CN103106632B - A kind of fusion method of the different accuracy three dimensional point cloud based on average drifting - Google Patents

Abstract

The invention discloses a method for fusion of three-dimensional point cloud data with different precisions based on mean drift. For two sets of three-dimensional point cloud data with different precision levels, the error distribution of low-precision point clouds is established by using high-precision point clouds, and then the error distribution of low-precision point clouds is established. The low-precision point cloud carries out mean drift to eliminate the drift error of the low-precision point cloud, thereby realizing the fusion of two sets of data information. The method includes: S1: Establish the topological structure information of the low-precision point cloud, including the neighborhood of each sample point Point set and unit normal vector; S2: Use the high-precision point cloud to perform density clustering on the low-precision point cloud, and determine the drift error of each sample point of the low-precision point cloud according to the clustering results; S3: Use the topology of the low-precision point cloud The structure information and the drift error determine the drift vector of each point of the low-precision point cloud, and drift the various points of the low-precision point cloud according to the drift vector to realize fusion. The method of the invention can realize the smoothing of small-amplitude noise while eliminating the drift error of the low-precision point cloud.

Description

A Fusion Method of Different Precision 3D Point Cloud Data Based on Mean Shift

technical field

The present invention belongs to the field of surface digital three-dimensional shape detection data processing. High-precision point cloud data is generally obtained through online contact detection of three-coordinate measuring machines or machine tools. Low-precision point clouds generally refer to three-dimensional points obtained by laser scanners or flexible joint arms. cloud data.

Background technique

With the increasing development of the manufacturing industry, the shape of the product is becoming more and more complex, and its development and design are facing many difficulties and challenges. In particular, the product shape modeling technology is facing more challenges. Complex curved surface parts represented by aviation blades and propellers have been widely used, and higher requirements have been put forward for the actual processing quality inspection, which has made great progress in the digital inspection of complex curved surfaces. The purpose of surface digital detection is to reflect the deviation between the actual measured object and its CAD model. The main process includes the acquisition of the point cloud of the actual measured surface, the positioning of the measurement point cloud and the CAD model, the calculation and evaluation of the surface error, and the measurement Results Uncertainty Analysis. Considering that the characteristics of complex surfaces are not obvious, and the quality and quantity of point clouds affect the surface location algorithm and shape error evaluation results, the measured points not only require high enough precision, but also have enough point clouds to be accurate. Reflect the shape feature information of the free-form surface. Usually, the measurement accuracy of the touch probe is an order of magnitude higher than that of the optical probe, but the efficiency of point cloud data collection by the touch probe is relatively low, so the method of multi-sensor combination measurement is considered, through the data The fusion technology realizes the acquisition and processing of the point cloud of the measured surface, which not only utilizes the high precision of contact detection, but also exerts the high efficiency of optical detection, and realizes the balance between the accuracy and efficiency of measurement point cloud data.

Obtain point cloud data of complex curved surfaces (such as aviation blade curved surfaces) through contact measurement (three-coordinate measuring machine or online in-situ detection of machine tools), with high detection accuracy, single-point measurement accuracy can reach within a few microns, and can evaluate the locality of complex curved surfaces error, but the scale of the point cloud can only reach hundreds of points at most, which is limited to reflect the overall shape of the complex surface, and it is impossible to redesign the surface with these high-precision point cloud data. In order to overcome the shortcomings of contact measurement, non-contact measurement came into being, which is mainly based on the basic principles in the fields of optics, magnetism, acoustics, etc., and reasonably converts the given physical analog quantity into the coordinate point of the sample surface . The non-contact measurement method has greatly improved the measurement efficiency. Some optical measuring machines can obtain tens of thousands of points in a few seconds, such as the laser scanner of the British 3DSCANNER company, the StereoSCAN portable device of the German Breuckmann company, etc., which greatly reduces the measurement process. Artificial planning, quickly collect a large number of data points on the entire surface of the sample while reducing the work of the measurement personnel, and can obtain more massive data including complex surfaces. However, although non-contact measurement can use calibration methods to improve measurement accuracy, due to surface defects such as roughness and ripples on the surface of the measured object and the impact of the resolution of the measurement system itself, sampling error, electrical noise, etc., the data sampling process It is inevitable to mix unreasonable noise points or isolated points, and its measurement accuracy can only reach tens of microns, and part of the accuracy is mainly caused by drift errors. The point cloud obtained by contact measurement has high precision and small point cloud scale, while the point cloud obtained by non-contact measurement has low precision and contains defects such as noise points, but it can reflect the overall shape of complex surfaces. Therefore, it is of great significance to use high-precision point clouds to realize the drift of low-precision point clouds and improve the accuracy level of low-precision point clouds. It is also one of the important links in detection data processing to ensure accurate evaluation of processing quality.

