3D Reconstruction Research Based On Ordered Images

Image-based3D reconstruction technology is one of the important researchdirection in the fields of computer vision, to a large number of images reconstruction,for noisy images, computational complexity and taking long time etc, resulting in thereconstruction has low model accuracy, low efficiency and bad visual effect. This thesisis to improve these problems, and takes shorter time as possible as to reconstruct modelwhich has sense of reality and high accuracy. This thesis research contents are:SfM(Structure from Motion) based on the basic principle of binocular stereo vision andmodel appearance recovery based on image texture information. Primarily, thediscussion and the research which it carrys out include three aspects:For the camera calibration and feature matching: in SfM it needs image initial focallength, distortion coefficient. Through the research of camera calibration principle itcompletes the internal parameters calibration experiment using camera self-calibrationmethod based on OpenCV; It choses a good SURF detector that has high robustness andefficient by comparing to SIFT and SURF detector performance. In the image matchingprocess, it puts forward an improved normalized weighted eight-points algorithmcomputing fundamental matrix, and verifys the effectiveness of the eight-pointsalgorithm by the experiment.For the SfM: It uses the variable adding images method to recovery scene depthbased on the triangulation method. It computes the scene deep by the choosing eligiblefirst image pairs, after that, it computes essential matrix by five-points algorithm andelects the method of eliminating incorrect solutions. It computes camera externalparameters in the adding images process; With the SfM process, it puts forward a kindof improved feature matching reconstruction strategy, and it confirms that the strategyhas a higher efficiency by the experiments.For the model visualization: It recovers model appearance using texturemapping(2D-3D) method based on image texture. The first step is to mesh the imagefeatures by Delaunay Triangularization Method. The second step is to project the imagetextures to3D model by the RGTM(Regional Grid Texture Mapping) Method, itincludes choosing the scene by the deep of scene, choosing best images by AngleMinimum Principle between the scene point to the camera position’s connection andscenery normal, and storing corresponding scene image texture. Finally, it shows thevisual model by the OpenGL.