Fast Reconstruction Of Dense 3D Scene Point Clouds

Structure from Motion and vSLAM algorithm aim to recover a sparse 3D structure and estimate camera poses precisely.These methods track features from series of related images and optimize 3D structure and camera poses in a nonlinear optimization which incorporates the geometric multi-view constrains.This nonlinear optimization problem is normally solved by using bundle adjustment.However,sparse point clouds models do not provide enough details to appreciate the underlying structure of the environment.There have been various efforts towards automated dense 3D reconstruction in the last few years.Automated dense 3D reconstruction facilitates scene understanding and has countless applications in different areas such as augmented reality,cultural heritage preservation,autonomous vehicles and robotics.In this article,in-depth research in image-based 3D reconstruction is carried out firstly,designed an image-based sparse reconstruction algorithm to get sparse point clouds model of the scenario.The main works of this part include:Firstly,study on the pinhole camera model and projection relation between images.Secondly,study on image feature extraction and match algorithm,track feature points with ASIFT which is intensive and robust.Thirdly,research on theory of multi-view geometric,explain the concept of epipolar geometry,essential matrix etc.Fourthly,a detailed design of image-based sparse point clouds reconstruction,research on the optimization algorithm of both scene structure and camera pose.Based on the accurate scene point clouds,go on with dense point clouds reconstruction.A comprehensive comparison is made between all dense reconstruction methods and employ PMVS,which works good currently.Then,an improvement in PMVS is made to better fit our application environment.As PMVS is computationally intensive,even if an improvement is made,the execution efficiency of the algorithm is still low.To satisfy the purpose of fast reconstruction of dense scene point clouds,an image-clustering based dense reconstruction algorithm is proposed.Our method divides input images in to clusters,and remove redundant images,then reconstruct every cluster independently with the improved PMVS.To obtain better reconstruction results and improve algorithm efficiency,view clustering requires each cluster to satisfy coverage and size constrain.After reconstruction of all clusters is done,then merge the MVS points to get full model of the scenario.However,to merge directly has some problem,and there is a post-processing step later to remove erroneous points.At last,since the clusters are independent,then operate them in parallel,which Improves efficiency.