WO2022099958A1

WO2022099958A1 - Head-face dimension classification method based on three-dimensional point cloud coordinates

Info

Publication number: WO2022099958A1
Application number: PCT/CN2021/080982
Authority: WO
Inventors: 冉令华; 钮建伟; 周玉霖; 刘静; 赵朝义; 张欣; 呼慧敏
Original assignee: 中国标准化研究院; 北京科技大学
Priority date: 2020-11-11
Filing date: 2021-03-16
Publication date: 2022-05-19
Also published as: CN112418030B; CN112418030A; AU2021105639A4

Abstract

A head-face dimension classification method based on three-dimensional point cloud coordinates, comprising: step 1: acquiring head-face three-dimensional point cloud data (1); step 2: defining a key parameter to obtain a radius at a key point (2); step 3: performing data processing to obtain a final head-face data model (3); and step 4: according to the final head-face data model, completing head dimension classification using principal component analysis (4), wherein step 3 comprises: step 31: denoising the acquired head-face three-dimensional point cloud data; step 32: for the denoised point cloud data, adjusting data model grid vertex coordinates; step 33: performing hole filling; and step 34: performing smoothing. According to the method, head-face shape information and curved surface information are comprehensively considered according to point cloud data obtained by three-dimensional scanning; by selecting a data processing template to complete a data processing procedure, data recovery and noise reduction effects are achieved, and the analysis efficiency is improved; in addition, principal component analysis is creatively performed on point cloud coordinates, so that the accuracy of head dimension classification is improved.

Description

A head and face shape classification method based on three-dimensional point cloud coordinates

technical field

The invention relates to the technical field of head type classification, in particular to a head and face type classification method based on three-dimensional point cloud coordinates.

Background technique

The measurement and observation items of the human head and face can reflect the morphological characteristics of the head and face, which are important indicators of human population genetics research. At the same time, the accuracy of head and face data and accurate size classification based on head and face data are also crucial to the design of head and face products. The traditional head and face features are mostly described by one-dimensional or two-dimensional data, so they can only represent the length, width, and circumference of the head and face. The shape and curve of the head and face are very complex, and neither one-dimensional or two-dimensional data can fully reflect it. The shape information and surface information between the measurement points of the human head and face surface. The size classification of human head based on one-dimensional or two-dimensional data has great inaccuracy.

At present, the domestic head shape classification is mainly based on GB/T 2428-1998 "Chinese Adult Head and Facial Dimensions" and GB/T 23461-2009 "Adult Male Head Shape Three-Dimensional Dimensions". GB/T 2428-1998 "Head and Facial Dimensions of Chinese Adults" is based on the measurement data of a small sample, the data measured in 1987-1988 is regressed, and the relationship between one-dimensional size data and a small amount of two-dimensional size data is provided, mainly focusing on two-dimensional size data. Dimensional graphic design application; GB/T 23461-2009 "Three-dimensional dimensions of adult male head shape" is based on the two-dimensional distribution of adult male head width and length and head height and length index, and only gives the three-dimensional size of Chinese adult male head shape. Use limitations. Chinese patent application CN102125323A discloses a method for making head shape of minors aged 4-12 based on the coverage rate of three-dimensional image feature parameters, which includes: a measuring step, scanning a number of minors aged 4-12 through a three-dimensional body scanner image, and then use 3D scanning software to extract feature parameters of the image; preprocessing step, preprocess the extracted feature parameters, that is, create and edit 4-12 minor human head size data files and detect and processing abnormal body size data, and remove the unqualified samples; in the step of size formulation, formulate a statistical table according to the characteristic parameters, and then calculate the coverage rate between the characteristic parameters according to the statistical table, and do not set the coverage rate less than 5‰ model, if it is greater than or equal to 5‰, set the model. Although this method extracts the head feature parameters of minors through three-dimensional scanning, the extracted feature parameters are still limited to the head length and head circumference, so the shape information of the head and face is not fully considered when classifying the head size. and surface information.