For the drift of the low-precision point cloud, the usual method is to search for the nearest point in the low-precision point cloud for each sample point of the high-precision point cloud, calculate the distance between these two points, and then calculate the average of all distances, As the magnitude of the drift vector of all points in the low-precision point cloud, the drift of the low-precision point cloud is realized. This method is easy to implement, but the size of the above-mentioned drift vector cannot truly reflect the drift error between the low-precision point cloud and the high-precision point cloud, so that some areas of the low-precision point cloud have over-drift or under-drift, and the improved accuracy is limited. In fact, the drift error of each sample point of the low-precision point cloud is treated equally, which cannot truly reflect the difference in drift in different regions of the low-precision point cloud.

In view of the defects of high and low precision point cloud data fusion, the high precision point cloud can be used to establish the error division of different areas of the low precision point cloud. The division process is realized by density clustering, and then the topological structure information of the low precision point cloud is used to establish each sample point Drift vector, and with the help of the information entropy model to optimize the selection of Gaussian weights, each sample point is drifted, and the filtering of low-precision point cloud and small-amplitude noise is naturally realized during the drift process.

In information theory, entropy is a measure of the degree of disorder of a system, which can be used to measure the amount of effective information contained in known data. Entropy is a measure of system uncertainty. The larger its value, the greater the uncertainty of the system, which is not enough to reflect the internal information of the system; on the contrary, the smaller the entropy value, the smaller the uncertainty of the system, which is enough to reflect the system’s inherent information. inner information. Therefore, in the process of optimizing the selection of Gaussian weights, the kernel estimation of the difference density of normal vectors is proposed through the normal vector information, and the information entropy model is established. Ensure reasonable drift of low-precision point clouds.

Contents of the invention

The purpose of the present invention is to propose a fusion method of three-dimensional point cloud data under two different precisions, and then use the high-precision point cloud to analyze the error of each point of the low-precision point cloud through the high-precision point cloud obtained by contact measurement, based on The principle of mean drift is to drift the low-precision point cloud to remove the drift error of the low-precision point cloud, and to achieve the smoothing of the low-precision point cloud with small amplitude noise during the drift process.

The specific technical scheme adopted to realize the object of the present invention is as follows:

A fusion method of 3D point cloud data with different precisions based on mean shift. For two sets of 3D point cloud data with different precision levels, the high-precision point cloud is used to perform mean shift on the low-precision point cloud to eliminate low-precision points. Cloud drift error, so as to realize the fusion of two sets of data information, the method specifically includes:

S1: Establish the topological structure information of the low-precision point cloud, including the neighborhood point set and unit normal vector of each sample point;

S2: Using the high-precision point cloud to perform density clustering on the low-precision point cloud, and determine the drift error of each sample point of the low-precision point cloud according to the clustering result;

S3: Using the topology information of the low-precision point cloud and the drift error to determine a drift vector of each point of the low-precision point cloud, and drifting the various points of the low-precision point cloud according to the drift vector, Achieve integration.

