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 kind of fusion method based on the three dimensional point cloud under the different accuracy of average drifting, for the three dimensional point cloud that two groups of accuracy classes are different, high-precision dot cloud is utilized to set up the error distribution of low precision point cloud, and then average drifting is carried out to low precision point cloud, eliminate the drift error of low precision point cloud, thus realizing the fusion of two groups of data messages, the method comprises: S1: the topology information setting up low precision point cloud, comprises each sample neighborhood of a point point set and per unit system arrow; S2: utilize high-precision dot cloud to carry out Density Clustering to low precision point cloud, determine the drift error of each sampling point of low precision point cloud according to cluster result; S3: utilize the topology information of low precision point cloud and described drift error to determine the drift vector of each sampling point of low precision point cloud, drifts about according to each sampling point of this drift vector to low precision point cloud, realizes merging.Method of the present invention, while the low precision point cloud drift error of elimination, can realize the fairing of noise by a small margin.

Description

Fusion method of three-dimensional point cloud data with different precisions based on mean shift

Technical Field

The invention belongs to the field of processing of curved surface digital three-dimensional shape detection data, high-precision point cloud data is generally obtained by online contact detection of a three-coordinate measuring machine or a machine tool, and low-precision point cloud data generally refers to three-dimensional point cloud data obtained by a laser scanner or a flexible joint arm.

Background

As the manufacturing industry develops, the product shape is complicated, and the development and design of the product shape are difficult and challenging, and in particular, the product shape modeling technology is more challenging. Complex curved surface parts represented by aviation blades, propellers and the like are widely applied, higher requirements are also put forward for actual processing quality detection, and the digital detection of the complex curved surface is greatly developed. The digital detection of the curved surface aims to reflect the deviation of an actual measured object and a CAD model thereof, and the main processes comprise the acquisition of the point cloud of the actual measured curved surface, the locating of the point cloud of the measured point and the CAD model, the calculation and the evaluation of the error of the curved surface and the uncertainty analysis of the measurement result. Considering the unobvious characteristics of the complex curved surface and the influence of the quality and quantity of the point clouds on the curved surface locating algorithm and the shape error evaluation result, the measured points are required to have high enough precision, and the quantity of the point clouds is enough to accurately reflect the shape characteristic information of the free curved surface. Under the normal condition, the measurement accuracy of the contact measuring head is one order of magnitude higher than that of the optical measuring head, but the efficiency of the contact measuring head for collecting point cloud data point by point is lower, so that the method of adopting multi-sensor combination measurement is considered, the acquisition and processing of the measured curved surface point cloud are realized through a data fusion technology, the high accuracy of the contact detection is utilized, the high efficiency of the optical detection is also exerted, and the balance of the accuracy and the efficiency of the measured point cloud data is realized.

The point cloud data of a complex curved surface (such as an aviation blade curved surface) is obtained through contact measurement (three-coordinate measuring machine or machine tool online in-situ detection), the detection precision is high, the single-point measurement precision can reach within several microns, the local error of the complex curved surface can be evaluated, but the point cloud scale can only reach hundreds of points at most, the point cloud scale is used for reflecting the overall appearance of the complex curved surface is limited, and the curved surface can not be redesigned through the high-precision point cloud data. In order to overcome the defects of contact measurement, non-contact measurement is carried out, and the method is mainly based on the basic principle in the fields of optics, magnetics, acoustics and the like and reasonably converts given physical analog quantity into a coordinate point on the surface of a sample piece. The non-contact measurement method greatly improves the measurement efficiency, and some optical measuring machines can obtain tens of thousands of points within seconds, such as laser scanners of British 3DSCANNER company, StereoSCAN portable equipment of Breuckmann company in Germany and the like, so that manual planning is greatly reduced in the measurement process, a large number of data points are rapidly collected on the surface of the whole sample piece, the work of measuring personnel is reduced, and mass data containing more complex curved surfaces can be obtained. However, although the non-contact measurement can adopt a calibration method to improve the measurement accuracy, because the surface of the measured object has surface defects such as roughness and ripples and the influence of the resolution, sampling error, electrical noise and the like of the measurement system, unreasonable noise points or isolated points are inevitably mixed in the data sampling process, the measurement accuracy of the non-contact measurement can only reach about tens of micrometers, and part of the accuracy is mainly caused by drift error. The point cloud obtained by contact measurement has high precision and small scale, while the point cloud obtained by non-contact measurement has low precision and contains defects such as noise points and the like, but can reflect the overall appearance of a complex curved surface. Therefore, the method has important significance for realizing the drift of the low-precision point cloud by utilizing the high-precision point cloud and improving the precision level of the low-precision point cloud, and is also one of important links for detecting data processing so as to ensure the accurate evaluation of the processing quality.