SUMMARY OF THE INVENTION

In order to solve the above technical problems, the present invention provides a head and face shape classification method based on three-dimensional point cloud coordinates, which comprehensively considers the shape information and surface information of the head and face, and realizes the classification of the head shape of the human body.

A head and face shape classification method based on three-dimensional point cloud coordinates, comprising the steps of:

Step 1: Collect 3D point cloud data of head and face;

Step 2: Define key parameters and get the radius of key points;

Step 3: Data processing to obtain the final head and face data model;

Step 4: According to the final head and face data model, use principal component analysis to complete the head size classification.

Preferably, in step 1, the head and face are scanned by a three-dimensional scanning device to obtain three-dimensional point cloud data of the head and face.

It is preferred that any of the above-mentioned schemes includes sub-steps in step 2:

Step 21: Extract the equidistant cross-sections of the human head and face;

Step 22: Read the coordinates of each key point of each cross section, and obtain the radius at each key point.

Preferably in any of the above solutions, in step 22, for each cross-section of the head and face, the number of the key points is defined as N+1, and the N+1 key points surround the center point of the cross-section and are located at the edge of the cross-section. On the curve, the first keypoint overlaps with the N+1th keypoint, and the connection between the ith keypoint and the center point of the cross section and the connection between the i+1th keypoint and the center point of the cross section. The included angle of the lines is 360°/N, i≥1 and i≤N.

Preferably, according to any of the above-mentioned schemes, according to the key parameters defined in step 2, define a method for layering the scanned point cloud data, and then proceed to step 3 by matching a suitable data processing template.

Preferably any of the above-mentioned schemes, the step 3 comprises:

Step 31: Denoise the collected 3D point cloud data of the head and face;

Step 32: Adjust the coordinates of each grid vertex of the data model for the denoised point cloud data;

Step 33: fill the hole;

Step 34: Smoothing.

Preferably in any of the above solutions, the step 31 includes determining the first type of noise point and the second type of noise point based on a distance calculation method.

Preferably in any of the above solutions, the first type of noise points are points that are inconsistent with the real data point distribution of the human head and are far away, and the method for determining the first type of noise points is: point cloud data on the human head Concentration, arbitrarily select a point, find the points in the neighborhood of the point whose distance exceeds the set threshold, and set the points outside these thresholds as the first type of noise points; the first type of noise points are removed during data preprocessing.

Preferably in any of the above solutions, the second type of noise point is a point where some data overlap, and the method for determining and deleting the second type of noise point is: for the registered human body point cloud, use the YZ plane to segment the head. It is divided into two parts. Find two points A ₀ and B ₀ with the smallest distance on the edge of the point cloud. The z-coordinates of the two points are represented by z _A and z _B respectively. If z _A ≥ z _B , set these two points as A new edge point; if z _A <z _B , set A ₁ and B ₁ to represent the next edge data point of A ₀ and B ₀ respectively, and calculate the distance d ₁ between A ₁ and B ₀ and between B ₁ and A ₀ The distance between d ₂ , if d ₁ ≥ d ₂ , set points A ₀ and B ₁ as new edge points, otherwise set A ₁ and B ₀ as new edge points; after the new edge point is determined, set it The data points between the original edge points are set as the second type of noise points, which are removed during data preprocessing.

It is preferred that any of the above-mentioned schemes, step 32 comprises the steps:

Step 321: According to the formula

Calculate the geometric center point coordinates of the data model, where n is the total number of grid vertices, and V _i is the position coordinates of the grid model vertices in the three-dimensional space;

Step 322: Calculate the translation transformation matrix M _m from the geometric center point to the coordinate origin, and obtain the maximum widths in the x, y, and z directions respectively, and then calculate the rotation transformation matrix M _r , the rotation transformation matrix M _r makes the y direction has a maximum width;

Step 323: Obtain an affine transformation matrix M according to the translation transformation matrix M _m and the rotation transformation matrix M _r , where M=M _m ×M _r ;

Step 324 : According to the formula V _iN =V _i ×M (i=0,1, _. The position coordinates of the adjusted vertices in three-dimensional space.

