Pose invariant geometrie feature exploring for dense 3D data registration

Basic and compound pose invariant geometric features are explored to conduct pre-alignment, down sampling, and outlier removal to improve the speed and accuracy of Iterative Closest Point (ICP) registration on dense 3D laser scanning data for metallic part modeling. Rectification of planar surfaces, which are segmented from point normals and matched on similarity of clustered viewpoint feature histogram (CVFH), achieves good pre-alignment without the need of point correspondence searching across two dense data sets. Salient points after scale-variant difference of normals (DoN) filtering are extracted to significantly down sample the dense data. Squared Euclidian distance between two point feature histograms (DoH) is incorporated into the error function of ICP for outlier removal to further enhance registration accuracy.