Research On The Uncertainty Of 3D Point Cloud Reconstruction Based On Image Sequences

Image contains not only color,size and other attribute information,also rich spatial information,such as position,shape and topological relation.The image can provide timely,comprehensive and reliable spatial data for GIS,thus becomes an important spatial data source of GIS.With the development and popularization of digital camera,cell phone and the other camera equipments,especially the resolution improvement of the image sensors,the spatial information included in the images becomes more abundant.Therefore,rapid extraction of spatial information and construction of 3D model from the image have gradually become the research trend of 3DGIS spatial data acquisition and update.In computer vision,there is currently more concern about extraction and construction 3D model based on the image,however less attentions are paid on the accuracy of 3D model.So,the uncertainty of 3D model reconstruction from images sequence should be studied based on the principle of binocular vision.The accuracy of 3D point cloud reconstruction from the sequence image has a direct influence on the accuracy of 3D model reconstruction based on the binocular vision.In this research,3D point cloud reconstruction process based on binocular vision was studied using images sequence as data source.Based on the above,the impact factors in 3D point cloud reconstruction process based on image sequences was analyzed,and the uncertainty model of 3D point cloud reconstruction and its quantitative method of uncertainty were studied with uncertainty theories.The contents and results of this research include the following parts:(1)The 3D point cloud reconstruction model based on images sequence is introduced.This research studies the key technology of 3D reconstruction based on the principle of binocular stere vision,and also evaluates the advantages and disadvantages of these technologies.Based on the above results,the SIFT feature matching,the RANSAC fundamental matrix estimation and 3D point cloud computed by essential matrix,technologies are chosen as fundamental methods of the uncertainty research of 3D point cloud.(2)The uncertainty model of SIFT feature matching is proposed.Firstly,based on the model of SIFT feature matching,the uncertainty theoretical model for SIFT feature matching is constructed using SIFT impact factors.And then,based on uncertainty theories such as probability theory,mathematical statistics and Monte Carlo method,the uncertainty of SIFT feature matching can be quantized.In this research,mismatch rate of feature points and transfer error of correct matching points are chosen as the evaluation uncertainty criteria of feature matching points.Through the analysis,transfer error mainly falls into the region that the image gradient changes small.And,Compared with the average transfer error,the impact factors of SIFT has a large effect on mismatch rate of feature points.(3)The measurement method for reliability of the fundamental matrix estimation using the geometry of the feature matching points is proposed.The geometry of the points has a great influence on the estimation of basic matrix,so Horizontal Dilution of Precision in satellite navigation and positioning system is introduced to study the influence of the points geometry on the fundamental matrix estimation,in order to improve the accuracy and reliability of the results.By constructing the mapping relation between the geometry of the points and the fundamental matrix estimation,the reliability of the fundamental matrix estimation using the geometry of the feature points is studied.The results indicate that the more evenly the feature matching points are distributed,the higher the reliability of the fundamental matrix estimation based on RANSAC is.(4)The uncertainty model of the fundamental matrix estimation based on RANSAC is proposed.Covariance matrix and Mahalanobis distance are used as the index of the fundamental matrix uncertainty.Based on that,Monte Carlo method is used to quantize the uncertainty of the fundamental matrix estimation based on RANSAC.And through the experiment we can see that it is unnecessary to consider the influence of impact factors of RANSAC on the uncertainty of fundamental matrix.(5)The uncertainty model of 3D point cloud reconstruction based on images sequence is proposed.Combined with the matrix differential theories,the uncertainty propagation of feature matching points,fundamental matrix and intrinsic parameters of camara,are analyzed using GUM,and the measurement of 3D points uncertainty during 3D point cloud reconstruction is implemented.Meanwhile,the shape of the uncertainty of 3D point cloud reconstruction based on images sequence is determined,which is a 3D ellipsoid whose uncertainty along different axis is related,and with bigger uncertainty along depth direction.