Image-based 3D Reconstruction And Rendering Of Buildings

3D reconstruction technology is a reflection of the real world in 3D model of a scene or object build. Recently, the smart phone has been widely used. The smart phone produce more and more pictures. Because people’s visual demand for 3D image becomes more and more strong, the 3D reconstruction technology which can get three-dimensional model receives widely attention. Among them, image-based3 D reconstruction techniques uses the two-dimensional image information contained in the photograph to reconstruct the three-dimensional model of the object. Because of its advantages of simple reconstruction process, convenient data acquisition, fast update, it has become the current research focus.In this paper, we researched image-based 3D reconstruction techniques,designed and implemented building image sequence-based 3D reconstruction. After entering a set of unordered images, image-based 3D reconstruction can output image information. The main steps include: image acquisition sequence, detection and matching image feature points, camera calibration, 3D reconstruction.During the acquisition phase sequence of images, using a mobile phone camera to take a group of images. Because the images captured by cell phone camera are too large, which may result in the subsequent reconstruction process is more time-consuming. We pre-process the images to zoom out.In the two-dimensional image feature point detection and matching stage, we use the SIFT algorithm, which has good invariance and strong matching ability.Firstly, we extract SIFT feature from each images which have been preprocessed.Then randomly select two adjacent images to match feature, locate the feature after the match points, and mark out all the feature descriptor in the original image, but there have been cases of false matches during the image matching process. This situation will affect the matching accuracy of the results and increase matches time.Therefore, we use random sample consensus algorithm to eliminate false match feature points to obtain a more accurate matching of SIFT feature points.In the camera calibration stage, we use Bundler to complete self-calibration of the camera. For matched feature points, Bundler system uses the algorithm based on the structure of motor recovery to estimate extrinsic parameters of all cameras,while obtaining sparse 3D point cloud.In the reconstruction phase of 3D point cloud, we use the multi-view stereo vision algorithms to recover sparse image acquired into a dense 3D point cloud.