The Research Of Multi-View 3D Reconstruction And Evaluation Algorithm

3D Reconstruction is an important branch of computer vision. It has a wide range of applications in daily life, medical field, military measurement and so on.3D Reconstruction method can be divided into active method and passive method. The active method requires the measurement equipment to transmit the laser or the structured light, so the coverage area is limited, and it is suitable for the high precision measurement of the small scene. The passive method directly uses the camera to shoot the pictures, then use the 2D images of different angles for 3D reconstruct. It is both be used in the reconstruction of small scene and large scene reconstruction, and the equipment is more simple, so the method is currently a research hotspot.The paper according to the principle of the Patch-based Multi-view Stereo Algorithm (PMVS), programming realized a multi-view reconstruction algorithms. The algorithm firstly extracting the feature points from the 2D views of many different angles, then matching the feature points, and according to the principle of triangulation to obtained the sparse 3D point cloud. Then through the photometric discrepancy to patch growing and obtain the dense 3D point cloud. This paper use the scene of the actual shooting and the standard library for 3D reconstruct., all obtained better reconstruction results.The paper conducted experimental comparative analysis of several feature points detection such as Harris, SUSAN, SIFT and SURF, and feature point descriptor operator, such as SIFT, SURF and Daisy, and proposed a mixture of detection and descriptor operators. Then selected a group of good combination after experiment, using the combination to improve the PMVS algorithm. Experiments show that the improved algorithm are improved in the number of point cloud, accuracy and completeness.The paper proposed a fast evaluation method. It is essential to evaluate the performance of a 3D reconstruction algorithm. Since different coordinates have been adopted by different algorithms, the reconstructed scenes have different sizes, positions, and poses. As a result, they could not be compared directly with those accurately measured values provided by the standard data bases. This paper's method using the projection matrixes provided by multi-view database to fast estimation the coordinate transformation matrix from 3D reconstruction coordinate system to database measurement coordinate system. The transformation matrix only needs a little optimization, even does not optimize, and can be applied to statistical evaluation of the accuracy and completeness of 3D reconstruction. This method can be widely used for quantitative assessment of 3D reconstruction algorithm.