A structure point cloud data multi-scale filtering method considering environmental dynamic influence

Abstract

The application discloses a structure point cloud data multi-scale filtering method considering environmental dynamic influence, which firstly adopts a 'K' nearest field method considering structure dynamic influence to perform dynamic filtering processing on original point cloud data, removes scattered outliers in the original point cloud data, and then divides the point cloud data into flat regions and mutation regions based on an improved principal component analysis algorithm (LMSR-PCA), adopts statistical filtering based on local surface fitting for the flat regions, and adopts spatial adaptive bilateral filtering for the mutation regions. The application is used to solve the problems that the existing point cloud data processing method cannot efficiently process massive point cloud data with complex curve characteristics, and cannot solve the problem of excessive smoothing caused by the loss of edge features in the noise reduction process of point cloud data with complex curve characteristics.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of noise reduction of three-dimensional laser scanning point cloud data, and particularly relates to a multi-scale filtering method for structural point cloud data considering dynamic influence of environment. BACKGROUND

[0002] Structures such as high-rise buildings, high-rise towers and long-span bridges are easily affected by the surrounding environment vibration due to their high height or large span and small self-rigidity. Although three-dimensional laser scanning technology can quickly and completely obtain the spatial three-dimensional information of the above monitoring objects, the structure at different positions will also have corresponding dynamic deformation due to the interaction of self-weight, live load and surrounding environment vibration, which will cause the point cloud data to be dispersed and the data amount of the point cloud data to be increased and the processing efficiency of the point cloud data to be reduced. The existing filtering method cannot quantitatively consider the deviation range caused by dynamic deformation when setting parameters, and is easy to identify the deviated point cloud caused by dynamic deformation as an outlier and remove it, resulting in distortion of the point cloud data and failure to well reflect the spatial deformation information of the structure.

[0003] In the aspect of point cloud partitioning, the document (Ren Y J, Li T Z, Xu J K, et al. Overall Filtering Algorithm for Multiscale Noise Removal From Point Cloud Data [J]. IEEE Access, 2020, 9: 110723-110734) calculates the surface variation factor of the point cloud data based on the principal component analysis method, and divides the point cloud data into flat areas and mutation areas. This method can efficiently estimate the normal vector of the point cloud, but when encountering point clouds with sharp features, the normal vector estimated by this method will be too smooth and cannot well reflect the true sharp features. See the document: Wang Xiaohui, Wu Lushen, Chen Hua-wei. Denoising of Scattered Point Cloud Data Based on Normal Vector Distance Classification [J]. Journal of Jilin University (Engineering Edition), 2020, 50(1): 278-288.

[0004] In the aspect of point cloud denoising, by analyzing the distribution characteristics of various structural point cloud data and noise point cloud, the point cloud data can be divided into flat areas and mutation areas. Among them, the curvature of the point cloud data in the flat area is small, the surface deviation degree is small, and the noise point cloud is mainly composed of isolated noise and other large-scale noise point clouds; the curvature of the point cloud data in the mutation area is large, and contains more spatial feature information, and the noise point cloud is mainly composed of clustered noise and other small-scale noise point clouds. Therefore, the noise point clouds in these different areas are called multi-scale noise.

[0005] For the large-scale noise distribution law of the flat area of the point cloud, the main requirement for noise reduction of the smooth area is to enhance the regional characteristics and improve the smoothness of the local surface. The commonly used filtering methods are median filtering and statistical filtering algorithm. Among them, the median filtering is good for eliminating data burr, but the filtering effect of mixed point cloud noise close to each other is not good, which may cause the discontinuity of the processed point cloud data; the statistical filtering algorithm removes outliers according to the density, and the effect of removing point cloud noise with large density difference is good. For the small-scale noise distribution law of the mutation area of the point cloud, the noise reduction requirement of the mutation area is to reduce noise while retaining point and edge detail features. The commonly used nonlinear filtering methods are Gaussian filtering and bilateral filtering. Among them, the Gaussian filtering has good smoothing effect, but the edge corner points are also smoothed; the bilateral filtering can achieve the effect of keeping the edge and reducing the noise, and to some extent, it makes up for the shortcomings of Gaussian filtering, but the filtering parameters need to be input manually, and it cannot achieve adaptive filtering.

[0006] It can be seen that for different types of noise point cloud, there are corresponding filtering algorithms, but the existing single noise reduction algorithm is strong in pertinence, and many parameters need to be input manually, which cannot meet the requirements of processing different scale noise points while ensuring high precision. For the massive point cloud data of the engineering structure with complex spatial characteristics, how to efficiently reduce the noise of the point cloud while accurately reflecting the spatial deformation information of the structure is a problem to be solved at present. SUMMARY

[0007] To solve the above problems, the present application provides a massive point cloud data noise reduction method for dynamic noise point identification, point cloud data partitioning and multi-scale noise dynamic removal, which is used to solve the problems that the existing point cloud data processing method cannot efficiently process massive point cloud data with complex curve characteristics, and cannot solve the problem of excessive smoothing caused by edge feature loss in the process of noise reduction of point cloud data with complex curve characteristics.

