Improvement of Voxel Down-Sampling Method in Point Cloud Registration

Abstract

In order to mitigate the influence of outliers in point clouds on voxel down-sampling results and the accuracy of point cloud registration in practical applications, the voxel down-sampling method in the Point Cloud Library (PCL) has been improved. First, the nearest-neighbor distance method is applied to search for the actual point that is closest to the calculated centroid point of a voxel volume. This point is then used as the reserved point to replace the voxel, which helps to preserve the local detailed features of the edge contour of the point cloud. Secondly, considering that the voxels located at the edges of a point cloud often contain a large number of outliers and these outliers are characterized by a sparse distribution, a constraint of a minimum number of points is imposed to effectively eliminate outliers in the edge area and greatly reduce their impact on the accuracy of point cloud registration. Down-sampling and registration experiments were conducted using Bunny point clouds provided by Stanford University. The experimental results demonstrate that the improved voxel down-sampling method effectively improves the accuracy and speed of point cloud registration.