Research On Some Key Techniques Of Image-based3D Reconstruction Of Close-range Objects

3D reconstruction of close-range objects is a complex integrated technology in digital photogrammetry and computer vision, and can be used for national production industries. This paper selected a number of key technologies in three typical problems,3D reconstruction technique based on structured light,3D reconstruction techniques based on silhouette, and registration of geometric model and multi-view images, for in-depth study. The specific research contents and results are as follows:(1) The key technologies of structured light scanning under single-frame projection. This paper completed the design and characterization analysis of the pseudo-random pattern projection, the analysis of global unique property of neighborhood window along the epipolar line direction, proposed a fast image matching method SEEM, including the seed point matching of epipolar images, the fast image matching based on region grow, the wrong point detecting method and the optimization of the matching algorithm. Through the experimental data, we presented an assumption that the correlation coefficient of two random sequences of length n obedience the normal distribution with the mean is0and the standard deviation is1/(?)n, under this assumption, using a11×11window in seed point image matching to ensure uniqueness in the entire epipolar line, use a5×5window in region grow image matching to ensure uniqueness in a small area. Once a incorrect matching occurs, the region growing algorithm will quickly reach the boundary depending on the nature of the pseudo-random projection pattern, which is the key to remove the incorrect matching point. When the scanning distance is about600mm and the baseline length is about276mm, the absolute accuracy of the point cloud is0.185mm, the relative accuracy is one of3200, and the image point accuracy is about0.3pixels. With the CPU is Core2T5850, RAM is2GB, the image matching algorithm achieved a speed of about400,000points per second.(2) Multi-view images based3D reconstruction using silhouette. This paper completed the semi-automatic extraction of objects’silhouette based on graph cuts, smooth of the silhouette image, quadtree forest compression and storage of silhouette image, generation of the voxel model and the surface mesh model, and optimization of the surface mesh model. Different weighting methods get different results of graph cuts, but no weighting method has obvious advantages. In order to minimize the workload of human interaction, we should try to choose the background that has large difference in brightness or color with the object. Because when generating the voxel model or the surface mesh model, we need all the silhouette images simultaneously read into memory (each image about10MB), using the quad-tree forest to store the silhouette images, binary image, can be reach the compression ratio of100-300times, and access quadtree leaf nodes need to search drill down from the root, the best situation is accessing one layer, the worst case is nine layers. For the rounded transition surface, the shooting angle of the image reconstruction of the arbitrary nature of its minimal effect on the subjective appropriately increase the number of images can be finer results. For larger plane, it is easy to form a triangular protrusion only when an image of a photographic center located just at or near the plane, we can eliminate this phenomenon, to be targeted to select viewing angle shooting. Recessed portion for the existence of this chapter methods from the theory and practice can not get a good reconstruction results; surface used for the experiment in this section are lack of (or partial lack of) textures, image matching methods can not be used to obtain a better reconstruction result, is not using structured light scanning (or other similar under the technical means) premise of this chapter is the most appropriate method.(3) Registration of the geometric model and multi-view images. This paper completed the coarse registration of the geometric model and multi-view images (absolute orientation), the fine registration based on mutual information, including OpenGL rendering method for generating renderd maps, the area selection of the geometric characteristics of significantly in renderd maps, the statistics of the gray joint histogram, and the optimization of the registration parameters using Powell’s method. In coarse registration, we need to differ the images into different groups based on the relative relationship between the object and the background, each group of images in the same coordinate system. Between image coordinates and geometric model group coordinates, between different images of different groups coordinate system with different registration passing methods, select corresponding points in the manual, you can choose different strategies according to the actual situation, and ultimately all of the images groups are registered to coordinate geometry coordinate system. In the fine registration, the illumination light in different directions OpenGL selected area will affect the final histogram registration result, the direction of light closer to the real image, the higher the accuracy of registration, the only statistical characteristics of the impact is greater than the geometric texture affected area will get better registration results. The object having the mirror reflection characteristics, the gradation of the image depends on the geometric changes in the curvature of the object, after the optimization based on the mutual information, the accuracy is improved registration parameters. In the next image is just a set of images of exceptional circumstances, rounding scaling factor, only six parameters are optimized, you can still get a fine registration results. The rough registration based on absolute orientation requires a small amount of of human interaction, each image group needs at least three pairs of corresponding points, each group spends within5minutes; while the fine registration based on mutual information also has acceptable time-spending.