Research On Multi-view Stereo Matching Algorithm Based On Unordered View

In recent years,3D reconstruction technology is widely used.Although a lot of equipment can directly extract depth information from the scene to rebuild the scenario model,these devices are not only easily affected by the natural environment,but have relatively complex operation and high equipment costs.Multi-view stereo matching algorithm based on unordered view can apply to the images data collected from the natural environment using some simple image-acquisition equipment,having a wide range of adaptation,therefore multi-view stereo matching algorithm based on unordered view is one of the particularly important research on 3D reconstruction.Recently there are a lot of excellent algorithms in the research of Multi-View Stereo matching(MVS),but there are general problems of long processing time and the low reconstruction precision.MVS algorithm is mainly studied in this thesis,some improvements are done towards processing time and reconstruction precision for current researches of MVS algorithm.Detailed studies and innovation works are as follows:1.This thesis firstly combines with the features of SIFT and SURF for view matching,improving the views matching precision.In the process of Structure from Motion(SFM),the combination of the two features improves the precision of the position parameters between cameras,at the same time increases the number of sparse points cloud,which not only increases the algorithm's stability,but reduces the probability of growing to error area in the process of regional growth.Therefore it can improve the accuracy of the eventual reconstruction model.2.This thesis studies the global view selection based on image level and local view selection based on pixel level.The proposed algorithm,firstly,adds decision conditions to the process of global view selection,and makes the number of the neighborhood candidate set adaptive,which not only boost the speed of the whole the reconstruction process,but improve the accuracy of the neighborhood view in the stereo matching process.In addition,this thesis improves the high accuracy of the local view selection by changing the epipolar lines weighting factor of the scoring function of local view selection.3.In the regional growth,the reconstruction of the MVS algorithm can get accurate results,still some errors at the border.According to research,MVS algorithm utilizes the fixed rectangular window shape,the window size affects the precision of the reconstruction.The proposed algorithm uses the adaptive window,so that the shape of the space plane projected is adaptive,which can capture enough information to improve the precision of the reconstruction.This thesis gives a brief introduction to the Poisson process below.In the experiment part,we select the templeRing data set in Middlebury database and some other image sets collected from natural environment,to testify the high performance of the proposed algorithm at running time and the reconstruction precision.