Multi-line structured light three-dimensional reconstruction frame point cloud real-time de-stacking and fusion method

Abstract

The application provides a multi-line structured light three-dimensional reconstruction frame point cloud real-time de-stacking and fusion method, and relates to the technical field of line structured light three-dimensional reconstruction. The distribution characteristics of the spatial position of the point cloud are utilized, a global grid structure is established through the position of the multi-view frame point cloud during three-dimensional reconstruction, the corresponding domain point set of each frame point cloud is solved, and the normal feature estimation is performed according to the distribution characteristics of the neighborhood points. The overlapping condition of the point cloud is detected according to the global grid structure, the cylindrical neighborhood structure of the overlapping frame point cloud is established, and the point cloud fusion is realized through the weighted sum of the point cloud vector direction of the target point and the fusion neighborhood point set and the normal vector feature. The algorithm can meet the system stability requirement in the three-dimensional reconstruction, improves the quality of the final point cloud model, greatly improves the measurement adaptability of the multi-line structured light three-dimensional reconstruction in the complex environment, and effectively improves the multi-line structured light three-dimensional reconstruction precision.

Description

Technical Field

[0001] This invention relates to the field of line structured light 3D reconstruction technology, and in particular to a method for real-time destacking and fusion of point clouds in multi-line structured light 3D reconstruction frames. Background Technology

[0002] Structured light scanning systems based on stereo vision are one of the most effective methods for acquiring 3D surface data of objects. They are widely used in reverse engineering, biomedicine, and cultural relic preservation, and are an important means of measuring workpiece dimensions and detecting defects. Like other 3D scanning devices, multi-line structured light scanners can quickly reconstruct object surfaces, but due to their limited scanning field of view, images of the object need to be acquired from different perspectives to reconstruct a complete 3D surface model. Therefore, during the scanning process, the point clouds reconstructed from different perspectives need to be registered with the point cloud in the global coordinate system. To make the reconstructed model more complete, multiple scans are performed at a certain location during reconstruction. The resulting laser point cloud model inevitably suffers from stacking, redundancy, and uneven distribution. Furthermore, when scanning and measuring large objects, the registration accuracy of multi-view point cloud data decreases with the number of registrations. Insufficient registration accuracy can lead to local layering lines in the point cloud model. Directly constructing a patch model from the obtained point cloud results in a large number of redundant patches on the model surface, making the surface too rough and with obvious splicing marks. To avoid the influence of multiple scans and stitching of point clouds at the same location and to obtain a uniform and smooth point cloud model, point cloud model fusion processing is required. Point cloud fusion is the foundation of 3D reconstruction. It can not only merge multi-layer point cloud data into single-layer point cloud data to reduce point cloud data redundancy, but also remove noise interference to obtain a model with unchanged features and a smooth surface.

[0003] Point cloud fusion can be divided into multi-view frame point cloud fusion processing and overall point cloud model post-processing. Overall point cloud model post-processing involves fusing the entire point cloud model after the scanning and reconstruction of the entire object under test to obtain a uniform, single-layer point cloud model. However, this post-processing method struggles to account for surface point cloud deviations during the reconstruction process, making it difficult to process point clouds with uneven surface thickness distribution and unable to effectively and accurately compensate for point cloud surface defects. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a real-time destacking and fusion method for point clouds in multi-line structured light 3D reconstruction frames.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A real-time destacking and fusion method for point clouds in multi-line structured light 3D reconstruction frames includes:

[0007] Acquire the laser point cloud data of the target to be reconstructed and preprocess it to obtain the preprocessed laser point cloud data;

[0008] An initial network structure is established based on the three-dimensional coordinates of the first frame of the laser point cloud data.

[0009] Using the point clouds of subsequent frames in the laser point cloud data, point cloud classification and network structure expansion are performed on the initial network structure to obtain the final network structure, and the common grid and the index of the frame point cloud and the total point cloud in the common grid are determined according to the final network structure;

[0010] Establish a corresponding domain network structure based on the target points in the public network;

[0011] Based on the corresponding domain network structure, the central grid of the target point and its eight neighboring grid point clouds in the x, y, and z axis directions are searched to form the neighborhood point set of the target point;

[0012] The normal vector is estimated based on the neighborhood point set and principal component analysis, and the bidirectionality of the normal is eliminated to obtain the optimized normal vector.

