Improved Feature Point Algorithm for 3D Point Cloud Registration

Abstract

This paper proposes a 3D surface registration algorithm based on the iterated closest point algorithm (ICP). The proposed algorithm uses the Scale-Invariant Feature Transform (SIFT) functions for initial alignment in combination with the K-Nearst Neighbor (KNN) algorithm for function comparison and the Iterative Closest Point (ICP) algorithm weighted for performing accurate registration. First, the point area properties are used for corresponding cloud point areas. Second, files with associated regions are classified to calculate the initial registration transformation matrix. Based on this combination, the correct matching points were extracted between the input data. The proposed registration approach is able to perform automatic registration without any assumptions about their initial positions. Experimental results using biomedical data (CT data) indicate the effectiveness of the proposed approach. Experimental results show that the proposed algorithm increases the number of correct function correspondences while reducing significantly corresponding errors compared to the original ICP and RPM algorithms.