| Analysis and 3D reconstruction of urban scene is a fundamental problem in computer vision and computer graphics. The oblique photography modeling costs low, run fast, and its texture model detail is rich, has been the mainstream of large scale urban reconstruction. Advanced multi-view stereo (MVS) workflow is widely used in oblique photography. Getting aerial images using unmanned aerial vehicle (uav) at first, and then using the structure from motion (SFM) for obtaining camera parameters and sparse 3D point cloud, next using Patch based multi view stereo (PMVS) to obtain dense point cloud and using Poisson to reconstruct the digital surface model (DSM), finally 3D model city would be got by texture mapping.But there are some problems to be solved in the processes. On the one hand, monomer segmentation is difficult, the amount of data is too huge, and building model is distortion and lack of sharp characteristics. On the other hand, triangular mesh does not have any semantic features that complex urban should have.In order to solve these problems, this paper introduces two approaches that reconstructs urban scene, taking dense triangular mesh that got by MVS workflow as input. In the second chapter introduce an approach of 3D reconstruction based on building contour. Our algorithm proceeds in two main steps, building extraction and 3D model reconstruction by polygon. Procedure of building extraction is, cutting from a certain height of the triangular mesh, getting the contour, and then getting the building by the contour. Procedure of 3D model reconstruction is, cutting the building from an initial height at a specify step, and gradually enhance the value of cutting height. So we could get the outer contour of building, and then streamline contour. Finally establish the 3D model. The experiments using two kind of datasets, a single building and large sale urban scene. Results show that the approach has a good performance, stability and expansibility while being high accuracy and efficiency.The third chapter introduce an approach of semantic three dimensional reconstruction based on feature extraction. Extracting the geometry feature of triangular mesh, classify the mesh by energy minimization, and then getting streamline model by planar clustering. Experimental results show that the classification approach is effective, the building models obtained are more concise and the surface is regulatory, with a high precision at the same time.The algorithms is a meaningful attempt of the current tilt photography typical problem, the research work and results have certain scientific value and will expand the application of urban 3D model and the development of smart city.Innovation of this paper are as follows:(1) Creatively proposed a new three-dimensional model streamlining processes and rules. (2) Creatively proposed a new method for semantic classification of three-dimensional model. |