[0008] The present application adopts the following technical scheme: a structure point cloud data multi-scale filtering method considering dynamic environmental influence, comprising the following steps:

[0009] Step one: based on the elastic center method, the displacement of the complex structure under the combined action of self weight, automobile load, crowd load, wind load, temperature, uneven settlement of support, etc. is calculated.

[0010] Step two: according to the maximum displacement calculated in step one under the combined action, the dynamic influence range of the structure action is determined, the optimal "K" field of different positions of the complex structure is obtained, the weight factor for measuring the balance of point cloud data under different actions is defined according to the distribution characteristics of point cloud data caused by different actions, and the structure point cloud data set is classified by considering the weight factor according to the principle of minority serving the majority, and the outliers in different point cloud partitions are dynamically removed.

[0011] Step three: Linear regression correction is performed on the principal component analysis algorithm, and then the point cloud data is subjected to weighted principal component analysis based on the improved principal component analysis method (LMSR-PCA), three eigenvalues corresponding to the feature matrix of the point cloud data are estimated, the ratio of the eigenvalue corresponding to the normal vector to the sum of the three eigenvalues is taken as the surface change factor, and the point cloud data is divided into flat areas and mutation areas.

[0012] Step four: For the point cloud noise of the flat area, a statistical filtering algorithm based on local surface fitting is used to remove large-scale noise in the flat area.

[0013] Step five: For the point cloud noise of the mutation area, a spatial adaptive bilateral filtering algorithm is used to remove small-scale noise in the mutation area.

[0014] Step six: Through multi-scale filtering processing of the engineering structure point cloud data, the processed point cloud data is finally obtained.

[0015] Further, the displacement in step one is:

[0016]

[0017] In the formula, F N , F Q , M represent the axial force, shear force and bending moment of the structure under the action of the influence; , F k , M represent the axial force, shear force and bending moment of the structure under the action of the influence; E is the elastic modulus; G is the shear modulus; I is the sectional moment of inertia; A is the sectional area; η is the sectional shape coefficient of shear strain; α is the linear expansion coefficient of the material; h is the distance between the top plate and the bottom plate of the component; is the support reaction; c k is the support displacement.

[0018] Further, the specific steps in step two are:

[0019] Let the original point cloud data set be A.

[0020] 1) Determine the dynamic influence parameters of the structure under different actions.

[0021] Based on the special structure form of the special-shaped arch tower, a field recognition filtering considering dynamic influence is proposed, the dynamic influence parameters under different load influences are analyzed, and the optimal "K" is found in the displacement influence interval, then the range of parameter K can be expressed as follows:

[0022]

[0023] 2) Determine the weight factor considering the unbalanced characteristics of point cloud data under different types of actions.

[0024] Under the action of self-weight, temperature and uneven settlement of support, the structure is in quasi-static state, at this time the point cloud data obtained by three-dimensional laser scanning is uniformly distributed, and the point cloud feature distribution is balanced. When under the action of vehicle load, crowd load and wind load, dynamic influence will be generated on the structure, at this time the point cloud data distribution has great difference and poor balance. For completely balanced point cloud data, the weight factor is 1, and with the decrease of the balance of point cloud data, the weight factor is reduced in turn until 0. Denote the different types of subsets in the original point cloud data set A as {d1, d2, …, d n}。

[0025] The d-type weight factor W[d i ,A] considering the unbalanced nature of the data set can be expressed as:

[0026]

[0027] In the formula, represents the ratio of the occurrence rate of d-type in the overall data set to the total number of all samples; λ i is an input parameter, which can be obtained by considering the nature of data distribution. Collect point cloud data sets under different actions, and perform local surface fitting, then λ i can be defined as:

[0028]

[0029] In the formula, suface[δ] is the number of point clouds on the fitted surface in the data set, δ i is the total number of point clouds in the data set, and the balance of point cloud data is measured by the ratio of the two.

[0030] 3) For a given original data set A, the KNN algorithm calculation criterion based on dynamic classification weighting can be expressed as:

[0031]

[0032] 4) Based on the “K” field recognition algorithm considering dynamic influence, the problem of large calculation amount of traditional KNN is improved, and the outlier points with large attribute difference are efficiently removed, and the point cloud data set B is obtained.

[0033] Further, the specific steps in step three are:

[0034] 1) Linear regression analysis is performed on the sample points p i {(x i ,y i ,z i ), i = 1, 2, 3…n} in the point cloud data set B, and the linear regression model is as follows:

[0035]

[0036] where e is the error i , f i and g i are normal distributions with mean zero and standard deviation σ1, σ2 and σ3, respectively.

