CN111882595A - Human body semantic feature extraction method and system - Google Patents

Abstract

The invention discloses a human body semantic feature extraction method and system in the technical field of human body measurement, which can accurately obtain human body semantic features and has high matching accuracy. Including: collecting 3D human body feature data to obtain a 3D human body model; preprocessing the 3D human body model, adjusting the orientation and center coordinates of the 3D human body model and placing the 3D human body model at the coordinate origin of the world coordinate system; performing shape segmentation on the 3D human body model And extract the skeleton features of the 3D human body model, based on the skeleton similarity template selection algorithm, select the template model and the 3D human body model from the constructed 3D human body semantic feature database according to the skeleton features of the 3D human body model for registration, and obtain the parameterized human body model. Model; use the NURBS curve to fit the feature sampling points on the parameterized human body model, calculate the length of the fitted feature curve, and obtain the human body semantic features of the three-dimensional human body model.

Description

A method and system for extracting human semantic features

technical field

The invention belongs to the technical field of human body measurement, and in particular relates to a method and system for extracting human semantic features.

Background technique

With the continuous development of the social economy and the continuous improvement of people's living standards, more and more people are pursuing a personalized lifestyle, and there is an increasing demand for customized products. Online shopping, virtual try-on, clothing customization, fitness industry, ergonomic design and other industries also hope to improve the technological content of products and increase the added value of services. One of the most important directions for improvement is to quantify human size and body shape. For example, online virtual try-on needs to generate a three-dimensional model that is exactly the same as the real body, and clothing customization needs to obtain the body shape data of customers in multiple dimensions.

The existing methods for obtaining human semantic features can be divided into contact methods and non-contact methods, and the traditional manual measurement method is the most common contact method. With the development of 3D scanning technology, the non-contact method will become the mainstream semantic feature extraction method in the future, but the extraction matching degree of human semantic features in the existing technology is poor, and the matching effect is not ideal.

SUMMARY OF THE INVENTION

In order to solve the deficiencies in the prior art, the present invention provides a method and system for extracting human body semantic features, which can accurately obtain human body semantic features and have high matching accuracy.

In order to achieve the above purpose, the technical scheme adopted in the present invention is: a method for extracting human body semantic features, comprising: collecting three-dimensional human body feature data to obtain a three-dimensional human body model; preprocessing the three-dimensional human body model, adjusting the orientation of the three-dimensional human body model, The center coordinate and let the 3D human model be at the coordinate origin of the world coordinate system; the shape segmentation of the 3D human model is performed and the skeleton features of the 3D human model are extracted. The template model is selected from the 3D human body semantic feature database for registration with the 3D human body model to obtain a parameterized human body model; the NURBS curve is used to fit the feature sampling points on the parameterized human body model, and the length of the fitted characteristic curve is calculated to obtain Human Semantic Features of 3D Human Models.

Further, the registration includes rigid body registration and non-rigid body registration;

The rigid body registration includes:

的有向包围盒

计算从三维人体模型D的有向包围盒D_B到模板模型

的有向包围盒

的仿射变换T；对三维人体模型D施加仿射变换T得到第一中间模型D′：Construct the directed bounding box D _B of the 3D human body model D and the template model

directed bounding box of

Compute the directed bounding box D _B from the 3D human model D to the template model

directed bounding box of

The affine transformation T of ; apply the affine transformation T to the three-dimensional human body model D to obtain the first intermediate model D′:

D′=TD (13)

The iterative closest point algorithm is used to calculate the optimal rigid body transformation parameters R and t from the first intermediate model D′ to the template model T, and then obtain the second intermediate model D″:

D″=RD′+t (14)

Among them, R is the linear transformation parameter, and t is the translation transformation parameter;

The non-rigid body registration includes:

向第二中间模型

变形，得到初步配准模型

Using the Laplacian mesh deformation algorithm to transform the template model

to the second intermediate model

Deformed to get a preliminary registration model

和第二中间模型D″上建立数据误差函数E_d和光滑度误差函数E_s，并最小化其误差之和，获得参数化的人体模型

in the preliminary registration model

Establish data error function Ed and smoothness error function _Es on the second intermediate model _D ″, and minimize the sum of their errors to obtain a parameterized human body model

