A 3D Reconstruction Method for Virtual Teaching Modeling of Ships

Abstract

This invention belongs to the technical field of virtual teaching modeling methods, and particularly relates to a 3D reconstruction method for ship virtual teaching modeling. It includes the following steps: establishing a point cloud data coordinate system, optimizing the feature points of the ship's spatial contour, optimizing the feature points and non-feature points, updating the point cloud data based on the aforementioned steps, and generating a ship scene model based on the Possion surface reconstruction method. This 3D reconstruction method for ship virtual teaching modeling is mainly used in the 3D reconstruction modeling of large equipment such as ships based on point cloud data. It improves problems such as point position offset and line direction errors in the point cloud generation process of various equipment within the ship, while optimizing the distribution of point cloud data during reconstruction. By reducing the density of low-feature point cloud data, the total amount of point cloud data is reduced, decreasing the load on system resources, while increasing the proportion of high-feature point clouds, optimizing the performance of point cloud data, and improving the display performance of the reconstructed image.

Description

Technical Field

[0001] This invention belongs to the technical field of virtual teaching modeling methods, and particularly relates to a three-dimensional reconstruction method for virtual teaching modeling of ships. Background Technology

[0002] As an emerging teaching method, scene reproduction and simulation technologies based on virtual technologies and equipment such as VR and AR can solve the problem of inconvenient physical teaching or simulation of large equipment such as ships. Through 3D modeling and scene reproduction, it can achieve better teaching results than traditional teaching methods in smaller teaching scenarios and at lower costs. To achieve a realistic reproduction of real physical equipment, structures, and environments, it is necessary to convert real physical information into visible data in the virtual scene through mathematical modeling, 3D reconstruction, and other methods. Among them, 3D reconstruction uses technologies such as laser scanning to obtain point cloud information of structural surfaces and reconstructs surfaces, lines, and other structures through a series of methods, effectively improving the efficiency of reconstruction and display effects. However, in the 3D reconstruction of equipment such as ships, due to the large number of devices in the relatively enclosed space of ships, more serious problems such as point cloud offset and structural line deformation occur compared to other situations. Summary of the Invention

[0003] The purpose of this invention is to provide a 3D reconstruction method for ship virtual teaching modeling, based on the current situation and practical needs, to improve the quality of 3D reconstruction of ship teaching scenarios and suppress point offset during the reconstruction process. This method enhances the matching degree between the reconstructed point cloud data and the actual physical structure.

[0004] To achieve the above objectives, the present invention adopts the following technical solution.

[0005] A 3D reconstruction method for virtual teaching modeling of ships includes the following steps:

[0006] Step 1: Establish a point cloud data coordinate system: Obtain a point cloud dataset of the ship scene based on point cloud scanning {p k}, where p k This represents the point cloud data acquired in the k-th scan. For point cloud data p... k The i-th point p inside k,i Its coordinates in the scanning station coordinate system L are defined as L. k,i The coordinates in the Earth coordinate system W are W k,i ;

[0007] Step 2: Optimization of ship spatial contour feature points

[0008] Obtain point cloud data of the ship scene, and establish surface point cloud datasets F:F based on continuous smooth surfaces within the point cloud data. i(x,y,z)=0; for the i-th smooth surface F in the surface point cloud dataset i :F i P, the m-th (m≤M) continuously differentiable point on (x,y,z)=0 m :f i (x m ,y m ,z m Given that ) = 0, solve for point P. m K neighborhood point set N pm :N pm ={p j (x j ,y j ,z j )},j=1,2....J;

[0009] Calculate point P m The K-neighborhood point set {p k The center coordinates {x0, y0, z0} of k = 1, 2, ..., K are:

[0010]

[0011] The covariance matrix C of the neighboring points in the K-neighborhood set can be further expressed as:

[0012]

[0013] The covariance matrix C is a positive semi-definite symmetric matrix. Find its minimum eigenvalue λ. min and the corresponding feature vector v min On the smooth curved surface F i Within, the smallest eigenvalue λ min and its corresponding feature vector v min That is, a smooth curved surface F i At point P m The normal vector; based on the aforementioned steps, determine point P respectively. m The normal vectors of each point in the K neighborhood point set are used to specify the threshold angle α between the normal vectors according to the simplification ratio. min ; Taking the normal vector corresponding to the center point {x0, y0, z0} of the K-neighborhood point set as the initial normal vector n0, and for any point p in the K-neighborhood point set... k Its nearest neighbor p k-1 The angle α between the normal vectors k , retain α k ≥a min point p k Create feature points representing the spatial outline of the ship and delete other points;

