Aerodynamic configuration point cloud gridding method based on point cloud distribution perception
Through adaptive bilateral filtering and distribution perception methods, combined with heuristics and deep learning methods, aerodynamic profile point clouds are gridded, solving the shortcomings in modeling complex surfaces and flat areas in the existing technology, and achieving more accurate and automated point cloud gridding.
Patent Information
- Application Number
- CN202411661581.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-20
- Publication Date
- 2025-03-14
AI Technical Summary
The prior art is difficult to accurately model local complex surfaces and flat areas during the process of aerodynamic profile point cloud gridding, and deep learning-based methods are difficult to capture local details.
The method based on point cloud distribution perception is adopted to preprocess point cloud data through adaptive bilateral filtering, the geometric structure complexity of local areas is judged, and the heuristic or deep learning methods are used to triangulate according to different regions.
It realizes more accurate gridding of aerodynamic point clouds, and can automatically analyze and improve the gridding results of point clouds, suitable for different types of data.
Smart Images
Figure CN119625215A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of point cloud processing and three-dimensional modeling, and in particular to an aerodynamic shape point cloud meshing method based on point cloud distribution perception. Background Art
[0002] Aerodynamic shape point cloud meshing means: for a given aerodynamic shape, point cloud data is collected, and the surface geometry of the aerodynamic shape is modeled through point cloud preprocessing, point cloud meshing and other methods.
[0003] Aerodynamic shape refers to the shape of an object moving in a gas medium, which is suitable for moving in the gas medium in order to reduce the resistance of the object moving in the gas medium.
[0004] Point cloud preprocessing means: for a given point cloud collection data, by analyzing and adjusting the data, to evaluate the error information such as noise that may be contained in the data, and to eliminate it as much as possible, so as to obtain a point cloud distribution that is more in line with our expectations. Generally speaking, the scale of aerodynamic shape point clouds is very large, which may contain tens of thousands to millions of points, which brings huge computational pressure and challenges for us to efficiently and accurately evaluate its error information.
[0005] Point cloud meshing means using a series of meshes to approximate point clouds. For a given 3D point cloud, its distribution in 3D space is analyzed and its corresponding geometric shape is eventually modeled as a triangular mesh. Due to the high uncertainty of the spatial distribution and density of point clouds, it is often difficult to perform accurate geometric analysis on locally dense point clouds during the meshing process, and it is also easy to miss the geometric surfaces corresponding to locally sparse point clouds.
[0006] Disadvantages of existing technology:
[0007] Most existing point cloud preprocessing algorithms use weighted averaging of local point clouds to reduce noise. However, due to the presence of geometrically complex local surfaces in aerodynamic shapes, the point clouds corresponding to these surfaces will be treated as noise, and their geometric structures will be destroyed by the noise reduction algorithms. This results in a lot of time and manpower required for manual adjustments.
[0008] Point cloud meshing algorithms are mainly divided into two categories: heuristic methods and deep learning-based methods. The triangular mesh obtained by the heuristic method can fit the given point cloud more accurately, but often there will be problems such as modeling omissions in locally sparse point clouds; and the deep learning-based method is trained on a large number of reasonable geometric shapes, so compared with the heuristic method, it can more completely model the complete geometric shape corresponding to the given point cloud. However, due to the limited resolution of the network's input point cloud, such methods often find it difficult to accurately model the local details corresponding to the point cloud, and can only obtain the approximate geometric shape corresponding to the point cloud. Summary of the invention
[0009] In order to overcome the defects in the above-mentioned prior art, the present invention provides an aerodynamic shape point cloud meshing method based on point cloud distribution perception, which can use different triangulated meshing methods for flat areas and complex areas in the point cloud, so as to have better performance on more different types of data.
[0010] To achieve the above object, the present invention adopts the following technical solutions, including:
[0011] A method for meshing aerodynamic shape point cloud based on point cloud distribution perception includes the following steps:
[0012] S1, input point cloud data;
[0013] S2, preprocessing the point cloud data;
[0014] S3, determining whether any local area in the point cloud is a flat area, if so, proceeding to step S4; if not, proceeding to step S5;
[0015] S4, using a heuristic point cloud meshing method to triangulate the point cloud of the local area, and output a triangulated mesh of the local area;
[0016] S5, using a deep learning point cloud meshing method to triangulate the point cloud of the local area, and outputting a triangulated mesh of the local area.