Further, the data error function _Ed and smoothness error function _Es are obtained by the following formulas:

的顶点个数，v′_i为第i个顶点的坐标，距离函数dist²()为变形后网格与目标网格中最近相容点的距离，T_i为第i个顶点对应的3×3变换矩阵；将法相量角度小于90°，欧式距离小于10cm的两个点，定义为相容点，其中距离最近的一个称为最近相容点，T_j为第j个顶点对应的3×3变换矩阵，

为模型

上的边，|| ||_F表示弗罗贝尼乌斯范数。where n represents the preliminary registration model

The number of vertices of , v′ _i is the coordinate of the ith vertex, the distance function dist ² () is the distance between the deformed mesh and the nearest compatible point in the target mesh, T _i is the 3× corresponding to the ith vertex 3 Transformation matrix; two points whose normal phasor angle is less than 90° and whose Euclidean distance is less than 10cm are defined as compatible points, and the one with the closest distance is called the closest compatible point, and T _j is the 3× 3 transformation matrices,

for the model

On the edge, || || _F denotes the Frobenius norm.

Further, the template selection algorithm of the skeleton similarity is specifically:

为三维人体模型D上的p骨节点的坐标，

为三维人体模型D上的q骨节点的坐标，

为模板模型

上的p骨节点的坐标，

为模板模型

上的q骨节点的坐标，

为向量夹角符号，

为权重因素，m＝1，2，3，4，L₁、L₂、L₃、L₄为四个权重级别，人体主干中的骨节点权重级别为L₁，上臂、大腿、头部的权重级别为L₂，下臂、小腿的权重级别为L₃，手部、脚部的权重级别为L₄，且

Among them, SE is the bone similarity parameter,

is the coordinate of the p-bone node on the three-dimensional human model D,

is the coordinate of the q-bone node on the three-dimensional human model D,

template model

The coordinates of the p-bones node on,

template model

the coordinates of the q-bone node on,

is the vector angle symbol,

is the weight factor, m=1, 2, 3, 4, L ₁ , L ₂ , L ₃ , and L ₄ are four weight levels, the bone node weight level in the main body of the human body is L ₁ , the upper arm, thigh, head The weight level is L ₂ , the weight level of the lower arm and the calf is L ₃ , the weight level of the hands and feet is L ₄ , and

Further, the construction method of the three-dimensional human body semantic feature database is: selecting the human body posture model in the open source three-dimensional human body database, and performing surface subdivision, setting the number of faces and vertices of the human body posture model, and obtaining the initial three-dimensional human body database. ; segment the human body pose model in the initial 3D human body database to form a 3D human body segmentation database; extract 3D human bones from the segmentation model in the 3D human body segmentation database to form a 3D human skeleton database; analyze the human bones in the 3D human skeleton database Semi-automatic labeling of semantic feature sampling points to form a 3D human body semantic feature database.

Further, the 3D human body semantic feature database incorporates the shape segmentation result of the 3D human body model generated in the process of extracting the human body semantic feature into the 3D human body segmentation database, and the skeleton features of the 3D human body model are included in the 3D human skeleton database, including the human body semantic features. The 3D human body model is incorporated into the 3D human body semantic feature database.

Further, the segmentation of the human body posture model in the initial three-dimensional human body database is specifically:

Perform supervoxel clustering on vertex data, and the result is an over-segmentation of vertex data, including:

1) The model space is divided into grids, and the point cloud data forms point clusters, which are called voxels. A spatial index is established for the voxels, and the octree algorithm is used to organize these voxels;

2) Select multiple voxels as seed voxels;

3) Calculate the similarity between voxels, and recursively fuse adjacent voxels with seed voxels to form supervoxels;

Judging the concave-convex relationship between adjacent supervoxels according to the preset judgment criteria;

The random sampling consensus algorithm is used to fit the cutting plane on the concave edge, and the human body pose model is segmented.

Further, the similarity between voxels is calculated, specifically:

为体素直径。Among them, ω _c is the weight factor of color difference, ω _s is the weight factor of distance difference, ω _n is the weight factor of normal difference, D _c (i, j) is the CIELab color difference, D _s (i, j) is Coordinate distance difference, D _n (i, j) is the normal difference,

is the voxel diameter.