[0014] Step 3: Optimization of Feature Points and Non-Feature Points

[0015] Based on the aforementioned covariance matrix C, establish a point P containing point distance weight coefficients. m K-domain covariance matrix C′ m : Where r m Let P be the point m Find the radius of the K-neighborhood; solve for the eigenvalues ​​λ′ of the covariance matrix C′. m1 ,λ′ m2 ,λ′ m3 ,λ′ m1 ≤λ′ m2 ≤λ′ m3 Then the rate of change of the normal vector corresponding to each point in the K-neighborhood on the surface is expressed as the eigenvalue λ′. m1 ,λ′ m2 ,λ′ m3 To ensure that the surface variations of the three-dimensional ellipsoidal surface composed of basic parameters are consistent, a matching threshold is constructed. Then for point P m If ε′ m If ≥ε′0, then point P m If it is a feature point, then it is a non-feature point;

[0016] Non-feature point cloud data constitutes the smooth portion of the 3D surface. By reducing the point cloud density of the smooth portion, the total amount of point cloud data in the entire 3D model can be effectively reduced without affecting key features such as the 3D surface contour, thus improving data processing performance. The specific processing steps are as follows:

[0017] Based on the aforementioned steps, the non-feature point cloud dataset {(x′) is determined. i ,y′ i ,z′ i Extract the maximum value x′ from the three coordinates. i,max ,y′ i,max ,z′ i,max and the minimum value of the three coordinates x′ i,min ,y′ i,min ,z′ i,min Construct a 3D bounding box ΔL, with side lengths of ΔL as follows: Where Δ i >0 (i = x, y, z) means that the side length margin is sufficient to ensure that all points are included;

[0018] Based on the voxel bounding method, the 3D bounding box ΔL is divided into N smaller bounding boxes Δl. n (n = 1, 2... N), the bounding box Δl of each cube n The side length is l i =n -1 L i (i = x, y, z); determine the bounding box Δl for each cube. ncenter point O n coordinates

[0019] Where (x) n,1 ,y n,1 ,z n,1 ) and (x n,2 ,y n,2 ,z n,2 ) are the bounding boxes of the cube Δl n The coordinates of the two endpoints of any diagonal line; calculate the bounding box Δl of each cube. n Points p in the middle n (x n ,y n ,z n ) and center point O n distance Extract and retain the data of points with the smallest distance, while deleting the bounding boxes Δl of each cube. n Other points in the process; repeat the above segmentation process until the preset requirements are met;

[0020] Step 4: Update the point cloud data based on the previous steps, and obtain valid point cloud data after necessary preprocessing; the preprocessing includes determining whether the average Euclidean distance from the point to the midpoint of the neighborhood satisfies a Gaussian distribution after obtaining the neighborhood through K-nearest neighbor search for possible outliers; if it does not satisfy the distribution, it is removed.

[0021] Step 5: Update the point cloud data and generate a ship scene model based on the Possion surface reconstruction method.

[0022] In a further improved or preferred implementation of the aforementioned 3D reconstruction method for virtual teaching modeling of ships, step 3 further includes point cloud direction adjustment and optimization, specifically: determining the coordinates of each point and calculating the normal direction based on the point cloud data determined in the aforementioned steps, and establishing a query sequence {p′0, p′1...p′} in ascending order of height values. n ...}; Transform the normal vectors of each point, and convert the normal vectors of each point... After translation, the normal vector can be expressed as: The normal vector of the first point p′0 in the query sequence is (0,0,1); the weight value of the specified point p′0 is 1, and the weight values ​​of the other points are updated sequentially according to the order of the query sequence. in This refers to the normal vector of the point corresponding to the minimum weight value among all updated points; during each update process, if... Then The direction is changed.

[0023] A further improvement or preferred implementation scheme for the aforementioned 3D reconstruction method for virtual teaching modeling of ships also includes a step for establishing a point cloud octree and its position index and gradient, specifically referring to: acquiring all point cloud data P all ={x i ,y i ,z i For each node i = 1, 2, 3, ..., n, a point cloud octree is constructed based on the aforementioned bounding box structure; each child node T at each level of the point cloud octree is... i For each group of bounding boxes of the same size; assuming the point cloud octree is N layers, then the side length of the minimum bounding box is l. min =L max ×2 -N ;

[0024] Assume point cloud data P all The index of the child node in the octree is q. m Then the description function used to represent the corresponding child node can be expressed as:

[0025]

[0026] Where m represents the point cloud data P all The level number of the child node in the octree; q mid Represents point cloud data P all The child node of the octree (Tree) corresponds to the midpoint of its bounding box; i = x, y, z; * indicates convolution operation; then the child node q m gradient of the indicator function t m Interpolate the weights for the child nodes.