As an improvement of the present invention, performing clustering and determining the drift error in the step S2 is specifically:

First, search the k points with the closest Euclidean distance of each sample point in the low-precision point cloud in the high-precision point cloud to form the k neighborhoods corresponding to each sample point in the high-precision point cloud in the low-precision point cloud, and then Calculate the projection points and normal vectors of each sample point in their respective k neighborhoods;

Secondly, using each projection point as a clustering center, clustering the low-precision point cloud by using density clustering to form a plurality of clustering units, wherein each projection point corresponds to a clustering unit;

Then, the drift error of all points within the range of each clustering unit is the same, that is, the drift error of the corresponding sample point.

As an improvement of the present invention, the drift error of all points within the range of each clustering unit can be expressed as:

${\mathrm{<span\; class="notranslate">\; \&Delta;</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; Lr</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; HR</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; Lr</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}$

Among them, if Indicates the k neighborhood corresponding to the sample point p _Hr in the low-precision point cloud, then Be a drift error, p _Hr is any sample point in the described high-precision point cloud P _H , is the sample point p _Hr at The projected foot point in , is the projected foot point The corresponding normal vector.

As an improvement of the present invention, in the step S3, the drifting of each sample point is carried out by the following formula:

p' _Li =p _Li -m _Li

In the formula, p _Li is any sample point in the low-precision point cloud, p′ _Li is the drifted point of sample point p _Li , and m _Li is the drift vector.

As an improvement of the present invention, the drift vector m _Li is obtained by the following formula:

${\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}{\mathrm{<span\; class="notranslate">\; \&Delta;</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}}{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}}$

In the formula, Q _Li is the neighborhood point set in the topology information of the sample point p _Li , w _Lij is the Gaussian weight of any point in the neighborhood point set Q _Li , n _Li is the normal vector of the sample point p _Li , n _Lij is the normal vector corresponding to the corresponding point q _Lij in the neighborhood point set Q _Li , Δ _Lij is the drift error Δ _Lij corresponding to the neighborhood point q _Lij , σ _n is the window width, reflecting the change of the normal vector of each point in the neighborhood .

As an improvement of the present invention, the Gaussian weight is calculated by the following formula:

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\frac{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; .</span>$

As an improvement of the present invention, the optimal value of the window width σ _n Determined by the following formula:

$\mathrm{<span\; class="notranslate">\; min</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; o</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>$

in,

${\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\frac{{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}}\mathrm{<span\; class="notranslate">\; ln</span>}\frac{{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; k</span>}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\frac{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; .</span>$

As an improvement of the present invention, before the topological structure information of the low-precision point cloud is established in step S1, coordinate registration may be performed on the point clouds of two different precisions to be fused, so as to transform them into the same coordinate system.

As an improvement of the present invention, the steps S1-S3 can be performed iteratively, wherein the iteration is based on the average value of the drift error Control is terminated when within a certain range, where the average of the Specifically:

As an improvement of the present invention, the high-precision point cloud data is acquired by triggering contact measurement, and the low-precision point cloud data is acquired by non-contact measurement.

The method of the present invention establishes the drift error model of the low-precision point cloud through the high-precision point cloud, and then establishes the drift vector of the low-precision point cloud, introduces the Gaussian weight of the unit normal vector information into the drift vector, and proposes the density of the normal vector difference Kernel estimation, and with the help of the minimum entropy principle of the information entropy model, the parameters of the Gaussian weights are optimized to obtain more effective Gaussian weights. This method can reduce the drift error of low-precision point cloud caused by measuring equipment, environmental interference and other factors, improve the accuracy, and thus describe the position information of the point cloud more accurately. The fusion method has the characteristics of high precision and accurate reflection of the shape characteristics of point cloud information, and achieves the effect of fusion.

