Research And Application Of Point Cloud Registration Algorithm For Disordered 3D Parts

Three-dimensional point cloud registration is mainly used in the automatic reconstruction of objects and scenes.The research of point cloud registration algorithms is of great significance to the research and development of high-end automation equipment guided by machine vision.The commonly used point-to-feature-based 3D target recognition method directly establishes a Point Pair Feature(PPF)relationship between two points,and establishes a hash table,and uses the principle of Hough voting to determine which position is likely to be correct.The poses are voted to obtain the correct registration result.This method avoids the problem of falling into the local optimum caused by the poor initial pose.However,it is necessary to establish a point-to-point relationship for all points in the point cloud,which results in a massive amount of calculation;because the pose obtained by voting is not optimized,wrong matching results may also be generated.Aiming at the existing problems of PPF,this paper improves the point cloud filtering and simplification algorithm in the point cloud and processing to ensure the smoothness and integrity of the input point cloud in the registration stage.The traditional ICP algorithm is performed in the pose optimization stage.Improved,adding KD tree to speed up the search of neighboring points and further improve the efficiency of the overall registration algorithm.The main research results of this paper are as follows:(1)Aiming at the problem of many noise points in the initial point cloud,this paper studies the three filtering algorithms of radius,statistics,and bilateral filtering.To achieve noise points in the point cloud,and improved bilateral filtering method is proposed.The threshold value in statistical filtering and the average threshold value is 1.5 times the value.The two-domain(including spatial domain and value domain)information in bilateral filtering is associated with constraints,which effectively improves the filtering efficiency of point clouds.For model filtering in actual scenes,the time is only 70% of the original bilateral filtering.(2)Aiming at the problem that the initial point cloud contains many redundant points,the curvature method and the voxel grid method in the point cloud simplification algorithm are studied.The traditional voxel mesh simplification method has a faster calculation speed,but the parameters need to be manually set in the process.Otherwise,the local features of the point cloud cannot be better preserved.The point cloud simplification algorithm based on curvature fitting is sound,but the calculation efficiency is low.This paper proposes a voxel grid algorithm based on adaptive factors,which simplifies the point cloud and keeps the point cloud while improving the simplification efficiency.With better shape features,the curvature criterion is used to further process the point cloud to better retain the local features of the point cloud.Experiments show that the simplification effect of the proposed algorithm is significantly better than that of the voxel grid method for the curvature method,and the simplification time is only 48% of the curvature fitting method.(3)Aiming at the traditional point-to-feature voting pose optimization problem,the classic Iterative Closest Point algorithm is studied,and an improved ICP algorithm based on KD tree is proposed to solve the problem of the original algorithm's slow search speed for nearby points.This method uses the K-D tree structure in the three-dimensional point cloud.By building a three-dimensional K-D tree on the model point cloud,the point cloud points to be registered are used as input nodes so that the time complexity of the ICP algorithm for finding the nearest neighbor points is reduced.Experimental results show that compared with the traditional PPF algorithm,the registration efficiency is improved by 53%,and the registration accuracy reaches 98%.