Applying an Artificial Neural Network to Building Reconstruction

Abstract

This paper highlights the uses of Centroid Neural Network for detecting and reconstructing the 3D rooftop model from aerial image data. High overlapping aerial images are used as an input to the method. The Digital Elevation Map (DEM) data and 2D lines are generated and then combined to form 3D lines. The core of the technique is a clustering process using Centroid Neural Network algorithm to classify these 3D lines into groups of lines that belong to the corresponding building areas. This work differs from the previous researches, as it affiliates 3D lines and corners - obtained by applying the Harris corner detector - to automatically extract accurate and reliable 3D rooftop information. The proposed approach is tested with the synthetic images generated from the Avenches dataset of the Ascona aerial images and gives an average error of 0.38m in comparison with the ground truth data. The experiment result proves the applicability and efficiency of the method in dealing with building reconstruction in complicated scenes.