Description of drawings

FIG. 1 is a schematic diagram of a process of density clustering of low-precision point clouds using high-precision point clouds in an embodiment of the present invention;

Fig. 2 is a schematic diagram of the neighborhood point set corresponding to each sample point of the high-precision point cloud in the low-precision point cloud and the corresponding drift error model in the embodiment of the present invention;

3 is a schematic diagram of the results of two sets of point clouds that are not in the same coordinate system through coarse moment of inertia matching and ADF stitching fine matching according to an embodiment of the present invention, wherein 3(a) is a schematic diagram of two sets of point clouds obtained by rough moment of inertia matching, Figure 3(b) is a schematic diagram of two sets of point clouds obtained by ADF fine matching;

Figure 4 is a schematic diagram of the chromatogram information of the error distribution before and after the drift of the low-precision point cloud according to the embodiment of the present invention (the darker the color, the greater the corresponding drift error), wherein 4(a) is the low-precision point cloud before drifting Schematic diagram (the average error is 0.0649mm), 4(b) is a schematic diagram of the low-precision point cloud after two drifts (the average error is 0.0266mm);

5 is a schematic diagram of the results of low-precision point cloud triangulation before and after drifting according to an embodiment of the present invention, wherein 5(a) is a schematic diagram of a triangular mesh before drifting, and 5(b) is a schematic diagram of a triangular mesh after drifting.

detailed description

The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. The following examples are illustrative only, and are not construed as limiting the present invention.

The method of this embodiment is aimed at two sets of three-dimensional point cloud data with different precision levels, using high-precision point clouds to analyze the drift error of low-precision point clouds, performing mean drift on low-precision point clouds, eliminating drift errors of low-precision point clouds, and Realize the smoothing of small-amplitude noise of low-precision point clouds, improve the accuracy level of low-precision point clouds, and realize the fusion of two sets of data information.

The point cloud in this embodiment is preferably described by a three-dimensional point cloud of the curved surface of the aviation blade. Among them, the high-precision point cloud data is generally preferably obtained by a three-dimensional coordinate measuring machine or an online contact detection of a machine tool. In this embodiment, the three-dimensional point cloud of the curved surface of the blade can be preferably obtained by a three-dimensional coordinate measuring machine (single-point accuracy can reach 0.002um), for example Scan 15 section lines, each line contains 20 points, a total of 300 points, as a high-precision point cloud. Low-precision point cloud generally preferably refers to 3D point cloud data obtained by non-contact measurement such as laser scanners or flexible articulated arms. In this embodiment, it is preferable to obtain massive point cloud data of the blade through the Hexagon flexible articulated arm (with a calibration accuracy of 0.03mm) as a low-precision point cloud.

In this embodiment, the high-precision point cloud data is recorded as _{PH and its number of points is N H ,} and the low-precision point cloud data is recorded as _PL and its number of points is N _L .

The method of this embodiment specifically includes the following processes:

1. Topology creation

Create the topology structure of the low-precision point cloud, specifically including establishing the neighborhood information and unit normal vector information of each point in the low-precision point cloud.

To create neighborhood information for the low-precision point cloud _PL , this embodiment may preferably use a three-dimensional grid method to establish the k-neighborhood information of each point p _Li of the low-precision point cloud _PL (i represents the point cloud sequence number, that is, the i-th point, i=1, 2,..., N _L , N _L represents the number of points in _PL ), that is, the k points closest to point p _Li in _PL (in this embodiment, the value of k is preferably k=8~ 20), the neighborhood is recorded as Q _Li ={q _Li1 ,q _Li2 ,…,q _Lik }(q _{Lij ∈PL} ,j=1,2,…,k).

Its establishment steps are:

(a) Rasterize the point cloud _PL to obtain the grid where each point p _Li is located and the 26 grids around the grid;

(b) Find all points in the adjacent grid, and calculate the distance from all points to point p _Li ;

(c) Select the nearest k points as the neighborhood of p _Li , and so on to obtain the k neighborhood of each sample point.

