A partitioned quadrilateral mesh reconstruction method for three-dimensional point cloud model
Patent Information
- Application Number
- CN202311782204.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-22
- Publication Date
- 2026-08-18
- Estimated Expiration
- 2043-12-22
AI Technical Summary
然而，为了保证重建的准确性，重建的网格模型通常较为复杂且网格为三角面篇，仍需要进一步处理才能进行实际使用
[0042] Step 6-6: Alternately implement steps 6-4 and 6-5 until the maximum number of iterations is reached or the difference in deformation energy E between two iterations is less than the threshold θ, to obtain a quadrilateral mesh model M that can well reflect the partitioning characteristics of the original model.
Smart Images
Figure CN117726772B_ABST
Abstract
Description
Technical fields:
[0001] This invention relates to the field of mesh reconstruction of three-dimensional point cloud models, and more particularly to a method for reconstructing a three-dimensional point cloud model into a quadrilateral mesh model with partitioning information. Background technology:
[0002] 3D point cloud reconstruction refers to recovering the 3D shape and structure of an object's surface from a 3D point cloud. This technology has wide applications in computer vision, computer graphics, and other fields. One common method is surface reconstruction, which connects discrete point cloud data to form a continuous 3D mesh or surface. In this process, techniques such as point cloud registration, voxelization, interpolation, and triangulation are typically applied to improve the accuracy and precision of the reconstruction. Another common 3D point cloud reconstruction technique is voxelization reconstruction, which maps the 3D point cloud into a 3D voxel mesh. By sampling and filling the point cloud within the voxel mesh, a dense 3D data representation can be generated, thus better describing the shape and structure of the object. In specific applications, such as architecture, cultural relic preservation, and industrial design, 3D point cloud reconstruction technology can help people quickly and accurately obtain 3D models of objects. Simultaneously, 3D point cloud reconstruction also plays a crucial role in fields such as autonomous driving and robot navigation, helping devices perceive and understand their surroundings.
[0003] 3D point clouds are datasets composed of a large number of 3D point coordinates, typically acquired by devices such as laser scanners, cameras, and depth sensors. However, raw 3D point cloud data may contain noise, missing data, and incompleteness. Therefore, 3D point cloud reconstruction is developing towards improving accuracy and robustness. This involves both processing noisy and missing data and restoring the good details and sharp features of the 3D point cloud. A major advantage of 3D point cloud reconstruction is its ability to quickly convert real-world objects into computer mesh models, significantly reducing modeling time and greatly improving work efficiency. However, to ensure reconstruction accuracy, the reconstructed mesh model is usually quite complex and contains triangular faces, requiring further processing before practical use. Summary of the Invention:
[0004] To fill the gap in existing technologies for directly reconstructing partitioned quadrilateral mesh models from 3D point cloud models, this invention provides a method for reconstructing partitioned quadrilateral meshes from 3D point cloud models. For the input 3D point cloud model, this method first extracts the corresponding central skeleton, then divides the central skeleton, and initializes a quadrilateral mesh model that accurately reflects the partitioning characteristics of the original model based on the divided skeleton. Automated mesh reconstruction is achieved by alternately implementing iterative mesh deformation based on minimizing the energy function E and adaptive local subdivision, resulting in a quadrilateral mesh model that approximates the 3D point cloud model and contains partitioning information.
[0005] A method for reconstructing partitioned quadrilateral meshes for 3D point cloud models includes the following steps:
[0006] Step 1: Input the 3D point cloud model P to be reconstructed, including Among them I p Let P be the set of vertex indices.
[0007] Step 2: Extract the central skeleton S from the 3D point cloud model P, which contains the vertex set. The set of edges that connect each pair of adjacent vertices. Among them I x Let X be the set of vertex indices.
[0008] For each vertex x of the central skeleton S in sequence i Principal component analysis is performed on ∈X to construct a minimum rotation bounding box based on the local coordinate system. Joint nodes x are obtained by detecting the length high proportion of the corresponding bounding boxes and classifying them according to a given proportion threshold. j-node ∈X.
[0009] Step 4: Detect whether there is an area abrupt change in the 3D point cloud model P to be reconstructed along the principal axis direction of the local coordinate system obtained based on principal component analysis. If so, construct the corresponding cutting plane to obtain the partition information of the 3D point cloud model P.
