Automatic Classification of LiDAR Point Cloud Data

Abstract

Road surveys use laser scanners to create 3D maps of the road and its surroundings, known as point clouds. This research addresses the automatic detection of moving vehicles within these point clouds. Moving vehicles distort the survey while it is being scanned, so Atritec AB, the company conducting these surveys, faced the tedious task of manually removing them from the resulting point clouds. This consequently became a problem of labeling every point in the cloud by category, tackled using machine learning methods that learn this labeling from examples. An end-to-end pipeline was developed around a model called RandLA-Net to carry out this point-by-point labeling, trained on a dataset created by manually annotating Atritec’s point cloud recordings. Moving vehicles were far rarer in the dataset than the background classes, an imbalance that required a training approach able to counterbalance it. This was disproportionately penalizing the model on certain classes to make up for their infrequency. Experiments were also conducted with other datasets in an attempt to supplement the native Atritec AB dataset. However, due to fundamental differences between the datasets, this was not to any avail. Experimenting with different model designs and iteratively adjusting their settings led to a pipeline that reached a score of 0.74 for the moving vehicle class on parts of the dataset the model had never seen before. On this scale, a score of 0 means none of the moving vehicles were correctly identified, while a score of 1 means every one of them was.