The k-neighborhood information Q _Li of each sample point p _Li of the low-precision point cloud _PL is established, and the normal vector information of the point cloud can be calculated. To this end, a 3×3 weighted covariance matrix is introduced as

in h is an impact factor, which can be taken as $h=\underset{1\&le;j\&le;k}{\max }\left\{{\left|\right|q_{\mathrm{Lij}}-{\stackrel{\&OverBar;}{q}}_{\mathrm{Li}}\left|\right|}_2\right\}/3.$

It can be seen from the above formula, is a symmetric positive semidefinite matrix, its eigenvalues λ _t (t=1, 2, 3) are all real numbers, and the corresponding eigenvectors n _t (t=1, 2, 3) are mutually orthogonal. Assuming λ ₁ ≤λ ₂ ≤λ ₃ , the size of the eigenvalue reflects the variation of the points in the neighborhood point set Q _Li in the directions of three eigenvectors n ₁ , n ₂ , and n ₃ respectively, then the point set Q _Li The least-squares fitting plane (or microcutting plane) of is the degenerate plane ∏ _p defined by n ₂ and n ₃ , and the corresponding normal vector is n ₁ .

Other similar methods can also be used to create neighborhood information, such as octree method and kd-tree method, but building a tree structure involves data encoding, and the process is cumbersome.

Before establishing the topological structure information of the low-precision point cloud _PL , coordinate registration can be performed on two point clouds with different precision to match them into the same coordinate system. Coordinate matching is specifically:

(a) During the measurement process, a standard calibration sphere can be set, and a certain point cloud information of the calibration sphere can be obtained at the same time during the measurement, and then the coordinate mapping relationship of the two sets of point clouds can be established through the pose of the calibration sphere, and the low-precision The point cloud _PL is matched to the coordinate system described by the high-precision point cloud _PH .

(b) If there is no calibration ball, the moment of inertia matching method can be used to roughly match the two sets of point clouds, and adjust the poses of the two to be roughly the same. The coarse moment of inertia matching method does not depend on whether there is a corresponding relationship between the two sets of point clouds. For any two sets of data, the pose can be adjusted conveniently and quickly. Transform the relationship to achieve rough matching from low-precision point clouds to high-precision point clouds. Next, perform fine matching on the two sets of point clouds.

Under the assumption that the single-point detection error of the high-precision point cloud P _H is very small, the low-precision point cloud is converted to the corresponding position of the high-precision point cloud through the point cloud matching algorithm. Considering the differences in location and quantity of point cloud data, it is preferable to use an adaptive distance function (ADF) algorithm for point cloud matching in this embodiment. The matching completed at this time is to transfer a small amount of high-precision point P _H to the location of the low-precision point cloud P _L , but the high-precision point cloud P _H more accurately expresses the position information on the surface of the measured object, so the high-precision The precision point P _H is fixed, and the low-precision point cloud P _L is transformed into the coordinate system of the high-precision point cloud P _H , so as to realize the coordinate matching of the two sets of point clouds.

2. Drift error model

Next, establish the error relationship between the low-precision point cloud and the high-precision point cloud through the definition of drift error, realize data fusion, and realize the low-precision point cloud through the high-precision point cloud (which is considered to accurately reflect the quality of the actual processed surface) Smoothing for small to medium amplitude noise.

The high-precision point cloud P _H reflects the ideal information of the processed surface, which is set to _SH , then the drift error of each sample point p _Li of the low-precision point cloud P _L is the vertical distance from the point p _Li to the curved surface; conversely, if If a surface S _L is constructed with a low-precision point cloud P _L , the drift error can be described as the error distribution of the high-precision point cloud P _H on the surface S _L. Obviously, due to the large number of low-precision point cloud _PL , the drift error is suitable for the latter description, but the error between high-precision point cloud _PH and low-precision point cloud _PL is established, so the error needs to be described in reverse, which can be The low-precision point cloud is divided into regions by using the high-precision point cloud, and the errors of the point sets in the same region are approximately equal. For this reason, the low-precision point cloud _PL needs to be divided into regions, and the density clustering method is used for the division.