For the drift of the low-precision point cloud, a method is usually adopted, in which a closest point in the low-precision point cloud is searched for each sample point of the high-precision point cloud, the distance between the two points is calculated, and then an average value of all the distances is calculated to be used as the magnitude of a drift vector of all the points of the low-precision point cloud, thereby realizing the drift of the low-precision point cloud. The method is easy to implement, but the size of the drift vector cannot truly reflect the drift error between the low-precision point cloud and the high-precision point cloud, so that certain areas of the low-precision point cloud have over drift or under drift, and the improvement precision is limited. In fact, the drift error of each sample point of the low-precision point cloud is treated equally, and the difference of the drift of different areas of the low-precision point cloud cannot be reflected really.

Aiming at the defects of high-precision point cloud data fusion and low-precision point cloud data fusion, error division of different areas of low-precision point cloud can be established by using the high-precision point cloud, the division process is realized by density clustering, then drift vectors of all sampling points are established by using topological structure information of the low-precision point cloud, and after Gaussian weights are optimized and selected by means of an information entropy model, drift is carried out on all the sampling points, and filtering of low-precision point cloud small-amplitude noise is naturally realized in the drift process.

In information theory, entropy is a measure of the degree of disorder of the system, and can be used to measure the amount of useful information contained in known data. The entropy is used as a measurement of system uncertainty, and the larger the value of the entropy is, the larger the uncertainty of the system is, and the information in the system cannot be sufficiently reflected; conversely, the smaller the entropy value, the less uncertainty of the system, which is sufficient to reflect the intrinsic information of the system. Therefore, in the optimization selection process of the Gaussian weight, the kernel estimation of the difference density of the normal vector is provided through the normal vector information, an information entropy model is established, the important parameters of the Gaussian weight are optimized and selected through the minimum entropy principle, and the appropriate drift vector is determined, so that the reasonable drift of the low-precision point cloud is ensured.

Disclosure of Invention

The invention aims to provide a fusion method of three-dimensional point cloud data under two different precisions, which comprises the steps of obtaining high-precision point cloud through contact measurement, analyzing errors of each point of low-precision point cloud by utilizing the high-precision point cloud, drifting the low-precision point cloud based on the principle of mean value drift to remove drift errors of the low-precision point cloud, and realizing fairing of low-precision point cloud with small amplitude noise in the drifting process.

The specific technical scheme adopted for achieving the purpose of the invention is as follows:

the utility model provides a three-dimensional point cloud data's fusion method under different precision based on mean shift, it is to two sets of three-dimensional point cloud data that the precision grade is different, utilizes wherein high accuracy point cloud to carry out mean shift to low accuracy point cloud, eliminates the drift error of low accuracy point cloud to realize the fusion of two sets of data information, and the method specifically includes:

s1: establishing topological structure information of the low-precision point cloud, including a neighborhood point set and a unit normal vector of each sample point;

s2: carrying out density clustering on the low-precision point cloud by using the high-precision point cloud, and determining the drift error of each sample point of the low-precision point cloud according to a clustering result;

s3: determining the drift vector of each sample point of the low-precision point cloud by using the topological structure information of the low-precision point cloud and the drift error, and drifting each sample point of the low-precision point cloud according to the drift vector to realize fusion.

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

firstly, searching k points with the nearest Euclidean distance of each sample point in the high-precision point cloud in the low-precision point cloud to form a k neighborhood of each sample point of the high-precision point cloud in the low-precision point cloud, and then calculating a projection point and a normal vector of each sample point in the respective k neighborhoods;

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

then, the drift errors of all the points in the range of each clustering unit are the same, namely, the drift errors of the corresponding sampling points are taken as the drift errors of the corresponding sampling points.

