Research On Registration Algorithm Of 3D Point Cloud Data

Three-dimensional(3D) reconstruction is an important topic in computer vision, and has broad application foreground. The 3D data of the target object must be obtained before conducting the 3D reconstruction of the target object. The 3D laser scanning technique is a newly-developed method which can be used to capture the 3D point cloud data of the target object quickly and accurately. However, it is impossible to obtain all the 3D point cloud data of a complex object in one time due to the restrictions caused by observing environment, instruments and the shape of the object. Therefore, first the 3D laser scanner should be used to capture the 3D point cloud data of the target from different perspectives. Then, splice the data acquired from different perspectives, and conduct the fusion of all the data in the same coordinate system, achieving the construction of the complete point cloud model. Registration of 3D point cloud data is a key technology in 3D modeling and it directly affects the final synthesized result and model precision. Nowadays, the most popular registration algorithm is the ICP algorithm, which acquires the optimal result iteratively. However, the convergence speed is rather slow and the ICP algorithm may not result in a global optimal result. In addition, when there is too much difference between the initial positions of the data sets, the ICP algorithm may result in a wrong result. Thus, in the process of the registration of 3D point cloud, feature point extraction, selecting matching point pairs and determining the initial transition matrix are necessary to ensure the accuracy before using the ICP algorithm.In this paper, the problems encountered in the process of the geometrical feature extraction and registration of point cloud data are studied, and the optimization of 3D point cloud data based on geometrical features and the ICP algorithm is realized. The main research is as follows. Firstly, fit the surface through the target point and its neighboring points and extract the feature points of the 3D point cloud data through the variation of the geometric curvature of surface to compose the feature point set. Secondly, select the initial matching point set from the feature point set based on the Hausdorff distance condition, and then obtain the accurate matching point set and the initial transformation matrix by using the Random Sample Consensus combined with the distance constraint condition to remove the error points in the initial matching point set. Thirdly, utilize the optimized solution of the objective function and the ICP algorithm to determine the final transition matrix of the rigid body, completing the registration of the 3D point cloud data. The experimental results show that the method studied in this paper can result in higher registration precision.