As shown in Figure 2, for each sample point p _Hr in the high-precision point cloud P _H (r is the serial number of the high-precision point cloud, r=1,2,..., N _H , N _H is the number of P _H points), search for its In the low-precision point cloud _PL , the k points with the closest Euclidean distance (k=8~20) are denoted as $Q_{\mathrm{Lr}}^h=\{q_{\mathrm{Lr}1}^h,q_{\mathrm{Lr}2}^h,...,q_{\mathrm{Lrk}}^h\}(q_{\mathrm{Lrj}}^h\&Element;P_L,j=\mathrm{1,2},...,k),$ The p _Hr is then calculated using the moving least squares method at projected foot point in Fa Yawei (direction and The average direction of each point in the point set is the same). Then with each projected point As the clustering center, the density clustering method is used to cluster the low-precision point cloud _PL to form N _H clustering units. The clustering process is shown in Figure 1, and each projection point corresponds to a clustering unit _CLr .

Then, the drift error of each clustering unit is analyzed, and the drift error model is established.

Low-precision point cloud P _L at point The approximate drift error (including sign) of all points within the determined cluster unit range C _Lr is That is, the drift error of each clustering unit is

${\mathrm{<span\; class="notranslate">\; \&Delta;</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; Lr</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; HR</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}& <span\; class="notranslate">\; (</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; Lr</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; HR</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; Lr</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; \&Greater\; Equal;</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; -</span>{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; Lr</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; HR</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}& <span\; class="notranslate">\; (</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; Lr</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; HR</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; Lr</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; )</span>\end{array}\right.$

From this, the overall average drift error (drift control error) of the low-precision point cloud _PL can be obtained as

${\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; \&Delta;</span>}}}_{\mathrm{<span\; class="notranslate">\; L</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}}\underset{\mathrm{<span\; class="notranslate">\; r</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; \&Delta;</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; |</span>$

In the formula, N _H represents the number of high-precision point cloud P _H.

According to the above definition, each clustering unit of the low-precision point cloud _PL corresponds to a drift error And the drift error of all points in the clustering unit is Establishing the above drift error model actually determines the drift error Δ _Li of each sample point p _Li (i=1, 2, . . . , N _L ) of the low-precision point cloud _PL .

3. Point cloud drift and smoothing

After the error distribution of the low-precision point cloud _PL is established, it can be drifted according to the topological neighborhood information of the low-precision point cloud. The drifting process is actually a smoothing process for the small-amplitude noise of the low-precision point cloud. According to the neighborhood point set Q _Li of each sample point p _Li (i represents the serial number of the low-precision point cloud _PL ) of the low-precision point cloud _PL , each neighborhood point q _Lij (normal vector is n _Lij ) corresponds to a drift Error Δ _Lij (j=1,2,…,k), then the drift vector of the i-th sample point p _Li is defined as

${\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}{\mathrm{<span\; class="notranslate">\; \&Delta;</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}}{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}}$

Among them, w _Lij is the Gaussian weight of each neighborhood point. Then the new sample point after the drift of the i-th sample point p _Li is

p' _Li =p _Li -m _Li

where the Gaussian weight of each neighborhood point is defined as

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\frac{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>$

Among them, n _Li is the normal vector of p _Li , n _Lij is the normal vector of the point q _Lij in the field Q _Li , σ _n reflects the change of the neighborhood normal vector, generally called the window width, and its value has an important influence on the weight , also plays a crucial role in the drift results. For this reason, consider the following steps to optimize the selection of the window width _σn :

(a) Use kernel estimation to define the difference density of normal vectors, reflecting the change information of normal vectors in the neighborhood point set Q _Li .

(b) According to the normal vector density and with the help of the existing information entropy theory, construct the information entropy model, and use the minimum entropy principle to optimize and select the window width σ _n .

(c) Determine the search range of the window width σ _n according to the change of the normal vector Q _Li of the neighborhood point set, and use the heuristic optimization theory to determine the optimal window width σ _n .

