Research On Image Matching And Three Dimensional Scene Structure Reconstruction Method

Reconstructing 3D models from 2D images is an important task in computer vision.It is widely used in areas such as education,medical care,entertament and public safety because of easy data acquisition,low cost and flexibility.Incremental Structure from Motion(SfM)is the most important way to reconstruct 3D models from unordered image set.It contains two key ingredients:image feature point matching and 3D structure recovery.Image feature point matching,which is the basis of 3D structure recovery,is evaluated by precision and the number of correct matches.However,they are ususall affected by repeated texture or the rotation,illumination and viewpoint difference between two images.Poor matching result will further deteriorate the 3D reconstruction performance.3D structure recovery computes camera poses and 3D point positions.But this problem is still challenging due to large data size,unknown overlapping and uneven distribution of unordered images.This paper studies both how to increase the number of correct matches while ensuring high matching precision and how to improve reconstruction speed while preserving reconstruction accuracy.First,a Non-Uniform Gaussian Mixture Model(NGMM)based image feature matching algorithm is proposed.It treats feature point sets on two images as the centers of Gaussian Mixture Model(GMM)and the data,respectively.Then a structure preserving spatial transformation which maximizes the posterior probability is computed.When constructing the GMM,NGMM assigns different weights to each component according to feature similarity between two points,resulting in a unified framework incorporating both spatial consistency and feature similarity constraints.On one hand,local feature information is used to guide the computation of spatial transformation between two point sets.On the other hand,global structure information can avoid locally optimal matching result.In this way,not only the artifacts caused by feature ambuity is reduced,but also the robustness against outliers,rotation and deformation is enhanced.Consequently,the proposed method can find more correct matches while preserving high accuracy.Next,a new coherent subspace model is proposed to find quasi-dense matches between two images.The discrete 2D disparity gradient model and local affine transformation model used by existing quasi-dense match expansion methods could not handle large rotation and non-smooth surface correctly.The new method will first compute feature descriptors for each pixel in the expanding window around a seed match.Then new coordinates of these points are generated by fusing both feature and coordinate information of these pixels via subspace mapping.Next,a coherent non-rigid transformation in the subspace is computed to find the matches between two point sets.This new model is more robust to image geometry difference,and can handle more complex scenes without the assumption of approximate planar scene surface.Besides,a preemptive matching scheme utilizing matches found in the earlier stage to guide the searching for the remaining matches is also proposed to further improve reliablility and effciency.Finally,a multiple starting points selection and data partitioning method for unordered image structure from motion is proposed.First of all,it finds several kernels and a starting point image in each of them at places where images are densely distributed.Each kernel,consisting of a group of connected,largely overlapping images,will be used to reconstruct a base model of the scene.Next,several image clusters are obtained by clustering all the non-kernel images according to their optimal reconstruction path length to the starting point images.Non-kernel images in an image cluster are called leaf images,which are further divided into several balanced leaf image clusters.Each leaf image cluster will be indenpendently added to the base model in parallel.The proposed starting point selecting method ensures that the scene is reconstructed from places with dense images to places with sparse images,avoiding large accumulative error caused by passing 3D structure via weak overlapping images.Data partitioning is more relevant to 3D reconstruction by considering factors such as starting point and reconstruction path.Since all kernels and leaf image clusters can be reconstructed in parallel,the speedup is remarkable.