Research On 3D Reconstruction Application System

3D construction for objects means to create mathematical models for 3D objects which are suitable for computer to represent and process. It is the base for data process, management and analysis in computer environment. It is also the key technology in creating virtual reality to express the external world. So 3D reconstruction for objects is the commom science problem and key technology in many domains, including computer aided geometric design (CAGD), computer graphics (CG), computer animation, computer vision, medical image processing, science calculation, virtual reality and digital media invention, etc.There are two kinds of approachs to generate 3D representation for objects. One way is to interact with geometric modeling software to generate 3D geometric models for objects according to manual design. The other way is to obtain range data from real objects. The former is more mature, and has rich software support, such as: 3DSMAX, Maya, AutoCAD and UG, etc. They describe objects with curves and surfaces defined by mathematic expressions. Generally, the latter approach is called a 3D reconstruction which means to resume 3D information of objects from their 2D projection images. It involves several function modules, such as data obtaining, data preprocessing, data cloud re-assembly and geometric feature analysis. Our work is to construct a whole 3D reconstruction application system, including 3D modeling module, 3D fragment re-assembly module, 3D face recognition module and 3D shaping module based on 2D editing. Simultaneously, we make a fusion of stereo vision and 3D reconstruction. Among them, the 3D reconstruction and range data point cloud feature-analysis seem to be more important which are addressed in this paper as key points.In this paper, we organize some research work done with support of "Research on key problems in 3D representation based on mass data" by the National Natural Science Funds and "Research on problems in accurate 3D representation during mass data processing" by the Shandong Province Natural Science Funds. The key research points include: the process analysis of 3D scanning, 3D fragments automatic reassembly, geometry invariant of 3D range point data cloud and the merge of stereo vision and panorama stitching, etc. Specifically, they are listed as follow.First, a whole structured white light based untouching 3D scanning system is implemented. It and can provide rich and accurate data for 3D reconstruction and other related work. Common lens has greate geometry distortion, which can affect the quality of scan data badly. And the edge detection algorithm will also determine the accuracy of range data. Here, an iterative method is carried out to perform the automatic lens calibration. Then a data-fitting based method is proposed to improve the accuracy of sub-pixel edge detection. Finally, the visualization of the data cloud is given.Second, we present an automatic re-assembly method for matching 3D fragments, which can compute the best match of a pair of broken pieces. To improve speed and accuracy, we give a torsion estimation method based on anti-noise section and a fixed constant time cost 3D overlapping test method. First, the contour curves of fragments are found, then the corner points. Contour curves are divided into several sub-contours. Second, the torsion sequences are obtained for the DDTW based matching work. Thus, the scale and translation matrixes can be computed to change the size and position of one fragment. Best matching result will be finally chosen according to matching test and extra evaluations. Experiment results show that this method is simple and robust, and can get proper matching results quickly.Third, the expression ignored invariant of 3D face scan data is discovered during the feature analysis on 3D range data point cloud. This makes it possible that face surface data with different expressions can share a uniform data presentation. This can help to identify human face no matter what expression he makes. Then a 3D human face database is setup and serves a practical face recognition system. These efforts provide new thoughts and fashions to identity recognition and security guarantee applications.Fourth, an automitic panorama stitching method is carried out. An image transformation based method is proposed to provide automatic stitching for fish-eye panorama pictures. Thus, with a fish-eye lens and a common digital camera, people can get a whole panorama picture easily, which can cover 180 degrees vertically and 360 degrees horizontally. Then the capture position is changed. So a pair of pictures with different view angle is taken. The stereo vision algorithm is also combined in the application to obtain the 3D range data cloud of the scene. This work is very useful for scene reservation and review.This research can also be regarded as a new branch of 3D reconstruction.Finally, to provide additional ability to manipulate 3D shape, we implement a 3D editing method base on 2D drawing. 3D shape can be defined in 2D domain in simple and general ways, and then create a mapping from 2D domain to 3D domain, thus the shape definition can be presented on the surface of the 3D geometry model as 3D shapes.