Research On Registration Of Multi-view Range Image From Binocular Stereo Measurement System

Multi-view measuring data from different views is necessary to get the complete 3D surface by Binocular Stereo Measurement System. The overlapings of the multi-view measuring data are then used to achieve measuring of global surface, which is the process of registration of multi-view meas-uring data. The methods of registration can be divided into coarse registration, fine registration of pair-wise and global registration. The coarse registration is used to get initial position of views for fine registration, and the global registration is the most important part to make sure measuring data can be used to practice. The binocular stereo measurement system which is called Recreator is devel-oped by our team to get the point clouds of 3D surface. This paper focuses on registrations of multi-view measuring data from Recreator to develop its ability of data processing, which includes a method of coarse registrations of feature points, a fine registration of pair-wise based on an improved ICP algorithm and an improved global registration algorithm based on the method of rigid body mo-tion.A method of coarse registration based on range image and SIFT algorithm is proposed here as registration of feature points. It translates the measurement data into the character image with curva-ture which is considered as characteristic. The character image reflects features of range image itself, and will not affected by environment. For measurement data without pixel information, a simulation scanning algorithm is given to translate the measurement data into the range image, which is based on initial data and keeps the normal changing same. Two different functions of feature measurement are discussed and compared in this paper to get a better registration result.About the fine registration of pair-wise based on an improved ICP algorithm, comparation among different measurement standards in different improved ICP algorithm and analysis between different ways to solve matrix for registrations are done to elicit the improved algorithm of this paper. The number used for registration, overlap ratio of multi-view measuring data, initial motion and the normal are studied to find how they will affect the registration accuracy, which are useful for applica-tion.An improved registration algorithm based on the method of rigid body motion to achieve the global registration of multi-view measuring data is proposed in global registration. It is deduced from fine registration of pair-wise by the model on data sets'forces and motions. The algorithm sets up the model on data sets'forces and motions which are caused by virtual springs via analyzing forces and motions of rigid body caused by real springs. That translates the issues from registration of data sets to motions of rigid body. Threshold of corresponding points is used in this paper to improve precision of global registration after discussing how to find the closest point with the measurement standards and how to compute the registration matrix and how to subsample points for registration. Compared with original algorithm, the accuracy of registration are much improved.Experiments on a few groups of multi-view measuring data are completed to verify the improved registration algorithm.