Preferably in any of the above solutions, the mesh model is a triangular mesh model.

Preferably, in any of the above-mentioned solutions, step 33 filling the hole is to re-sample the data model after the adjustment of the grid vertex coordinates for the situation that the top of the head data is missing in the process of collecting the three-dimensional point cloud data of the head and face, and reconstruct the data model. .

Preferably in any of the above-mentioned schemes, step 33 specifically includes:

Step 331: Mark the long axis and short axis of the missing area of the overhead data;

Step 332: define a group of planes whose normal direction is parallel to the long axis, and the group of planes intersects with the top surface of the head to form a group of parallel plane slice curves;

Step 333 : each plane slice curve intersects the scan curve formed by the layered 3D point cloud data of the head and face obtained by scanning to obtain two intersection points, and use these intersection points as part of the data points of the re-fitting curve, perform data interpolation, and obtain a complete The head and face data model.

In any of the above solutions, preferably, in step 34, the position of each vertex on the layered scanning curve formed after data interpolation is adjusted to achieve smoothing, and then obtain the final head and face data model.

Preferably in any of the above-mentioned schemes, the step 4 includes:

Step 41: Assume that in the final head and face data model, the point set of the sample point cloud data is P=(p ₁ ,p ₂ ,...,p _n ) ^T , where p _i =(x _i ,y _i ,z _i ) ^T , n is the number of sample points;

Step 42: Select the three-dimensional coordinates x _i , y _i , and _zi of the sample point p _i as three indicators, denoted as X ₁ , X ₂ , X ₃ ;

Step 43: Search the target point pt = (x _t , y _t , z _t ^{) T} _for its neighborhood point set P _tn = (p ₁ , p ₂ ,..., p _k ) ^T , where k is the number of points in the neighborhood number, calculate the distance d _i from all neighbor points to the target point and its mean

Step 44: Seek the linear combination of 3 indicators X ₁ , X ₂ , X ₃

to satisfy the condition

to get the matrix

where i=1,2,3.

Step 45: Calculate the three eigenvectors of the matrix C ^3×3 , which are expressed as λ ₀ , λ ₁ , and λ ₂ in order from small to large, and then obtain the local feature descriptor of the target point

Step 46: Establish a principal component analysis panel according to the principal component analysis result, and complete the size classification.

The head and face shape classification method based on three-dimensional point cloud coordinates of the present invention comprehensively considers the shape information and surface information of the head and face according to the point cloud data obtained by three-dimensional scanning; The noise reduction effect improves the analysis efficiency; at the same time, the principal component analysis is creatively carried out for the point cloud coordinates, which improves the accuracy of the head size classification.

Description of drawings

FIG. 1 is a schematic flowchart of a preferred embodiment of a method for classifying head and face shapes based on three-dimensional point cloud coordinates according to the present invention.

FIG. 2 is a schematic diagram of key points defined for a certain cross section in the embodiment shown in FIG. 1 according to the three-dimensional point cloud coordinate-based head and face shape classification method of the present invention.

3 is a schematic diagram of filling holes for missing data on the top of the head of the embodiment shown in FIG. 1 according to the method for classifying head and face shapes based on three-dimensional point cloud coordinates according to the present invention.

Fig. 4 is a schematic diagram of a complete head and face data model obtained after filling holes according to the embodiment as shown in Fig. 1 of the head and face shape classification method based on three-dimensional point cloud coordinates of the present invention.

5 is a schematic diagram of the smoothing method of the embodiment shown in FIG. 1 according to the method for classifying head and face shapes based on three-dimensional point cloud coordinates according to the present invention.

FIG. 6 is a schematic diagram of the head shape classification of the embodiment shown in FIG. 1 according to the three-dimensional point cloud coordinate-based head and face shape classification method of the present invention.

Detailed ways

In order to better understand the present invention, the present invention will be described in detail below with reference to specific embodiments.

