Optimizing point-cloud neighborhood calculation for classification

The increasing global availability of dense point clouds provides the potential to better capture complex environments and their changes, e.g., forests state and growth, river erosion processes, city planing, etc. An important process for turning point clouds into useful datasets is their classification. The quality of classified point clouds relies on four critical processing steps: neighborhood definition, feature extraction and selection, quality of training data and classification model. Determining the optimal neighborhood for each point is essential for capturing local information, fast calculation, enhancing feature richness, and improving the quality of downstream processes.

We propose a novel method for constructing neighborhoods for geometric feature calculation using a kd-tree-based region-growing approach. We construct neighborhoods by selectively adding points guided by local point connectivity, normal orientations, and distance from the seed point. In particular, the local connectivity is determined by a nearest-neighbor graph, parameterized to connect only points belonging to the same object. Following this graph, points are added iteratively to the neighborhood if the angular difference between their normal orientations lies below a locally derived tolerance threshold. The growing process is limited by the distance from the seed point. The new neighborhoods maximize information gain while minimizing boundary crossings between classes, e.g., ground to wall (normal orientation) or branches to buildings (connectivity constraint). Our results demonstrate that this approach outperforms the classical spherical neighborhood and good classification results are more resilient to changes of the neighborhood size. 

Our analysis focus on the classification of urban areas, including ground, building, vegetation and other classes. We evaluate performance using datasets from various platforms, including airborne, mobile, and UAV systems and across different areas such as Berlin, Potsdam, and Paris. The effects of sensor characteristics and point-cloud densities are investigated as well as the improvement of individual features.