The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLIII-B2-2021
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B2-2021, 115–122, 2021
https://doi.org/10.5194/isprs-archives-XLIII-B2-2021-115-2021
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B2-2021, 115–122, 2021
https://doi.org/10.5194/isprs-archives-XLIII-B2-2021-115-2021

  28 Jun 2021

28 Jun 2021

ROAD TYPE CLASSIFICATION OF MLS POINT CLOUDS USING DEEP LEARNING

Q. Bai1,2, R. C. Lindenbergh1, J. Vijverberg2, and J. A. P. Guelen2 Q. Bai et al.
  • 1Dept. of Geoscience and Remote Sensing, Delft University of Technology, The Netherlands
  • 2Cyclomedia Technology

Keywords: mobile mapping, point clouds, semantic segmentation, road type, local feature aggregation, deep learning

Abstract. Functional classification of the road is important to the construction of sustainable transport systems and proper design of facilities. Mobile laser scanning (MLS) point clouds provide accurate and dense 3D measurements of road scenes, while their massive data volume and lack of structure also bring difficulties in processing. 3D point cloud understanding through deep neural networks achieves breakthroughs since PointNet and arouses wide attention in recent years. In this paper, we study the automatic road type classification of MLS point clouds by employing a point-wise neural network, RandLA-Net, which is designed for consuming large-scale point clouds. An effective local feature aggregation (LFA) module in RandLA-Net preserves the local geometry in point clouds by formulating an enhanced geometric feature vector and learning different point weights in a local neighborhood. Based on this method, we also investigate possible feature combinations to calculate neighboring weights. We train on a colorized point cloud from the city of Hannover, Germany, and classify road points into 7 classes that reveal detailed functions, i.e., sidewalk, cycling path, rail track, parking area, motorway, green area, and island without traffic. Also, three feature combinations inside the LFA module are examined, including the geometric feature vector only, the geometric feature vector combined with additional features (e.g., color), and the geometric feature vector combined with local differences of additional features. We achieve the best overall accuracy (86.23%) and mean IoU (69.41%) by adopting the second and third combinations respectively, with additional features including Red, Green, Blue, and intensity. The evaluation results demonstrate the effectiveness of our method, but we also observe that different road types benefit the most from different feature settings.