Example 1

As shown in Figure 1, a method for classifying head and face shape based on three-dimensional point cloud coordinates includes the steps:

Step 1: Collect 3D point cloud data of head and face;

Step 2: Define key parameters and get the radius of key points;

Step 3: Data processing to obtain the final head and face data model;

In step 1, the head and face are scanned with a three-dimensional scanning device to obtain three-dimensional point cloud data of the head and face.

In this embodiment, the number of point clouds obtained for the entire head is between 38,420 and 56,200, and the number of point clouds on the upper head (more than two ears) is between 15,223 and 24,221. The number of point clouds in the forward π/2 range) is between 2809 and 3977.

Step 2 includes sub-steps:

Step 21: Extract the equidistant cross-sections of the human head and face;

Step 22: Read the coordinates of each key point of each cross section, and obtain the radius at each key point. In step 22, for each cross-section of the head and face, the number of the key points is defined as N+1, and the N+1 key points surround the center point of the cross-section and are located on the edge curve of the cross-section. The first key point Overlapping with the N+1th key point, the angle between the line connecting the i-th key point and the center point of the cross section and the line connecting the i+1-th key point and the center point of the cross section is 360°/ N, i≥1 and i≤N.

The human head is an uneven surface body. If the cross-sections of the head are extracted at equal intervals, the cross-sections of different parts have the characteristics of different perimeters, different center points, and different arc curvatures, but the adjacent cross-sections have different characteristics. There is a strong similarity and coherence, especially in the arc shape of the cross section. By defining key points, the characteristics of the cross-section of the human head can be represented by the key points, and the head can be divided into multiple contour layers according to the key points.

In this embodiment, the head is extracted with equal intervals of 2mm. As shown in Figure 2, the center point of the cross-section is defined as the origin of the plane coordinate axis, and 61 key points are defined to represent the cross-section of the human head. features, 61 key points are connected in turn in four quadrants, overlapping end to end, that is, key point 1 and key point 61 are the same point, increasing clockwise; key points 1 (61), 16, 31, 46 are respectively associated with the plane The X and Y axes of the coordinate axes intersect, and the key point 1 and the X axis intersect in the positive direction of the X axis. In the clockwise direction, the key point 7 forms an angle of 36° with the X axis, and so on, and the key point 16 intersects the Y axis. In the negative direction of the Y axis, the key point 31 and the X axis intersect in the negative direction of the X axis, and are symmetrical with the key point 1 about the Y axis, the key point 46 intersects with the positive direction of the Y axis, and the key point 61 and the positive X axis. At an angle of 360°, it overlaps with the key point 1; the key point 1 and the key point 31 are the lateral center points of the lateral section of the human head, and the key point 16 and the key point 46 are the front and rear center points of the lateral section of the human head; 61 key points are connected to the origin, the contour of the cross section is divided into 60 segments, the arc center angle corresponding to the arc length of each segment is about 6°, and the arc sides of the arc length are the distances from two adjacent key points to the origin. .

According to the key parameters defined in step 2, define the method of layering the scanned point cloud data, and then proceed to step 3 by matching the appropriate data processing template.

The initial point cloud data of the human head obtained by non-contact measurement is scattered, has holes, noise points and is not smooth, so it is necessary to process the point cloud data.

The step 3 includes:

Step 31: Denoise the collected 3D point cloud data of the head and face;

Step 33: fill the hole;

Step 34: Smoothing.

Affected by the scanning environment and system hardware equipment, noise is inevitably generated during the measurement process. The points far away from the actual data points of the human body are called the first type of noise points, and the points where part of the data overlaps are called the second type of noise. point. For the first and second types of noise points, the selection deletion method is adopted, that is, it is judged whether the human body data is a noise point, and if so, it is directly deleted.