[0017] Preferably, in step S2, the point cloud data is preprocessed by adaptive bilateral filtering, as shown below:
[0018] S21, for the point to be subjected to adaptive bilateral filtering, i.e., the target point, find the nearest k neighboring points of the target point in the point cloud, and calculate the normal vector of the target point using the target point and the k neighboring points; the specific calculation method is as follows:
[0019] First calculate the covariance matrix C of the target point with respect to the neighborhood points:
[0020]
[0021] in, is the direction vector from the target point to the jth neighboring point, j = 1, 2, ..., k; is the centroid of k neighborhood points;
[0022] Then perform eigenvalue decomposition on the covariance matrix C to obtain three eigenvalues λ 1 ,λ 2 ,λ 3 , and get the three eigenvalues λ 1 ,λ 2 ,λ 3 The corresponding eigenvectors v 1 ,v 2 ,v 3 ; Among them, λ 1 ≥λ 2 ≥λ 3 ;
[0023] Eigenvector v 3 It is the normal vector of the target point;
[0024] S22, calculate the curvature c of the target point:
[0025]
[0026] S23, based on the distance d between the target point and each neighboring point j , obtain the component of the direction vector from the target point to each neighboring point in the normal direction of the target point, and accumulate the component in the normal direction of the target point, and use the accumulated value as the target displacement size d of the target point along the normal direction target ; Then according to the target displacement size d target Get the update step length d of the target point along the normal direction update for:
[0027] d update =c*d target
[0028] The target point is updated according to the step size d update Update the spatial position along the normal direction to complete the adaptive bilateral filtering of the target point;
[0029] S24, performing adaptive bilateral filtering on each point in the point cloud in the manner of steps S21-S23 to complete preprocessing of the point cloud data.
[0030] Preferably, in step S3, whether any local area in the point cloud is a flat area is determined by:
[0031] The curvature of each point in the local area is calculated to determine whether the curvature variation range exceeds a set range. If it does not exceed the set range, the local area is a flat area, otherwise the local area is not a flat area.
[0032] Preferably, in step S4, the heuristic point cloud meshing method based on Poisson's equation is specifically as follows:
[0033] S41, after preprocessing in step S2, the normal and position of each point in the point cloud are obtained, and the vector field V is constructed using the point cloud, and the position and normal of the vector field V are recorded at each point;
[0034] S42, the forward value of each point in the vector field V along the normal direction is recorded as 0, and the backward value along the normal direction is recorded as 1, so as to obtain a binary scalar field B, which satisfies that the values inside the field are all 0 and the values outside the field are all 1, that is, the values inside the object corresponding to the point cloud are all 0, and the values outside the object are all 1;
[0035] S43, estimate the scalar field B and minimize the energy term:
[0036]
[0037] Where E(B) represents the energy term of the scalar field B; represents the vector field to be solved; f gs represents Gaussian function; p represents integral variable; Represents the gradient field, using the Gaussian function f gs As a smoothing filter, the filtered scalar field B*f gs With continuous gradient field
[0038]
[0039] in, represents the gradient at the spatial point p; f gs (pq) represents the relative position coordinates of spatial point q relative to spatial point p, and the corresponding Gaussian function value. The farther the spatial point q is from the spatial point q, the smaller the Gaussian function value is, and vice versa; Indicates the surface of an object; Represents the normal vector at point q on the surface of the object; It means to find the first-order gradient;
[0040] Since the energy term E(B) involves integration, we convert E(B) into Poisson's equation by finding the divergence of E(B), that is, the second-order gradient:
[0041]
[0042] Where ΔB represents the second-order gradient of the energy term E(B); Δ represents the second-order gradient;
[0043] S44, using the Marching Cubes algorithm, extracts the boundary of 0 and 1 from the scalar field B as the surface of the object corresponding to the point cloud, and represents the surface as a triangular mesh data structure.
