Research On 3D Point Cloud Automatic Stitching Algorithm

With the continuous upgrading of computer technology and the development of hardware equipment, the process data ability of computer has been continuously strengthened,making it possible to handle the information such as three-dimensional data that cannot be process before. Thus also promote the development of three dimensional data processing technology and extend it applies to daily life practice. For example: 3D film production,medical image processing, reverse engineering, heritage protection work, 3D printing equipment manufacturing, and so on. Among them reverse engineering technology is one of the widest used technologies and it is also one of the essential technologies of modern industrial design and manufacturing.Point cloud registration technology is an important and difficult part of reverse engineering. The fit degree between point clouds will directly affect the accuracy of the overall three-dimensional model, therefore point cloud registration technology become a hot spot in the field of academic research. For the point cloud acquisition device lens angle and the shape of the object itself limited, measure equipment is difficult to obtain 360 degrees total data of the object. To obtain whole three dimensional point cloud data, it is usually practice for many times form different points of view and through the point cloud registration algorithm to splice each view cloud data together for a complete point cloud data. The essence of point cloud registration is to move point cloud under different coordinate systems of same object to a common coordinate system. The point cloud registration algorithm can be dividing into manual registration and automatic registration, and commonly we referred to point cloud registration algorithm is the automatic registration algorithm. Automatic registration algorithm can be divided into local registration algorithm and global registration algorithm,the global registration algorithm is usually highly registration precision and a better anti-noise performance. But it takes much more time and often require two point cloud has a better initial position. The representative algorithm of global registration algorithm is ICP algorithm and genetic algorithm. Local algorithm is fast but less precision, commonly used algorithms are a class of algorithms based on geometric features. Both algorithms have their advantages and corresponding defect. The current mature algorithms that are algorithms combine with two types of algorithms, that is, coarse registration to provide an initial value by the local algorithm and precision matching is responsible for improving the final registration accuracy.This article begin with the basic knowledge for the point cloud registration, detailed derivation this content used in this article, such as singular value method algorithms and four elements theory for calculating the transformation matrix, various rigid body transformation as well as kd-tree search method that used most frequent in the point cloud data processing.The classical coarse registration algorithm and the fine registration algorithms are analyzed and discussed in this thesis. The detail derivation of each algorithm is given and the applicable scope and limitations of various registration methods has pointed out in the article.After a large number of excellent algorithms are summarized, this thesis designs a new point cloud registration method that based on surface feature. This method selected key points based on multi-neighborhood curvature difference value, which can remove noise points effectively and provide high quality key points for the subsequent process. Curved surface fitting coefficient is used as point cloud feature descriptors which can access a large amount of local point cloud surfaces information quickly and has the advantage of highly distinguish.After the initial registration, distance between the correct correspondence relation and the error correspondence relation have large difference. So we can use the character to eliminate the error correspondence by set a threshold value, which obtained a good result.At last, this thesis program used C++ language. Experiments are carried out by using the Standford 3D point cloud model, which verifies the correctness and efficiency of the algorithm proposed in this thesis.