In the step 31, the first type of noise point and the second type of noise point are determined based on the distance calculation method. The method of determining the first type of noise point is: find the point closest to the point in the neighborhood of a point and calculate the distance between the two points, if the distance between the two points is greater than the set threshold, then the The point is the first type of noise point; delete the point cloud data of this point. The method for determining and deleting the second type of noise points is as follows: the second type of noise points are points in the point cloud where part of the data overlaps. For the registered human head point cloud data, use the YZ plane to divide the head into two parts, find two points A ₀ and B ₀ with the smallest distance on the edge of the point cloud, and use z _A for the z coordinates of the two points. , z _B said. If z _A ≥ z _B , set these two points as new edge points; if z _A <z _B , set A ₁ and B ₁ to represent the next edge data points of A ₀ and B ₀ respectively, and calculate A ₁ and B 0 The distance d ₁ between B ₀ and the distance d ₂ between B ₁ and A ₀ , if d ₁ ≥ d ₂ , set points A ₀ and B ₁ as new edge points, otherwise set points A ₁ and B ₀ is a new edge point; after the new edge point is determined, the data point between it and the original edge point is set as the second type of noise point, and it is deleted during data preprocessing. In this embodiment, the threshold value may be set as the theoretical spacing value during laser scanning, that is, the interlayer spacing during laser scanning of the human head during the data collection process. Since the coordinate systems of the multiple head point cloud data obtained by scanning are not completely consistent, all data must be converted to the same coordinate system, which requires data registration of the human head point cloud; during registration, For each sample, the z-axis direction is from the origin to the top of the head, the y-axis direction is from the origin to the tip of the nose, the direction of the x-axis is consistent with the direction of the cross product of the z-axis and the y-axis, and the vector connecting the sample centroid and the tip of the nose is defined, Based on this vector, all head samples are rotated around the z-axis so that the nose tip is aligned in the same direction. The next head data point refers to the point with the next smallest distance on the edge of the scan line corresponding to the front and back of the head.

The 3D point cloud data obtained by scanning is initially read in at any position and in any direction. The data model at this time cannot be displayed in the system correctly, and it will bring a lot of inconvenience to the future processing process. Therefore, after denoising the data model obtained by scanning, the coordinates of each mesh vertex in the model are adjusted, and then the template is matched.

Step 32 includes the steps:

Step 321: According to the formula

In this embodiment, the positive direction of the x-axis is defined as the left side of the data model, and the negative direction is defined as the right side of the data model. The mesh model used is a triangular mesh model.

Define the triangular mesh model as a two-tuple M=(K, V), where V={v ₀ , v ₁ ,..., v _m }, V _i ∈ R ³ represents the mesh model vertex in three-dimensional space The position coordinates of ; K is a simplicial complex representing the mesh topology, each simplicial complex contains a set of simplex {v ₀ }, {v ₀ , v ₁ }, {v ₀ , v ₁ , v ₂ }, They are mesh vertices, connecting edges and triangular patches defined on ^R3 , respectively.

The triangular mesh model consists of two parts: geometric information and topological information. The basic topological entities that express the shape include:

1) Vertex. The positions of vertices are represented by three-dimensional (geometric) points. Points are the most basic elements of a triangular mesh model, and other elements are directly or indirectly formed by points.

2) Edge (Edge). An edge is the intersection of two adjacent faces, and the direction of the edge is from the start vertex to the end vertex.

3) Loop. A ring is a closed boundary composed of ordered, directed edges. The start point of each edge in the ring coincides with the end point of the previous side, and the end point coincides with the start point of the latter side, forming a closed loop with all sides in the same direction. The ring is divided into inner and outer directions. The edges in the inner ring are connected in a clockwise direction, and the edges in the outer ring are connected in a counterclockwise direction.

4) Face. A polygon is a triangular area enclosed by a closed loop. The surface also has directionality. The direction of the surface is determined by the boundary ring. The normal vector of the surface enclosed by the outer ring boundary is outward, which is called the forward surface; the normal vector of the surface enclosed by the inner ring boundary is inward, which is called the reverse surface. In the triangular mesh model, each face of the model surface is the forward face, and the inner face is the reverse face.