[0010] Step 5: Based on joint node x j-node The minimum bounding box of revolution and the partition information corresponding to the cutting plane are used to construct an initial quadrilateral mesh model M0 with partition information.
[0011] Step 6: Construct the deformation energy function Where E e E is the deformation energy term. s For the smoothing term, ω e and ω s The corresponding weights are assigned; and the mesh iterative deformation and adaptive local subdivision steps based on minimizing the energy function E are implemented alternately to realize automated mesh reconstruction for 3D point cloud models, resulting in a quadrilateral mesh model M that can well reflect the partitioning characteristics of the original model.
[0012] In step 2, the method for extracting the central skeleton S from the 3D point cloud model P is as follows:
[0013] Step 2-1: Randomly extract a subset of points from the 3D point cloud model P.
[0014] Step 2-2: Based on each vertex The adjacent vertex information is used to project the points in the sample sub-point set to obtain a point cloud model with an approximate central skeleton.
[0015] Steps 2-3: Alternately perform downsampling, smoothing, and re-alignment steps to obtain the final central skeleton S.
[0016] In step 3, some vertices x of the central skeleton S are... i ∈X is classified as a joint node x j-node The method for ∈X is:
[0017] Step 3-1: For the central axis skeleton vertex x i and the set of adjacent vertices N(x) within a local range i Principal component analysis is performed to construct a local coordinate system.
[0018] Step 3-2: Construct the corresponding central axis skeleton vertex x based on the local coordinate system i The smallest bounding box of revolution, such that x i and N(x) i All are inside the enclosure.
[0019] Step 3-3: Analyze the lengths λ1, λ2, and λ3 (λ1>λ2>λ3) of the three orthogonal sides of the minimum bounding box of revolution. If If it is less than a given threshold, typically set to 0.85, then x will be... i and N(x) i Classified as joint nodes x j-node ∈X.
[0020] Step 3-4: Repeat steps 3-1 to 3-3 until all joint nodes x j-node All values ∈X were detected and classified.
[0021] In step 4, the method for constructing the cutting plane to obtain the model partition information is as follows:
[0022] Step 4-1: Select each vertex x of the central skeleton i The maximum distance to the nearest neighbor vertex is used as the interpolation length L.
[0023] Step 4-2: Along each vertex x of the central axis skeleton i In the local coordinate system, along the x-axis, a set of points is uniformly selected within the distance range [-L, L] to form a local point set. Where X i All vertices are in x i The local coordinate system and the partitioning attribute are related to x i Maintain consistency.
[0024] Step 4-3: For each Along x i A family of scanning planes perpendicular to the x-axis is constructed in the local coordinate system, and discrete points p located on each scanning plane are detected. i ∈P to average distance
[0025] Step 4-4: If the average distance between adjacent scan planes and There are mutations, in A cutting plane is created at the current location to mark the new partition information.
[0026] Steps 4-5: Repeat steps 4-2 to 4-4 until all cutting planes are created to reflect the partitioning information of the 3D point cloud model P.
[0027] In step 5, the method for constructing the initial quadrilateral mesh model M0 with partitioning information is as follows:
[0028] Step 5-1: For each vertex x of the central skeleton S i Add noise coordinates to the vertex set X.
[0029] Step 5-2: Using the central skeleton as the distance target, construct an unsigned distance field based on the discrete voxel model representation, and assign voxels according to whether they contain joint nodes x. j-node They are divided into two categories.
[0030] Step 5-3: Extract all voxels with a distance value of 0, and use a ray casting algorithm to extract the initial quadrilateral mesh model M0 = {F, V} with partitioning information, where The mesh faces of the initial mesh. Let v be the grid vertices of the initial grid, where each vertex v i Each has corresponding partition information p(v) i ).
[0031] Step 5-4: Based on the Laplacian operator, smoothly collapse the initial mesh M0 to an isosurface with a distance value of 0.
[0032] Step 5-5: If vertex v i If a vertex is shared by three faces, merge it with the nearest vertex that is shared by four faces.
[0033] Steps 5-6: Move vertex v i The average position mapped to the vertices of its first-order neighborhood.
[0034] Steps 5-7: Repeat steps 5-5 to 5-6 until all vertices have been checked.
[0035] Steps 5-8: By cutting the plane, and each vertex v i Partition information p(v) i Add the partition information to the initial quadrilateral mesh model M0.
