Point Cloud Registration Based On ICP Algorithm Research

Nowadays,reverse engineering has been widely used in various fields due to its high development efficiency and ability to reflect the surface information of objects.As the core technology,point cloud registration has always been a hot topic of research.In the practical application of point cloud registration,due to the influence of the scanning angle problem on the surface of the object,such as occlusion,light intensity and noise background,we have to scan the scanned object many times and scan the point cloud after each angle.The data is registered to unify the point clouds in different reference coordinate systems into the same coordinate system.Point cloud registration is generally divided into initial registration and precise registration.However,there are still two problems in the registration process: on the one hand,the initial registration process using the FPFH(Fast Point Feature Histograms)algorithm requires manual attempts to select the neighborhood radius of the point cloud.Guarantee efficiency and effectiveness.On the other hand,although the traditional ICP(Iterative Closest Points)algorithm can basically complete the registration process,it takes a long time and is easy to fall into local optimization.In response to these problems,this paper has made some improvements to the current algorithm.The specific research contents are as follows:(1)This paper introduces several initial registration algorithms in detail.By comparison,the FPFH algorithm is selected for initial registration.When using the FPFH algorithm,it is necessary to manually try to select the neighborhood radius.The number,size and shape of the point cloud will affect the choice of the neighborhood radius.The different neighborhood radius registration effects are not the same,so choose a suitable one.The neighborhood radius is a key step in this algorithm.Based on the method of estimating the cloud density of the point,this paper proposes an algorithm based on the point cloud arc length density to automatically match the neighborhood radius.By counting the arc length density of different point clouds,and selecting multiple neighborhood radiuses for these point clouds,The Huber function calculates the error of each neighborhood radiusand preserves the neighborhood radius when the error is minimal.The above process is repeated,and a plurality of point clouds of different densities are sequentially tested to obtain an array of multiple sets of neighborhood radius and arc length density,and finally a function curve is fitted by using these arrays.If the unknown point cloud needs to be registered,only the arc length density can be calculated to automatically obtain the appropriate neighborhood radius according to the function curve,thereby eliminating the time for manually selecting the neighborhood radius and improving the efficiency of the whole process.(2)The ICP algorithm is the most widely used algorithm in the precise registration process of point clouds,but it also has some problems.When there are more point cloud data,the registration process takes a long time,and the ICP algorithm sometimes falls into local optimization.This paper proposes improvements to the precise registration process for these two issues.In this paper,we use the downsampling to streamline the point cloud,use the MLS(Moving Least Squares)algorithm to smooth the point cloud data,and use the RANSAC(Random Sample Consesus)algorithm to remove the noise of the point cloud,calculate the ISS(Intrinsic shape signatures)feature points of the point cloud data,and based on these feature point pairs.Accurate registration.This method greatly improves the efficiency of registration while ensuring the accuracy of registration.(3)This paper uses the point cloud in Stanford University's point cloud library to verify the improved algorithm,and uses the scanning device to collect data from the real sculpture,and uses the improved algorithm to register the data to test the actual application effect of the improved algorithm.The experimental results show that the proposed algorithm can be applied not only to the point cloud in the point cloud database,but also to the actual scene,which has certain application value.