Research On Model Registration Method Based On 3D Point Cloud Data Features

In recent years,with the progress of industrial technology and the rapid development of 3D scanning technology,the acquisition of complete data model has become a major research hotspot,and point cloud automatic registration technology is also one of the hot issues that people have been keen to study.This paper focus on key point detection,feature description and other methods,combined with the idea of "first coarse,then fine",to carry out research on 3D point cloud data registration methods.First of all,in view of the point cloud data acquisition and calibration of stereo vision system is very critical step in the setting of parameters in calibration directly affects the accuracy of measurement,based on this,this article put forward to reprojection error,uniformity of the point cloud and scanning accuracy three parameters for the concept of point cloud data quality evaluation index,and explore the influence factors of 3D scanning system calibration.Secondly,the key technology in 3D point cloud registration: key point extraction and feature description are studied.Firstly,the key point detection algorithms of SIFT,ISS and Harris are studied,and analyzes the extraction effect of these three key point extraction algorithms based on the 3D point cloud data models of different objects,so as to summarize the applicability of key point extraction algorithms based on the different characteristics of the models.Then,aiming at the above three key point extraction algorithms,combining three different feature descriptors of FPFH,SHOT and 3DSC,and combining with RANSAC algorithm to form different point cloud coarse registration processes,the above point cloud model was coarse registration,and the influence of different key point extraction methods and feature descriptors on the accuracy of point cloud coarse registration was studied.Finally,the initial pose of the point cloud model was obtained based on the coarse registration algorithm with different collocations,and then the ICP registration algorithm was changed to carry out the final registration of the point cloud.Thus,the influence of the acquisition of the initial pose on the precision of the fine registration algorithm was analyzed.The focus of this topic is the research based on the influence of point cloud data features on registration accuracy.Different data features and their descriptions have different effects on the rough registration accuracy.Experiments show that SIFT+FPFH algorithm has the highest accuracy but the slowest speed.ISS+FPFH algorithm is slightly less accurate than SIFT+FPFH algorithm,but it is faster 3?4 times than SIFT+FPFH algorithm.Harris+SHOT algorithm is the fastest,but the registration accuracy is low.ISS+ SHOT algorithm is close to its efficiency,but the effect will be better than it.Secondly,for different fine registration algorithms,the initial position obtained by coarse registration has different effects on them.The classical ICP algorithm requires the high precision of coarse registration as far as possible,so that the posture of point cloud and target point cloud after fine registration can be closer and the registration accuracy can be higher.However,LLSICP algorithm has a low requirement on the initial pose,and the registration accuracy can reach a high level within a small error range.