3D Reconstruction Of Large Scale Environments Using Optical Images

3D reconstruction of large scale environments is the hotspot in virtual reality technology. Reconstruction using optical images is the most significant means, but many problems have not been solved in theory. With the development of 3D reconstruction technology, the requirements for higher quality models and automatic system are getting higher and higher. In the paper, we firstly present a method based on image segmentation and SIFT detector to achieve the disparity range automatically, we use its result as the input of an iterative method which is an improvement of the least squares for solving the limitation of the hypothesis of the front parallel surface in Graph Cut. To build a unified scene model that contains multiple subsets, we present a novel method for registration of 3D scene reconstructions in different scales. The main work and innovations include the following aspects:1. The stereo matching algorithm based on Graph Cut theory produces disparity images from two or more images of the same scene in different visual angles, from which we can obtain the depth information of the scene, which can bring critical advantages to a very wide spectrum of visual application domains, such as 3-D reconstruction and vision measurement. However, the choice of the disparity range is often overlooked, the paper presents an algorithm based on image segmentation and SIFT detector to achieve the disparity range automatically, and the result is applied to the stereo matching algorithm. The experimental results show that the proposed algorithm can obtain the disparity range accurately.2. To solve the limitation of the hypothesis of the front parallel surface in Graph Cut, we present an iterative method of the improvement of the least squares. The reconstruction of teddy's 3D point cloud shows that the proposed algorithm can improve the accuracy of the disparity of flat and curved pla ne.3. According to the principle of similarity reconstruction, the scale of each reconstruction produced by Bundler and PMVS system is different from the scale of real-world scenes. So we need to normalize the models' scales before the registration, then unifying models to the same coordinates system. We normalize models' scales based on the constraint of the camera center. The scaling can be confirmed by the ratio between two average distances computed by different corresponding points to their own point cloud centers. We select the sharing camera center as the corresponding points for its high accuracy.4. We use Cayley transform to fit the matrix of coordinates transformation for the models in normalization scales. The result is compared with the ICP arithmetic, the reconstruction and registration for Notre Dame de Paris model and tank model show the effectiveness of the algorithm in scale-normalized model and point clouds' merging. The approach is an extending way for image-based modeling, which can be applied to the registration for scenes in different scales.