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

${\mathrm{\Δ}}_r^H=(q_{\mathrm{Lr}}^H-p_{\mathrm{Hr}})\·n_{\mathrm{Lr}}^H$

wherein, ifRepresents a sample point p_HrIn the corresponding k neighborhood in the low-precision point cloud, thenFor drift error, p_HrFor the high-precision point cloud P_HAny one of the plurality of the sampling points,is a sample point p_HrIn thatThe projected foot drop point in (1) is,for projecting foot-hanging pointsThe corresponding normal vector.

As a refinement of the present invention, the step S3 of shifting the various points is performed by the following formula:

p′_Li＝p_Li-m_Li

in the formula, p_LiIs any sample point, p 'in the low-precision point cloud'_LiIs a sample point p_LiPoint after the shift, m_LiIs a drift vector.

As an improvement of the invention, the drift vector m_LiObtained by the following formula:

$m_{\mathrm{Li}}=\frac{\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{w}_{\mathrm{Lij}}{\mathrm{\Δ}}_{\mathrm{Lij}}{n}_{\mathrm{Lij}}}{\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{w}_{\mathrm{Lij}}}$

in the formula, Q_LiIs a sample point p_LiSet of neighborhood points, w, in the topology information_LijSet Q for the neighborhood point_LiGaussian weight of any point in, n_LiIs a sample point p_LiNormal vector of (1), n_LijFor a neighborhood set of points Q_LiMiddle corresponding point q_LijCorresponding normal vector, Δ_LijIs a neighborhood point q_LijCorresponding drift error delta_Lij，σ_nAnd the window width reflects the normal vector change condition of each point in the neighborhood.

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

$w_{\mathrm{Lij}}=\exp (-\frac{{\left|\right|n_{\mathrm{Li}}-n_{\mathrm{Lij}}\left|\right|}_2^2}{2{\mathrm{\σ}}_n^2}).$

as an improvement of the invention, the window width σ is_nOptimum value of (2)Determined by the following formula:

$\min \left(E_{\mathrm{Li}}\right)=E_{\mathrm{Li}}\left({\mathrm{\σ}}_n^o\right),$

wherein,

$E_{\mathrm{Li}}\left({\mathrm{\σ}}_n\right)=-\underset{j=1}{\overset{k}{\mathrm{\Σ}}}\frac{f_{\mathrm{Li}}\left(q_{\mathrm{Lij}}\right)}{G_{\mathrm{Li}}}\ln \frac{f_{\mathrm{Li}}\left(q_{\mathrm{Lij}}\right)}{G_{\mathrm{Li}}}(G_{\mathrm{Li}}=\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{f}_{\mathrm{Li}}\left(q_{\mathrm{Lij}}\right))$

$f_{\mathrm{Li}}=\frac{1}{k}\underset{j=1}{\overset{k}{\mathrm{\Σ}}}\exp (-\frac{{\left|\right|n_x-n_{\mathrm{Lij}}\left|\right|}_2^2}{2{\mathrm{\σ}}_n^2}).$

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 two point clouds to be fused with different precisions, so as to convert the two point clouds to the same coordinate system.

As an improvement of the invention, said steps S1-S3 may be iteratively repeated, wherein the iteration is performed with the average value of said drift errorControl is terminated within a range wherein said average valueThe method specifically comprises the following steps:

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

According to the method, a drift error model of the low-precision point cloud is established through the high-precision point cloud, then a drift vector of the low-precision point cloud is established, Gaussian weights of unit normal vector information are introduced into the drift vector, density kernel estimation of normal vector difference is provided, parameters of the Gaussian weights are optimized and selected by means of the minimum entropy principle of the information entropy model, and more effective Gaussian weights are obtained. The method can reduce the drift error of the low-precision point cloud caused by factors such as measuring equipment, environmental interference and the like, and improve the precision, thereby more accurately describing the position information of the point cloud. The fusion method has the characteristics of high precision and accurate reflection of the point cloud information morphology characteristics, and achieves the fusion effect.

