A fast registration method for 3D point cloud

Abstract

The invention discloses a quick registration method for three-dimensional point clouds. The quick registration method comprises the steps of: obtaining a three-dimensional reference point cloud A and a point cloud B to be registered; building a projected coordinate system of the point cloud A, and performing perspective projection on the point cloud A to obtain a two-dimensional point cloud A'; subdividing the point cloud A' into a plurality of triangles based on a triangulation algorithm; performing perspective projection on the point cloud B in the projected coordinate system to obtain a two-dimensional point cloud B'; judging whether points in the point cloud B' fall into the triangles, and thereby obtaining a plurality of fall points o; working out an interpolation coefficient of each fall point o with three peaks of the corresponding triangle where the fall point o is located; working out the coordinate value of the corresponding point q of each fall point o in the point cloud A according to the interpolation coefficient and a trigonometric interpolation algorithm so as to obtain point pairs <qn, pn> composed of the points q and the corresponding points p of the fall points o in the point cloud B, wherein n is not less than 1; working out the minimal rigid body transformation value T of the point pairs <qn, pn>; and updating the coordinate value of each point in the point could B to be registered based on the transformation value T.

Description

technical field [0001] The invention relates to the technical field of computer vision and pattern recognition, in particular to a fast registration method for three-dimensional point clouds. Background technique [0002] The 3D image acquisition equipment can only obtain the 3D data of one side of the object at a time. In order to obtain the overall 3D data of the object, it is necessary to obtain 3D data from multiple angles and register these 3D data. The conventional point cloud registration process is as follows: first, extract 3D data from different perspectives through sampling, and then, based on some algorithms, such as iterative closest point (ICP) algorithm, etc., through calculation, establish the connection between points in the starting point cloud , and finally, based on these connections, a complete three-dimensional image model is finally established by iterating layer by layer on the three-dimensional data of an initial position. [0003] The traditional I...

AI Technical Summary

Problems solved by technology

[0003] The traditional ICP algorithm compares the distance between the points in the point cloud to be registered and each point in the reference point cloud one by one to find the closest point, so as to establish the connection between points in the starting point cloud. This calculation method The calculation speed is too slow, the efficiency is too low, and there is also a large error matching problem. Therefore, many researchers have proposed many improvement methods based on it. For example, "Image Vision Compute" 1992 No. 10 In the article "Object modeling by registration of multiple range images" published in Volume 3 of the issue, it is proposed to use the distance in the direction of the normal vector instead of the point-to-point distance as the matching evaluation function, so that the points in the two views do not need to correspond one-to-one. However, this method needs to solve the nonlinear least squares problem, and the calculation efficiency is still low; for another example, "Iterative point matching for registration of free-form curves and surfaces" published in Volume 13 of the 12th issue of "International Journal of Computer Vision" in 1994 "In the article, a method of speeding up the search for the nearest point by modifying the KD tree is proposed. However, the establishment of a spatial retrieval structure similar to the KD tree needs to consume additional space and time, so it can only be performed in the prior preprocessing. Therefore, it is only applicable to static three-dimensional point clouds, and is not suitable for dynamic three-dimensional point cloud acquisition process; for another example, a kind of point cloud registration algorithm based on topological features is disclosed in CN 103150747A, which extracts the topological feature points ( Boundary feature points and prominent feature points) are selected as registration elements to establish the connection between point clouds. However, there is also low registration efficiency, which cannot meet the requirements for fast and instant registration of continuously obtained 3D point clouds.

Images