[0013] Based on the optimized normal vector, the redundant points of the target point are determined and the redundant points are removed to obtain the laser point cloud data after removal.

[0014] A cylindrical neighborhood of the target point is established based on the removed laser point cloud data.

[0015] The target point is optimized and iterated based on the cylindrical region to obtain a fused point cloud.

[0016] Preferably, the step of establishing the initial network structure based on the three-dimensional coordinates of the first frame of the laser point cloud data includes:

[0017] Based on the three-dimensional coordinates of the first frame point cloud, calculate the maximum and minimum coordinates of the first frame point cloud to obtain the spatial bounding box of the first frame point cloud.

[0018] Calculate the average point spacing of the point cloud;

[0019] Calculate the grid side length based on the average point spacing;

[0020] The initial network structure is obtained by numbering the center coordinates of the scalable mesh side length and the bounding box.

[0021] Preferably, the step of using subsequent frame point clouds from the laser point cloud data to perform point cloud classification and network structure expansion on the initial network structure to obtain the final network structure and determining the common grid based on the final network structure includes:

[0022] Calculate the minimum Euclidean distance from the point cloud of subsequent frames to the center point of the initial network structure;

[0023] Based on the minimum Euclidean distance of the center point, determine whether the point cloud of the subsequent frame exceeds the range of the initial network structure, and obtain the judgment result. If the judgment result is yes, then the current corresponding point cloud is stored. If the judgment result is no, then the network structure of the current point cloud is expanded to obtain a common network.

[0024] Preferably, the step of estimating the normal vector based on the target point using a neighborhood point set and principal component analysis, and then eliminating the bidirectionality of the normal vector to obtain the optimized normal vector, includes:

[0025] Calculate the centroid of the preset points in the domain point set based on the target point;

[0026] Construct a corresponding covariance matrix based on the centroid and preset points in the neighborhood point set;

[0027] The covariance matrix is ​​decomposed using singular values ​​to obtain eigenvalues ​​and eigenvectors;

[0028] The normal vector of the target point is obtained based on the eigenvalues ​​and eigenvectors.

[0029] The bidirectionality of the normal vector of the target point is eliminated to obtain the optimized normal vector.

[0030] Preferably, the step of determining redundant points of the target point based on the optimized normal vector and removing redundant points to obtain the removed laser point cloud data includes:

[0031] The corresponding direction vector is calculated based on the optimized normal vector and the normal vector corresponding to the preset point in the neighborhood point set.

[0032] The angle between the direction vector and the optimized normal vector and the normal vector corresponding to the preset point is calculated to obtain the first angle and the second angle.

[0033] Redundant points are determined and removed based on the angle difference between the first included angle and the second included angle to obtain the removed laser point cloud data.

[0034] Preferably, establishing the cylindrical neighborhood of the target point based on the removed laser point cloud data includes:

[0035] The normal vector of the target point in the removed laser point cloud data is corrected by using the redundant point set corresponding to the neighborhood point set to obtain the corrected normal vector.

[0036] Calculate the projection points of the target point along the approximate corrected normal vector in the positive and negative directions, and determine the center coordinates of the upper and lower surfaces of the cylinder;

[0037] The cylinder neighborhood is determined based on the preset cylinder parameters and the center coordinates.

[0038] Preferably, the step of optimizing and iterating the target point based on the cylindrical region to obtain the fused point cloud includes:

[0039] Determine redundant reference points;

[0040] The target point is optimized and iterated based on the redundant reference points to obtain the fusion point position;

[0041] The fusion point cloud is obtained based on the location of the fusion point.

[0042] Preferably, the expression for the redundant reference point is:

[0043]

[0044] Where, p i For the target point, p j For neighborhood points, v ij Let d be the direction vector. i With d j v ij In p i and p j The forward projection length, p, in the normal direction c This is a redundant reference point.

