CN107341825A - A kind of method for simplifying for large scene high-precision three-dimensional laser measurement cloud data - Google Patents

Abstract

The invention relates to a method for simplifying point cloud data of high-precision three-dimensional laser measurement in large scenes, which can support rapid simplification of point cloud data in large scenes and high-precision three-dimensional laser scanning, while maintaining key features in point cloud data, and belongs to three-dimensional construction field of mold technology. The invention uses the uniform grid method to evenly space the scattered point cloud, establishes the grid index corresponding to the point cloud data, uses the spatial position to quickly find the K neighborhood of the data point; extracts the point cloud feature point according to the projection residual value, and uses The surface variation value divides the point cloud data into regions; the original point cloud is simplified by using the region division and the surface variation value of data points, and finally the simplified point cloud data is obtained. The method of the present invention can quickly simplify the high-precision scanning data with a data volume greater than 100 million points, and the execution speed is fast. While effectively reducing the data capacity, it can maintain key feature points in the scene, which is beneficial to later three-dimensional modeling, etc. work carried out.

Description

A simplified method for high-precision 3D laser measurement point cloud data in large scenes

technical field

The invention relates to a method for simplifying point cloud data of high-precision three-dimensional laser measurement in large scenes, which can support rapid simplification of point cloud data in large scenes and high-precision three-dimensional laser scanning, while maintaining key features in point cloud data, and belongs to three-dimensional construction field of mold technology.

Background technique

With the continuous improvement of measurement accuracy of 3D scanners, the collected point clouds are becoming denser and contain rich details. However, the huge point cloud data brings a lot of inconvenience to subsequent processing, storage, display and transmission. If it is processed directly, it will inevitably take up a lot of hardware resources and time, and not all data points need to be used for subsequent processing. Too dense point cloud data will affect the quality of 3D object reconstruction during the visualization process. The existing processing methods mainly include space-based point cloud simplification methods, normal-based point cloud simplification methods, surface change-based point cloud simplification methods and hybrid simplification methods. The space-based point cloud simplification method uses octree or spatial grid to divide the point cloud space, and replaces the points in the subdivided space with one point. The execution speed is fast, but the point cloud features are seriously missing; based on The point cloud simplification method based on normal error takes into account the local geometric features of the point cloud, but it is prone to holes; the point cloud simplification method based on the degree of surface change can effectively control the distribution of point cloud and data points, but the execution speed is slow; The hybrid simplification method uses an octree to divide the point cloud space according to the number of points and the degree of surface change, and only retains the data points closest to the center of gravity of the point set in the leaf nodes. This method executes faster, but it is difficult to fully preserve the geometry in the point cloud feature.

Contents of the invention

Aiming at the large scene, high-precision 3D laser scanning point cloud data is huge, direct storage takes up a lot of space, and cannot be directly transmitted, 3D modeling, display and other shortcomings, the present invention provides a simplified method for 3D laser measurement point cloud data , can support the rapid simplification of large-scene, high-precision 3D laser scanning point cloud data, effectively reduce data redundancy, while maintaining key features and details in point cloud data.

A technical scheme of a simplified method for three-dimensional laser measurement point cloud data is as follows.

1) Data point field search

(1) Point cloud space division

The uniform grid method is used to spatially partition the scattered point cloud. First read the scattered point cloud data, obtain the maximum and minimum values of the data point set on the X, Y, and Z coordinate axes, and establish a cuboid bounding box containing all data points parallel to the coordinate axes. Divide the bounding box into uniform grids, and establish the grid index corresponding to the point cloud data. A grid can contain multiple data points.

(2) Construct the enclosing sphere

Construct the enclosing sphere with the data points in the point cloud as the center of the sphere, and use a linear linked list to record all the data points in the enclosing sphere except the center of the sphere. Use the spatial position relationship between the grids to quickly construct the enclosing sphere, calculate the cube grid index number where the point p is located and the 26 adjacent cube grids according to the coordinate value of the point p, and calculate each data point in the grid separately The distance from point p, if the distance is less than or equal to the radius R of the enclosing sphere, then add this point to the linear linked list corresponding to the enclosing sphere. Repeat the above steps until each data point in the point cloud corresponds to a bounding sphere.

(3) K field search

For a data point p, it is sufficient to find the K data points closest to the point p from the enclosing sphere with the point p as the center. For boundary points and isolated noise points, there may be less than K neighbor data points in the enclosing sphere. Marking such data points requires special treatment in the surface variation calculation step. If the actual number of neighbor data points is less than K 20%, the data point can be considered as a noise point, and it will no longer participate in subsequent calculations after marking.

2) Calculation of surface variation

For data point p, calculate the real eigenvalues of the covariance matrix of point p according to the K data points closest to point p, and the eigenvector corresponding to the smallest eigenvalue is the normal vector of point p. The surface change at point p is the sum of the distances between the data points in the K neighborhood of point p and the tangent plane of point p.