Drawings

FIG. 1 is a schematic diagram of a process of performing density clustering on low-precision point clouds by using high-precision point clouds according to an embodiment of the invention;

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

fig. 3 is a schematic diagram illustrating results of coarse matching and fine matching of ADF splicing of two groups of point clouds not in the same coordinate system according to an embodiment of the present invention, where 3(a) is a schematic diagram of two groups of point clouds obtained by coarse matching of the moment of inertia, and fig. 3(b) is a schematic diagram of two groups of point clouds obtained by fine matching of the ADF;

fig. 4 is a chromatogram information schematic diagram of error distribution before and after drift of the low-precision point cloud in the embodiment of the present invention (where the darker the color is, the larger the corresponding drift error is), where 4(a) is a schematic diagram of the low-precision point cloud before drift (average error is 0.0649mm), and 4(b) is a schematic diagram of the low-precision point cloud after two drifts (average error is 0.0266 mm);

fig. 5 is a schematic diagram of a result of triangulation of a low-precision point cloud before and after drift in the embodiment of the present invention, where 5(a) is a schematic diagram of a triangular mesh before drift, and 5(b) is a schematic diagram of a triangular mesh after drift.

Detailed Description

The invention is described in further detail below with reference to the figures and the specific embodiments. The following examples are illustrative only and are not to be construed as limiting the invention.

According to the method, for two sets of three-dimensional point cloud data with different precision levels, the drift error of the low-precision point cloud is analyzed by using the high-precision point cloud, the mean value drift of the low-precision point cloud is carried out, the drift error of the low-precision point cloud is eliminated, the smoothness of small-amplitude noise of the low-precision point cloud is realized, the precision level of the low-precision point cloud is improved, and therefore the fusion of two sets of data information is realized.

The point cloud in this embodiment is preferably described as a three-dimensional point cloud of the aerial blade surface. The high-precision point cloud data is preferably obtained through online contact detection of a three-coordinate measuring machine or a machine tool, in this embodiment, three-dimensional point cloud of a blade curved surface can be preferably obtained through the three-coordinate measuring machine (the single-point precision can reach 0.002um), for example, 15 section molded lines are scanned, each molded line contains 20 points, and the total number of the molded lines is 300 points, and the three-dimensional point cloud is used as the high-precision point cloud. The low-precision point cloud is preferably three-dimensional point cloud data obtained by non-contact measurement of a laser scanner or a flexible joint arm and the like. In this embodiment, mass point cloud data of the blade may be preferably acquired through a Hexagon flexible joint arm (with a calibration accuracy of 0.03mm) as a low-accuracy point cloud.

In this embodiment, the high-precision point cloud data is recorded as P_HThe point number is N_HLow precision point cloud data is noted as P_LThe point number is N_L。

The method of the embodiment specifically includes the following steps:

1. topology creation

And (3) creating a topological structure of the low-precision point cloud, specifically including the steps of creating neighborhood information and unit normal vector information of each point in the low-precision point cloud.

For low-precision point cloud P_LCreating neighborhood information, the embodiment can preferably adopt a three-dimensional grid method to establish a low-precision point cloud P_LEach point p_LiK neighborhood information (i is a point number indicating a point cloud, i.e., the ith point, i is 1,2, …, N)_L，N_LRepresents P_LNumber of points) of P_LMiddle distance point p_LiThe nearest k points (k in the embodiment is preferably 8-20), and the neighborhood is marked as Q_Li＝{q_Li1,q_Li2,…,q_Lik}(q_Lij∈P_L,j＝1,2,…,k)。

The establishing steps are as follows:

(a) rasterizing P point cloud_LObtaining each point p_LiThe grid and 26 grids around the grid;

(b) finding all points in the close grid, calculating all points to point p_LiThe distance of (d);

(c) selecting the k points with the nearest distance as p_LiAnd so on, k neighborhood for each sample point can be found.

Establishes a low-precision point cloud P_LEach sample point p_LiK neighborhood information Q of_LiTo this end, a weighted covariance matrix of 3 × 3 is introduced as

Whereinh is an influencing factor, which may be taken as $h=\underset{1\≤j\≤k}{\max }\left\{{\left|\right|q_{\mathrm{Lij}}-{\stackrel{\‾}{q}}_{\mathrm{Li}}\left|\right|}_2\right\}/3.$

According to the formula, the compound has the advantages of,is a symmetric semi-positive definite matrix with a characteristic value lambda_t(t is 1,2, 3) is a real number, and corresponds to a feature vector n_t(t is 1,2, 3) are orthogonal to each other. Let λ₁≤λ₂≤λ₃The magnitude of the characteristic value reflects the neighborhood point set Q_LiIn three feature vectors n₁,n₂,n₃The magnitude of the variation in direction is the point set Q_LiThe least squares fitting plane (or micro-tangent plane) of n₂And n₃Defined degradation plane pi_pAnd corresponding to a normal vector of n₁。

Other similar methods such as an octree method and a kd-tree method can be adopted for creating neighborhood information, but the process of constructing a tree structure involves data encoding and is complicated.