[0027] Its beneficial effects are as follows:

[0028] The 3D reconstruction method for virtual teaching modeling of ships proposed in this application is mainly used in the 3D reconstruction modeling of large equipment such as ships based on point cloud data. It is used to improve problems such as point position offset and line direction error in the point cloud generation process of various equipment in the ship, and at the same time optimize the distribution of point cloud data in the reconstruction process. By reducing the density of low feature quantity point cloud data, the total amount of point cloud data is reduced, the load on system resources is reduced, and the proportion of high feature quantity point cloud is increased, thereby optimizing the performance of point cloud data and improving the display performance of the reconstructed image. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the sampling principle of voxel encirclement method. Detailed Implementation

[0030] The present invention will be described in detail below with reference to specific embodiments.

[0031] This invention is mainly used to provide a method for three-dimensional modeling and reconstruction of various types of equipment in a confined space in scenarios such as virtual teaching and training, so as to realize the virtual reproduction of real scenes, especially the three-dimensional reconstruction method for modeling the interior of large equipment such as various ships.

[0032] Step 1: Establish a point cloud data coordinate system: Obtain a point cloud dataset of the ship scene based on point cloud scanning {p k}, where p k This represents the point cloud data acquired in the k-th scan. For point cloud data p... k The i-th point p inside k,i Its coordinates in the scanning station coordinate system L are defined as L. k,i The coordinates in the Earth coordinate system W are W k,i ;

[0033] Point cloud data serves as a data bridge between the actual physical structure and the virtual 3D model. The coordinates and relative positions of each point within the point cloud data are necessary conditions for reconstructing the physical structure. Therefore, it is necessary to establish a corresponding coordinate system and generate the original point cloud data using specialized equipment or instruments such as laser scanners and depth cameras.

[0034] In visual information, the edge contour of an object's surface is a core feature attribute representing the structure and an important part of point cloud data processing. However, since point cloud data is discrete, the contour edges are not clearly represented, leading to display anomalies and other problems. Therefore, this application includes the following steps for optimization:

[0035] Step 2: Optimization of ship spatial contour feature points

[0036] Obtain point cloud data of the ship scene, and establish surface point cloud datasets F:F based on continuous smooth surfaces within the point cloud data. i (x,y,z)=0;

[0037] For the i-th smooth surface F in the surface point cloud dataset i :F i P, the m-th (m≤M) continuously differentiable point on (x,y,z)=0 m :f i (x m ,y m ,z m Given that ) = 0, solve for point P. m K neighborhood point set N pm :

[0038] N pm ={p j (x j ,yj ,z j )},j=1,2....J;

[0039] Calculate point P m The K-neighborhood point set {p k The center coordinates {x0, y0, z0} of k = 1, 2, ..., K are:

[0040]

[0041] The covariance matrix C of the neighboring points in the K-neighborhood set can be further expressed as:

[0042]

[0043] The covariance matrix C is a positive semi-definite symmetric matrix, and its minimum eigenvalue λ can be solved. min and the corresponding feature vector v min On the smooth curved surface F i Within, the smallest eigenvalue λ min and its corresponding feature vector v min That is, a smooth curved surface F i At point P m The normal vector; based on the aforementioned steps, determine point P respectively. m The normal vectors of each point in the K neighborhood point set are used to specify the threshold angle α between the normal vectors according to the simplification ratio. min ; Taking the normal vector corresponding to the center point {x0, y0, z0} of the K-neighborhood point set as the initial normal vector n0, and for any point p in the K-neighborhood point set... k Its nearest neighbor p k-1 The angle α between the normal vectors k , retain α k ≥α min point p k Create feature points representing the spatial outline of the ship and delete other points;

[0044] Step 3: Optimization of Feature Points and Non-Feature Points

[0045] Based on the aforementioned covariance matrix C, establish a point P containing point distance weight coefficients. m K-domain covariance matrix C′ m : Where r m Let P be the point m The radius of the K-domain;

[0046] Find the eigenvalues ​​λ′ of the covariance matrix C′. m1 ,λ′ m2 ,λ′ m3 ,λ′ m1 ≤λ m2 ≤λm3 Then the rate of change of the normal vector corresponding to each point in the K-neighborhood on the surface is expressed as the eigenvalue λ′. m1 ,λ′ m2 ,λ′ m3 To ensure that the surface variations of the three-dimensional ellipsoidal surface composed of basic parameters are consistent, a matching threshold is constructed. Then for point P m If ε′ m If ≥ε′0, then point P m If it is a feature point, then it is a non-feature point;

