Research On Dense 3D Reconstruction From Unordered Images

Due to the advantages of non-contact,low-cost and being able to measure multiple points simultaneously,image-based 3D measuring technique has been widely applied in large-scale structural deformation measurement and large-scale motion measurement.However,these applications are only limited to the measuring a number of sparse points,acturally,denser reconstruction would be beneficial to both applications.On one hand,dense 3D reconstruction can provide more valid observations for large scale structural deformation measurement.On the other hand,it can provide more intuitive representation of the objects in large scale motion measurement.Besides,as the popularization of smartphone,and as the price of camera decreases,it is more convenient now for people to obtain digital images,also with the development of 3D printing technology,the desire of people for acquiring complete 3D scene model from readily obtained images has been stimulated,this also brings broad prospect in civil applications for image-based dense 3D reconstruction technique.According to whether images are sequential or not,the image-based dense 3D reconstruction technique can be further divided into sequential-image-based and unordered-image-based dense 3D reconstruction techniques.The latter is more generalized.However,because there is no time sequence information can be used,unordered images often face more stringent application conditions,such as a wider image baseline and more serious occlusions,which brings more challenging problems.In this thesis,three basic problems,i.e.,orientation of undered images,simultaneous optimization of both scene structure and orientation and dense 3D reconstruction,are studied in order to improve the accuracy and completeness of dense reconstruction results and the automation and robustness of the whole reconstruction process.Aiming at the problem of orientation of undered images,researches on control-point-based,coded-target-based and image-feature-based methods are respectively conducted in this thesis.Aiming at the problem of simultaneous optimization of both scene structure and orientation,a depth-based sparse bundle adjustment method is studied.Aiming at the problem of dense 3D reconstruction,a multi-view stereo method is studied.The main innovations of this thesis are as follows:1.Research on control-point-based engineering-practical single station calibration method is conducted.6 kinds of control-point-based engineering-practical initial value calibration methods and 2 kinds of parameter optimization methods are summarized and studied.A general calibration procedure and software is designed,and calibration algorithm selection guide analysis is conducted.2.Research on automatic network orientation method based on simple coded target is conducted.Aiming at the shortcomings of the existing coded targets,a simple coded target based on circular features is designed,and corresponding image detection and recognition algorithms are studied;Then automatic network orientation is completed through an image clustering process based on relative orientation.3.Research on automatic network orientation method based on natural image features is conducted.First,a robust procedure for acquiring one-to-one feature matches is designed;then a feature tracking algorithm based on global ordering of match numbers is proposed;A incremental Sf M procedure is designed,and automatic network orientation is completed based on the found feature tracks during the Sf M procedure;As for outlier handling,a semi-global and incremental semi-global outlier handling methods are proposed.4.A new depth-based sparse bundle adjustment method is proposed.A new object representation model based on depth relative to reference image is proposed,and sparse bundle adjustment methods based on this model are studied,corresponding sparse structures of the normal equation and covariance matrix are exploited.5.An accurate and occlusion-robust multi-view stereo method is proposed.The support window model is based on an approximate 3D support plane described by a depth and two per-pixel depth offsets.For the visibility estimation,the multi-view constraint is initially relaxed by generating separate support plane maps for each support image using a modified Patch Match algorithm.Then the most likely visible support image,which represents the minimum visibility of each pixel,is extracted via a discrete Markov Random Field model and it is further augmented by parameter clustering.Once the visibility is estimated,multi-view optimization taking into account all redundant observations is conducted to achieve optimal accuracy in the 3D surface generation for both depth and surface normal estimates.Finally,multi-view consistency is utilized to eliminate any remaining observational outliers.The proposed method is experimentally evaluated using well-known Middlebury datasets,and results obtained demonstrate that it is amongst the most accurate of the methods thus far reported via the Middlebury MVS website.Moreover,the new method exhibits a high completeness rate.