Establishing a low-precision point cloud P_LBefore the topological structure information, the coordinate registration can be firstly carried out on two point clouds with different precisions so as to match the point clouds to the same coordinate system. The coordinate matching specifically comprises the following steps:

(a) during measurement, a standard calibration ball can be set, certain point cloud information of the calibration ball is obtained during measurement, and then a coordinate mapping relation of two groups of point clouds is established according to the position and the attitude of the calibration ballSystem, i.e. low precision point cloud P_LMatching to high precision point cloud P_HThe coordinate system described.

(b) If the calibration ball is not available, the two groups of point clouds can be roughly matched by adopting an inertia moment matching method, and the poses of the two groups of point clouds are adjusted to be approximately consistent. The rough matching method of the moments of inertia does not depend on whether two groups of point clouds have corresponding relations or not, pose adjustment can be conveniently and quickly carried out on any two groups of data, the moments of inertia of the two groups of point clouds are calculated firstly, and then the transformation relation between the two moments of inertia is calculated, so that rough matching from low-precision point clouds to high-precision point clouds is realized. And then, carrying out fine matching on the two groups of point clouds.

Assuming a high precision point cloud P_HAnd under the condition of small single point detection error, converting the low-precision point cloud into a position corresponding to the high-precision point cloud through a point cloud matching algorithm. In consideration of the difference of the positions and the number of the point cloud data, the point cloud matching based on an Adaptive Distance Function (ADF) algorithm is preferably adopted in the present embodiment. The matching is now done with high accuracy and a small number of points P_HTransfer to low precision point cloud P_LAt the location, but with high precision point cloud P_HMore accurately express the position information on the surface of the measured object, so that the high-precision point P is required_HFixedly, and low-precision point cloud P_LPerforming coordinate inverse transformation to obtain high-precision point cloud P_HAnd the coordinate system realizes the coordinate matching of the two groups of point clouds.

2. Drift error model

And then establishing an error relation between the low-precision point cloud and the high-precision point cloud through the definition of the drift error to realize data fusion, and realizing the fairing of small-amplitude noise in the low-precision point cloud through the high-precision point cloud (which is considered to accurately reflect the quality condition of an actually processed curved surface).

High precision point cloud P_HReflecting ideal information of the machined curved surface, and setting S as_HThen low precision point cloud P_LEach sample point p_LiIs the point p_LiA vertical distance to the curved surface; against passing throughTo, if the point P is to be pointed with low precision_LConstructing a curved surface S_LIn this case, the drift error can be described as a high-precision point cloud P_HAt the curved surface S_LUpper error distribution. Apparently, due to the low precision of the point cloud P_LThe drift error is suitable to be described by the latter method when the number is large, but the high-precision point cloud P is established_HRelatively low precision point cloud P_LThe error of (2) is required to be described reversely, the low-precision point cloud can be divided into areas by using the high-precision point cloud, and the errors of point sets in the same area are approximately equal. For this purpose, a low-precision point cloud P is required_LAnd carrying out region division by adopting a density clustering method.

As shown in fig. 2, for a high precision point cloud P_HIn each sample point p_Hr(r is a high-precision point cloud number, r is 1,2, …, N_H，N_HIs P_HPoint number), searching it at low precision point cloud P_LAnd recording the k points (k is 8-20) with the nearest middle-ohm-scale distance as $Q_{\mathrm{Lr}}^H=\{q_{\mathrm{Lr}1}^H,q_{\mathrm{Lr}2}^H,...,q_{\mathrm{Lrk}}^H\}(q_{\mathrm{Lrj}}^H\∈P_L,j=\mathrm{1,2},...,k),$ Then p is calculated by using a moving least square method_HrIn thatProjected foot drop point inThe normal vector is(direction andthe average direction of the points in the point set is consistent). Then using each projection pointFor clustering center, density clustering method is adopted to carry out low-precision point cloud P_LClustering is carried out to form N_HEach clustering unit, the clustering process is shown in FIG. 1, each projection point corresponds to a clustering unit C_Lr。

