Efficient Classification of Airborne LiDAR Point Clouds using PointNet++ with Scanning Line Segmentation Makoto Ishiwata(a*), Masafumi Nakagawa(a)
a) Shibaura Institute of Technology, Japan
*ah21040[at]shibaura-it.ac.jp
Abstract
Recently, various methods have been proposed for classifying point cloud features using machine learning. However, machine learning requires a high-performance computing environment. Representative examples include the outdoor segmentation datasets, such as SemanticKITTI and nuScenes. Processing both of these datasets requires multiple GPUs for processing due to their large size. This results in an extremely long learning time, making them impossible to handle them on commonly used computers. Therefore this study verifies the feasibility of classifying point cloud features using machine learning with a small dataset. We used an aerial LiDAR point clouds consisting of ordered points along scanning lines, that acquired in dense urban areas in Tokyo. We subjected approximately 2,000 points to point interpolation to align the point intervals. Then, we performed RANSAC plane estimation to correct distortion. Next, we performed DBSCAN and convex hull processing to label surfaces and prepare the dataset. We used a computer equipped with only a built-in GPU and PointNet++ for learning. Moreover, we evaluated the accuracy of the processing using point clouds from other areas. The overall accuracy rate was 0.63, and the accuracy rate was higher for the flat and consistent ground surfaces than for other surface features. This method offers a new approach to processing the classification features in airborne LiDAR point clouds by generating a 3D model from 2D cross-sections. Additionally, we demonstrated that high-precision maps could be more easily created from point clouds using machine learning, even without a high-performance processing environment.
Keywords: 3D Mapping, LiDAR, point cloud classification, machine learning