[0036] Step 6: The method for achieving automated mesh reconstruction by alternately implementing iterative mesh deformation based on minimizing the energy function and adaptive local subdivision is as follows:
[0037] Step 6-1: Using the 3D point cloud model P as the distance target, construct an unsigned distance field based on the discrete voxel model representation.
[0038] Step 6-2: Define the deformation energy To measure the vertex v after performing one deformation step i The deformation, where v i 'v' is obtained by performing a deformation step i coordinate.
[0039] Step 6-3: Define the smoothing term in Where N1(v) i ) for v i The index set of the first-order neighborhood vertices, n1(v i ) is N1(v i The number of vertices.
[0040] Step 6-4: Minimize the energy function A robust mesh deformation can be implemented, transforming each vertex v i Move a distance εD(v) along the normal of the corresponding mesh model surface i ), where ε is the desired step size, D(v) i The distance field value at the current position of the vertex is sampled. Each mesh deformation process is performed by the weight ω. e and ω s Control is needed to maintain a balance between the degree of detail restoration and the rate of deformation.
[0041] Step 6-5: If the side length of a certain mesh face in the currently deformed mesh model exceeds the longest expected side length l, perform an adaptive local subdivision in the local region formed by the mesh face and its adjacent mesh faces.
[0042] Step 6-6: Alternately implement steps 6-4 and 6-5 until the maximum number of iterations is reached or the difference in deformation energy E between two iterations is less than the threshold θ, to obtain a quadrilateral mesh model M that can well reflect the partitioning characteristics of the original model. Attached image description:
[0043] Figure 1 This is a flowchart of a method for reconstructing a 3D point cloud model into a quadrilateral mesh with partitions, according to the present invention.
[0044] Figure 2 The input is a 3D point cloud model of a dog.
[0045] Figure 3 Example of a 3D point cloud model of a "dog" and its central skeleton.
[0046] Figure 4 For the minimum revolved bounding box and joint node x j-node Classification examples
[0047] Figure 5 Example of a case where a cutting plane needs to be created.
[0048] Figure 6 The initial quadrilateral mesh model M0 for the 3D point cloud model "dog"
[0049] Figure 7 Reconstructing the quadrilateral mesh model M of the 3D point cloud model "dog"
[0050] Figure 8 Comparison results of reconstructed partitioned quadrilateral mesh model M and input 3D point cloud model P Detailed implementation method:
[0051] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The following implementation does not constitute a limitation of the present invention.
[0052] This invention discloses a method for directly reconstructing a partitioned quadrilateral mesh model from a 3D point cloud model, the flowchart of which is shown below. Figure 1 As shown, the specific implementation steps are as follows:
[0053] Step 101: Input the 3D point cloud model to be reconstructed. Among them I p Let P be the set of vertex indices.
[0054] like Figure 2 As shown, the input 3D point cloud model contains a point set P, where each point p i It includes x-axis, y-axis, and z-axis coordinate data. The 3D point cloud model can be acquired by devices such as laser scanners, cameras, and depth sensors, and the file is saved in PLY format or other easily readable interface formats.
[0055] Step 102: Extract the central skeleton from the above 3D point cloud model to reflect the topological structure of the 3D point cloud model. For any 3D point cloud model, the specific extraction method is as follows:
[0056] like Figure 3 As shown, a subset of points is randomly extracted from the 3D point cloud model P. Based on each vertex The adjacent vertex information is used to project the points in the sample sub-point set to obtain an approximate central skeleton point cloud model. By alternately performing downsampling, smoothing, and re-alignment steps, the final central skeleton S is obtained.
[0057] Step 103, extract the joint nodes x in the central skeleton. j-node The method is as follows: The region with abrupt directional changes in the central skeleton is identified to complete the region division of the input 3D point cloud model.
[0058] For the central axis skeleton vertex x i and the set of adjacent vertices N(x) within a local range i Principal component analysis is performed to construct a local coordinate system.
[0059] like Figure 4 As shown, the corresponding central axis skeleton vertex x is constructed based on the local coordinate system. i The smallest bounding box of revolution, such that x i and N(x) i All are inside the bounding box. Analyze the lengths λ1, λ2, and λ3 (λ1>λ2>λ3) of the three orthogonal edges of the minimum bounding box of revolution. If... If it is less than a given threshold, typically set to 0.85, then x will be... i and N(x) i Classified as joint nodes x j-node ∈X.