[0044] Preferably, in step S5, the deep learning point cloud meshing method based on local geometric perception is specifically as follows:
[0045] S51, for any position x in space, generate observation areas with different orientations, and take the L points closest to the position x in each observation area as observation points to form a point set of the observation area;
[0046] S52, using the point cloud encoder PointNet Encoder to extract features from the point set of the observation area as the point cloud distribution feature f of the observation area i ; Average the positions of each observation point in the observation area to obtain the average position coordinate c i ; Average the distances between each observation point and position x in the observation area to obtain the average distance value d i ; Wherein, the subscript i represents the i-th observation area;
[0047] S53, the point cloud distribution feature f of each observation area i , average position coordinate c i and the average distance value d i Merge as the observation features of the observation area, and obtain the observation features in different observation directions;
[0048] S54, the observation features in different observation directions are decoded by the self-attention module and the space occupancy decoder, and the decoding outputs the space occupancy probability, which is used to describe the relative position relationship between the current position x and the surface of the object corresponding to the point cloud;
[0049] S55, divide the space into M×M×M grid regions, and obtain the space occupancy probability of the center position of each grid region according to the method of steps S51-S54, so as to obtain a space occupancy probability matrix O with a shape of M×M×M p ;
[0050] S56, space occupancy probability matrix O p Perform binarization processing, transform the element values exceeding the set threshold into 1, and transform the element values not exceeding the set threshold into 0, so as to obtain a binary space occupancy matrix O;
[0051] S57, extracting the triangular mesh corresponding to the surface of the object by using a marching cubes algorithm, thereby obtaining a final result.
[0052] A readable storage medium stores a computer program, which, when executed, implements the above-mentioned aerodynamic shape point cloud meshing method based on point cloud distribution perception.
[0053] An electronic device comprises a processor, a memory and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the aerodynamic shape point cloud meshing method based on point cloud distribution perception is implemented.
[0054] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the above-mentioned aerodynamic shape point cloud meshing method based on point cloud distribution perception.
[0055] The advantages of the present invention are:
[0056] (1) The present invention aims to perform point cloud preprocessing operations before meshing the aerodynamic shape point cloud. Specifically, it first accurately determines whether the geometric structure corresponding to any local area in the point cloud is complex, thereby separating flat areas from geometrically complex areas, and taking different degrees of noise reduction operations on different areas.
[0057] (2) In order to effectively determine the geometric structure complexity of a local point cloud, the present invention proposes an adaptive bilateral filtering algorithm to quickly estimate the complexity of the underlying geometric structure corresponding to a point cloud with arbitrary distribution.
[0058] (3) The present invention proposes an adaptive point cloud meshing algorithm that combines the advantages of heuristic methods and deep learning methods, which is used to automatically perform more accurate aerodynamic shape surface modeling based on the geometric complexity of the point cloud.
[0059] (4) The present invention can automatically analyze a given point cloud and improve the results of a subsequent point cloud meshing algorithm through a preprocessing operation based on adaptive bilateral filtering.
[0060] (5) The present invention can use different triangulation methods for flat areas and complex areas in the point cloud, thereby achieving better performance on more different types of data. BRIEF DESCRIPTION OF THE DRAWINGS
[0061] Figure 1 The present invention is a flow chart of a method for meshing aerodynamic shape point cloud based on point cloud distribution perception.
[0062] Figure 2A schematic diagram of the neighborhood search.
[0063] Figure 3 Schematic diagram of the field-based object surface.
[0064] Figure 4 Schematic diagram of the observation information obtained by observing point clouds at different locations.
[0065] Figure 5 Network architecture diagram for the space occupancy probability prediction network. DETAILED DESCRIPTION
[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0067] Depend on Figure 1 As shown, a method for meshing aerodynamic shape point cloud based on point cloud distribution perception includes the following steps:
[0068] S1, input point cloud data.
[0069] S2, preprocessing the point cloud data by adaptive bilateral filtering.
[0070] S3, determine whether the geometric structure corresponding to any local area in the point cloud is complex. If it is not complex, it is a flat area, then go to step S4; if it is complex, it is a complex area, then go to step S5.
[0071] S4, using a heuristic point cloud meshing method to triangulate the point cloud of the local area, and outputting a triangulated mesh of the local area.
[0072] S5, using a deep learning point cloud meshing method to triangulate the point cloud of the local area, and outputting a triangulated mesh of the local area.
[0073] In step S2, the adaptive bilateral filtering method of the point cloud is different from the traditional bilateral filtering method. Before the point cloud is gridded, the present invention first quickly analyzes the spatial distribution of the point cloud, and then performs adaptive bilateral filtering based on the analysis results.