[0037] where, where, n is the number of all sample points in the point cloud data set A. The minimum value of the least square-based residual error can be expressed as:

[0038]

[0039] where κ xi = u i - [x i1 , x i2 … x ik ][θ1, θ2… θ k ] T ;

[0040] κ yi = v i - [y i1 , y i2 … y ik ][θ1', θ2'… θ k '] T

[0041] κ zi = w i - [z i1 , z i2 … z ik ][θ1", θ2"… θ k "] T

[0042] The residual square median estimation model is established:

[0043]

[0044] where med r xi 2 is the median of the residual squares of any subset of the x-axis coordinates and the other remaining sample points in the point cloud data set A, and the minimum value thereof is selected as the estimation model, med r yi 2 and med r zi 2 Similarly, r xi , r yi and r zi are calculated.

[0045] Let

[0046] 2) In weighted principal component analysis, the weighting factors are assumed to be a sampling point p, and the distance between the sampling point and its neighboring points is denoted by c. i This indicates that the median residual r is used here. i Alternate distance d i Then the weighting factor α i The calculation formula can be rewritten as:

[0047]

[0048] 3) The weighted average of the coordinates of adjacent points is:

[0049]

[0050] In the formula, p i This indicates the coordinate information of adjacent points.

[0051] 4) Weighted covariance matrix calculation:

[0052]

[0053] After calculating the weighted covariance matrix of the point cloud data, list the characteristic polynomial of the matrix and calculate the three eigenvalues ​​of the sample points in the point cloud dataset B.

[0054] 5) Different regions are distinguished based on different surface variation factors in point cloud data:

[0055] The ratio of the eigenvalue of the sample point's normal vector to the sum of its three eigenvalues ​​is taken as the surface variation factor, denoted as σ(p):

[0056]

[0057] In the formula, λ0 is the eigenvalue corresponding to the normal vector of the sample point, and λ1 and λ2 are the two remaining eigenvalues ​​of the sample point. The average surface change factor of all sample points in the "K" neighborhood is used. express:

[0058]

[0059] Based on the relationship between the surface variation factor of the sample points and the average surface variation factor, noise at different scales can be distinguished. When the sampling point σ(p) is less than When the sample point σ(p) is greater than a certain value, the point belongs to the flat region B1, and noise points within the flat region are large-scale noise; conversely, if the sample point σ(p) is greater than a certain value, the noise point belongs to the flat region B1. Then the point belongs to the mutation region B2, and the noise points in the mutation region are small-scale noise.

[0060] Furthermore, the specific steps in step four are as follows:

[0061] The point cloud data flat region B1 is composed of m sub-sample sets S j (j = 1, 2, …, m) is taken as one of the sample sets S j {(x i ,y i ,z i ), i = 1, 2, 3…l}, local surface fitting is performed on the sample region, and the local quadratic surface fitting model equation of the flat region sub-sample set is obtained:

[0062] φ(x,y)=β0+β1x+β2y+β3x 2 +β4xy+β5y 2

[0063] In the formula, β0, β1, β2, β3, β4, β5 are unknown parameters of the local quadratic surface fitting model equation, and for a large amount of point cloud data, V T PV = min is obtained.

[0064] The normal distance h i of all sample points in the region to the local quadratic fitting surface is calculated, the distribution state of the data points is determined by using Gaussian distribution, the weighted average of the average distance between all sample points and the surface is obtained, the mean u is obtained, and finally the weighted average of the average distance between all sample points and the surface and the average distance q between sample points is obtained. The standard deviation σ. In the statistical filtering algorithm, the threshold parameter λ is set as the multiple of the standard deviation, and the calculation formula is as follows:

[0065]

[0066] c i ′=u+λσ

[0067] In the flat region, statistical analysis is performed, points with average distance outside the standard range are defined as outlier noise points, and are removed from the data to obtain the processed point cloud data set B′.

[0068] Further, the specific steps in step five are:

[0069] 1) Introduce the concept of spatial outlier factor, calculate the spatial distance between the sample points and the domain points, and take it as a scalar to calculate the average value μ and the standard deviation σ of the domain scalar. Due to the random uncertainty of the noise distribution of the actual engineering point cloud data, it is assumed that the scalar obeys the normal distribution, and the spatial outlier points are removed. At this time, the spatial outlier points contain domain points with scalar greater than μ+λσ. Define ε = 0 when the domain point is a spatial outlier point, and vice versa ε = 1.

[0070] 2) Definition of spatial adaptive parameters:

[0071] The definition of the adaptive spatial feature weight function is as follows:

[0072]

[0073] wherein γ s is a spatial feature weight function, used to measure the influence degree of the field point on the sampling point in the spatial domain; γ x is an x-direction feature weight function, used to measure the influence degree of the field point on the sampling point in the x-direction feature domain; γ y is a y-direction feature weight function, used to measure the influence degree of the field point on the sampling point in the y-direction feature domain; γ z is a z-direction feature weight function, used to measure the influence degree of the field point on the sampling point in the z-direction feature domain; η s is a spatial domain standard deviation; η x is an x-direction feature standard deviation; η y is a y-direction feature standard deviation; η z is a normal direction feature standard deviation.