Further, the judgment criterion is determined by a line connecting the centroids of adjacent supervoxels and the normal direction on the centroids and the common adjacent supervoxels of the adjacent supervoxels.

A human body semantic feature extraction system includes a processor and a storage device, wherein a plurality of instructions are stored in the storage device for the processor to load and execute the steps of the foregoing method.

Compared with the prior art, the beneficial effects achieved by the present invention are as follows: the present invention performs shape segmentation on the three-dimensional human body model and extracts bone features, selects a template model from the constructed three-dimensional human body semantic feature database for registration, and obtains a parameterized human body model. Human body model, and use the NURBS curve to fit the feature sampling points on the parameterized human body model to obtain the human body semantic features of the three-dimensional human body model. The extraction method of human body semantic features follows the basic knowledge of human anatomy and can accurately obtain human body semantic features. The matching effect is good; at the same time, the constructed 3D human body semantic feature database can incorporate the input 3D human body model into the database, thus broadening the topology structure of the 3D human body semantic feature database.

Description of drawings

1 is a schematic flowchart of constructing a three-dimensional human body semantic feature database in a method for extracting human body semantic features provided by an embodiment of the present invention;

2 is a schematic flowchart of a method for extracting human semantic features according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a topology structure of a human body semantic feature extraction system provided by an embodiment of the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings. The following examples are only used to illustrate the technical solutions of the present invention more clearly, and cannot be used to limit the protection scope of the present invention.

Example 1:

As shown in Figure 1, a method for constructing a three-dimensional human body semantic feature database includes: selecting a human body pose model in an open source three-dimensional human body database, and performing surface subdivision, setting the number of faces and vertices of the human body pose model, and obtaining Initial 3D human body database; segment the human body pose model in the initial 3D human body database to form a 3D human body segmentation database; extract 3D human skeletons from the segmentation models in the 3D human body segmentation database to form a 3D human skeleton database; The human skeleton is semi-automatically annotated with semantic feature sampling points to form a 3D human semantic feature database.

The three-dimensional human body database containing multiple human body poses is used as the basic database. This method takes the MPI database as an example, removes the models with too complex poses in the MPI database, and performs surface subdivision. 25790 model to obtain the initial 3D human body database.

A 3D human body shape segmentation algorithm based on supervoxel clustering and the concavity and convexity of adjacent voxels, performs shape segmentation on the model in the initial 3D human body database, and divides the model into head, upper arm, lower arm, hand, For chest, waist, buttocks, thighs, calves, feet, follow these steps:

(a) Perform super-voxel clustering on vertex data, and the result is over-segmentation of vertex data. The specific steps are:

1) The model space is divided into grids, and the point cloud data forms point clusters, which are called voxels. A spatial index is established for the voxels, and the octree algorithm is used to organize these point clusters;

2) Select multiple voxels as seed voxels among many voxels;

3) Calculate the similarity between voxels. The seed voxels are recursively fused with adjacent voxels to form supervoxels. The calculation formula of voxel similarity is:

为体素直径；D_c(i，j)为CIELab颜色差异，D_s(i，j)为坐标距离差异，D_n(i，j)为法线差异，其计算公式分别为：Among them, ω _c is the weight factor of color difference, ω _s is the weight factor of distance difference, ω _n is the weight factor of normal difference, color difference ω _c , distance difference ω _s and normal difference ω _n are used to control the vertex The influence of color difference, spatial distance and concave-convexity in similarity calculation;

is the voxel diameter; D _c (i, j) is the CIELab color difference, D _s (i, j) is the coordinate distance difference, and D _n (i, j) is the normal difference. The calculation formulas are:

D _c (i, j)=ΔA=||V(i)-V(j)|| _CIELab (2)

where V(i) is the CIELab color of the ith voxel, and V(j) is the CIELab color of the jth voxel;