The window width of the normal vector change can be determined according to the neighborhood point set Q _Li of the point p _Li , assuming that the local space range determined by the neighborhood Q _Li is Ω _Li , then for Its normal vector is n _x , then combined with kernel estimation, the normal vector difference density can be defined as

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; k</span>}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\frac{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 8</span>}<span\; class="notranslate">\; ~</span>\mathrm{<span\; class="notranslate">\; 20</span>}<span\; class="notranslate">\; )</span>$

Through the above definition of normal vector density, information entropy is introduced to determine the optimal window width σ _n . In information theory, entropy is a measure of system uncertainty, and the greater the entropy, the greater the uncertainty of the system. For the kernel estimation of the normal vector density, in the local area Ω _Li determined by the neighborhood Q _Li of the point p _Li , if the density function values everywhere are approximately equal (the normal vector density will lose its meaning at this time), the data distribution If the uncertainty is the largest, it has the largest entropy; on the contrary, if the value of the density function is very asymmetric (at this time, it can reflect the internal change of the normal vector information), then the uncertainty is the smallest and the entropy is the smallest. Therefore, the effectiveness of kernel density estimation can be measured by the concept of normal vector density estimation entropy. According to the definition of information entropy, for a point set Q _Li of k points, the kernel estimate of each point q _Lij ∈ Q _Li (j=1,2,L,k) is f _Li (q _Lij ), then the estimated entropy E _Li (σ _n ) can be defined as

${\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\frac{{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}}\mathrm{<span\; class="notranslate">\; ln</span>}\frac{{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; Lij</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; Li</span>}}}$

in, $G_{\mathrm{Li}}=\underset{j=1}{\overset{k}{\mathrm{\&Sigma;}}}{f}_{\mathrm{Li}}\left(q_{\mathrm{Lij}}\right)$ is the standardization factor.

Apparently, E _Li (σ _n ) is a one-variable function about σ _n , and the change is as follows: when σ _n →0, the value of the normal vector density function tends to The estimated entropy E _Li of the normal vector density is the maximum value, that is, E _Li = ln(k); as σ _n gradually increases from 0 to ∞, the estimated entropy decreases gradually at the beginning and is somewhere reaches the minimum value, and then increases gradually; when σ _n →∞, E _Li reaches the maximum value again. at an optimal value Take the minimum It can be considered as the optimal window width of the Gaussian weight, which can reflect the intrinsic information of the normal vector difference.

Choose an optimal value Theoretically, it can be searched in the interval range σ _n ∈ (0,+∞), but the interval range has no meaning of search actually. For the point set Q _Li , the normal vector n _Lij of each sample point can be obtained by averaging and unitizing the normal vectors of all point sets in the neighborhood to obtain the mean value change direction as Compute for each normal vector n _Lij Select the maximum value v _max and the minimum value v _min , then the variation range of the normal vector is approximately [v _min ,v _max ], so σ _n →0 can be used replace (if then take ); so σ _n →+∞ can be used Instead, the range of the search interval is ${\mathrm{\&sigma;}}_{\mathrm{no}}^o\&Element;[\frac{v_{\min }}{10},10v_{\max }].$

In the present invention, iterations (2) to (5) can be repeated multiple times to obtain a better fusion effect, wherein the iteration control parameters are set as low-precision point clouds The average error of is used as the parameter of error control, so that multiple iterations are performed until the average drift error is controlled within a certain range, and the iteration is stopped;

According to the low-precision point cloud and high-precision point cloud described above, rough matching is performed first, as shown in Figure 3(a), and then the ADF stitching method is used for fine matching, and the matching result is shown in Figure 3(b). The error distribution information of the low-precision point cloud is established, and its error chromatogram is shown in Figure 4(a), and the point cloud error after two drifts is shown in Figure 4(b). The smoothing of the small-amplitude noise of the low-precision point cloud has been achieved during the drifting process, and the results before and after the smoothing are shown in Figure 5(a) and (b).

