A Research On Multi-view Stereo Reconstruction Algorithm Based On Random Search And Quasi-dense Matching

With the growing up and development of computer vision technology,the image-based 3D reconstruction technology has been paid close attention by scholars and it has been becoming a research hotspot gradually.Compared with device-based 3D reconstruction method and geometric modelling method,the image-based 3D reconstruction method only needs image information of the reconstruction scene.This method has been widely applied in live entertainment,historic preservation,medical field and urban planning for the case of its advantages with low cost,easy operation and so on.In this paper,explore on the patch-based multi-view stereo(PMVS)reconstruction algorithm with the best comprehensive performance in recent years is carried out deeply,two problems of PMVS are focused on,one is local optimization in the parameter space and the other is over-sparse phenomenon of initial point cloud.The geometrical parameter optimization for spatial patch is an important part of PMVS algorithm,and the conjugate gradient method used in PMVS is easy to plunge into local optimum.With the global optimization ability of the random search algorithm,it can effectively solve nonlinear and multimodal optimization problem.Therefore,the random search algorithms are employed to improve optimization performance of the spatial patch geometric parameter of PMVS in this paper.Firstly,the principles of classical random search algorithms are introduced,and three optimization models of geometrical parameter for spatial patch are constructed based on particle swarm optimization(PSO),differential evolution and genetic algorithm respectively.By comparison and analysis on optimization performance,the PSO-PMVS algorithm based on particle swarm optimization is proposed.Simulation results show that the reconstruction results of PSO-PMVS have an improvement in completeness and accuracy compared with the classical PMVS algorithm in the relevant databases.Due to the over-sparse and rough accuracy of initial point cloud,the PMVS algorithm causes the accumulation of reconstruction errors in the expansion of point cloud.In this paper,a 3D reconstruction framework is proposed which is based on quasi-dense matching,neighborhood optimization of 3D parameter space and random search optimization.By propagating sparse matching in neighborhood,the quasi-dense matching can overcome the problem that the initial patch is few and improve the reconstruction quality of the initial point cloud.The neighborhood optimization of 3D parametric space uses the neighborhood information constraint between the spatial patches to complete the initial optimization of the geometric parameters of the patch,at that time it can fill the residual point cloud holes of quasi-dense matching.Finally,the quality of dense point cloud is further improved by random search and view propagation,the visibility consistency constraint and local continuity constraint are used to filter the exterior points,outliers,and redundant points to complete the whole reconstruction.Compared with PMVS algorithm,the simulation results show that the reconstruction framework proposed in the paper has an effect improvement in the reconstruction performance and visual effect.