Model-free Dense Stereo Reconstruction Creating Realistic 3D City Models

Abstract

In this paper we describe a framework for fully automatic and model-free generation of accurate and realistic 3D city models using multiple overlapping aerial images. The underlying DSM is computed by dense image matching, using a robustified Census transform as cost function. To further reduce the noise of mismatches, we afterwards minimize a global energy functional incorporating local smoothness constraints using variational methods. Due to the convexity of the framed problem, the solution is guaranteed to converge towards the global energy minimum and can be efficiently implemented on GPU using primal-dual algorithms. The resulting point cloud is then being triangulated, local planarity constraints are exploited to reduce the number of vertices and finally a multi-view texturing is applied. The quality of the DSM and the 3D Model is evaluated on a complex urban environment, using reference data generated by laser scanning (LiDAR).