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

Low precision point cloud P_LAt the point ofThe determined clustering unit range C_LrApproximate drift error (signed) for all points within isI.e. drift error per cluster unit of

${\mathrm{\&Delta;}}_r^H=\left\{\begin{array}{cc}{\left|\right|q_{\mathrm{Lr}}^H-p_{\mathrm{Hr}}\left|\right|}_2& ((q_{\mathrm{Lr}}^H-p_{\mathrm{Hr}})\&CenterDot;n_{\mathrm{Lr}}^H\&GreaterEqual;0)\\ -{\left|\right|q_{\mathrm{Lr}}^H-p_{\mathrm{Hr}}\left|\right|}_2& ((q_{\mathrm{Lr}}^H-p_{\mathrm{Hr}})\&CenterDot;n_{\mathrm{Lr}}^H<0)\end{array}\right.$

From this, a low-precision point cloud P can be obtained_LThe global average drift error (control error of drift) of

${\stackrel{\&OverBar;}{\mathrm{\&Delta;}}}_L=\frac{1}{N_H}\underset{r=1}{\overset{N_H}{\mathrm{\&Sigma;}}}\left|{\mathrm{\&Delta;}}_r^H\right|$

In the formula, N_HRepresenting a high-precision point cloud P_HThe number of the cells.

According to the above definition, a low-precision point cloud P_LEach clustering unit corresponds to a drift errorAnd drift errors of all points in the clustering unitThe drift error model is established, and the low-precision point cloud P is actually determined_LEach sample point p of_Li(i＝1，2,…,N_L) Drift error Δ of_Li。

3. Point cloud drift and fairing

Establishing a low-precision point cloud P_LAfter the error distribution, the drift can be performed according to the topological neighborhood information of the low-precision point cloud, and the drift process is actually an fairing process of the low-precision point cloud with small amplitude noise. From low-precision point clouds P_LEach sample point p_Li(i denotes a low-precision point cloud P_LSerial number of) of the neighborhood points Q_LiEach neighborhood point q_Lij(n is the normal vector)_Lij) Corresponding to a drift error delta_Lij(j ═ 1,2, …, k), then the ith sample point p_LiIs defined as

$m_{\mathrm{Li}}=\frac{\underset{j=1}{\overset{k}{\mathrm{\&Sigma;}}}{w}_{\mathrm{Lij}}{\mathrm{\&Delta;}}_{\mathrm{Lij}}{n}_{\mathrm{Lij}}}{\underset{j=1}{\overset{k}{\mathrm{\&Sigma;}}}{w}_{\mathrm{Lij}}}$

Wherein, w_LijIs the gaussian weight of each neighborhood point. The ith sample point p_LiThe new sample point after the drift is

p′_Li＝p_Li-m_Li

Wherein the Gaussian weight of each neighborhood point is defined as

$w_{\mathrm{Lij}}=\exp (-\frac{{\left|\right|n_{\mathrm{Li}}-n_{\mathrm{Lij}}\left|\right|}_2^2}{2{\mathrm{\&sigma;}}_n^2})(j=\mathrm{1,2},...,k)$

Wherein n is_LiIs p_LiNormal vector of (1), n_LijIs the field Q_LiMidpoint q_LijNormal vector of, σ_nThe change condition of the neighborhood normal vector is reflected, generally called window width, and the value of the window width has important influence on the weight and plays an important role in a drift result. For this purpose, consider the following procedure for window width σ_nOptimizing and selecting:

(a) defining normal vector difference density by using kernel estimation and reflecting neighborhood point set Q_LiAnd (4) change information of an inner normal vector.

(b) According to the normal vector density, an information entropy model is constructed by means of the existing information entropy theory, and the window width sigma is optimized and selected by utilizing the minimum entropy principle_n。

(c) Normal vector Q according to neighborhood point set_LiDetermining the window width σ_nDetermining the optimal window width sigma by adopting a heuristic optimization theory_n。

The window width of the normal vector variation can be based on the point p_LiNeighborhood point set Q of_LiDetermine, set neighborhood Q_LiThe determined local spatial range is omega_LiThen pairIts normal vector is n_xThen, in combination with the kernel estimate, a normal vector difference density of

