Research On 3D Point Cloud Automatic Registration Based On Local Geometric Features

With the development of software and hardware technology,the application of 3D reconstruction technology is more and more widely.We can obtain 3D point cloud through a variety of ways,including multi view dense matching,lidar and handheld 3D laser scanner.However,in the process of 3D point cloud acquisition,we often need to change the station for several times to obtain the 3D point cloud data of the whole target area.The premise of the application of piecewise point clouds is to transform the point clouds scanned by multiple stations into a unified coordinate system,that is,to register the point clouds in different coordinate systems.However,the high density of 3D point cloud will increase the fuzziness of extracting corresponding points between point cloud sequences,resulting in higher outliers of corresponding points,which affects the accuracy of registration parameters.In addition,scholars have proposed a variety of point cloud registration algorithms,but how to effectively combine these algorithms and use the best registration strategy to achieve the accurate registration of 3D point cloud sequences in different scenes is also the difficulty of point cloud registration research.Based on the above situation,this paper makes an in-depth study.The main research contents and achievements of this paper are as follows:(1)Research on point cloud registration strategy.In view of the different effects of point cloud downsampling,initial registration and precise registration on the registration results,by combining different point cloud downsampling,initial registration and precise registration methods,four point cloud automatic registrations are compared Finally,an effective point cloud registration strategy(SACIA algorithm+ICP algorithm based on point-to-surface)was obtained.Using this strategy for automatic point cloud registration,the accuracy can reach about 3 cm in outdoor scenes.The accuracy of point cloud registration in indoor scenes can reach about 6mm.(2)Research on initial point cloud registration method based on sampling consistency.Aiming at the problem that the SACIA algorithm uses the similarity of the point fast feature histogram(FPFH)to screen the corresponding points,the outlier rate is too high and the strategy is used to cause the poor initial registration robustness of the point cloud.According to the corresponding point pair in the point cloud registration For geometric features with equal distances,a SACIA point cloud registration algorithm with additional isometric constraints is proposed,which improves the accuracy and robustness of the SACIA algorithm.After experimental comparison,it is found that the accuracy of the indoor scene is within 4cm.(3)Research on point cloud registration method based on curvature invariant geometric features.Aiming at the disadvantage of the SACIA algorithm's poor registration effect of point clouds in large scenes such as indoor and outdoor,according to the geometric characteristics of the local curvature of the point cloud in the three-dimensional point cloud registration,a point cloud registration based on the curvature invariant feature is proposed The method uses the DoG operator to extract the key points of curvature from the down-sampled point cloud,and uses the 4-PCS algorithm to achieve the initial point cloud registration,which improves the accuracy and robustness of the point cloud registration.Finally,the point-to-surface iterative nearest neighbor algorithm is used to accurately register the point clouds,and a complete point cloud automatic registration scheme is obtained,and the registration accuracy of the point cloud in the indoor scene is better than 3 cm,and the accuracy of the outdoor scene is about 5cm.