Among them, d _i is the coordinate of the ith voxel, d _i =[x _i , y _i , z _i ], _xi is the x coordinate of the ith voxel, y _i is the y coordinate of the ith voxel, z _i is the z coordinate of the ith voxel, d _j is the coordinate of the j th voxel, d _j =[x _j , y _j , z _j ], x _j is the x coordinate of the j th voxel, and y _j is the j th individual The y coordinate of the voxel, z _j is the z coordinate of the jth voxel;

为第i个体素法向的x坐标，

为第i个体素法向的y坐标，

为第i个体素法向的z坐标，n_j为第j个体素的质心法向，

为第j个体素法向的x坐标，

为第j个体素法向的y坐标，

为第j个体素法向的z坐标；Among them, n _i is the normal direction of the i-th individual quality heart,

is the x coordinate of the i-th voxel normal,

is the y coordinate of the i-th voxel normal,

is the z coordinate of the normal direction of the i-th voxel, n _j is the centroid normal direction of the j-th voxel,

is the normal x coordinate of the jth voxel,

is the y coordinate of the jth voxel normal,

is the z coordinate of the jth voxel normal;

(b) Judging the concave-convex relationship between adjacent supervoxels. The edge within a concave voxel is called a concave edge. If the adjacent supervoxels are more "concave", they should be cut to determine the concave-convex relationship. There are two criteria. Criterion 1 is based on the normal relationship between the line connecting the centroids of adjacent supervoxes and the centroid. Criterion 2 considers the common adjacent supervoxels on the basis of criterion 1. For criterion 1, formula (5) can be used. Express:

CC(sv _i , sv _j )=CC _b (sv _i , sv _j )∧CC _b (sv _i ,sv _c )∧CC _b (sv _j ,sv _c ) (5)

Among them, CC(sv _i , sv _j ) is the first criterion of concavity and convexity, CC _b (sv _i , sv _j ) is the concavo-convex relationship between the supervoxel sv _i and the supervoxel sv _j , CC _b (sv _i , sv j ) sv _c ) is the concave-convex relationship between the supervoxel sv _i and the super-voxel sv _c , CC _b (sv _j , sv _c ) is the basic concave-convex relationship between the super-voxel sv _j and the super-voxel _svc ;

The calculation of CC _b (sv _i , sv _j ) is shown in formula (6):

为向量夹角符号，计算公式为：Among them, Ni is the centroid normal of the ith supervoxel, Nj is the centroid normal of the jth supervoxel, β _T is the offset, and β(N _i , N _j ) is the _{ith supervoxel} The angle between the voxel and the j-th supervoxel centroid normal,

is the symbol of the included angle of the vector, and the calculation formula is:

的计算公式为：

The calculation formula is:

Among them, X _i is the centroid coordinate of the supervoxel sv _i , X _j is the centroid coordinate of the supervoxel sv _j , and the second criterion can be defined as formula (7):

Among them, SC(sv _i , sv _j ) is the second criterion of concavity and convexity, and the calculation method of θ(sv _i , sv _j ):

Calculation method of θ _T (β(N _i , N _j )):

α＝0.25，β_off＝25°；in,

α = 0.25, β _off = 25°;

To sum up, the concave-convexity judgment method is shown in formula (10):

conv(sv _i , sv _j )=CC(sv _i , sv _j )∧SC(sv _i , sv _j ) (10)

Among them, conv(sv _i , sv _j ) is the concave-convexity between the ith supervoxel and the jth supervoxel;

(c) According to the concave-convex relationship judged in step (b), a random sampling consensus algorithm is used to fit the segmentation plane on the concave edge, and the human body pose model is segmented to form a three-dimensional human body segmentation database. The human body models in the database are divided into 16 meaningful parts;

The human body joints are located at the connection part of the rigid body part, and the boundary center of the two sub-regions is used as the skeleton node, and the skeleton nodes are connected according to the anatomical relationship of the joints to form the human skeleton; this operation is performed on all the models in the human body segmentation database to obtain a three-dimensional Human Skeleton Database.

Semantic feature expansion is a semi-automatic operation. Manually mark the feature sampling points of a model, and all models can use the same sampling point; the feature curves are fitted to the feature sampling points of each model, and the 3D human body semantic feature database is obtained. .