[0074] The adaptive bilateral filtering method of point cloud is as follows:
[0075] S21, since the sampling point cloud of aerodynamic shape generally does not contain point-by-point normal information, the point cloud is firstly oriented with a globally consistent normal.
[0076] Specifically, for a given point cloud, first find k neighboring points in the point cloud for each point in the point cloud. Figure 2 As shown in (a) in FIG, for the target point (red point), in this embodiment, when k=6, the six nearest points surrounded by the curve are finally selected as the neighborhood points of the target point through distance judgment, and the normal vector of the target point is calculated from the above-mentioned total of seven points. Figure 2 In the figure, the red point is the point currently being processed, i.e., the target point, and the black point is the point near the target point.
[0077] Calculate the covariance matrix C of the target point with respect to the neighborhood points:
[0078]
[0079] in, is the direction vector from the target point to the jth neighboring point, j = 1, 2, ..., k; is the centroid of the k neighborhood points, that is, the average coordinates of the k neighborhood points in space.
[0080] By calculating the above formula, we can get a 3×3 covariance matrix C. By performing eigenvalue decomposition on the covariance matrix C, we can get its corresponding three eigenvalues λ 1 ≥λ 2 ≥λ 3 , and 3 eigenvalues λ 1 ,λ 2 ,λ 3 The three corresponding eigenvectors v 1 ,v 2 ,v 3 .
[0081] Eigenvalue λ 1 ,λ 2 ,λ 3 The size of reflects the local geometric characteristics of the point cloud: 1 The largest value means that the data changes the most in this direction; 3 The smallest, which means that the data changes the least in this direction. Therefore, the eigenvector v corresponding to the smallest feature 3 It corresponds to the normal vector of the target point.
[0082] S22, through experiments, it is found that if only the estimated normal information is relied upon, it is difficult to distinguish the geometric complexity corresponding to the surfaces of different regions. In order to effectively represent the geometric complexity, the present invention further estimates the curvature c of each point in the point cloud:
[0083]
[0084] The curvature c has a clear physical meaning: the smaller the curvature c, the flatter and closer to a plane the local surface of the point cloud near the point is; otherwise, the more distorted it is and the more complex the geometry is.
[0085] S23, after steps S21 and S22, the point-by-point normal and curvature that can describe the local geometric information and geometric complexity are obtained, as shown in Figure 2 As shown in (b), the present invention is based on the distance d between the target point and each neighboring point j , accumulate the component of the direction vector from the target point to each neighboring point in the normal direction of the target point, and use the accumulated value as the target displacement size d of the target point along the normal direction target ; Then according to the target displacement size d target The update step length of the target point along the normal direction is:
[0086] d update =c*d target
[0087] The target point is updated according to the step size d update The spatial position is updated along the normal direction to complete the adaptive bilateral filtering of the target point.
[0088] S24, after the spatial positions of all points in the point cloud are updated, an adaptive bilateral filtering of the point cloud is completed.
[0089] In step S3, it is determined whether the geometric structure corresponding to any local area in the point cloud is complex. The determination method is:
[0090] The curvature of each point in the local area is calculated to determine whether the curvature variation range exceeds a set range. If it does not exceed the set range, the local area is a flat area, otherwise the local area is not a flat area.
[0091] The local area is the set area size.
[0092] In step S4, the heuristic point cloud meshing method based on Poisson's equation is as follows:
[0093] S41, such as Figure 3 As shown in (a) in FIG. 1 , the updated point cloud and its normal vector are obtained through preprocessing in step S2; Figure 3 As shown in (b), the point cloud is regarded as a sparse vector field V, and the position and normal of the vector field V are recorded at each point.
[0094] S42, the present invention hopes to model and estimate the complete information of the entire field through point cloud information, and further construct a scalar field F, which satisfies that the values in the scalar field F are strictly reduced along the direction in the vector field V, and the values at the position of each point in the corresponding point cloud in the scalar field F are all 0. In this way, the surface corresponding to the point cloud can be obtained by extracting the 0-isosurface in the scalar field F. However, since the representation of the scalar field F is not unique, it is difficult to solve it. To this end, the present invention simplifies the scalar field F, requiring it to contain only two values 0 and 1, and at each point in the point cloud, it is 0 forward along the normal and 1 backward, thereby obtaining a binary scalar field B, such as Figure 3 As shown in (c) in the figure, the scalar field B satisfies that the inside of the object corresponding to the point cloud is 0 and the outside of the object is 1. In this way, the surface corresponding to the point cloud can be obtained by quickly finding the boundary between 0 and 1, as shown in Figure 3 As shown in (d) in .
