Scan Registration using the Normal Distributions Transform with Region Growing Clustering for Point-Sampled 3D Surfaces

The normal distributions transform (NDT) registration algorithm often converges to a local minimum if there is a large initial transformation error. In order to improve convergence properties, a novel region growing clustering NDT (RGC-NDT) algorithm is proposed, which replaces volumetric divisions and removes discontinuities in the cost function at the voxel boundaries during the optimization step. A set of computationally efficient multi-scale difference of locally salient feature vectors is computed for each point in the point cloud using principal components analysis (PCA). Adjacent points around the seed point in the local neighborhood are aggregated into clusters according to the region membership criterion based on the similarity of these feature vectors. The normal distributions transform is then computed for each cluster and points within the region are represented as a probability density function (PDF) for distribution-to-distribution matching. Scan registration results from indoor laser scans validate the improvement of the basin of convergence for NDT algorithm over existing methods.