order to obtain high accuracy estimation, a discriminative descriptor is introduced for point-pair features. The object model is a set of point pair descriptors computed from CAD model. The voting process is performed on a local area of each key-point to boost the performance. Due to the simplicity of descriptor, a matching threshold is defined to enable the robustness of the algorithm. A clustering algorithm is defined for grouping similar poses together. Best pose candidates will be selected for refining and final verification will be performed. The robustness and accuracy of our approach are demonstrated through experiments. Our approach can be compared to state-of-the-art algorithms in terms of recognition rates. These high accurate poses especially useful for robot in manipulating objects in the factory. Since our approach does not use color feature, it is independent to light conditions. The system give accurate pose estimation even when there is no light in the area.