From Image Sequence to 3D Reconstructed Model

Abstract

Obtaining 3D model based on images by simple devices is convenient and low-cost, meanwhile the model generation is automatic. The paper proposes a method to extract the 3D point cloud of object surface by a sequence of images. To obtain sparse 3D point clouds, a highly precise method of camera self-calibration is proposed. The self-calibration method is based on the bundle adjustment and uses a localglobal hybrid iterative optimization. Meanwhile, we propose a neighboring image matching strategy to solve the problem in the multiple image matching, which can improve the matching speed and preserve the matching accuracy. Then, dense 3D point clouds can be obtained by our improved Patchedbased Multi-View Stereo(PMVS) algorithm. Finally, we adopt 3D mesh construction method based on Possion distribution. The experimental results show that our algorithm increases the number of triangulation patches by 2.5% 27.9%, reduces the operation time by 5.0% 20.7%, and decreases the cross and mismatch of triangular patches in most cases. It is convenient and efficient to print out 3D models with richer details.