In this embodiment, there are two data sources in the constructed 3D human body semantic feature database. One is to expand the open source 3D human body database according to the aforementioned method; absorb. In this embodiment, the template model in the 3D human semantic feature database is marked with semantic feature sampling points for fitting the semantic feature curve, the semantic feature sampling points marked on the template model are located near the actual position of the semantic feature, and the semantic feature curve is a cubic curve for locating semantic features.

Embodiment 2:

Based on the three-dimensional human body semantic feature database constructed in the first embodiment, this embodiment provides a human body semantic feature extraction method, as shown in FIG. 2 and FIG. 3 , including: collecting three-dimensional human body feature data to obtain a three-dimensional human body model; Perform preprocessing, adjust the orientation and center coordinates of the 3D human model and make the 3D human model at the coordinate origin of the world coordinate system; perform shape segmentation on the 3D human model and extract the skeleton features of the 3D human model, and a template selection algorithm based on skeletal similarity , according to the skeletal features of the 3D human body model, select the template model from the constructed 3D human body semantic feature database for registration with the 3D human body model, and obtain the parameterized human body model; use the NURBS curve to fit the feature sampling points on the parameterized human body model , calculate the length of the characteristic curve after fitting, and obtain the human body semantic features of the three-dimensional human body model.

The template selection algorithm based on skeleton similarity calculates the pose similarity between the 3D human body model and the template model in the 3D human body semantic feature database, and selects the model with the highest degree of pose similarity as the template model; the details are as follows:

为三维人体模型D上的p骨节点的坐标，

为三维人体模型D上的q骨节点的坐标，

为模板模型

上的p骨节点的坐标，

为模板模型

上的q骨节点的坐标，

为向量夹角符号，

为权重因素，m＝1，2，3，4，L₁、L₂、L₃、L₄为四个权重级别，人体主干中的骨节点权重级别为L₁，上臂、大腿、头部的权重级别为L₂，下臂、小腿的权重级别为L₃，手部、脚部的权重级别为L₄，且

Among them, SE is the bone similarity parameter,

is the coordinate of the p-bone node on the three-dimensional human model D,

is the coordinate of the q-bone node on the three-dimensional human model D,

template model

The coordinates of the p-bones node on,

template model

the coordinates of the q-bone node on,

is the vector angle symbol,

is the weight factor, m=1, 2, 3, 4, L ₁ , L ₂ , L ₃ , and L ₄ are four weight levels, the bone node weight level in the main body of the human body is L ₁ , the upper arm, thigh, head The weight level is L ₂ , the weight level of the lower arm and the calf is L ₃ , the weight level of the hands and feet is L ₄ , and

The template selection algorithm based on skeletal similarity selects the template model and the 3D human body model from the constructed 3D human body semantic feature database according to the skeleton features of the 3D human body model for registration, and obtains a parameterized human body model. The registration adopts rigid body registration and Non-rigid registration algorithm;

Rigid body registration includes:

的有向包围盒

计算从三维人体模型D的有向包围盒D_B到模板模型

的有向包围盒

的放射变换T；对三维人体模型D施加放射变换得到第一中间模型D′；Based on the initial alignment of the directed bounding box, let the two models overlap as much as possible in the global coordinate system: construct the directed bounding box D _B of the 3D human model D and the template model

directed bounding box of

Compute the directed bounding box D _B from the 3D human model D to the template model

directed bounding box of

The radiation transformation T is applied to the three-dimensional human body model D to obtain the first intermediate model D′;

D′=TD (13)

的最优刚体变换参数，进而获取第二中间模型D″：Based on the final alignment of the iterative closest point, the rotation matrix and translation matrix between the two models are calculated, and the spatial coordinates of the template model are adjusted according to the rotation matrix and the translation matrix: the iterative closest point algorithm is used to calculate the first intermediate model D' to the template model

The optimal rigid body transformation parameters of , and then obtain the second intermediate model D":

D″=RD′+t (14)

Among them, R is the linear transformation parameter, and t is the translation transformation parameter;

Non-rigid registration includes:

