Comparison of NDT and ICP Method’s point cloud registration performance

Abstract

The point cloud registration approaches are the key problem for three-dimensional reconstruction in reverse engineering, computer vision, cultural heritage and other fields. The Iterative Closest Point (ICP) is widely used in registering of point clouds in various fields of application. Furthermore, the performance of the Normal Distribution Transform (NDT) method directly depends on the selected cube cell size for the data. Choosing the optimum cell size is a challenging problem and there is no proved way for all cases. However, NDT has several advantages over ICP for data storage and speed. The main purpose of the study is to investigate the performance of NDT and ICP algorithms on point cloud registration. For this purpose, a sample dataset was used for comparative assessment. In the study, the fine registration analysis carried out for two different initial distances between point clouds as 10 cm and 5 cm. According to the results, NDT algorithm produced slightly lower root mean square error (RMSE) value for 10 cm initial alignment distances than ICP method while the ICP method produce lower RMSE value for 5 cm initial alignment distance. However, the calculated mean distances between the point clouds after registration demonstrate that the NDT method provides better results than the ICP method for this test data.