Scan registration using segmented region growing NDT

Abstract

The Normal Distributions Transform (NDT) scan registration algorithm divides a point cloud using rectilinear voxel cells, then models the points within each cell as a set of Gaussian distributions. A nonlinear optimization is performed in order to register the scans, however the voxel-based approach results in ill-defined cost function derivatives as points cross cell boundaries. In this work, a Segmented Region Growing NDT (SRG-NDT) variant is proposed, which first removes the ground points from the scan, then uses natural features in the environment to generate Gaussian clusters for the NDT algorithm. The removal of the ground points is shown to significantly speed up the scan registration process with negligible effect on the registration accuracy. By clustering the remaining points, the SRG-NDT approach is able to model the environment with fewer Gaussian distributions compared with the voxel-based NDT methods, which allows for a smooth and continuous cost function that guarantees that the optimization will converge. Furthermore, the use of a relatively small number of Gaussian distributions allows for a significant improvement in run-time. Experiments in both urban and forested environments demonstrate that the SRG-NDT approach is able to achieve comparable accuracy to existing methods, but with an average decrease in computation time over ICP, G-ICP, and NDT, of 90.1%, 95.3%, and 94.5%, respectively.