[0045] The present invention discloses the following technical effects:

[0046] This invention provides a real-time destacking and fusion method for multi-line structured light 3D reconstruction frame point clouds, comprising: acquiring laser point cloud data of the target to be reconstructed and preprocessing it to obtain preprocessed laser point cloud data; establishing an initial network structure based on the 3D coordinates of the first frame point cloud in the laser point cloud data; classifying and expanding the network structure using subsequent frame point clouds in the laser point cloud data on the initial network structure to obtain a final network structure, and determining a common grid and the indices of frame point clouds and the total point cloud in the common grid based on the final network structure; and establishing a corresponding neighborhood network structure based on the target points in the common network. Based on the corresponding neighborhood network structure, the central grid of the target point and its eight neighboring grids in the x, y, and z axes are searched to form a neighborhood point set of the target point. The normal vector of the target point is estimated using the neighborhood point set and principal component analysis, and bidirectional normal elimination is performed to obtain an optimized normal vector. Redundant points of the target point are determined based on the optimized normal vector and removed to obtain the removed laser point cloud data. A cylindrical neighborhood of the target point is established based on the removed laser point cloud data. The target point is then iteratively optimized based on the cylindrical neighborhood to obtain a fused point cloud. This invention, based on the spatial distribution characteristics of point clouds, establishes a global mesh structure and a cylindrical fusion neighborhood, and proposes a real-time fusion algorithm for multi-view frame point clouds in multi-line structured light 3D reconstruction based on normal constraints and point cloud direction vector weighting. This algorithm performs denoising and classification preprocessing on multi-line structured light point clouds. A global spatial grid structure is established based on the bounding box of the first frame point cloud. Subsequent frame point clouds are classified and expanded using the Euclidean distance of the point clouds within the spatial structure of the first frame. The relative position of each frame point cloud within the cumulatively reconstructed global point cloud is obtained based on the global grid structure. Points in a single frame point cloud that are classified into the existing point cloud grid are then fused. A 3×3×3 grid structure is established during the fusion process. The current frame point cloud and the cumulatively reconstructed global point cloud within the common grid are obtained, and the resulting global point cloud is used as the feature estimation neighborhood set for the points to be fused. Finally, a spatial cylindrical structure is established within the neighborhood set for the points to be fused, and point cloud normal fusion based on point cloud direction vector guidance is achieved within the cylindrical neighborhood. This invention effectively reduces the resource consumption of point cloud fusion algorithms and improves fusion efficiency. It enables uniform point cloud distribution in the reconstructed object point cloud model and has a good adjustment effect on the overall point cloud density. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of a method for real-time destacking and fusion of point clouds in multi-line structured light 3D reconstruction frames according to the present invention;

[0049] Figure 2 This is a schematic diagram of the neighborhood point cloud search of the central grid 3×3×3 structure of the present invention;

[0050] Figure 3 This is a schematic diagram illustrating the determination of the target point redundant point set and the establishment of the cylinder neighborhood in this invention.

[0051] Figure 4 This is a schematic diagram illustrating the point cloud effect before and after fusion during the artifact scanning and reconstruction process according to the present invention.

[0052] Figure 5 This is a schematic diagram of the point cloud effect before and after fusion during cylindrical scan reconstruction according to the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] like Figure 1 As shown, this invention provides a real-time destacking and fusion method for point clouds in multi-line structured light 3D reconstruction frames, including:

[0056] Step 100: Obtain the laser point cloud data of the target to be reconstructed and perform preprocessing to obtain the preprocessed laser point cloud data;

[0057] Specifically, a binocular multi-line structured light 3D reconstruction system is used to reconstruct and acquire laser data, and the multi-line structured light point cloud is preprocessed for denoising and classification:

[0058] Point cloud preprocessing includes denoising and classification preprocessing of the multi-line structured light point cloud obtained by scanning; in view of the problem that the extracted laser line point cloud is irregularly arranged and has a small number of isolated points and low-point clusters in a single laser line, Euclidean clustering algorithm is used to remove point clouds with too few points in the point cloud clusters and retain point clouds with high continuity.