The topological entities of the triangular mesh model satisfy the following relationships:

1) A triangle can only intersect with other triangles on its sides;

2) Each inner edge can only have two adjacent triangles;

3) A boundary edge can only have one adjacent triangle.

The 3D scanning obtains layered data points, which cannot completely cover the entire human head. Therefore, it is necessary to resample the data model after the grid vertex coordinates are adjusted to reconstruct the human head model, that is, perform step 33 to fill the hole.

As shown in Figure 3, Figure 3(a) is a schematic diagram of the top position of the data model after the grid vertex coordinates are adjusted. It can be seen that from the top of the head, the missing data is surrounded by a blank area that is approximately an ellipse.

First execute in step 33:

Step 331: Mark the long axis and short axis of the missing area of the overhead data, as shown in Figure 3(a); then execute:

Step 332: Define a set of planes whose normal direction is parallel to the long axis, and this set of planes intersects with the top surface of the head to form a set of parallel plane slice curves, as shown in Figure 3(b); finally execute:

In this embodiment, step 333 is performed by means of a cubic B-spline curve and an interpolation curve masking method.

Each slice curve intersects the scan curve formed by the original layered data points to obtain two intersection points, which are used as part of the data points of the re-fitted curve to perform data interpolation. Interpolate m+1 data points between two intersection points of the same plane slice curve, and the interpolated data points are denoted as Q ₀ , Q ₁ , . . . , Q _m . Then the cubic B-spline curve equation of m+1 interpolated data points Qi ( _i =0, 1, ..., m) can be written as:

where D _j is the control vertex of the curve, and B _j,3 (u) is the B-spline basis function of the cubic curve defined on the nodal vector U=[u ₀ , u ₁ , . . . , u _n+4 ]. According to the requirements of endpoint interpolation and curve definition domain, the endpoint condition of multiple nodes with four nodes at both ends of the domain is adopted, namely: u ₀ =u ₁ =u ₂ =u ₃ =0, u _n+1 =u _n+2 = u _n+3 =un ₊₄ =1. For the data point Q _i (i=0, 1, .

Correspondingly, the node value in the definition domain is obtained

The first point P ₀ (0)=Q ₀ =D ₀ , the end point P _m (1)=Q _m =D _n+1 , define the domain of the curve

The node values inside the equation are substituted into the equation in turn, and the interpolation conditions should be satisfied, that is:

The above formula contains a total of m+1=n-1 equations, which is not enough to determine that it contains n+1 unknown control vertices, and two additional equations given by boundary conditions are added. In order to ensure that the curve after interpolation is continuous with the original curve C1, the first and last points of the curve are the same as the original curve, so that the endpoint conditions of the interpolation curve can be obtained:

P' ₀ (0)=3(D ₁ -D ₀ )=3D ₁ -3Q ₀

_P'm (1)=3( _Dn ₊₁ _-Dn )=3Qm- _3Dn

For a quasi-uniform B-spline curve, since it has a fixed basis function coefficient matrix, the coefficient matrix in the interpolation calculation is also constant. For the quasi-uniform cubic B-spline curve, it can be rewritten as the following matrix form:

Solving the system of linear equations can find all the unknown control vertices. So far, the quasi-uniform cubic B-spline curve represented by the equation system is completely determined.

The surface data generated by the interpolation curve masking method is used to supplement the missing data at the top of the human head model to obtain a complete head and face data model, as shown in Figure 4.

In step 34, the position of each vertex on the layered scanning curve formed by the data interpolation is adjusted to achieve the purpose of smoothing, and then the final head and face data model is obtained.

In this embodiment, the Laplacian smoothing method is used to smooth the curve. The principle is shown in Figure 5. The algorithm adjusts the position of each vertex according to the geometric information around the vertex to achieve the purpose of smoothing, and then obtains the final result. The head and face data model. The Laplacian smoothing algorithm is as follows:

The amount of point cloud data in the final head and face data model is very large. In order to find out the most important components and structures in the point cloud data, remove the influence of redundant data on the size division, the final head and face data model point cloud The data is subjected to principal component analysis to perform data dimensionality reduction and reveal the simple structure of the data.

