Research On Parallel Algorithm Of Image Registration And Bundle Adjustment In Multi-view 3D Reconstruction

In recent years,with the development and maturation of computer vision,multi-view reconstruction with its low cost and convenient operation,has been widely used in digital city,medical imaging,virtual reality and other fields.Image feature extraction,feature matching and bundle adjustment are several key steps in multi-view 3D reconstruction,but there are also some problems.Firstly,the commonly used SIFT algorithm for feature extraction is not suitable for all scenes.Secondly,taking the huge advantages of GPU in computing capability and memory bandwidth,the multi-view 3D reconstruction algorithm based on GPU-accelerated become a research hotspot.However,the feature matching algorithm migration to the GPU platform is less,and because of the different hardware architecture,bundle adjustment parallel algorithms are greatly restricted by the reliability of parallel algorithm and the limited GPU memory.For this reason,the following studies have been carried out :(1)This paper presents a parallel algorithm of image registration which combines HarrisLaplace with SIFT descriptor,using improved Harris-Laplace to detect feature points of the image,which is invariant to brightness,rotation and scale change,then the feature points are described by SIFT descriptor.At the stage of feature matching,using birdirectional matching algorithm to get rough matching points,then epipolar constraint is used to get precise matching points.On the basis of analyzing the parallelism of the algorithm,a new CPU_GPU co-processing approach is put forward.Taking task character and transmission time into consideration,the task is reasonably divided between CPU and GPU,thus improving the efficiency of whole algorithm.(2)BA problems are often linearized by LM,then PCG algorithm is used to solve the normal equation.PCG makes it possible to solve normal equation without ever explicitly forming coefficient matrix in memory.Taking advantage of the characteristics,the procedure of PCG is broken down into a series of simpler matrix-vector products involving the jacobian.Then adding a filtering step in the preprocessing of BA parallel algorithm,excluding the wrong point caused by numerical conversion.On the premise of ensuring accuracy,making best use of high GFLOPs of GPU.Finally,based on the analysis of the relation between the jacobian transpose matrix and the original matrix,a parallel algorithm of bundle adjustment is proposed.With redesigning the parallelization of operation involving jacobian transpose matrix.There is no need to store the transpose of the matrix of the Jacoby when solving the bundle adjustment problem.From the experimental results,it turned out that GPU-accelerated image registration algorithm which combines Harris-Laplace with SIFT descriptor cannot only significantly increase efficiency,but obtain matches with high accuracy.Parallel algorithm of bundle adjustment proposed in this paper is able to achieve a good speedup.At the same time,the consumption of GPU memory is also significantly reduced.Benefit from the parallel processing capability of the GPU,the real-time processing ability of multi-view 3D reconstruction can be met very well.