Research On Registration Algorithm Of Three-dimensional Point Cloud

With the development of 3D scanning technology, the application of 3D Reconstruction technique gradually extended from the field of robot navigation to reverse engineering, virtual reality, and medicine. 3D point cloud registration is the key in the process of the reconstruction. Due to the measurement equipment, the shape of object and the affects of environment, the point cloud data acquired by each measurement is only part of the object. In order to obtain the overall point cloud data, multiple measurements from different angles are required, but point cloud data obtained by each measurement is in respective coordinate systems. The process of obtaining a complete point cloud data by adjusting many point clouds from multiple different coordinate systems to the unified coordinate system is the point cloud registration.This paper focuses on registration algorithms for 3D point cloud data, and proposes initail registration and precise registration method with high precision. The main tasks are as follows:First, aimed at the existing initial registration algorithms have the problems of lower precision and the limited scope of application, this paper presents a initail registration method based on geometrical features and RANSAC. The algorithm establishs the concept of the curvature value of local neighborhood, and extracts key points of two point clouds based on this geometric feature. The process of registration is based on RANSAC. In each sample, this algorithm uses FPFH to search corresponding points, and improves the accuracy of the correspondence relation according to the invariant constraints in rigid transformation. After multiple samples, this algorithm chooses the optimal transformation relying on the degree of consistency of two point clouds. This algorithm also uses multi-threading mechanism to accelerate the estimation process of geometric features, which can improve the efficiency of the algorithm.Then, aimed at the existing precise registration algorithms have the problems of slowly convergence and easily falling into local extremum, this paper proposes a precise registration method based on the geometric features of local neighborhood and the iterative thought. The algorithm gives a method for calculating the subset point cloud, which is located in the overlap region of the two point clouds, and the precise registration progress is executed in this two subsets. The process of precise registration uses iterative thought. In the process of iteration, this algorithm gradually adjusts the position between two point clouds. In each iteration process, this algorithm search initial corresponding points according to geometric features and the nearest point search method, and combines 5-dimensional description and invariant constraints in the rigid transformation to remove the wrong points in correspondence relation, which can improve the accuracy of correspondence relation of two point clouds and speed the rate of convergence.Finally, this paper verifies the validity of the registration algorithm proposed by this paper from registration errors, the algorithm efficiency and the speed of convergence,and shows the experimental results of registration algorithm.