$f_{\mathrm{Li}}=\frac{1}{k}\underset{j=1}{\overset{k}{\mathrm{\&Sigma;}}}\exp (-\frac{{\left|\right|n_x-n_{\mathrm{Lij}}\left|\right|}_2^2}{2{\mathrm{\&sigma;}}_n^2})(k=8~20)$

By the definition of the normal vector density, the information entropy is introduced to determine the optimal window width sigma_n. In information theory, entropy is a measure of the uncertainty of a system, and the larger the entropy, the larger the uncertainty of the system. For the kernel estimation of the normal vector density, point p_LiNeighborhood Q_LiThe determined local region omega_LiIn, if the density function value at each positionApproximately equal (at this time, the normal vector density loses meaning), and has the maximum entropy when the uncertainty of the data distribution is the maximum; on the contrary, if the density function value is very asymmetric (at this time, the inherent change situation of the normal vector information can be reflected), the uncertainty is minimum, and the entropy is minimum. Thus, the effectiveness of the kernel density estimation can be measured by the concept of normal vector density estimation entropy. According to the definition of information entropy, the point set Q of k points_LiOf each point q_Lij∈Q_LiThe kernel of (j ═ 1,2, L, k) is estimated as f_Li(q_Lij) Then estimate the entropy E_Li(σ_n) Can be defined as

$E_{\mathrm{Li}}\left({\mathrm{\&sigma;}}_n\right)=-\underset{j=1}{\overset{k}{\mathrm{\&Sigma;}}}\frac{f_{\mathrm{Li}}\left(q_{\mathrm{Lij}}\right)}{G_{\mathrm{Li}}}\ln \frac{f_{\mathrm{Li}}\left(q_{\mathrm{Lij}}\right)}{G_{\mathrm{Li}}}$

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

Obviously, E_Li(σ_n) Is about sigma_nThe change is as follows: when sigma is_nThe value of the normal vector density function approaches to the value of → 0Normal vector density estimation entropy E_LiIs a maximum value, i.e. E_LiLn (k); with a_nFrom a gradual increase of 0 to ∞, the estimated entropy is initially gradually reduced and somewhereReaching the minimum value and then gradually increasing; when sigma is_n→ ∞ time, E_LiAgain reaching a maximum value. At a certain optimum valueTake the minimum value The window width can be considered as the optimal window width of the Gaussian weight, and can reflect intrinsic information of the normal vector difference.

Selecting an optimum valueCan theoretically be in the interval range sigma_n∈ (0, + ∞) search, but the range of intervals has virtually no meaning to search_LiNormal vector n of each sample point_LijAveraging and unitizing normal vectors of all the point sets in the neighborhood to obtain a mean value with a change direction ofFor each normal vector n_LijComputingSelecting the maximum value v_maxAnd a minimum value v_minThen the variation range of the normal vector is approximately [ v ]_min,v_max]σ of_n→ 0 availableSubstituted (if)Then get) (ii) a Therefore sigma_nAvailable → + ∞Instead, the search interval ranges ${\mathrm{\&sigma;}}_n^o\&Element;[\frac{v_{\min }}{10},10v_{\max }].$

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

according to the low-precision point cloud and the high-precision point cloud described above, rough matching is performed first, as shown in fig. 3(a), and then fine matching is performed by applying an ADF stitching method, and the matching result is shown in fig. 3 (b). And (3) establishing error distribution information of the low-precision point cloud, wherein an error chromatogram of the low-precision point cloud is shown in fig. 4(a), and point cloud errors after two times of drift are shown in fig. 4 (b). Smoothing of low-amplitude noise of low-precision point clouds has been achieved during the drift process, and the results before and after smoothing are shown in fig. 5(a) and (b).

