Multiview Rasterization of Street Cross-sections Acquired with Mobile Laser Scanning for Semantic Segmentation with Convolutional Neural Networks

Abstract

Although point-based architectures are making inroads into point cloud semantic segmentation with Neural Networks, the rasterization remains a useful alternative when computational resources are limited. This paper presents a method for semantically segmenting a street point cloud by generating multiple views and using a Convolutional Neural Network. The method starts by segmenting the street into cross-sections, then the point cloud is rotated to generate three views that are rasterized into images and classified with a ResNet 18. Finally, the pixel labels are back-projected to the original point cloud, and each point is assigned the mode of the three classes received. The method was tested in a real case study, obtaining an accuracy of 88.77%. The method has the same disadvantages in terms of sample distribution as point-based Neural Networks, but the training is more efficient.