[0095] Figure 3 In the figure, the black dots are spatial points in the point cloud, the blue arrows are the normal vectors of each point, the yellow arrows are the spatial orientation distribution of the vector field, and the blue curve is the object surface corresponding to the point cloud.
[0096] S43, the present invention hopes that the gradient of the scalar field B satisfies the zero vector everywhere except the surface position, and is approximately the surface normal of the corresponding position near the surface, so that the scalar field B can be estimated by constraining the gradient of the scalar field B at each point in the point cloud to be the estimated normal of the point, that is, minimizing the following energy term E(B):
[0097]
[0098] Where E(B) represents the energy term of the scalar field B; represents the vector field to be solved; f gs represents Gaussian function; p represents integral variable; represents the gradient field; since the scalar field B contains only two values 0 and 1, the modulus of its gradient at the boundary between 0 and 1 tends to infinity, so the present invention uses the Gaussian function f gs As a smoothing filter, the filtered scalar field B*f gs With continuous gradient field
[0099]
[0100] in, represents the gradient at the spatial point p; f gs(pq) represents the relative position coordinates of spatial point q relative to spatial point p, and the corresponding Gaussian function value. The farther the spatial point q is from the spatial point q, the smaller the Gaussian function value is, and vice versa; Indicates the surface of an object; Represents the normal vector at point q on the surface of the object; It means to find the first-order gradient;
[0101] Since the energy term E(B) involves an integral, which is usually difficult to calculate, the present invention converts E(B) into a Poisson equation by finding the divergence of E(B) (i.e., finding the second-order gradient):
[0102]
[0103] Where ΔB represents the second-order gradient of the energy term E(B); Δ represents the second-order gradient;
[0104] This equation can be effectively calculated by discretizing the first-order and second-order gradients to solve the scalar field B.
[0105] S44, through the widely used Marching Cubes algorithm, extracts the boundaries of 0 and 1 from the scalar field B as the surface of the object, and finally the surface is represented as a triangular mesh data structure.
[0106] Although the point cloud triangulation method based on Poisson reconstruction can process point clouds in most cases and obtain a reasonable and as accurate as possible triangulated mesh representation of the object surface, when the point cloud data contains a large amount of obvious noise or the distribution of the point cloud is uneven, the results obtained by this heuristic method will often reconstruct the noise as well, or lose the object surface corresponding to the sparse position of the point cloud, resulting in serious degradation or even errors in the reconstruction results.
[0107] Therefore, the present invention also designs a point cloud meshing method based on deep learning, which makes full use of the ability of deep learning to align a large amount of prior information on the surface distribution of objects with the point cloud distribution. Taking into account that the estimated point cloud normal may be inaccurate in some extreme cases, only the given point cloud is used as the only input of the neural network. Specifically, the design of the neural network is mainly based on the observation that for a given point cloud, when it is observed at different spatial positions and along different directions, the spatial distribution of the points in the observed point cloud has obvious differences, so this difference can help us judge the relative position relationship between the spatial position and the object. For example, Figure 4 As shown, Figure 4 In the figure, the black dots represent the point cloud, the two red triangles represent different observation positions, and the blue and green sectors represent randomly selected observation directions. Figure 4By observing the point cloud along different directions shown by the blue sector at the left red triangle in the figure, it can be found that the position is surrounded by the point cloud, so it is likely that the position is near the surface of the object; while for the red triangle, the observation results along the green sector indicate that the points in the point cloud are far away from the position, so the position is likely to be inside or outside the object.
[0108] In step S5, the deep learning point cloud meshing method based on local geometric perception is specifically as follows:
[0109] S51, taking two-dimensional space as an example, Figure 5 As shown, for any position x in the space, firstly, an observation area with different orientations is generated for it (corresponding to a fan-shaped area in two-dimensional space and a cone-shaped area in three-dimensional space), and the relative position coordinates of the L points closest to the position x in each observation area are recorded. The L points closest to the position x in each observation area are used as observation points to constitute a point set of the observation area. In this embodiment, L=12. If it is less than L, it is filled with the position {0,0,0}.