[0074] 3) The spatial adaptive bilateral filtering moving equation can be expressed as:

[0075] p i ′=p i +β x *k xi +β y *k yi +β z *k zi

[0076] wherein p i is the spatial position of the point cloud sampling point, p i ′ is the spatial position of the sampling point after moving once, β x is an x-direction bilateral filtering factor, k xi is an x-direction vector at the sampling point p i , β y and β z are the same. The spatial adaptive bilateral filtering factors β x , β y and β z are respectively expressed as:

[0077]

[0078] The point cloud data of the mutation region B2 is subjected to spatial adaptive bilateral filtering denoising processing to obtain a processed point cloud data set B″.

[0079] Further, the specific steps in step six are:

[0080] The K nearest neighbor method considering the structure dynamic influence, the improved principal component analysis method (LMSR-PCA), the statistical filtering algorithm based on local surface fitting and the spatial adaptive edge filtering algorithm are combined, the double processing of large-scale and small-scale noise in the point cloud data set B is realized, the edge characteristics of the point cloud data can be well reserved during noise reduction, the point cloud data distortion is avoided to the maximum extent, and finally the complete processed point cloud data set C is obtained.

[0081] The beneficial effects of the present application are:

[0082] In the present application, firstly, the K nearest neighbor method considering the structure dynamic influence is adopted to perform dynamic filtering processing on the original point cloud data, and the scattered outliers in the original point cloud data are removed, then the point cloud data is divided into flat regions and mutation regions based on the improved principal component analysis algorithm (LMSR-PCA), the statistical filtering based on local surface fitting is adopted for the flat regions, and the spatial adaptive bilateral filtering is adopted for the mutation regions. The following beneficial effects are achieved:

[0083] 1. An improved point cloud classification algorithm considering the structure dynamic influence parameters is proposed, different weight factors considering the unbalanced properties of the data set under different actions are defined based on the point cloud data under different actions. Compared with the traditional KNN classification algorithm, the improved method can dynamically remove the discrete points in the structure point cloud data, and the classification efficiency and classification accuracy of the original massive point cloud data are improved.

[0084] 2. An improved principal component analysis method (LMSR-PCA) using linear regression correction is proposed, which solves the problem that the normal vector estimated by the traditional principal component analysis method is too smooth when encountering point cloud with sharp features.

[0085] 3. A statistical filtering algorithm based on local surface fitting is proposed to enhance the local features of flat regions and improve the smoothness of local surfaces. A spatial adaptive bilateral filtering algorithm is also proposed to realize spatial filtering parameter adaptation, and the edge detail characteristics are maintained while noise reduction and smoothing are achieved.

[0086] Compared with a single point cloud noise reduction algorithm, the multi-scale dynamic filtering algorithm can identify and filter noise of different scales, can efficiently remove noise points of different scales, and can significantly improve the point cloud quality, while the edge characteristics of the point cloud data can be well reserved during noise reduction, and the point cloud data distortion is avoided. BRIEF DESCRIPTION OF DRAWINGS

[0087] Figure 1 The flowchart of the embodiment of the present application is shown in the figure.

[0088] Figure 2 The improved point cloud classification algorithm considering the structure dynamic influence parameters of the present application is shown in the figure.

[0089] Figure 3 A schematic diagram of the vector estimation of the improved principal component analysis method based on linear regression correction of the application;

[0090] Figure 4 A schematic diagram of the improved statistical filtering method based on local surface fitting of the application;

[0091] Figure 5 A schematic diagram of the spatial adaptive bilateral filtering method of the application. DETAILED DESCRIPTION

[0092] In order for those skilled in the art to better understand the application, the application will be further described in detail below in combination with the drawings and the embodiments of the application.

[0093] As Figures 1-5 shown, the embodiment is a noise reduction processing of bridge tower point cloud data in the bridge completion stage of a special-shaped steel arch single-tower cable-stayed bridge tower, the special-shaped bridge tower is composed of two left and right "heart" type cable tower transverse inwardly inclined and combined, and the scanner is arranged on the bridge deck, wherein the X axis is the longitudinal bridge direction, the Y axis is the transverse bridge direction, and the Z axis is the bridge height direction.

[0094] A multi-scale filtering method considering the dynamic influence of complex structures, which specifically performs the following steps:

[0095] First step: calculate the displacement of the spatial special-shaped arch structure under the combined action of self-weight, automobile load, crowd load, wind load, temperature, and uneven settlement of support based on the elastic center method. Here, the combination of multiple loads is collectively referred to as action, and the corresponding action combination is: action basic combination = 1.1 x (self-weight x 1.2 + uneven settlement of support x 1 + automobile load x 1.4 + crowd load x 1.4 + wind load x 1.1 + temperature x 1.4). Reference standard: General Code for Design of Highway Bridges and Culverts JTG D60-2015.

[0096]

[0097] In the formula, F N , F Q , and M represent the axial force, shear force, and bending moment of the structure under the action of the action, respectively; , F k , and M represent the axial force, shear force, and bending moment of the structure under the action of the action, respectively; E is the elastic modulus; G is the shear modulus; I is the sectional moment of inertia; A is the sectional area; η is the sectional shape coefficient of shear strain; α is the linear expansion coefficient of the material; h is the distance between the top plate and the bottom plate of the component; is the support reaction force; c k is the support displacement.