[0047] Non-feature point cloud data constitutes the smooth portion of the 3D surface. By reducing the point cloud density of the smooth portion, the total amount of point cloud data in the entire 3D model can be effectively reduced without affecting key features such as the 3D surface contour, thus improving data processing performance. The specific processing steps are as follows:

[0048] Based on the aforementioned steps, the non-feature point cloud dataset {(x′) is determined. i ,y′ i ,z′ i Extract the maximum value x′ from the three coordinates. i,max ,y′ i,max ,z′ i,max and the minimum value of the three coordinates x′ i,min ,y′ i,min ,z′ i,min Construct a 3D bounding box ΔL, with side lengths of ΔL as follows: Where Δ i >0 (i = x, y, z) means that the side length margin is sufficient to ensure that all points are included;

[0049] like Figure 1 As shown, the three-dimensional bounding box ΔL is divided into N smaller bounding boxes Δl based on the voxel bounding method. n (n = 1, 2... N), the bounding box Δl of each cube n The side length is l i =n -1 L i (i = x, y, z);

[0050] Determine the bounding box Δl of each cube respectively. n center point O n coordinates

[0051] Where (x) n,1 ,y n,1 ,z n,1 ) and (x n,2 ,y n,2 ,z n,2 ) are the bounding boxes of the cube Δl nThe coordinates of the two endpoints of any diagonal;

[0052] Calculate the bounding box Δl for each cube separately. n Points p in the middle n (x n ,y n ,z n ) and center point O n distance Extract and retain the data of points with the smallest distance, while deleting the bounding boxes Δl of each cube. n The other points in;

[0053] Repeat the above segmentation process until the preset requirements are met;

[0054] Specifically, this application also includes steps for optimizing point cloud vector directions and adjusting curve representation capabilities. Specifically, based on the point cloud data determined in the aforementioned steps, the coordinates of each point are determined and the normal direction is calculated. A query sequence {p′0, p′1...p′} is established in ascending order of height values. n ...};

[0055] Transform the normal vectors of each point. After translation, the normal vector can be expressed as: Then the normal vector of the first point p′0 in the query sequence is (0,0,1);

[0056] The weight of the specified point p′0 is 1. The weights of the other points are updated sequentially according to the order of the query sequence. in It refers to the normal vector of the point corresponding to the minimum weight value among all updated points;

[0057] During each change process, if Then Change the direction;

[0058] It also includes steps for building the point cloud octree and its position indices and gradients, specifically:

[0059] Obtain all point cloud data P all ={x i ,y i ,z i For each node i = 1, 2, 3, ..., n, a point cloud octree is constructed based on the aforementioned bounding box structure; each child node T at each level of the point cloud octree is... i For each group of bounding boxes of the same size; assuming the point cloud octree is N layers, then the side length of the minimum bounding box is l. min =L max ×2 -N ;

[0060] Assume point cloud data P all The index of the child node in the octree is q. m Then the description function used to represent the corresponding child node can be expressed as:

[0061]

[0062] Where m represents the point cloud data P all The level number of the child node in the octree; q mid Represents point cloud data P all The midpoint of the bounding box corresponding to the child node of the octree; i = x, y, z; * indicates convolution operation;

[0063] Then child node q m gradient of the indicator function t m Interpolation weights for child nodes;

[0064] Step 4: Update the point cloud data based on the previous steps, and obtain valid point cloud data after necessary preprocessing; the preprocessing includes determining whether the average Euclidean distance from the point to the midpoint of the neighborhood satisfies a Gaussian distribution after obtaining the neighborhood through K-nearest neighbor search for possible outliers; if it does not satisfy the distribution, it is removed.