Step 4 includes:

Step 43: Search the target point p _t =(x _t , y _t , z _t ) ^T for its neighborhood point set P _tn =(p ₁ , p ₂ ,...,p _k ) ^T , where k is the number of points in the neighborhood number, calculate the distance d _i from all neighbor points to the target point and its mean

Step 44: Seek the linear combination of 3 indicators X ₁ , X ₂ , X ₃

to satisfy the condition

to get the matrix

where i=1,2,3.

In step 45, the LFD is used as the local feature description dimension of a point. The smaller the LFD value, the smaller the change in the local area where the point is located in an observation direction of the three-dimensional space; the larger the LFD, the smaller the local area to which the point belongs. The more obvious the change of morphological characteristics. Through the principal component analysis of the main change patterns of the head and face of all subjects, the results show that the spatial shape of the human head and face is composed of a small number of basis vectors.

In this embodiment, the principal component analysis result in step 46 is used to build a principal component analysis panel as shown in FIG. 6 , which contains one probability ellipse, and two lines divide the probability ellipse into four units. Based on the size classification, four types of head models are established: small, short/wide, long/narrow and large, where small, long/narrow, short/wide and large represent the head models located in

units

1, 2, 3 and 4, respectively. population sample. According to the statistical determination of the sample analysis results, the slope of the line dividing 1, 2 and 3, 4 units is about 0.40767.

Example 2

According to the results of the principal component analysis, a more refined principal component analysis panel can be established to conduct a more detailed analysis of the size, such as dividing the head size into 8 medium or even more medium; Sort by size. The results of the classification can provide a basis for the design of head and face products, such as helmets, masks, goggles, protective masks and other product design models.

It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the foregoing embodiments describe the present invention in detail, those skilled in the art should understand: The recorded technical solutions are modified, or some or all of the technical features thereof are equivalently replaced, and these replacements do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the present invention.

Claims

A head and face shape classification method based on three-dimensional point cloud coordinates, comprising the steps of:

Step 1: Collect 3D point cloud data of head and face;

Step 2: Define key parameters and get the radius of key points;

Step 3: Data processing to obtain the final head and face data model;

Step 4: According to the final head and face data model, use principal component analysis to complete the head size classification; it is characterized in that: the step 3 includes:

Step 31: Denoise the collected 3D point cloud data of the head and face;

Step 32: Adjust the coordinates of each grid vertex of the data model for the denoised point cloud data;

Step 33: fill the hole;

Step 34: Smoothing.
The head and face shape classification method based on three-dimensional point cloud coordinates as claimed in claim 1, is characterized in that: in step 2, comprises sub-steps:

Step 21: Extract the equidistant cross-sections of the human head and face;

