Research On Registration Algorithm Of Point Cloud Data

With the development of computer-aided design technology, reverse engineering technology which can construct digital model via physical model is more and more widely used because of its unique charm, at the same time, hardware which is becoming more perfect has provided sufficient technical support for the operation of digital model. In order to complete the data model of the objects, we need to identify a suitable coordinate transformation, after which data points collected from various perspective directions can be measured in a uniform coordinate system and form a complete Point Cloud Data, then the visualization operation is convenient, this is called Point Cloud Data Registration. The related technology has a widely range of applications in reverse engineering, surface quality testing and virtual reality.Point Cloud Registration includes manual registration, the registration rely on machines and automatic registration. Usually we are talking about refers the last one. The automatic registration technology is achieved through a certain algorithms or statistical regularity, computing the dislocation of two point clouds, so as to achieve an automatic registration result. In essence, it is to make a coordinate transformation between different data point clouds, in order to get the overall data model. The sticking point of the problem is how to confirm the transformation parameters R (rotation matrix) and T (translation vector), so that the distance between each group of 3D data measured in two perspective directions is minimum after transformation.Registration algorithm can be divided into the overall registration and local registration based on registration process, the advantage and disadvantage are discussed from the speed and convergence, a series of improving is introduced.First, the objective function based on ICP algorithm are improved, the point-to-point distance is replaced by point-to-surface distance. Thus a good convergence is gained. But the process of finding the nearest point on the surface increases the time complexity of the algorithm. Then, this paper presents a local registration method based on normal characteristic, improvement of iterative algorithm, local registration based on the normal characteristics, the corresponding registration points are gained through HCI method, then the precise registration points are searched in a local area, thus the unknown parameter is reduced to 1, that is the rotation around the normal direction. The algorithm gains the minimum of the objective function, the entire algorithm is simple and fast, the result is accurate.In addition, this paper proposes a local-selection registration algorithm based on multi-scale geometry describer, the geometric characteristics is measured through 3 kinds of geometry describer, and the multi-scale conception is introduced to measure the geometric characteristics in a larger area, thus the two-faced problem is solved effectively. The corresponding points and partial scope is more precise, the validity of the registration algorithm is guaranteed.Finally, this paper proposes a reverse registration algorithm based on curvature, first using hash function to carve up and quantify the curvature of the points, and then using the curvature value to gain the points, the pre-registration is completed using these points, then the accurate registration is put up. Experiments shows that the algorithm is more effective.Since the inherent defects of registration algorithm, the paper could not make an improvement that is rapid and accurate, compromise between speed and precision is needed, in fact, the requirements for accuracy is far higher than that speed, therefore, how to further improve the accuracy of registration, and make an acceptable time limit, will be the focus of our future study.