[0098] Second step: Analyze the dynamic influence parameters of the structure under different actions to determine the optimal "K" field at different positions of the spatial special-shaped arch structure. Analyze the distribution characteristics of point cloud data under different action types, calculate the weight factor considering the unbalanced characteristics under the corresponding action, dynamically identify and classify the point cloud in the optimal K field, and remove the outliers of different types of point cloud data based on the improved KNN classification algorithm. Let the original point cloud data set be A.

[0099] 1) Determine the dynamic influence parameters of the structure under different actions.

[0100] Based on the special structure form of the special-shaped arch tower, a field recognition filtering considering dynamic influence is proposed. The dynamic influence parameters under different load effects are analyzed, and the optimal "K" is found in the displacement influence interval. The range of parameter K can be expressed as follows:

[0101]

[0102] 2) Determine the weight factor considering the unbalanced characteristics of point cloud data under different types of actions.

[0103] Under the actions of self-weight, temperature and uneven settlement of support, the structure is in quasi-static state, and the point cloud data obtained by three-dimensional laser scanning is uniformly distributed, and the point cloud feature distribution is balanced. When under the actions of vehicle load, crowd load and wind load, the structure will be dynamically affected, and the point cloud data distribution has large difference and poor balance. For completely balanced point cloud data, the weight factor is 1, and with the decrease of the balance of point cloud data, the weight factor decreases in turn until 0. Let the different types of subsets in the original point cloud data set A be {d1, d2, …, d n}。

[0104] The d-type weight factor W[d i ,A] considering the unbalanced nature of the data set can be expressed as:

[0105]

[0106] In the formula, represents the ratio of the occurrence rate of d-type in the whole data set to the total number of all samples; λ i is an input parameter, which can be obtained by considering the nature of data distribution. Collect the point cloud data set under different actions, and perform local surface fitting, then λ i can be defined as:

[0107]

[0108] In the formula, suface[δ] is the number of point clouds on the fitted surface in the data set, δ i is the total number of point clouds in the data set, and the balance of point cloud data is measured by the ratio of the two.

[0109] 3) For a given original data set A, the KNN algorithm based on dynamic classification weighting criterion can be expressed as:

[0110]

[0111] 4) Based on the "K" field identification algorithm considering dynamic influence, the traditional KNN calculation amount is improved, and the attribute difference of outliers is efficiently removed, and the point cloud data set B is obtained.

[0112] Step 3: First, linear regression analysis is performed on the point cloud data set B to solve the sample point residual median, and then based on the improved principal component analysis algorithm (MSR-PCA), weighted principal component analysis is performed on the point cloud data to estimate the characteristic matrix corresponding to the three characteristic values of the point cloud data. The ratio of the characteristic value corresponding to the normal vector to the sum of the three characteristic values is taken as the surface change factor, and the point cloud data is divided into flat area and mutation area.

[0113] 1) Linear regression analysis is performed on the sample points p i {(x i ,y i ,z i ),i=1,2,3…n} in the point cloud data set B, and the linear regression model is as follows:

[0114]

[0115] In the formula, the errors e i , f i and g i are normally distributed with mean zero and standard deviation σ1, σ2 and σ3.

[0116] Wherein, In the formula, n is the number of all sample points in the point cloud data set A. The residual minimum value based on least squares can be expressed as:

[0117]

[0118] In the formula, κ xi =u i -[x i1 ,x i2 …x ik ][θ1,θ2…θ k ] T ;

[0119] κ yi =v i -[y i1 ,y i2 …y ik][θ1′,θ2′…θ k ′] T

[0120] κ zi =w i -[z i1 ,z i2 …z ik ][θ1″,θ2″…θ k ″] T

[0121] The residual square median estimation model is established:

[0122]

[0123] In the formula, med r xi 2 is the median of the residual square of any subset of the x-axis coordinates in the point cloud data set A and the other remaining samples, and the minimum value is selected as the estimation model, med r yi 2 and med r zi 2 Similarly, r xi , r yi and r zi are calculated.

[0124] Let

[0125] 2) Weight factor in weighted principal component analysis, assuming that there is a sampling point p, the distance between the sampling point and the adjacent point is represented by c i , here the residual median r i is used instead of the distance d i , then the calculation formula of the weight factor a i can be rewritten as:

[0126]

[0127] 3) The weighted average of the coordinate values of the adjacent points is:

[0128]

[0129] In the formula, p i represents the coordinate information of the adjacent points.

[0130] 4) The weighted covariance matrix is calculated:

[0131]

[0132] After the weighted covariance matrix of the point cloud data is calculated, the characteristic polynomial of the matrix is listed, and three eigenvalues of the sample points in the point cloud data set B are calculated.