[0060] Step 104, at joint node x j-node The method for detecting abrupt changes in area of a 3D point cloud model along a specific direction in the local coordinate system and constructing cutting planes to partition the 3D point cloud model is as follows:
[0061] like Figure 5 As shown, select each vertex x of the central axis skeleton. i The maximum distance to the nearest neighbor vertex is used as the interpolation length L, along x. i In the local coordinate system, along the x-axis, several points are uniformly selected within the distance range [-L, L] to form a partition attribute that is related to x. i Consistent local point set
[0062] For each Along x i A family of scanning planes perpendicular to the x-axis is constructed in the local coordinate system, and discrete points p located on each scanning plane are calculated. i ∈P to average distance If the average distance between adjacent scan planes and There are mutations, in A cutting plane is created at the current location to mark the new partition information.
[0063] Step 105, based on joint node x j-node The specific method for constructing the initial quadrilateral mesh model M0 with partitioning information using the cutting plane is as follows:
[0064] like Figure 6 As shown, with the central skeleton as the distance target, an unsigned distance field based on the discrete voxel model is constructed. All voxels with a distance value of 0 are extracted. An initial quadrilateral mesh model M0 with partition information is obtained by using a ray casting algorithm. The initial mesh M0 is then collapsed into a smooth isosurface based on the Laplacian operator.
[0065] Step 106, based on the initial mesh generated above, the specific method for achieving automated mesh reconstruction by alternately implementing mesh iterative deformation based on minimizing the energy function and adaptive local subdivision is as follows:
[0066] By minimizing the energy function A robust mesh deformation can be implemented, determined by weight ω. e and ω s Control is needed to maintain a balance between the degree of detail restoration and the rate of deformation.
[0067] If the side length of a mesh face in the currently deformed mesh model exceeds the longest expected side length l, an adaptive local subdivision is performed within the local region formed by that mesh face and its adjacent mesh faces. The subdivision uses the Catmull-Clark algorithm, where the newly generated vertex v... new partition value p(v) new It is derived from the weighted average of the partition values of its original point set.
[0068] Alternating between iterative mesh deformation based on minimizing the energy function and adaptive local subdivision, the process continues until a pre-defined maximum number of iterations is reached or the energy difference between two iterations is less than a threshold θ. This yields a quadrilateral mesh model that effectively reflects the partitioning characteristics of the original model. Figure 7 As shown.
[0069] Step 107: This method ultimately outputs a quadrilateral mesh model that effectively reflects the morphological characteristics of the original 3D point cloud model, while also exhibiting good accuracy and robustness. Figure 8 As shown.
[0070] A typical embodiment of the present invention is as follows:
[0071] 1. For example Figure 2 The 3D point cloud model selected for processing in this example is a "dog".
[0072] 2. Import the above three-dimensional point cloud model into the test program written in C++ according to the method of this invention.
[0073] 3. In the test program, ε is set to 0.0125, l to 0.05, and the iteration termination condition is θ < 0.001.
[0074] Figure 7 A reconstructed quadrilateral mesh model M of a 3D point cloud model of a "dog" is presented. (From...) Figure 8 It can be seen that the quadrilateral mesh reconstructed by the present invention can firstly maintain the good details of the original point cloud model, and secondly, it has partitions that conform to the original shape characteristics.
[0075] The above description is merely a preferred embodiment of the present invention. It should be noted that various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made without departing from the principle of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for reconstructing partitioned quadrilateral meshes for 3D point cloud models, characterized in that, Includes the following steps: Step 1: Input the 3D point cloud model to be reconstructed ; Step 2: From the 3D point cloud model Extracting the central skeleton , containing a set of vertices The set of edges that connect each pair of adjacent vertices. ; Step 3: Sequentially assemble the central axial frame. vertex Principal component analysis is performed to construct a minimum rotation bounding box based on the local coordinate system, and joint nodes are obtained by classification according to a pre-defined high-proportion threshold of the bounding box length. ; Step 4: Detect the 3D point cloud model to be reconstructed along the principal axis of the local coordinate system. If there is an area abrupt change, then construct a cluster of cutting planes to obtain the partitioning information of the input model; Step 5: Based on the joint nodes The minimum revolved bounding box and the cutting plane clusters construct the initial quadrilateral mesh model with partitioning information. ; Step 6: Construct a deformation energy function that includes deformation energy and a smoothing term. ,in, For deformation energy, To measure the vertex after performing a deformation step Deformation, of which To obtain by performing a deformation step coordinate; For smoothing terms, ,in ,in for The index set of first-order neighborhood vertices, for The number of vertices; and The weights for the mesh deformation process are assigned; and the process is alternately implemented based on minimizing the energy function. The iterative deformation and adaptive local subdivision steps enable automated mesh reconstruction for 3D point cloud models, resulting in a quadrilateral mesh model that effectively reflects the partitioning characteristics of the original model. .