Claims (8)

1. A method for fusion of 3D point cloud data with different precisions based on mean drift. For two sets of 3D point cloud data with different levels of precision, the high-precision point cloud is used to perform mean shift on the low-precision point cloud to eliminate low The drift error of the precision point cloud, so as to realize the fusion of two sets of data information, the method specifically includes: S1: Establish the topological structure information of the low-precision point cloud, including the neighborhood point set and unit normal vector of each sample point; S2: Using the high-precision point cloud to perform density clustering on the low-precision point cloud, and determine the drift error of each sample point of the low-precision point cloud according to the clustering result; S3: Using the topology information of the low-precision point cloud and the drift error to determine a drift vector of each point of the low-precision point cloud, and drifting the various points of the low-precision point cloud according to the drift vector, achieve integration; Wherein, performing clustering and determining the drift error in the step S2 is specifically: First, search the k points with the nearest Euclidean distance of each sample point in the low-precision point cloud in the high-precision point cloud to form the k neighborhood of each sample point, and then calculate the k-neighborhood points of each sample point in the respective k neighborhood Projection points and normal vectors in ; Then, using each projection point as a clustering center, clustering the low-precision point cloud by using density clustering to form a plurality of clustering units, wherein each projection point corresponds to a clustering unit; Finally, the drift error of all points of the low-precision point cloud within the range of each cluster unit is the same, that is, the drift error of the corresponding low-precision point cloud sample point; wherein, the drift error of all points within the range of each cluster unit Can be expressed as: ${\mathrm{<span\; class="notranslate">\; \&Delta;</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; r</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; r</span>}}^{\mathrm{<span\; class="notranslate">\; h</span>}}$ Among them, p _Hr is any sample point in the high-precision point cloud P _H , Represents the k neighborhood of the sample point p _Hr , is the drift error, is the sample point p _Hr in the k neighborhood The projected foot point in , is the projected foot point The corresponding normal vector. 2. method according to claim 1, is characterized in that, in described step S3, carry out drift to each sample point and carry out by following formula: p' _Li =p _Li -m _Li In the formula, p _Li is any sample point in the low-precision point cloud, p′ _Li is the drifted point of sample point p _Li , and m _Li is the drift vector. 3. method according to claim 2, is characterized in that, described drift vector m _Li obtains by following formula: ${\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; \&Delta;</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}}{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}}$ In the formula, Q _Li is the neighborhood point set corresponding to the sample point p _Li in the low-precision point cloud, w _Lij is the Gaussian weight of any neighborhood point q _Lij in the neighborhood point set Q _Li , n _Li is the sample point p The normal vector of _Li , n _Lij is the normal vector corresponding to the neighbor point q _Lij , and Δ _Lij is the drift error Δ _Lij corresponding to the neighbor point q _Lij . 4. method according to claim 3, is characterized in that, described Gaussian weight is calculated by following formula: ${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; ||</span>}{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; |</span>{<span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; )</span>$ Among them, σ _n is the window width. 5. method according to claim 4, is characterized in that, the optimal value of described window width σ _n Determined by the following formula: $\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; o</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>$ in, ${\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; -</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}\frac{{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}}}\mathrm{<span\; class="notranslate">\; l</span>}\mathrm{<span\; class="notranslate">\; no</span>}\frac{{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; q</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}}}$ ${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; k</span>}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\frac{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; |</span>{<span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; )</span>$ in, is the normalization factor, E _Li (σ _n ) is the estimated entropy, which is a one-variable function about the window width σ _n , n _x is the normal vector of any element x in the local space range Ω _Li determined by the neighborhood point set Q _Li . 6. The method according to any one of claims 1-5, characterized in that, before step S1 establishes the topological structure information of the low-precision point cloud, the two kinds of point clouds with different precisions to be fused can be first coordinately matched standard to transform them into the same coordinate system. 7. The method according to any one of claims 1-5, wherein the steps S1-S3 can be repeated iteratively, wherein the number of iterations is controlled so that the average value of the drift error Within a certain range, the average of the N _H represents the number of high-precision point clouds _PH . 8. The method according to any one of claims 1-5, wherein the high-precision point cloud data is acquired by triggering contact measurement, and the low-precision point cloud data is acquired by non-contact measurement.