[0065] Step 5: Update the point cloud data and generate a ship scene model based on the Possion surface reconstruction method.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A three-dimensional reconstruction method for virtual teaching modeling of a marine vessel, characterized in that, Comprising the following steps: Step 1: Establish a point cloud data coordinate system: Obtain a point cloud dataset of the ship scene based on point cloud scanning. ,in Indicates the first The point cloud data acquired in this scan, for point cloud data The first Points Define its coordinate system in the scanning station. The coordinates below are In Earth coordinate system The coordinates below are ; Step 2, optimization processing of feature points of ship space profile Obtain point cloud data of the ship scene, and establish surface point cloud datasets based on continuous smooth surfaces within the point cloud data. For the first point cloud dataset in the surface, A smooth curved surface The m-th continuously differentiable point If m < M, solve for the point of Domain point set : ; computing point of the field point set center coordinates in : ， ， ； Further obtained Covariance matrix of the points in the field point set is expressed as: ； covariance matrix is a positive semi-definite symmetric matrix, the smallest eigenvalue and the corresponding eigenvector are solved On a smooth surface the smallest eigenvalue and the corresponding eigenvector are the normal vector at point ​ Based on the aforementioned steps, points are determined respectively. of The normal vectors of each point in the neighborhood point set are used to specify the threshold angle between the normal vectors according to the simplification ratio. ;by The center point of the domain point set The corresponding normal vector is the initial normal vector. ,Sure Any point in the domain point set its nearest neighbor The angle between the normal vectors , retain point Create feature points representing the spatial outline of the ship and delete other points; Step 3, optimization processing of feature points and non-feature points Based on the aforementioned covariance matrix Establish points containing point distance weighting coefficients of Domain covariance matrix : ;in For point of Neighborhood radius; Solving the covariance matrix eigenvalues , ;but The rate of change of the normal vectors corresponding to each point in the domain on the surface is expressed as eigenvalues. To ensure that the surface variations of the three-dimensional ellipsoidal surface composed of basic parameters are consistent, a matching threshold is constructed. , Then for point ,like Then point If it is a feature point, then it is a non-feature point; The non-feature point cloud data constitutes a smooth part in the three-dimensional curved surface, and by reducing the point cloud density of the smooth part, the total amount of point cloud data of the entire three-dimensional model is effectively reduced without affecting the key features of the three-dimensional curved surface profile, and the data processing performance is improved; the specific processing steps are as follows: Determine the non-feature point cloud data set based on the foregoing steps , extract the three coordinate maximum value and the three coordinate minimum value Establish a three-dimensional bounding box , the edge length of the three-dimensional bounding box is respectively ; wherein refers to the edge length margin to ensure that all points are included, ; based on a voxel-based bounding method partitioned into small bounding boxes , ; each cubic bounding box with an edge length of ; Determine the bounding box of each cube separately center point coordinates ; wherein are the coordinates of two end points of an arbitrary diagonal of the cuboid bounding box respectively; the distance between each point in each cuboid bounding box and the center point is calculated respectively ; the point data with the smallest distance is extracted and reserved, while other points in each cuboid bounding box are deleted; the above segmentation process is executed in a loop until the preset requirement is met; Step 4, update the point cloud data based on the foregoing steps, and obtain effective point cloud data after necessary preprocessing; the preprocessing includes that for possible outliers, the average Euclidean distance from the point to the points in the neighborhood is determined after the neighborhood is obtained through K-neighbor search whether it satisfies the Gaussian distribution, and if not, it is removed; Step 5, update the point cloud data, and generate a ship scene model based on the Possion surface reconstruction method.

2. The three-dimensional reconstruction method for virtual teaching modeling of a marine vessel according to claim 1, characterized in that, Step 3 also includes point cloud orientation adjustment and optimization, specifically: determining the coordinates of each point and calculating the normal direction based on the point cloud data determined in the preceding steps, and establishing a query sequence in ascending order of height values. Transform the normal vectors of each point. After translation, the normal vector is expressed as: Then query the first point in the sequence. The normal vector is (0,0,1); specify the point The weight of the first point is 1, and the weights of the other points are updated sequentially according to the query sequence. ;in It refers to the normal vector of the point corresponding to the minimum weight value among all updated points; In each update process, if then the direction of is converted.

3. The three-dimensional reconstruction method for virtual teaching modeling of a marine vessel according to claim 2, characterized in that, Also included are steps for establishing an octree of point clouds and a position index and gradient thereof, specifically referring to: obtaining all point cloud data , establishing an octree of point clouds based on the aforementioned bounding box structure ; each sub-node of each level of the octree of point clouds corresponds to a number of bounding boxes of the same size; assuming that the established octree of point clouds is layers, the side length of the smallest bounding box is ; Assume the point cloud data The octree in which the point cloud data is located The position index corresponding to the child node is The description function for representing the corresponding child node is represented as: ； wherein representing point cloud data octree in which corresponding to the level of the child node; representing point cloud data octree in which midpoint of the bounding box corresponding to the child node; ; representing performing convolution operation; then the child node indicate the gradient of the function ; interpolation weight for the child node.

Images