Invariant feature point based ICP with the RANSAC for 3D registration

Abstract

To reduce the computation time and improve the convergence of Iterative Closest Point (ICP) in automatic 3D data registration, the Invariant Feature Point based ICP with the RANSAC(IFP-ICPR), which uses the modified surface curvature estimation for point extraction and embeds the RANSAC in ICP iteration, is proposed. The proposed IFP-ICPR utilizes the radius of estimated sphere for invariant feature point extraction, which is more accurate to extract crease and corner points than the surface variance method. Then the extracted invariant feature points are used in ICP to reduce the computation time. In every iteration of ICP, the RANSAC is embedded to remove the outliers and the convergence of ICP is guaranteed. Point extraction experimental results with simulated cube data show that, compared to surface variance method, the modified invariant feature point extraction algorithm improves the correct ratio of point extraction by 20%. Overall 3D registration experiments with simulated and real reconstructed 3D data show that the proposed IFP-ICPR converges to good solution and computation time is one more orders magnitude less than the compared algorithms. © 2011 Asian Network for Scientific Information.