2. The method for reconstructing a partitioned quadrilateral mesh for a 3D point cloud model according to claim 1, characterized in that, In step 2, from the 3D point cloud model Extracting the central skeleton The method is as follows: Step 2-1: From the 3D point cloud model Randomly extract a subset of points samples And based on each vertex The adjacent vertex information is used to obtain a point cloud model with an approximate central skeleton. Step 2-2: Alternately perform downsampling, smoothing, and re-alignment steps to obtain the target central skeleton. .
3. The method for reconstructing a partitioned quadrilateral mesh for a 3D point cloud model according to claim 1, characterized in that, In step 3, the central frame is... Some vertices Classified as joint nodes The method is as follows: Step 3-1: For the vertices of the central axis skeleton and the set of adjacent vertices within a local range Perform principal component analysis and construct a local coordinate system. The smallest revolving bounding box, such that and All are inside the enclosure; Step 3-2: Calculate the lengths of the three orthogonal sides of the smallest revolute bounding box. , , Perform analysis; if If it is less than a given threshold, then and Classified as joint nodes .
4. The method for reconstructing a partitioned quadrilateral mesh for a 3D point cloud model according to claim 1, characterized in that, In step 4, the method for constructing the cutting plane to obtain the model partition information is as follows: Step 4-1: Select each vertex of the central skeleton The maximum distance to the nearest neighbor vertex is used as the interpolation length. Along the local coordinate system In the axial direction, Within a distance range, several points are uniformly selected to form the partition attributes. Consistent local point set ; Step 4-2: For each along local coordinate system Construct a cluster perpendicular to the axial direction The scanning plane of the axis is used to detect discrete points located on each scanning plane. arrive average distance If the average distance between adjacent scan planes and There are mutations, in A cutting plane is created at the current location to mark the new partition information.
5. The method for reconstructing a partitioned quadrilateral mesh for a 3D point cloud model according to claim 1, characterized in that, In step 5, an initial quadrilateral mesh model with partitioning information is constructed. The method is as follows: Step 5-1: Using the central skeleton as the distance target, construct an unsigned distance field based on a discrete voxel model representation, and extract all distance values. The voxels are used to extract an initial quadrilateral mesh model with partitioning information through a ray casting algorithm. ; And based on the Laplace operator, the initial mesh is initialized. It collapses into a smooth isosurface; Step 5-2: Based on the distribution of the cutting plane clusters and each vertex Partition information Add partitioning information to the initial quadrilateral mesh model. .
6. The method for reconstructing a partitioned quadrilateral mesh for a 3D point cloud model according to claim 1, characterized in that, In step 6, the method based on minimizing the energy function is implemented alternately. The method for achieving automated mesh reconstruction through iterative mesh deformation and adaptive local subdivision is as follows: Step 6-1: Using a 3D point cloud model For the distance target, an unsigned distance field based on a discrete voxel model is constructed; Step 6-2: Define the deformation energy term as To measure the vertex after performing a deformation step The deformation is defined as the smoothing term. By minimizing the energy function A robust mesh deformation can be implemented, and each mesh deformation process is weighted. and Control is needed to maintain a balance between the degree of detail restoration and the rate of deformation; Step 6-3: If the side length of a certain mesh face in the currently deformed mesh model exceeds the longest expected side length. An adaptive local subdivision is performed within the local area formed by the grid surface and its adjacent grid surfaces; Step 6-4: Alternately perform steps 6-2 and 6-3 until the pre-defined maximum number of iterations or the deformation energy between two iterations is reached. Difference less than threshold Stop at a certain time to obtain a quadrilateral mesh model that can well reflect the partitioning characteristics of the original model.