Kolloquium Geoinformatik Classification of 3D Point Clouds using Deep Neural Networks

Per-point classification (semantic labeling) is an important step in processing topographic 3D point clouds. Current methods often rely on hand-crafted attributes to describe local point neighbourhood relations, feeding the resulting feature vectors to a state-of-the-art classifier. For classification tasks, deep neural networks (DNNs) have recently outperformed most traditional approaches. Since point clouds are unordered and irregular sets of tuples in space, the use of DNNs on point clouds has mostly been limited to strongly regularized (i.e. voxelized or rasterized) point cloud representations and their attributes. A novel method developed by Qi et al. (2017) allows the direct input of point clouds to a DNN. A feature describing a subset of points is hereby calculated using a commutative aggregation function. The commutative property solves issues with unordered input to DNNs. This method is adapted for topographic Airborne Laser Scanning (ALS) point clouds. While the neighbourhood definitions (on different scales) are required as input, the individual features describing the local point neighbourhoods are learned by the DNN in the training phase. The trained network is further evaluated in terms of robustness w.r.t. point density, distribution pattern and penetration rate. Also, attributes known to aid in classification (e.g. principal component analysis, echo broadening) can be added to see if the network further profits from this information or if the information is already learned inherently. First tests based on ALS data from the federal district of Vorarlberg (Austria) yielded an overall accuracy of 76.9%. In forested areas, this accuracy increases up to 98.6 %. Further improvements of the DNN are work-in-progress and are expected to overcome the sub-optimal classification of buildings, which are often misclassified as vegetation, as well as improve overall classification.