Research On Key Technologies Of Dense 3D Reconstruction Based On Image Sequence

With the development of computer vision and hardware devices,3D reconstruction has been widely used in many fields,such as medicine,film and television,games and virtual reality,and is playing an increasingly important role in life.Many excellent algorithms have emerged in response to the growing demand for different 3D geometric models,but the existing solutions cannot meet all the requirements.For example,user-driven modeling,simple but long and error-prone,3D scanners are precise but expensive to use.Among them,the method of 3D model reconstruction from a series of 2D images has been widely used due to its simple implementation,low cost and suitability for the reconstruction of most scenes.The sparse point cloud obtained by the image-based 3D reconstruction is too small to meet the needs of the follow-up work.Therefore,it becomes the research focus to obtain as many uniform dense point clouds as possible.The sparse point cloud obtained by image-based 3D reconstruction provides too little information to meet the needs of follow-up work.Therefore,it becomes the research focus to obtain as many uniform dense point clouds as possible.Although there are many existing 3D dense reconstruction algorithms,many problems and shortcomings still exist in some application scenarios.This paper studies the techniques and algorithms involved in the process of dense reconstruction from multiple images.In-depth research on the PMVS algorithm with better performance and wide application,combined with previous research experience,the following achievements have been achieved:1.Study on the feature detection algorithms,studies the commonly used algorithms according to the different types of features,then analyze and compare their advantages and disadvantages.An improved SIFT feature detection algorithm is proposed to solve the problem that feature points are missing or sparse in texture sparse regions: before the threshold filtering the bilinear interpolation of the extreme point is calculated according to the original SIFT algorithm to make the extreme points selected which gray value is less than the threshold where the local gray value changes significantly and increase the number of features.2.Study on 3D dense reconstruction algorithms,and study the correlation technology of PMVS algorithm in detail.In the initial feature matching stage,a candidate matching point selection strategy is proposed to convert the two-dimensional projection distance into three-dimensional space distance,so as to improve the accuracy.In the phase of patch expansion,a patch optimization method is proposed.In this method,local geometric information is used to modify the normal vector of the patch,so that it can better fit the model surface,and the accuracy can be improved as much as possible without affecting the global features.3.Through the research of camera calibration,feature point extraction and matching,sparse reconstruction,dense reconstruction and other methods,the system that can complete dense point cloud reconstruction through image sequence is realized.