The present invention adopts the uniform sampling method to estimate the average value of the surface variation of the point cloud, uniformly extracts 20% of the grids from the grids divided by the point cloud, calculates the surface variation of each point in the grid respectively, and then calculates the average value, Get an estimate of the point cloud surface variation. For data points with less than K neighbor data points in the enclosing sphere, the actual number of neighbor points is used for calculation.

3) Point cloud area division

(1) Extract feature points

Calculate the projection residual value of all domain points of data point p, and take the maximum value as the residual value of point p. Calculate the residual value of all data points, and take the average value as the judgment threshold t of the feature point. If the projection residual value of the data point p is greater than the threshold t, the point is considered as a feature point.

(2) Regional division

Sort all data points according to the surface variation value from small to large, take out the data points in turn, create a new division area, find the K field of the data point, establish a seed set, and calculate the surface variation angle between the data point and the data point in the K field If the angle difference is less than the threshold 0.25, add the data point to the current divided area; if the angle difference is less than the threshold 0.3, add the data point to the seed set. When the seed set is empty, start a new division area, and add the established area to the area queue. When all the data has been processed, the region division of the point cloud is completed.

4) Point cloud simplification

Mark all point cloud data as preserved. The point cloud data of the area is sequentially taken out from the area queue, 2000 feature points are randomly selected, and the surface variation difference between the feature points is calculated. If the average value of the surface variation difference is less than 0.12, the area is treated as a plane, using The uniform sampling method calculates the bounding box of the point cloud data and divides it evenly. The data point set g in each grid is replaced by the center point, and the original data point is marked as deleted; otherwise, the data point p is sequentially taken out from the regional data point set. If point p is a non-feature point and marked as reserved, calculate the surface variation difference between the domain point of point p and point p, if it is greater than the threshold 0.6 and mark the point as reserved, then mark the point as deleted. When all the divided areas are processed, delete the data points marked as deleted, and the simplified point cloud data can be obtained.

The method of the present invention is suitable for large scenes and high-precision three-dimensional laser scanning point cloud data. Compared with the prior art, the beneficial effect of the present invention is that it can quickly simplify the high-precision scanning data with a data volume greater than 100 million points, the execution speed is fast, and the key features in the scene can be maintained while effectively reducing the data capacity point, which is conducive to the development of 3D modeling and other work in the later stage.

detailed description

1) Data point field search

(1) Point cloud space division

The uniform grid method is used to spatially partition the scattered point cloud. First read the scattered point cloud data, obtain the maximum and minimum values of the data point set on the X, Y, and Z coordinate axes, and establish a rectangular parallelepiped bounding box containing all data points parallel to the coordinate axis. The side length of the bounding box is , , . Divide the bounding box into side length is A uniform grid of , taking M as an example, . Establish the grid index corresponding to the point cloud data. The coordinate value of the fake point is (x, y, z), and the corresponding grid index number (i, j, k) is: , , . A raster can contain multiple data points.

(2) Construct the enclosing sphere

Construct a surrounding sphere with the data points in the point cloud as the center of the sphere, the radius of the surrounding sphere is R=0.15, and use a linear linked list to record all data points in the surrounding sphere except the center of the sphere. The spatial position relationship between the grids can be used to quickly construct the bounding sphere, and the cube grid index number where the point p is located can be calculated according to the coordinate value of the point p. The data points contained in the bounding sphere with the point p as the center of the sphere are not only located In the grid, it may also be located in the 26 cubic grids adjacent to the cube grid where the point p is located. Assuming that the index number of the grid where the point p is located is (i, j, k), then the index number of the adjacent grid respectively , respectively calculate the distance between each data point in the grid and point p, if the distance is less than or equal to the radius R of the enclosing sphere, then add the point to the linear linked list corresponding to the enclosing sphere. Repeat the above steps until each data point in the point cloud corresponds to a bounding sphere.

(3) Field search

For a data point p, it is sufficient to find the K data points closest to the point p from the enclosing sphere with the point p as the center. For boundary points and isolated noise points, there may be less than K neighbor data points in the enclosing sphere. Marking such data points requires special treatment in the surface variation calculation step. If the actual number of neighbor data points is less than K 20%, the data point can be considered as a noise point, and it will no longer participate in subsequent calculations after marking.

2) Calculation of surface variation

For data point p, according to the K data points closest to point p, the covariance matrix of point p is defined in formula 1.

(Formula 1)

in . C is a symmetric positive semi-definite matrix, which has 3 real eigenvalues and satisfies , The corresponding eigenvectors are , which is the normal vector of point p. The surface variation at point p is calculated by Equation 2.

(Formula 2)

Indicates the distance between the jth data point in the neighborhood of point p and the tangent plane of point p.

