Automatic Segmentation of 3D Laser Point Clouds by Ellipsoidal Region Growing

We present and evaluate two variants of an algorithm for simultaneously segmenting and modeling a mixed-density unstructured 3D point cloud by ellipsoidal (Gaussian) region growing. The base algorithm merges initial ellipsoids into larger ellipsoidal segments with a minimum spanning tree algorithm. The variants differ only in the merge criterion used—a threshold on a generalised distance measure defined on the merge candidates. The first variant (shape-distance) considers the relative shape, orientation and position of the ellipsoids, and can grow regions across missing or sparse data, whilst the second (density-distance) attempts to maintain a good fit to the data by setting a minimum sample density threshold on the merged ellipsoid. Adjusting the threshold in each case changes the quality and degree of segmentation achieved. The threshold parameter is tuned by minimising Akaike’s Information Criterion (AIC) with respect to the threshold value. Experiments show that thresholds selected in this way lead to low complexity models and are stable across different environments. The shape-distance measure segments large-scale structures more readily than the density-distance measure, but leads to higher AIC scores, and higher model complexity.