Step 22: Read the coordinates of each key point of each cross section, and obtain the radius at each key point; in step 22, for each head and face cross section, define the number of the key points as N+1, N +1 keypoints around the center point of the cross section, on the edge curve of the cross section, the first keypoint overlaps with the N+1th keypoint, the line connecting the ith keypoint and the center point of the cross section The angle between the line connecting the i+1th key point and the center point of the cross section is 360°/N, i≥1 and i≤N.
The method for classifying head and face shapes based on three-dimensional point cloud coordinates as claimed in claim 2, characterized in that: according to the key parameters defined in step 2, a method for layering the scanned point cloud data is defined, and then by matching the appropriate data processing template, go to step 3.
The method for classifying head and face shape based on three-dimensional point cloud coordinates according to claim 1, wherein the step 31 includes determining the first type of noise points and the second type of noise points based on a distance calculation method.
The method for classifying head and face shape based on three-dimensional point cloud coordinates according to claim 4, wherein the first type of noise points are points that are inconsistent with the real data point distribution of the human head and are far away, and determine the first type of noise points. The method of one type of noise point is: in the point cloud data set of the human head, select a point arbitrarily, find the points in the neighborhood of this point whose distance exceeds the set threshold, and set the points outside these thresholds as the first type of noise points ; Remove the first type of noise points during data preprocessing.
The head and face shape classification method based on three-dimensional point cloud coordinates according to claim 4, wherein: the second type of noise points are points where some data in the point cloud overlap; for the registered human head points For cloud data, use the YZ plane to divide the head into two parts, and find two points A 0 and B 0 with the smallest distance on the edge of the point cloud. The z-coordinates of the two points are represented by z A and z B respectively. If z A ≥ z B , set these two points as new edge points; if z A <z B , set A 1 and B 1 to represent the next edge data points of A 0 and B 0 respectively, and calculate the distance between A 1 and B 0 The distance d 1 and the distance d 2 between B 1 and A 0 , if d 1 ≥ d 2 , set points A 0 and B 1 as new edge points, otherwise set A 1 and B 0 as new edge points After the new edge point is determined, the data point between it and the original edge point is set as the second type of noise point, and it is deleted during data preprocessing.
The head and face shape classification method based on three-dimensional point cloud coordinates as claimed in claim 1, is characterized in that: step 32 comprises the steps:

Step 321: According to the formula
Calculate the geometric center point coordinates of the data model, where n is the total number of grid vertices, and V i is the position coordinates of the grid model vertices in the three-dimensional space;

Step 322: Calculate the translation transformation matrix M m from the geometric center point to the coordinate origin, and obtain the maximum widths in the x, y, and z directions respectively, and then calculate the rotation transformation matrix M r , the rotation transformation matrix M r makes the y direction has a maximum width;

Step 323: Obtain an affine transformation matrix M according to the translation transformation matrix M m and the rotation transformation matrix M r , where M=M m ×M r ;

Step 324 : According to the formula V iN =V i ×M (i=0,1, . The position coordinates of the adjusted vertices in three-dimensional space.
The head and face shape classification method based on three-dimensional point cloud coordinates as claimed in claim 1, is characterized in that: step 33 specifically comprises:

Step 331: Mark the long axis and short axis of the missing area of the overhead data;

Step 332: define a group of planes whose normal direction is parallel to the long axis, and the group of planes intersects with the top surface of the head to form a group of parallel plane slice curves;

Step 333 : each plane slice curve intersects the scan curve formed by the layered 3D point cloud data of the head and face obtained by scanning to obtain two intersection points, and use these intersection points as part of the data points of the re-fitting curve, perform data interpolation, and obtain a complete The head and face data model.
The method for classifying head and face shapes based on three-dimensional point cloud coordinates according to claim 1, wherein in step 34, the position of each vertex on the layered scanning curve formed after data interpolation is adjusted to achieve smoothness , and then obtain the final head and face data model.
The head and face shape classification method based on three-dimensional point cloud coordinates as claimed in claim 9, it is characterized in that: described step 4 comprises:

Step 41: Assume that in the final head and face data model, the point set of the sample point cloud data is P=(p 1 ,p 2 ,...,p n ) T , where p i =(x i ,y i ,z i ) T , n is the number of sample points;

Step 42: Select the three-dimensional coordinates x i , y i , and zi of the sample point p i as three indicators, denoted as X 1 , X 2 , X 3 ;

Step 43: Search the target point p t =(x t , y t , z t ) T for its neighborhood point set P tn =(p 1 , p 2 ,...,p k ) T , where k is the number of points in the neighborhood number, calculate the distance d i from all neighbor points to the target point and its mean

Step 44: Seek the linear combination of 3 indicators X 1 , X 2 , X 3

to satisfy the condition

to get the matrix
where i=1,2,3.

Step 45: Calculate the three eigenvectors of the matrix C 3×3 , which are expressed as λ 0 , λ 1 , and λ 2 in order from small to large, and then obtain the local feature descriptor of the target point

Step 46: Establish a principal component analysis panel according to the principal component analysis result, and complete the size classification.