Semantic segmentation of aerial laser scanning point clouds and centering of global neighborhood elevation
DOI:
https://doi.org/10.31449/upinf.100Keywords:
aerial laser scanning, deep learning, point clouds, semantic segmentationAbstract
Aerial laser scanning point clouds are an important data source in many geospatial applications such as
digital terrain model generation or asset mapping of critical infrastructure. Semantic segmentation can
be used in majority of point cloud processing pipelines. Current state-of-the-art methods for semantic
segmentation and classification are based on deep learning. Quality of semantic segmentation depends
also on neighbourhood selection and elevation centring. In this paper, we propose and evaluate different
methods for elevation centring. Experiments on ISPRS 3D Semantic Labelling show that use of a minimal
elevation centring increases results by nearly two percents.