In this paper, we propose a surface-based reconstruction technique which simplifies the process of scene reconstruction and increase reliability. Integrationg grid coding and intensity image analysis methods, discontinuities are classified and stable surfaces which make up the skeleton structure of a room (i.e., wall, ceiling, floor, etc.) are extracted. Local models are then reconstructed under the planar assumption, and merged into the global model. The experimental results show that local errors do not radically alter the adjacency relationships of detected surfaces, so robust reconstruction can be achieved.