The present invention adopts the uniform sampling method to estimate the average value of the surface variation of the point cloud, uniformly extracts 20% of the grids from the grids divided by the point cloud, calculates the surface variation of each point in the grid respectively, and then calculates the average value, Get an estimate of the point cloud surface variation , n is the number of data points contained in the 20% grid. For data points with less than K neighbor data points in the enclosing sphere, the actual number of neighbor points is used for calculation.

3) Point cloud area division

(1) Extract feature points

Calculate the projection residual value of all domain points of data point p, and take the maximum value as the residual value of point p. Calculate the residual value of all data points, and take the average value as the judgment threshold t of the feature point. If the projection residual value of the data point p is greater than the threshold t, the point is considered as a feature point.

(2) Regional division

Step 1: Sort all data points in descending order according to the surface variation value, and record them in dataset A.

Step 2: Select the data point p with the smallest surface variation value in the data set A and put it into the seed set s for operation. Create a set of regions Ri, set to empty.

Step 3: When the seed set is not empty, take out the seed q in the seed set, and find the neighbor point set N(q) of the seed q.

Step 4: Calculate the surface variation angle difference between the neighbor point qi and the seed point q , if the current neighbor point is not marked and the surface variation angle difference is less than the threshold 0.25, put the current neighbor point into the region set Ri and mark the data point as having joined the region state, and remove it from the data set A.

Step 5: If the surface variation value of the current neighbor point is less than the threshold 0.3, add the current neighbor point to the seed set s.

Step 6: If the neighbor point set N(q) is not empty, return to step 4.

Step 7: When the seed set s is not empty, return to step 3.

Step 8: Record the newly divided region Ri into the region division queue.

Step 9: When there are still data points in data point set A, return to step 2 and clear the seed set.

4) Point cloud simplification

Step 1: Mark all data points in the point set as reserved.

Step 2: Take the point cloud data of an area sequentially from the area queue, randomly select 2000 feature points, and calculate the surface variation difference between feature points.

Step 3: If the average value of the surface variation difference is less than 0.12, the area is treated as a plane, and the uniform sampling method is used to calculate the bounding box of the point cloud data, according to the side length The bounding box is divided, and the data point set g in each grid is replaced by a new center point, namely , the original data points are marked as deleted. Otherwise, go to step 4.

Step 4: Take out data point p sequentially from the regional data point set. If point p is a non-feature point and marked as reserved, calculate the surface variation difference between the domain point of point p and point p, if it is greater than the threshold 0.6 and mark the point as reserved, then mark the point as deleted.

Step 5: If there is still data in the area point cloud, go to step 4.

Step 6: If there are still zones in the zone queue, go to step 2.

Step 7: Delete the data points marked for deletion to obtain the simplified point cloud data.

Claims (1)

1. a kind of technical scheme of method for simplifying for large scene high-precision three-dimensional laser measurement cloud data is as follows：

（1）Space division is carried out to dispersion point cloud using uniform cube-algorithm first, establishes grid index corresponding to cloud data；With Data point in point cloud surrounds ball for centre of sphere structure, and all data points surrounded in ball in addition to the centre of sphere are recorded with linear linked list； For each data point, K nearest data point of range points is found out using surrounding in ball, establishes K fields；

（2）For each data point, according to the normal vector of K fields calculating point, the K that the curved surface variation at defining point p is point p is adjacent In domain between data point and point p section apart from sum, take the curved surface variation average value of uniform sampling method estimation point cloud；When Data point of the neighbor data point less than K in ball is surrounded, is calculated with actual neighbors points；

（3）For each data point of a cloud, the projection residual errors value in the K fields of the point is calculated, point p is used as using maximum therein Residual values, using the average value of the residual values of all data points as the judgment threshold of characteristic point, if the projection residual errors of data point Value is more than threshold value, then judges that data point is characterized a little；All data points are sorted from small to large by curved surface variation value, taken out successively Data point, new zoning is established, the curved surface variation differential seat angle of data point and data point in K fields is calculated, if differential seat angle is small In threshold value, then data point is added current zoning；If differential seat angle is less than threshold value, data point is added seed set；When Seed set is space-time, starts new zoning, and the region having built up is added in area queue；At all data When reason finishes, the region division for putting cloud is completed；

（4）All cloud datas are labeled as reserved state；

（5）Take out the cloud data in region successively from area queue, randomly choose characteristic point, the curved surface calculated between characteristic point becomes Divide difference, if the average value of curved surface variation difference is less than threshold value, plane treatment is pressed in the region, and letter is carried out using uniform sampling method Change and be labeled as deleting by legacy data point；Otherwise concentrated from area data point and take out data point successively, calculate the K necks of data point The curved surface variation difference of domain point and data point, determine to retain or delete；When all zonings are disposed, deletion is labeled as The data point of deletion, you can the cloud data after being simplified.