[0133] 5) Different regions are distinguished based on different surface variation factors in point cloud data:

[0134] The ratio of the eigenvalue of the sample point's normal vector to the sum of its three eigenvalues ​​is taken as the surface variation factor, denoted as σ(p):

[0135]

[0136] In the formula, λ0 is the eigenvalue corresponding to the normal vector of the sample point, and λ1 and λ2 are the two remaining eigenvalues ​​of the sample point. The average surface change factor of all sample points in the "K" neighborhood is used. express:

[0137]

[0138] Based on the relationship between the surface variation factor of the sample points and the average surface variation factor, noise at different scales can be distinguished. When the sampling point σ(p) is less than When the sample point σ(p) is greater than a certain value, the point belongs to the flat region B1, and noise points within the flat region are large-scale noise; conversely, if the sample point σ(p) is greater than a certain value, the noise point belongs to the flat region B1. Then the point belongs to the mutation region B2, and the noise points in the mutation region are small-scale noise.

[0139] Step 4: For the flat region B1 in the point cloud data, a statistical filtering algorithm based on local surface fitting is used to remove large-scale noise.

[0140] Suppose that the flat region B1 of the point cloud data consists of m subsets S j The set S is composed of (j = 1, 2, ..., m), and one of them is selected as the sample set S. j {(x i ,y i ,z i For the sample region {i = 1, 2, 3…l}, perform local surface fitting on the flat region subset to obtain the local quadratic surface fitting model equation:

[0141] φ(x,y)=β0+β1x+β2y+β3x 2 +β4xy+β5y 2

[0142] In the formula, β0, β1, β2, β3, β4, and β5 are the unknown parameters of the local quadratic surface fitting model equation. For large amounts of point cloud data, these parameters can be determined based on V. T PV = min is obtained.

[0143] Calculate the normal distance h from all sample points within the region to the local quadratic fitted surface. i, the distribution state of data points is determined by Gaussian distribution, the average distance between all sample points and the surface is weighted and averaged to obtain the mean u, and finally the average distance between all sample points and the surface and the average distance between sample points are weighted and averaged to obtain the standard deviation σ. In the statistical filtering algorithm, the threshold parameter λ is set as the multiple of the standard deviation, and the calculation formula is as follows:

[0144]

[0145] c i ′=u+λσ

[0146] Statistical analysis is performed in the flat area, points with average distance outside the standard range are defined as outlier noise points, and are removed from the data to obtain the processed point cloud data set B′.

[0147] Step 5: Based on the problem of sharp sensitivity of the traditional bilateral filtering formula in the area with dramatic changes in features, a new type of spatial adaptive bilateral filtering algorithm is proposed to solve this problem.

[0148] 1) Introduce the concept of spatial outlier factor, calculate the spatial distance between the sampling point and the domain point, and use it as a scalar to calculate the average value μ and the standard deviation σ of the domain scalar. Due to the random uncertainty of the noise distribution of the actual engineering point cloud data, it is assumed that the scalar obeys the normal distribution, and the spatial outlier points are removed. At this time, the spatial outlier points contain domain points with scalar greater than μ+λσ. Define ε=0 when the domain point is a spatial outlier point, and ε=1 otherwise.

[0149] 2) Definition of spatial adaptive parameters:

[0150] The definition of adaptive spatial feature weight function is as follows:

[0151]

[0152] In the formula, γ s is the spatial feature weight function, which is used to measure the influence degree of the domain point on the sampling point in space; γ x is the x-direction feature weight function, which is used to measure the influence degree of the domain point on the sampling point in the x-direction feature domain; γ y is the y-direction feature weight function, which is used to measure the influence degree of the domain point on the sampling point in the y-direction feature domain; γ z is the z-direction feature weight function, which is used to measure the influence degree of the domain point on the sampling point in the z-direction feature domain. η s is the spatial domain standard deviation; η x is the x-direction feature standard deviation; η y is the y-direction feature standard deviation; η z is the normal direction feature standard deviation.

[0153] 3) The spatial adaptive bilateral filtering moving equation can be expressed as:

[0154] p i ′=p i +β x *k xi +β y *k yi +β z *k zi

[0155] where p i is the spatial position of the sampling point of the point cloud, p i ′ is the spatial position of the sampling point after moving once, β x is the bilateral filtering factor in the x direction, k xi is the x direction vector of the sampling point p i , β y and β z , and β x , β y and β z are the same.

[0156]

[0157] The point cloud data of the mutation region B2 is subjected to spatial adaptive bilateral filtering denoising processing to obtain the processed point cloud data set B″.

[0158] Sixth step: combined with the consideration of the structure dynamic influence "K" nearest field method, improved principal component analysis method (LMSR-PCA), local surface fitting statistical filtering algorithm and spatial adaptive edge filtering algorithm, the double processing of large-scale and small-scale noise in the point cloud data set B is realized, while the edge characteristics of the point cloud data are well preserved, the distortion of the point cloud data is avoided to the greatest extent, and finally the complete processed point cloud data set C is obtained.

