Roof Type Classification Using Deep Convolutional Neural Networks on Low Resolution Photogrammetric Point Clouds From Aerial Imagery
Maria Axelsson, Ulf Söderman, Andreas Berg, Thomas Lithen
Abstract
Three-dimensional (3D) reconstruction of buildings is an active research area with applications in e.g. city planning, environmental simulations, and city navigation. Automatic 3D building reconstruction methods based on point clouds from laser scanning or methods based on high resolution dense photogrammetric point clouds are common in the literature. In applications where large land areas need to be covered regularly it is not practical to use laser scanning or acquire images with high resolution and large image overlaps. In these applications the reconstructed photogrammetric point cloud has low resolution with less building details. We present a method where the most common roof types are classified using a deep convolutional neutral network (CNN) pre-trained using RGB data in this challenging type of data. In addition' a method for roof height estimation for each roof type is presented to support automatic 3D building reconstruction using model building shapes. Results are shown for a low resolution dense photogrammetric point cloud generated using multi-view stereo reconstruction of standard overlapping aerial images from nationwide data collection. The method is intended to support automated generation of a nationwide 3D landscape model.
BibTeX
@inproceedings{icassp2018_rooftypeclassifi,
title = {Roof Type Classification Using Deep Convolutional Neural Networks on Low Resolution Photogrammetric Point Clouds From Aerial Imagery},
author = {Maria Axelsson and Ulf Söderman and Andreas Berg and Thomas Lithen},
booktitle = {ICASSP 2018},
year = {2018}
}