Efficient and Globally Optimal Multi View Dense Matching for Aerial Images

A variety of applications exist for aerial 3D reconstruction, ranging from the production of digital surface models (DSMs) and digital terrain models (DTMs) to the creation of true orthophoto and full 3D models of urban scenes that can be visualized through the web. In this paper we present an automated end-to-end workflow to create digital surface models from large scale and highly overlapping aerial images. The core component of our approach is a multi-view dense matching algorithm that fully exploits the redundancy of the data. This is in contrast to traditional two-view based stereo matching approaches in aerial photogrammetry. In particular, our solution to dense depth estimation is based on a multi-view plane sweep approach with discontinuity preserving global optimization. We provide a fully automatic framework for aerial triangulation, image overlap estimation and dense depth matching. Our algorithms are designed to run on current graphics processing units (GPUs) which makes large scale processing feasible at low cost. We present dense matching results from a large aerial survey comprising 3000 aerial images of Graz and give a detailed performance analysis in terms of accuracy and processing time.