LAUSR

This paper proposes a supervised probabilistic framework for building extraction. Multiple features from both images and point clouds can be fused in this framework to achieve global optimal building extraction. Basically, it adopts conditional random fields (CRFs) to discriminatively model the ground scene and its observed data, and formulates building extraction as a pixel-labeling problem. Color, edge, and height explored from both kinds of data are fused into the association potential and interaction potential of CRFs in order to achieve a global optimal labeling. Furthermore, it develops Gaussian mixture model and height range model for color distribution and height distribution, respectively. These models facilitate parameter learning and model specification from training data. Furthermore, it constructs regular interaction potentials such that global optimization can be solved efficiently by a graph cut algorithm. With the proposed framework, buildings can be extracted automatically and efficiently if training data are annotated in advance. As demonstrated by the experiments and their evaluations, the proposed approach outperforms state of the art techniques and allows further improvement by incorporating additional data or features.