[0110] S52, using a point cloud encoder (PointNetEncoder) to extract features from the point set of the observation area as the point cloud distribution feature f of the observation area i ; Average the positions of each observation point in the observation area to obtain the average position coordinate c i ; Average the distances between each observation point and position x in the observation area to obtain the average distance value d i ; where the subscript i represents the i-th observation area.
[0111] S53, the point cloud distribution feature f of each observation area i , average position coordinate c i and the average distance value d i Merge as the observation features of the observation area, and obtain the observation features in different observation directions;
[0112] S54, finally, the spatial occupancy probability is decoded through the self-attention module (SelfAttention) and the spatial occupancy decoder (OCC Decoder) based on the multi-layer perceptron. The spatial occupancy probability is described as a floating point number between 0 and 1, which is used to describe the relative position relationship between the current position x and the surface corresponding to the given point cloud.
[0113] S55, divide the space into M1×M2×M3 grid regions, in this embodiment, M1=M2=M3=64, and obtain the space occupancy probability of the center position of each grid region according to the method of steps S51-S54, thereby obtaining a space occupancy probability matrix O with a shape of M1×M2×M3p .
[0114] S56, space occupancy probability matrix O p A binarization process is performed to transform the element values exceeding the set threshold into 1, and the element values not exceeding the set threshold into 0, thereby obtaining a binary space occupancy matrix O.
[0115] S57, extracting the triangular mesh corresponding to the surface of the object by using a marching cubes algorithm, thereby obtaining a final result.
[0116] The moving cube algorithm is: place a small cube in the space occupancy probability matrix, each time only focus on the occupancy probability distribution in the small cube, and get the corresponding triangular mesh patch. Specifically, when determining the corresponding triangular mesh patch, the algorithm will use the values in the space occupancy matrix to find the boundary between 0 and 1, which is the triangular mesh corresponding to the surface of the object; then move the cube to obtain the triangular mesh patches in all areas, thus forming the final complete triangular mesh. This method can ensure that the triangular patches obtained in each small cube have a connection relationship with the triangular patches in the position area of any adjacent cube, and can form a complete and continuous triangular mesh.
[0117] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for meshing aerodynamic shape point cloud based on point cloud distribution perception, characterized in that: The following steps are involved: S1, input point cloud data; S2, preprocessing the point cloud data; S3, determining whether any local area in the point cloud is a flat area, if so, proceeding to step S4; if not, proceeding to step S5; S4, using a heuristic point cloud meshing method to triangulate the point cloud of the local area, and output a triangulated mesh of the local area; S5, using a deep learning point cloud meshing method to triangulate the point cloud of the local area, and outputting a triangulated mesh of the local area.
2. The aerodynamic shape point cloud meshing method based on point cloud distribution perception according to claim 1, characterized in that: In step S2, the point cloud data is preprocessed by adaptive bilateral filtering, as shown below: S21, for the point to be subjected to adaptive bilateral filtering, i.e., the target point, find the nearest k neighboring points of the target point in the point cloud, and calculate the normal vector of the target point using the target point and the k neighboring points; the specific calculation method is as follows: First calculate the covariance matrix C of the target point with respect to the neighborhood points: in, is the direction vector from the target point to the jth neighboring point, j = 1, 2, ..., k; is the centroid of k neighborhood points; Then perform eigenvalue decomposition on the covariance matrix C to obtain three eigenvalues λ1, λ2, λ3, and obtain the eigenvectors v1, v2, v3 corresponding to the three eigenvalues λ1, λ2, λ3 respectively; where λ1≥λ2≥λ3; The feature vector v3 is the normal vector of the target point; S22, calculate the curvature c of the target point: S23, based on the distance d between the target point and each neighboring point j , obtain the component of the direction vector from the target point to each neighboring point in the normal direction of the target point, and accumulate the component in the normal direction of the target point, and use the accumulated value as the target displacement size d of the target point along the normal direction target ; Then according to the target displacement size d target Get the update step length d of the target point along the normal direction update for: d update =c*d target The target point is updated according to the step size d update Update the spatial position along the normal direction to complete the adaptive bilateral filtering of the target point; S24, performing adaptive bilateral filtering on each point in the point cloud in the manner of steps S21-S23 to complete preprocessing of the point cloud data.
