Object Recognition and Pose Estimation Based on Combined Use of Projection Histograms and Surface Fitting

Abstract

In this paper we present a configurable object recognition and locating system for 3D point cloud sensors. The objects are recognized based on cylindrical projection histograms and after the object is recognized, the initial pose of the object is computed based on the eigenvectors of the modelled and measured 3D point clusters. The optimal 6 degree of freedom pose is estimated by fitting the CAD-model surfaces to the measured 3D-points, where the model surfaces and 3D points are correlated based on the modelled and measured eigenvectors. The novelty of our system is the combination of reliable histogram based object recognition and accurate CAD-based pose estimation in the object recognition system with configurability options according to application requirements and point cloud properties.