Research On PatchMatch Based Multi-View 3D Reconstruction

Multi-view 3D reconstruction uses pictures of a static target with different perspectives to estimate the 3D shape and surface texture of the target.Compared with active methods such as laser scanning and structured light,it has the advantages of low cost,easy operation and wide application scenarios.In the past ten years,with the success of Structure from Motion algorithm and the development of hardware equipments(such as the popularity of high-precision digital cameras and the improvement of computing power),the research and application of multi-view 3D reconstruction have made great development.For example,we can use the pictures on the internet to do a sparse reconstruction in city-scale.Despite this,the robustness,completeness and accuracy of the related algorithms remain to be improved to adapt to more complex applica-tion scenarios,such as the reconstruction of the target that has repetitive texture or is textureless.In this thesis,we study on the related technologies of multi-view 3D reconstruction with unordered image set as input,and improve some related algorithms which improve the robust-ness,completeness and accuracy of 3D reconstruction.The main research achievements are as follows:·A GPU-accelerated SIFT algorithm is implemented,which solves the problem of time-consuming keypoints extraction in SIFT.The experiments show that the GPU-accelerated SIFT algorithm greatly accelerates the keypoints extraction without compromising accu-racy.At the same time,compared with the SiftGPU algorithm and CudaSift algorithm also implemented by GPU,our realization has the advantages of higher speed or higher accuracy.·A hybrid SFM algorithm with high robustness and precision is proposed and imple-mented.First,we get the global rotations of the cameras by rotation average.Then,under the constraint of global rotations,incremental SFM is performed.The algorithm alleviates the problem of cumulative errors in incremental SFM while retaining the high robustness of incremental SFM.·Improve the patch-based multi-view stereo(MVS)algorithm.Using the plane of 3D points,we can refine the position of dense 3D points and performe the holes filling,which improve the realness and completeness of the reconstruction result.