[0059] Step 200: Establish an initial network structure based on the three-dimensional coordinates of the first frame of the laser point cloud data;

[0060] Step 300: Using the subsequent frame point clouds in the laser point cloud data, perform point cloud classification and network structure expansion on the initial network structure to obtain the final network structure, and determine the common grid and the index of the frame point cloud and the total point cloud in the common grid according to the final network structure;

[0061] Step 400: Establish the corresponding domain network structure based on the target points in the public network;

[0062] Step 500: Based on the corresponding domain network structure, search the central grid where the target point is located and its eight neighboring grid point clouds in the x, y, and z axis directions to form the neighborhood point set of the target point;

[0063] Step 600: Estimate the normal vector of the target point based on the neighborhood point set and principal component analysis, and eliminate the bidirectionality of the normal vector to obtain the optimized normal vector;

[0064] Step 700: Determine the redundant points of the target point based on the optimized normal vector and remove the redundant points to obtain the laser point cloud data after removal;

[0065] Step 800: Establish a cylindrical neighborhood of the target point based on the removed laser point cloud data;

[0066] Step 900: Optimize and iterate the target point based on the cylindrical region to obtain a fused point cloud.

[0067] Furthermore, such as Figure 2 As shown, the step of establishing the initial network structure based on the three-dimensional coordinates of the first frame of the laser point cloud data includes:

[0068] Based on the three-dimensional coordinates of the first frame point cloud, calculate the maximum and minimum coordinates of the first frame point cloud to obtain the spatial bounding box of the first frame point cloud.

[0069] Calculate the average point spacing of the point cloud;

[0070] Calculate the grid side length based on the average point spacing;

[0071] The initial network structure is obtained by numbering the center coordinates of the scalable mesh side length and the bounding box.

[0072] Specifically, the bounding box is solved based on the spatial location of the point cloud in the first frame, and a global spatial mesh structure is established;

[0073] Subsequent frame point clouds are classified into meshes and expanded based on the Euclidean distance of the point cloud in the spatial structure of the first frame point cloud.

[0074] The specific method for establishing a global mesh structure for the first frame point cloud is as follows:

[0075] Based on the 3D coordinates of the point cloud in the first frame, the maximum and minimum coordinates of the point cloud are calculated, and the spatial bounding box of the point cloud is obtained:

[0076] V={(x min ,y min ,y min ),(x max ,y max ,y max )};

[0077] The average distance d is calculated by setting parameters such as grid side length, total number of points in the point cloud, and coordinates based on the average distance between adjacent points in the point cloud. avg Take d avg The integer multiple of k is used as the grid side length;

[0078] Based on the determined point cloud side length L, determine the minimum bounding box of the point cloud in the x, y, and z axes of space, and number it as V(v) with the coordinates of the grid center. i ,v j ,v k ).

[0079] Furthermore, the step of using subsequent frame point clouds from the laser point cloud data to perform point cloud classification and network structure expansion on the initial network structure to obtain the final network structure, and determining the common grid based on the final network structure, includes:

[0080] Calculate the minimum Euclidean distance from the point cloud of subsequent frames to the center point of the initial network structure;

[0081] Based on the minimum Euclidean distance of the center point, determine whether the point cloud of the subsequent frame exceeds the range of the initial network structure, and obtain the judgment result. If the judgment result is yes, then the current corresponding point cloud is stored. If the judgment result is no, then the network structure of the current point cloud is expanded to obtain a common network.

[0082] Specifically, determine each 3D point p in the current frame point cloud. i (x i ,y i ,z i ) and the coordinates of the center of the global grid structure (v i ,v j ,v k The minimum Euclidean distance d) i The distance within the cloud for frame points satisfies d i Points < L are classified to the grid center coordinates as (v i ,v j ,v k In the grid index;

[0083] Iterate through all existing point cloud meshes, expanding and storing the mesh structure for points that do not satisfy all mesh center distance constraints; let the spatial bounding box coordinates of the first frame be... The spatial bounding box coordinates of subsequent frames are To ensure the continuity and uniformity of the spatial grid structure, the minimum coordinates of the first frame's bounding box will be used. This is the starting point of the spatial grid.

[0084] Furthermore, the step of estimating the normal vector based on the target point using the neighborhood point set and principal component analysis, and then eliminating the bidirectionality of the normal vector to obtain the optimized normal vector, includes:

[0085] Calculate the centroid of the preset points in the domain point set based on the target point;

[0086] Construct a corresponding covariance matrix based on the centroid and preset points in the neighborhood point set;

[0087] The covariance matrix is ​​decomposed using singular values ​​to obtain eigenvalues ​​and eigenvectors;

[0088] The normal vector of the target point is obtained based on the eigenvalues ​​and eigenvectors.