向第二中间模型

变形，得到初步配准模型

Coarse Registration Based on Laplacian Mesh Deformation Algorithm: Using Laplacian Mesh Deformation Algorithm to

to the second intermediate model

Deformed to get a preliminary registration model

和第二中间模型D″上建立数据误差函数E_d和光滑度误差函数E_s，并最小化其误差之和，获得参数化的人体模型

数据误差函数E_d和光滑度误差函数E_s通过以下公式获得：Fine registration based on data error and smoothness error: in a preliminary registration model

Establish data error function Ed and smoothness error function _Es on the second intermediate model _D ″, and minimize the sum of their errors to obtain a parameterized human body model

The data error function _Ed and smoothness error function _Es are obtained by the following formulas:

的顶点个数，v′_i为第i个顶点的坐标，距离函数dist²()为变形后网格与目标网格中最近相容点的距离，T_i为第i个顶点对应的3×3的变换矩阵；将法相量角度小于90°，欧式距离小于10cm的两个点，定义为相容点，其中距离最近的一个称为最近相容点；T_j为第j个顶点对应的3×3变换矩阵，

为模型

上的边，|| ||_F表示弗罗贝尼乌斯范数。where n represents the preliminary registration model

The number of vertices of , v′ _i is the coordinate of the ith vertex, the distance function dist ² () is the distance between the deformed mesh and the nearest compatible point in the target mesh, T _i is the 3× corresponding to the ith vertex The transformation matrix of 3; the two points whose normal phasor angle is less than 90° and whose Euclidean distance is less than 10cm are defined as compatible points, and the one with the closest distance is called the closest compatible point; T _j is the 3 corresponding to the jth vertex ×3 transformation matrix,

for the model

On the edge, || || _F denotes the Frobenius norm.

The result of the registration is a parameterized human body model; the parameterized human body model is marked with characteristic sampling points, and the NURBS curve is used to fit these characteristic sampling points, and the length of the fitted characteristic curve is calculated to obtain the human body semantic features of the three-dimensional human body model. In the process of human body semantic feature extraction, the 3D human body semantic feature database incorporates the shape segmentation results of the 3D human body model generated in the process of human semantic feature extraction into the 3D human body segmentation database, and the skeleton features of the 3D human body model into the 3D human skeleton database, including The 3D human body model of human semantic features is incorporated into the 3D human semantic feature database.

Embodiment three:

Based on the human body semantic feature extraction method provided in the second embodiment, this embodiment provides a human body semantic feature extraction system, including a processor and a storage device. The storage device stores multiple instructions for the processor to load and execute the second embodiment. the steps of the method.

The above are only the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the technical principles of the present invention, several improvements and modifications can be made. These improvements and modifications It should also be regarded as the protection scope of the present invention.

Claims (10)

1. a human body semantic feature extraction method, is characterized in that, comprises: Collect 3D human body feature data to obtain 3D human body model; Preprocess the 3D human body model, adjust the orientation and center coordinates of the 3D human body model, and place the 3D human body model at the coordinate origin of the world coordinate system; Perform shape segmentation on the 3D human body model and extract the skeleton features of the 3D human body model. The template selection algorithm based on the skeleton similarity selects the template model from the constructed 3D human body semantic feature database according to the skeleton features of the 3D human body model and matches the 3D human body model. Accurate to obtain a parameterized human body model; Use the NURBS curve to fit the feature sampling points on the parameterized human body model, calculate the length of the fitted feature curve, and obtain the human body semantic features of the 3D human body model. 2. The human body semantic feature extraction method according to claim 1, wherein the registration comprises rigid body registration and non-rigid body registration; The rigid body registration includes: Construct the directed bounding box D _B of the 3D human body model D and the template model

directed bounding box of

Compute the directed bounding box D _B from the 3D human model D to the template model

directed bounding box of

The affine transformation T of ; apply the affine transformation T to the three-dimensional human body model D to obtain the first intermediate model D′:

D′=TD (13) Using the iterative closest point algorithm to calculate the first intermediate model D' to the template model

The optimal rigid body transformation parameters R and t are obtained, and then the second intermediate model D″ is obtained: D″=RD′+t (14) Among them, R is the linear transformation parameter, and t is the translation transformation parameter; The non-rigid body registration includes: Using the Laplacian mesh deformation algorithm to transform the template model

to the second intermediate model

Deformed to get a preliminary registration model

in the preliminary registration model

Establish data error function Ed and smoothness error function _Es on the second intermediate model _D ″, and minimize the sum of their errors to obtain a parameterized human body model