[0159] In summary, in order to solve the problems of low processing efficiency of complex spatial feature point cloud data of engineering structure and data distortion caused by noise reduction while losing edge detail features, through the above specific technical solutions, the dynamic displacement of the complex structure under the combined action of load and external factors is calculated first, a field recognition method considering the dynamic influence of the structure is proposed, the weight factor considering the unbalanced nature of the data set under different actions is defined, the dynamic outlier elimination of the point cloud data is carried out by using the method, then the improved principal component analysis method (LMSR-PCA) is used to divide the processed point cloud data into flat area and mutation area, the statistical filtering algorithm of local surface fitting is used for large size noise in the flat area, and the spatial adaptive bilateral filtering algorithm is used for small scale noise in the mutation area, so that the processing efficiency of the point cloud data can be reasonably improved, and the double processing of large scale and small scale noise is realized, the edge detail features of the point cloud are well preserved while the noise is reduced to a great extent, the data distortion is avoided, and the real deformation information of the structure is well reflected.

[0160] It should be noted that, for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should also be considered within the scope of protection of the present application.

Claims

1. A method for multi-scale filtering of structure point cloud data considering environmental dynamic impact, characterized in that: The method comprises the following steps: Step one: based on the elastic center method, the displacement of the complex structure under the combined action of self weight, automobile load, crowd load, wind load, temperature and uneven settlement of support is calculated; Step two: the specific steps are: Record the original point cloud data set as A; 1) determine the dynamic influence parameters of the structure under different actions; Based on the structure form, a field recognition filtering considering dynamic influence is proposed, the dynamic influence parameters under different load influences are analyzed, and the optimal "K" is found in the displacement influence interval, and the range of the parameter K is represented as follows: 2) determine the weight factor considering the unbalanced characteristics of point cloud data under different types of actions; Under the action of self-weight, temperature and uneven settlement of support, the structure is in quasi-static state, at this time the point cloud data obtained by three-dimensional laser scanning is uniformly distributed, and the point cloud feature distribution is balanced; while under the action of vehicle load, crowd load and wind load, the structure will be dynamically affected, at this time the point cloud data distribution has great difference and poor balance; for the completely balanced point cloud data, the weight factor is 1, with the decrease of the balance of the point cloud data, the weight factor decreases in turn until 0; record different types of subsets in the original point cloud data set A as {d1, d2, …, d n} d-class weight factors W[d] that take into account the imbalanced nature of the dataset i A] are represented as: In the formula, represents the ratio of the occurrence rate of d class to the total number of all samples in the whole data set; λ i is an input parameter, which is obtained by considering the properties of data distribution; collect point cloud data sets under different actions, and perform local surface fitting, then λ i is defined as: In the formula, the number of point clouds on the fitting surface in the surface [delta] data set, delta i The balance of the point cloud data is measured by the ratio of the total number of point clouds in the data set to the number of point clouds on the fitting surface. 3) for a given original data set A, the KNN algorithm calculation criterion based on dynamic classification weighting is represented as: 4) based on the "K" field recognition algorithm considering dynamic influence, the problem of large calculation amount of traditional KNN is improved, and the outlier points with large attribute difference are efficiently removed, and a point cloud data set B is obtained; Step three: the principal component analysis algorithm is linearly regressed, and then based on the improved principal component analysis method, the point cloud data is weighted principal component analyzed, three characteristic values corresponding to the feature matrix of the point cloud data are estimated, the ratio of the characteristic value corresponding to the normal vector to the sum of the three characteristic values is taken as the surface change factor, and the point cloud data is divided into flat regions and mutation regions; Step four: for the point cloud noise of the flat region, a statistical filtering algorithm based on local surface fitting is used to remove large-scale noise in the flat region; Step five: for the point cloud noise of the mutation region, a spatial adaptive bilateral filtering algorithm is used to remove small-scale noise in the mutation region; Step six: through multi-scale filtering processing of the engineering structure point cloud data, the processed point cloud data is finally obtained.

2. The method of claim 1, wherein: The displacement in step one is: where F N , F Q , M represent the axial force, shear force and bending moment of the structure under the action of the influence, respectively; , F Q , M represent the axial force, shear force and bending moment of the structure under the action of the influence, respectively; E is the elastic modulus; G is the shear modulus; I is the sectional moment of inertia; A is the sectional area; η is the sectional shape coefficient of shear strain; α is the linear expansion coefficient of the material; h is the distance between the top plate and the bottom plate of the component; is the support reaction force; c k is the support displacement.