[0089] The bidirectionality of the normal vector of the target point is eliminated to obtain the optimized normal vector.

[0090] Specifically, based on the global grid structure, the relative position of each frame's point cloud in the cumulatively reconstructed global point cloud is obtained, and the grid with overlapping point clouds and the neighboring grid point clouds of the overlapping grids are obtained.

[0091] The normal vector of the neighborhood point cloud is estimated based on the neighborhood point set of the target point in a single frame point cloud, and the bidirectionality of the normal vector is eliminated.

[0092] The method for obtaining the relative position of the current frame point cloud in the cumulatively reconstructed global point cloud and its neighboring point clouds is as follows: After the current grid is added to the point cloud data of a certain frame, if the point cloud of subsequent frames is also merged into this grid, then this grid may have redundant point clouds, which are regarded as the region to be fused; a 3×3×3 structure of the central grid is established with the grid of the region to be fused as the center; starting from the central grid, the eight neighboring grids are searched layer by layer in the positive and negative directions of the x-axis, y-axis, and z-axis to form the nearest voxel grid structure, and the total point cloud within the 3×3×3 grid is taken as the target point p. i (x i ,y i ,z i The neighborhood normal feature estimation point cloud set

[0093] Mark the global grid structure and the search results of the central grid neighborhood, the point cloud set of the i-th frame is The set of point clouds in a frame point cloud that belong to the common grid of existing point clouds is: The newly added point cloud set in the frame point cloud is

[0094] The total point cloud set accumulated by the 3D reconstruction is S global During neighborhood search, the target point p of the frame point cloud within the common grid is determined. i (x i ,y i ,z i The corresponding total point cloud point set located within the same grid. And obtained based on the 3×3×3 structure of the central grid A collection of point clouds that satisfy the following relationships.

[0095] The specific methods for estimating the normal features and eliminating the bidirectionality of the normal vector in the point cloud of the target point's neighborhood are as follows:

[0096] Estimate the target point p using Principal Component Analysis (PCA) i (x i ,y i ,z i The set of neighboring points The characteristics of the normal vector are used to determine Normals to each point in the middle; first, take the target point p i (x i ,y i ,z i Using as the center, calculate the points p in its k-neighborhood. j (x j ,y j ,z j The center of mass of ) Then, a covariance matrix M is constructed within the k-neighborhood based on the neighboring points and the centroid, and the covariance matrix M is a symmetric matrix; finally, singular value decomposition is used to obtain the eigenvalues ​​λ of M. r and eigenvector v r The eigenvector v0 corresponding to the smallest eigenvalue λ0 of matrix M is p. i The normal vector;

[0097]

[0098] right The initially estimated normal vector undergoes bidirectional elimination. During multi-line structured light 3D reconstruction, the object being reconstructed always faces the scanning device, and each frame of laser point cloud extracted from the laser image is always along the optical center direction of the camera in the camera coordinate system. Therefore, the angle β between the obtained laser point cloud coordinates and the normal vector of the point cloud on the surface of the object being measured has only three possibilities: greater than 90°, less than 90°, and equal to 90°. For the target point p...i (x i ,y i ,z i The normal vector corresponding to ) is n i (x i ,y i ,z i Correction is performed; the point cloud normal vector n is defined. i For β ≥ 90°, when β < 90°, let n i =-n i This enables fast global redirection of the normal vector.

[0099] Furthermore, the step of determining redundant points of the target point based on the optimized normal vector and removing redundant points to obtain the removed laser point cloud data includes:

[0100] The corresponding direction vector is calculated based on the optimized normal vector and the normal vector corresponding to the preset point in the neighborhood point set.

[0101] The angle between the direction vector and the optimized normal vector and the normal vector corresponding to the preset point is calculated to obtain the first angle and the second angle.

[0102] Redundant points are determined and removed based on the angle difference between the first included angle and the second included angle to obtain the removed laser point cloud data.

[0103] Specifically, the redundancy of the point cloud is estimated based on the point cloud normal vector. Redundant points of the target point cloud are determined by the projection of the direction vectors of the target point and its nearest neighbor points and the normal vectors of the nearest neighbor points in the vertical direction. The cylindrical neighborhood structure of the target point is solved based on the approximate normal direction of the target point.