Claims (8)

1. The utility model provides a three-dimensional point cloud data's fusion method under different precision based on mean shift, it is to two sets of three-dimensional point cloud data that the precision grade is different, utilizes wherein high accuracy point cloud to carry out mean shift to low accuracy point cloud, eliminates the drift error of low accuracy point cloud to realize the fusion of two sets of data information, and the method specifically includes:

s1: establishing topological structure information of the low-precision point cloud, including a neighborhood point set and a unit normal vector of each sample point;

s2: carrying out density clustering on the low-precision point cloud by using the high-precision point cloud, and determining the drift error of each sample point of the low-precision point cloud according to a clustering result;

s3: determining a drift vector of each sample point of the low-precision point cloud by using the topological structure information of the low-precision point cloud and the drift error, and drifting each sample point of the low-precision point cloud according to the drift vector to realize fusion;

wherein, the clustering and determining the drift error in the step S2 specifically include:

firstly, searching k points with the nearest Euclidean distance of each sample point in the high-precision point cloud in the low-precision point cloud to form a k neighborhood of each sample point, and then calculating a projection point and a normal vector of each sample point in the respective k neighborhoods;

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

finally, the drift errors of all the points of the low-precision point cloud in the range of each clustering unit are the same, namely the drift errors are used as the drift errors of the corresponding low-precision point cloud sampling points; wherein, the drift error of all points in the range of each clustering unit can be represented as:

${\mathrm{\&Delta;}}_r^H=(q_{Lr}^H-p_{Hr})\&CenterDot;n_{Lr}^H$

wherein p is_HrFor high-precision point cloud P_HAny one of the plurality of the sampling points,represents a sample point p_HrThe k neighborhood of the field (c),in order to be able to drift the error,is a sample point p_HrIn k neighborhoodThe projected foot drop point in (1) is,for projecting foot-hanging pointsThe corresponding normal vector.

2. The method of claim 1, wherein the step S3 of shifting the sample points is performed according to the following formula:

p′_Li＝p_Li-m_Li

in the formula, p_LiIs any sample point, p 'in low-precision point cloud'_LiIs a sample point p_LiPoint after the shift, m_LiIs a drift vector.

3. The method of claim 2, wherein said drift vector m_LiObtained by the following formula:

$m_{Li}=\frac{\underset{j=1}{\overset{k}{\&Sigma;}}{w}_{Lij}{\mathrm{\&Delta;}}_{Lij}{n}_{Lij}}{\underset{j=1}{\overset{k}{\&Sigma;}}{w}_{Lij}}$

in the formula, Q_LiFor sample points p in low-precision point clouds_LiCorresponding set of neighborhood points, w_LijFor a neighborhood set of points Q_LiAny neighborhood point q_LijGaussian weight of, n_LiIs a sample point p_LiNormal vector of (1), n_LijIs a neighborhood point q_LijCorresponding normal vector, Δ_LijIs a neighborhood point q_LijCorresponding drift error delta_Lij。

4. The method of claim 3, wherein the Gaussian weight is calculated by the following formula:

$w_{Lij}=\exp (-\frac{\mathrm{||}{n}_{Li}-n_{Lij}|{|}_2^2}{2{\mathrm{\&sigma;}}_n^2})$

wherein σ_nIs the window width.

5. Method according to claim 4, characterized in that said window width σ_nOptimum value of (2)Determined by the following formula:

$min\left(E_{Li}\right)=E_{Li}\left({\mathrm{\&sigma;}}_n^o\right),$

wherein,

$E_{Li}\left({\mathrm{\&sigma;}}_n\right)=-\underset{j=1}{\overset{k}{\&Sigma;}}\frac{f_{Li}\left(q_{Lij}\right)}{G_{Li}}ln\frac{f_{Li}\left(q_{Lij}\right)}{G_{Li}}$

$f_{Li}=\frac{1}{k}\underset{j=1}{\overset{k}{\&Sigma;}}\exp (-\frac{\left|\right|n_x-n_{Lij}|{|}_2^2}{2{\mathrm{\&sigma;}}_n^2})$

wherein,as a normalization factor, E_Li(σ_n) To estimate entropy, it is related to the window width σ_nA unitary function of n_xFor a neighborhood set of points Q_LiDetermined local spatial extent omega_LiThe normal vector of any element x in (a).

6. The method according to one of claims 1 to 5, wherein before the topology information of the low-precision point cloud is established in step S1, the two point clouds with different precisions to be fused are co-registered to be transformed into the same coordinate system.

7. The method according to one of claims 1 to 5, wherein the steps S1-S3 are performed iteratively, wherein the number of iterations is controlled such that the average value of the drift error isWithin a certain range, the average valueN_HRepresenting a high-precision point cloud P_HThe number of the cells.

8. The method according to one of claims 1 to 5, characterized in that the high-precision point cloud data are acquired by triggering contact measurement and the low-precision point cloud data are acquired by non-contact measurement.