Roads in modern cities facilitate different types of users, including car drivers, cyclists, and pedestrians. These different users often have a designated section of the road to operate on. Road management, e.g., by municipalities, needs to take this sectioning into account, preferably in an efficient way. Mobile laser scanning (MLS) point clouds provide accurate and dense three-dimensional (3D) measurements of road scenes, showing strong mapping capabilities, although their massive data volume and lack of structure still bring difficulties in automatic processing. Methods for the automatic classification of road surface types are still largely lacking, and the existing methodology did not consider the potential of MLS point clouds yet. In recent years, point cloud understanding through deep neural networks has achieved breakthroughs. However, perceiving large-scale point clouds by deep learning depends on aggregating local features and progressive downsampling, to extract rich contextual information. As a consequence, low-level features that reveal details in point clouds may not be well preserved, possibly resulting in ambiguous delineation in point cloud classification. For road mapping, inaccurate classification of points on road boundaries hinders the generation of high-quality map products. Some existing deep learning methods propose to mitigate the fuzzy classification near boundaries, either by utilizing refinement for network predictions, or by indirectly modifying neighboring weights when summarizing local information. Approaches to achieving a satisfying overall performance, while maintaining accurate delineation, still need investigation. In this study, we propose a novel approach for road type classification of MLS point clouds in dense urban areas based on a deep neural network. We follow the main architecture of RandLA-Net, a point-wise neural network designed for large-scale point cloud processing. To alleviate the ambiguous delineation of point cloud classification, we propose two strategies. The first is to refine predictions of RandLA-Net by conditional random field. In the second strategy, we incorporate boundary constraints in the network by introducing a novel distance label for each road surface point to represent the distance to its closest boundary. The distance prediction task is combined with road type classification by adding another branch in RandLA-Net to formulate multi-task learning. Through experiments, we show that 3D point cloud semantic segmentation by deep learning is applicable for road type classification. Also, the multi-task learning strategy is verified to be more effective in improving the delineation performance. Using MLS point clouds acquired from 5 German cities (Hamburg, Delmenhorst, Bremerhaven, Hannover, and Oldenburg), we classify road points separately into different usages (sidewalk, cycling path, rail track, parking area, motorway, green area, and island without traffic) and materials (cobblestone, asphalt, plates, unpaved, and railway). When adopting Hannover and Oldenburg for testing, and the other three cities for training, we obtain a mean intersection over union of 46.1% for usage type and 52.0% for material type with the multi-task learning strategy and input features (x, y, z, R, G, B, intensity), outperforming the original RandLA-Net by approximately 4%. Moreover, from the point cloud classification results, we achieve lightweight polygon representations of road objects in different types through post-processing, which is demonstrated to perform better than an image semantic segmentation-based solution quantitatively and qualitatively.

Semantic segmentation of aerial point clouds with high accuracy is significant for many geographical applications, but is not trivial since the data is massive and unstructured. In the past few years, deep learning approaches designed for 3D point cloud data have made great progress. Pointwise neural networks, such as PointNet and its extensions, show their ability to process 3D point clouds, especially in classification and semantic segmentation. In this work, we implement DGCNN (Dynamic Graph CNN), which combines PointNet with Graph CNN, and extend its semantic segmentation application from indoor scenes to an aerial point cloud dataset: The Current Elevation File Netherlands (AHN), which was produced by airborne laser scanners for the whole Netherlands. Point clouds from the iteration AHN3 are classified into four classes: ground, building, water and others (including vegetation, railways, etc). Moreover, DGCNN splits the input point cloud into regular blocks before operating on it and processes each block independently, which limits the effective range (receptive field) of the network to some extent. Thus, the second aim of this work is to investigate the impact of the effective range on the performance of DGCNN by adjusting two crucial parameters: the block size and the neighborhood size k in k-NN graphs. It turns out that enlarging the block size or k helps to improve the overall accuracy of DGCNN, but cannot ensure better segmentation results from each individual class. With the block size 50 m and k=20, the most balanced F1 scores for all classes and an overall accuracy of 93.28% are achieved. Based on the evaluation for each setting with a certain block size and k, we also manage to further improve the overall accuracy to 93.51% by combining smaller-scale (with block size 30 m) and larger-scale (with block size 50 m) segmentation results, with k=20.

Classification of aerial point clouds with high accuracy is significant for many geographical applications, but not trivial as the data are massive and unstructured. In recent years, deep learning for 3D point cloud classification has been actively developed and applied, but notably for indoor scenes. In this study, we implement the point-wise deep learning method Dynamic Graph Convolutional Neural Network (DGCNN) and extend its classification application from indoor scenes to airborne point clouds. This study proposes an approach to provide cheap training samples for point-wise deep learning using an existing 2D base map. Furthermore, essential features and spatial contexts to effectively classify airborne point clouds colored by an orthophoto are also investigated, in particularly to deal with class imbalance and relief displacement in urban areas. Two airborne point cloud datasets of different areas are used: Area-1 (city of Surabaya—Indonesia) and Area-2 (cities of Utrecht and Delft—the Netherlands). Area-1 is used to investigate different input feature combinations and loss functions. The point-wise classification for four classes achieves a remarkable result with 91.8% overall accuracy when using the full combination of spectral color and LiDAR features. For Area-2, different block size settings (30, 50, and 70 m) are investigated. It is found that using an appropriate block size of, in this case, 50 m helps to improve the classification until 93% overall accuracy but does not necessarily ensure better classification results for each class. Based on the experiments on both areas, we conclude that using DGCNN with proper settings is able to provide results close to production.@en