3. The aerodynamic shape point cloud meshing method based on point cloud distribution perception according to claim 1 or 2, characterized in that: In step S3, it is determined whether any local area in the point cloud is a flat area. The determination method is: The curvature of each point in the local area is calculated to determine whether the curvature variation range exceeds a set range. If it does not exceed the set range, the local area is a flat area, otherwise the local area is not a flat area.
4. The aerodynamic shape point cloud meshing method based on point cloud distribution perception according to claim 1 or 2, characterized in that: In step S4, the heuristic point cloud meshing method based on Poisson's equation is as follows: S41, after preprocessing in step S2, the normal and position of each point in the point cloud are obtained, and the vector field V is constructed using the point cloud, and the position and normal of the vector field V are recorded at each point; S42, the forward value of each point in the vector field V along the normal direction is recorded as 0, and the backward value along the normal direction is recorded as 1, so as to obtain a binary scalar field B, which satisfies that the values inside the field are all 0 and the values outside the field are all 1, that is, the values inside the object corresponding to the point cloud are all 0, and the values outside the object are all 1; S43, estimate the scalar field B and minimize the energy term: Where E(B) represents the energy term of the scalar field B; represents the vector field to be solved; f gs represents Gaussian function; p represents the integral variable; Represents the gradient field, using the Gaussian function f gs As a smoothing filter, the filtered scalar field B*f gs With continuous gradient field in, represents the gradient at the spatial point p; f gs (pq) represents the relative position coordinates of spatial point q relative to spatial point p, and the corresponding Gaussian function value. The farther the spatial point q is from the spatial point q, the smaller the Gaussian function value is, and vice versa; Indicates the surface of an object; Represents the normal vector at point q on the surface of the object; It means to find the first-order gradient; Since the energy term E(B) involves integration, we convert E(B) into Poisson's equation by finding the divergence of E(B), that is, the second-order gradient: Where ΔB represents the second-order gradient of the energy term E(B); Δ represents the second-order gradient; S44, using the Marching Cubes algorithm, extracts the boundary of 0 and 1 from the scalar field B as the surface of the object corresponding to the point cloud, and represents the surface as a triangular mesh data structure.
5. The aerodynamic shape point cloud meshing method based on point cloud distribution perception according to claim 1 or 2, characterized in that: In step S5, the deep learning point cloud meshing method based on local geometric perception is specifically as follows: S51, for any position x in space, generate observation areas with different orientations, and take the L points closest to the position x in each observation area as observation points to form a point set of the observation area; S52, using a point cloud encoder to extract features from the point set of the observation area as the point cloud distribution feature f of the observation area i ; Average the positions of each observation point in the observation area to obtain the average position coordinate c i ; Average the distances between each observation point and position x in the observation area to obtain the average distance value d i ; Wherein, the subscript i represents the i-th observation area; S53, the point cloud distribution feature f of each observation area i , average position coordinate c i and the average distance d i Merge as the observation features of the observation area, and obtain the observation features in different observation directions; S54, the observation features in different observation directions are decoded by the self-attention module and the space occupancy decoder, and the decoding outputs the space occupancy probability, which is used to describe the relative position relationship between the current position x and the surface of the object corresponding to the point cloud; S55, divide the space into M1×M2×M3 grid regions, and obtain the space occupancy probability of the center position of each grid region according to the method of steps S51-S54, thereby obtaining a space occupancy probability matrix O with a shape of M1×M2×M3 p ; S56, space occupancy probability matrix O p Perform binarization processing, transform the element values exceeding the set threshold into 1, and transform the element values not exceeding the set threshold into 0, so as to obtain a binary space occupancy matrix O; S57, extracting the triangular mesh corresponding to the surface of the object by using a marching cubes algorithm, thereby obtaining a final result.
6. A readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed, the method for meshing aerodynamic shape point cloud based on point cloud distribution perception as described in any one of claims 1 to 5 is implemented.
7. An electronic device, characterized in that: It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aerodynamic shape point cloud meshing method based on point cloud distribution perception as described in any one of claims 1 to 5.
8. A computer program product, characterized in that It includes a computer program / instruction, which, when executed by a processor, implements a method for aerodynamic shape point cloud meshing based on point cloud distribution perception as described in any one of claims 1 to 5.