Research On Point Clouds Registration In 3D Model Reconstruction Pipeline

Three-dimensional reconstruction is an important topic in computer vision, which has wide application prospects. Typically, in order to obtain the 3D model of an object, it is necessary to acquire the surface data from different views as the limitations of the observing. Therefore, the registration among individual views is particularly important in 3D model reconstruction pipeline, as it is directly related to the accuracy and automation of 3D reconstruction.Nowadays, the most popular registration algorithm is the ICP algorithm, which utilizes iterative method to approach the best result. But the convergence speed is still slow and the ICP algorithm may not converge to a global minimum. When there is much difference between the initial configuration of the data sets, the ICP algorithm may finds a wrong result. Therefore, the initial alignment of data sets is a key point in 3D registration.This paper focus on 3D point clouds registration algorithms in the pipeline of 3D model reconstruction. We select the RANSAC algorithm based on affine invariant as the coarse registration algorithm. Based on the algorithm, we introduce a simple method in extracting the 4-Points congruent sets, and improve the method in finding the consensus sets to a higher alignment speed. Experimental results show that the algorithm is fast and robust. During the process of ICP registration, we choose different strategies in various stages of iteration and compose a simple and fast ICP algorithm. Experimental results show that, within a relatively short number of iterations, the algorithm can get a high accuracy. When the ICP algorithm combines with RANSAC algorithm, the registration is high-speed and robust.In the experimental method, we developed and implemented the registration software based on RANSAC algorithm and ICP algorithm. The software is easy to use and will give us a strong support in the theoretical research.