[0104] Furthermore, such as Figure 3-5 As shown, establishing the cylindrical neighborhood of the target point based on the removed laser point cloud data includes:

[0105] The normal vector of the target point in the removed laser point cloud data is corrected by using the redundant point set corresponding to the neighborhood point set to obtain the corrected normal vector.

[0106] Calculate the projection points of the target point along the approximate corrected normal vector in the positive and negative directions, and determine the center coordinates of the upper and lower surfaces of the cylinder;

[0107] The cylinder neighborhood is determined based on the preset cylinder parameters and the center coordinates.

[0108] The method for approximating the normal direction of the target point cloud based on the normal vector characteristics of the neighborhood point set and the weighted Euclidean distance of the point cloud is as follows:

[0109] The normal vector n of the target point is estimated using the PCA algorithm.i and neighboring point p j normal vector n j ; Calculate each neighboring point p j and p i Direction vector v j Find v respectively j With n i and n j The included angles θ and α are considered, and points whose angle difference satisfies a certain threshold are all regarded as p. i The criteria for determining redundant points are as follows.

[0110]

[0111] The method for estimating the circular domain of redundant point sets based on the normal vector characteristics of neighboring point sets is as follows:

[0112] Based on the point set within the redundant neighborhood of the point cloud, the target point p i normal vector n i Make corrections;

[0113] To calculate the projection points of the target point along the approximate normal in the positive and negative directions, first normalize the normal direction vector of the target point, and calculate the projection distance based on the height H of the cylinder, thereby obtaining the coordinates of the center of the upper and lower surfaces of the cylinder.

[0114] Find the point cloud set based on the parameters used to build the cylinder. Point set located inside a cylinder The target point is located at the center of mass of the cylinder, and the circles on the upper and lower surfaces have the same normal vector as the target point. calculate The direction vector of each point in the target point Distance d from the set j , and d j Then, calculate the angle θ between the direction vector and the target point's normal vector. Based on the set cylinder radius R and height H, when the distance of the point cloud in the target point's normal and tangential directions is less than R and H respectively, it is determined to be a point inside the cylinder. The determination formula is as follows.

[0115]

[0116] Furthermore, the step of optimizing and iterating the target point based on the cylindrical region to obtain the fused point cloud includes:

[0117] Determine redundant reference points;

[0118] The target point is optimized and iterated based on the redundant reference points to obtain the fusion point position;

[0119] The fusion point cloud is obtained based on the location of the fusion point.

[0120] Specifically, redundant reference points in the neighborhood are determined, the direction vectors of each point in the neighborhood set of the cylinder and the target point are calculated, and the angle between the direction vector and the normal vector of the target point is calculated. The neighborhood point with the smallest angle is used as the redundant reference point of the target point in the corresponding neighborhood.

[0121] Sliding iterations are performed in the direction between the target point and the reference point to find the point with the minimum cumulative error of redundant point iterations, which is then used as the fusion point.

[0122] The specific method for determining redundant reference points in the neighborhood is as follows:

[0123] Determine the nearest reference point of the target point within the redundant neighborhood, and calculate the target point p. i With neighboring point p j Direction vector v ij Find v ij In p i and p j Forward projection length d in the normal direction i With d j ,|d i -d j The smallest neighboring point is the reference point p of the target point. c ;

[0124]

[0125] Where, p i For the target point, p j For neighborhood points, v ij Let d be the direction vector. i With d j v ij In p i and p j The forward projection length, p, in the normal direction c This is a redundant reference point.

[0126] Find the direction vectors between the reference point and the target point, and slide the target point along the direction vectors iteratively. When the sum of the vector angle differences between the redundant point set and the non-redundant points in the neighborhood is minimized, it indicates that the current iteration position has similar spatial distribution characteristics to the neighborhood position. The position of this point is taken as the new point after fusion.