3. The human body semantic feature extraction method according to claim 2, wherein the data error function E _d and the smoothness error function E _s are obtained by the following formula:

where n represents the preliminary registration model

The number of vertices of , v _i ′ is the coordinate of the ith vertex, the distance function dist ² () is the distance between the deformed mesh and the nearest compatible point in the target mesh, and T _i is the 3× corresponding to the ith vertex 3 Transformation matrix; two points whose normal phasor angle is less than 90° and Euclidean distance less than 10cm are defined as compatible points, and the one with the closest distance is called the closest compatible point; T _j is the 3× corresponding to the jth vertex 3 transformation matrices,

for the model

On the edge, |||| _F denotes the Frobenius norm. 4. human body semantic feature extraction method according to claim 1, is characterized in that, the template selection algorithm of described skeleton similarity, is specially:

Among them, SE is the bone similarity parameter,

is the coordinate of the p-bone node on the three-dimensional human model D,

is the coordinate of the q-bone node on the three-dimensional human model D,

template model

The coordinates of the p-bones node on,

template model

the coordinates of the q-bone node on,

is the vector angle symbol,

is the weight factor, m=1, 2, 3, 4, L ₁ , L ₂ , L ₃ , and L ₄ are four weight levels, the bone node weight level in the main body of the human body is L ₁ , the upper arm, thigh, head The weight level is L ₂ , the weight level of the lower arm and the calf is L ₃ , the weight level of the hands and feet is L ₄ , and

5. human body semantic feature extraction method according to claim 1, is characterized in that, the construction method of described three-dimensional human body semantic feature database is: Select the human body pose model in the open source 3D human body database, perform surface subdivision, set the number of facets and vertices of the human body pose model, and obtain the initial 3D human body database; Segment the human body pose model in the initial 3D human body database to form a 3D human body segmentation database; Extracting 3D human bones from the segmentation model in the 3D human body segmentation database to form a 3D human skeleton database; The human skeleton in the 3D human skeleton database is semi-automatically marked with semantic feature sampling points to form a 3D human body semantic feature database. 6. The human body semantic feature extraction method according to claim 5, wherein the three-dimensional human body semantic feature database incorporates the shape segmentation result of the three-dimensional human body model generated in the human body semantic feature extraction process into the three-dimensional human body segmentation database, The skeleton features of the 3D human body model are included in the 3D human skeleton database, and the 3D human body model containing the human body semantic features is included in the 3D human body semantic feature database. 7. human body semantic feature extraction method according to claim 5, is characterized in that, the described human body posture model in the initial three-dimensional human body database is segmented, specifically: Perform supervoxel clustering on vertex data, and the result is an over-segmentation of vertex data, including: 1) The model space is divided into grids, and the point cloud data forms point clusters, which are called voxels. A spatial index is established for the voxels, and the octree algorithm is used to organize these voxels; 2) Select multiple voxels as seed voxels; 3) Calculate the similarity between voxels, and recursively fuse adjacent voxels with seed voxels to form supervoxels; Judging the concave-convex relationship between adjacent supervoxels according to the preset judgment criteria; The random sampling consensus algorithm is used to fit the cutting plane on the concave edge, and the human body pose model is segmented. 8. human body semantic feature extraction method according to claim 7, is characterized in that, calculates the similarity between voxels, is specifically:

Among them, ω _c is the weight factor of color difference, ω _s is the weight factor of distance difference, ω _n is the weight factor of normal difference, D _c (i, j) is the CIELab color difference, D _s (i, j) is Coordinate distance difference, D _n (i, j) is the normal difference,

is the voxel diameter. 9. The human body semantic feature extraction method according to claim 7, wherein the judgment criterion consists of the connecting line of adjacent super-voxels and the normal direction on the centroid and the common adjacent super-voxels of adjacent super-voxels. Voxels are determined. 10. A human body semantic feature extraction system, characterized by comprising a processor and a storage device, wherein the storage device stores a plurality of instructions for the processor to load and execute any one of claims 1-9. steps of the method described.

Images