3. The method of claim 2, wherein: The specific steps in step three are: 1) Linear regression analysis is performed on the sample points p in the point cloud dataset B i {(x i ,y i ,z i ), i = 1, 2, 3…n}, and the linear regression model is as follows: where the error e i , f i , and g i are normally distributed with zero mean and standard deviations σ1, σ2, and σ3, respectively. wherein, in which, n is the number of all sample points in the point cloud dataset A; the minimum value of the residual based on the least square is represented as: wherein κ xi = u i - x i1 , x i2 … x ik [θ1, θ2… θ k ] T ; K yi = v i - [y i1 , y i2 … y ik ][θ1', θ2'... θ k '] T k zi = w i - z i1 , z i2 … z ik [θ1", θ2"... θ k "] T Establish a residual square median estimation model: In the formula, med r xi 2 For any subset of x-axis coordinates in the point cloud dataset A and the median of the residual squares of the other remaining samples, the minimum value is selected as the estimated model, med r yi 2 And med r zi 2 Similarly, r xi , r yi , and r zi are calculated. Let 2) Weight factor in weighted principal component analysis, assuming a sampling point p, the distance between the sampling point and the adjacent point is denoted as c i , here the residual median r i is used to replace the distance c i , then the calculation formula of the weight factor α i is rewritten as: 3) the weighted average of the coordinate values of adjacent points is: In the formula, p i represents coordinate information of adjacent points; 4) the weighted covariance matrix is: After calculating the weighted covariance matrix of the point cloud data, the characteristic polynomial of the matrix is listed, and three characteristic values of the sample points in the point cloud data set B are calculated; 5) different regions are distinguished according to different surface change factors of the point cloud data: Take the ratio of the characteristic value of the sample point normal vector to the sum of the three characteristic values as the surface change factor, denoted as sigma (p): In the formula, λ0is the eigenvalue corresponding to the normal vector of the sample point, λ1and λ2are the remaining two eigenvalues of the sample point; the average surface variation factor of all sample points in the "K" neighborhood is represented by ​ According to the relationship between the surface variation factor of the sample point and the average surface variation factor, different scale noises are distinguished; when the sample point σ(p) is less than , the point belongs to the flat region B1, and the noise point in the flat region is large scale noise; otherwise, if the sample point σ(p) is greater than , the point belongs to the mutation region B2, and the noise point in the mutation region is small scale noise.

4. The method of claim 3, wherein: The specific steps in step four are: A point cloud data flat region B1 is composed of m sub-sample sets S j j = 1, 2,..., m, and one of them is taken as a sample set S j {(x i ,y i ,z i ), i = 1, 2, 3,..., l} is locally fitted with a surface, and a local quadratic surface fitting model equation of the flat region sub-sample set is obtained. φ(x, y) = β0+ β1x + β2y + β3x 2 +β4xy+β5y 2 In the formula, β0, β1, β2, β3, β4, and β5 are unknown parameters of a local quadratic surface fitting model equation. In view of the huge amount of point cloud data, V T PV = min Calculate the normal distance h of all sample points in the region to the local quadratic fitting surface i , the distribution state of the data points is determined by Gaussian distribution, the average distance between all sample points and the surface is weighted average, the mean u is obtained, and finally the average distance between all sample points and the surface and the average distance q between sample points are weighted average, and the standard deviation σ is obtained; in the statistical filtering algorithm, the threshold parameter λ is set as the multiple of the standard deviation, and the calculation formula is as follows: c i ′= u + λσ In the flat region, statistical analysis is carried out, points with average distance outside the standard range are defined as outlier points, and are removed from the data to obtain the processed point cloud data set B'.

5. The method of claim 4, wherein: The specific steps in step five are: 1) introduce the concept of spatial outlier factor, calculate the spatial distance between the sampling point and the field point, and take it as a scalar to calculate the average value mu and the standard deviation sigma of the field scalar; Due to the random uncertainty of the noise distribution of the actual engineering point cloud data, it is assumed that the scalar obeys the normal distribution, and the spatial outlier points are removed; At this time, the spatial outlier points contain the field points whose scalar is greater than mu+lambda sigma; Define epsilon=0 when the field point is a spatial outlier point, and vice versa epsilon=1; 2) definition of spatial adaptive parameter: The definition of adaptive spatial feature weight function is as follows: In the formula, γ s is a spatial feature weight function for measuring the influence degree of the domain point on the sampling point in space; γ x is an x-direction feature weight function for measuring the influence degree of the domain point on the sampling point in the x-direction feature domain; γ y is a y-direction feature weight function for measuring the influence degree of the domain point on the sampling point in the y-direction feature domain; γ z is a z-direction feature weight function for measuring the influence degree of the domain point on the sampling point in the z-direction feature domain; η s is a spatial domain standard deviation; η x is an x-direction feature standard deviation; η y is the y-direction characteristic standard deviation; η z is the normalised standard deviation of the features; 3) the moving equation of the spatial adaptive bilateral filtering is represented as: p i ′=p i +β x *k xi +β y *k yi +β z *k zi where p i is the spatial position of a point cloud sample, p i ′ is the spatial position of the sample after moving once, β x is the bilateral filtering factor in x direction, k xi is the vector in x direction at sample p i , β y and β z Similarly, the spatial adaptive bilateral filtering factors β x , β y and β z are represented as: The point cloud data of the mutation region B2 is subjected to spatial adaptive bilateral filtering denoising to obtain a processed point cloud data set B".

6. The method of claim 5, wherein: The specific steps in the step six are: In combination with the "K" nearest neighbor method considering the structural dynamic influence, the improved principal component analysis method, the statistical filtering algorithm of local surface fitting and the spatial adaptive edge filtering algorithm, the double processing of large-scale and small-scale noise in the point cloud data set B is realized, the edge features of the point cloud data are well preserved during the denoising, the distortion of the point cloud data is avoided to the greatest extent, and finally the complete processed point cloud data set C is obtained.

Images