[0127] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0128] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for real-time destacking and fusion of point clouds in multi-line structured light 3D reconstruction frames, characterized in that, include: Acquire the laser point cloud data of the target to be reconstructed and preprocess it to obtain the preprocessed laser point cloud data; An initial network structure is established based on the three-dimensional coordinates of the first frame of the laser point cloud data. Using the point clouds of subsequent frames in the laser point cloud data, point cloud classification and network structure expansion are performed on the initial network structure to obtain the final network structure, and the common grid and the index of the frame point cloud and the total point cloud in the common grid are determined according to the final network structure; Establish a corresponding neighborhood network structure based on the target points in the public grid; Based on the corresponding neighborhood network structure, the central grid of the target point and its eight neighboring grid point clouds in the x, y, and z axis directions are searched to form the neighborhood point set of the target point. The normal vector is estimated based on the neighborhood point set and principal component analysis, and the bidirectionality of the normal is eliminated to obtain the optimized normal vector. Based on the optimized normal vector, the redundant points of the target point are determined and the redundant points are removed to obtain the laser point cloud data after removal. A cylindrical neighborhood of the target point is established based on the removed laser point cloud data. The target point is optimized and iterated based on the neighborhood of the cylinder to obtain a fused point cloud; The process of determining redundant points of the target point based on the optimized normal vector and removing redundant points to obtain the removed laser point cloud data includes: The corresponding direction vector is calculated based on the optimized normal vector and the normal vector corresponding to the preset point in the neighborhood point set. The angle between the direction vector and the optimized normal vector and the normal vector corresponding to the preset point is calculated to obtain the first angle and the second angle. Redundant points are determined and removed based on the angle difference between the first included angle and the second included angle to obtain the removed laser point cloud data.

2. The method for real-time destacking and fusion of point clouds in multi-line structured light 3D reconstruction frames according to claim 1, characterized in that, The step of establishing an initial network structure based on the three-dimensional coordinates of the first frame of the laser point cloud data includes: Based on the three-dimensional coordinates of the first frame point cloud, calculate the maximum and minimum coordinates of the first frame point cloud to obtain the spatial bounding box of the first frame point cloud. Calculate the average point spacing of the point cloud; Calculate the grid side length based on the average point spacing; The initial network structure is obtained by numbering the center coordinates of the scalable mesh side length and the bounding box.

3. The method for real-time destacking and fusion of point clouds in multi-line structured light 3D reconstruction frames according to claim 1, characterized in that, The step of using subsequent frame point clouds from the laser point cloud data to perform point cloud classification and network structure expansion on the initial network structure to obtain the final network structure, and determining the common grid based on the final network structure, includes: Calculate the minimum Euclidean distance from the point cloud of subsequent frames to the center point of the initial network structure; Based on the minimum Euclidean distance of the center point, determine whether the point cloud of the subsequent frame exceeds the range of the initial network structure, and obtain the judgment result. If the judgment result is yes, then the current corresponding point cloud is stored. If the judgment result is no, then the network structure of the current point cloud is expanded to obtain a common grid.

4. The method for real-time destacking and fusion of point clouds in multi-line structured light 3D reconstruction frames according to claim 1, characterized in that, The process of estimating the normal vector of the target point based on the neighborhood point set and principal component analysis, and then eliminating the bidirectionality of the normal vector to obtain the optimized normal vector includes: Calculate the centroid of the preset points in the neighborhood point set based on the target point; Construct a corresponding covariance matrix based on the centroid and preset points in the neighborhood point set; The covariance matrix is ​​decomposed using singular values ​​to obtain eigenvalues ​​and eigenvectors; The normal vector of the target point is obtained based on the eigenvalues ​​and eigenvectors. The bidirectionality of the normal vector of the target point is eliminated to obtain the optimized normal vector.

5. The method for real-time destacking and fusion of point clouds in multi-line structured light 3D reconstruction frames according to claim 1, characterized in that, The step of establishing a cylindrical neighborhood of the target point based on the removed laser point cloud data includes: The normal vector of the target point in the removed laser point cloud data is corrected by using the redundant point set corresponding to the neighborhood point set to obtain the corrected normal vector. Calculate the projection points of the target point along the approximate corrected normal vector in the positive and negative directions, and determine the center coordinates of the upper and lower surfaces of the cylinder; The cylinder neighborhood is determined based on the preset cylinder parameters and the center coordinates.

6. The method for real-time destacking and fusion of point clouds in multi-line structured light 3D reconstruction frames according to claim 1, characterized in that, The step of optimizing and iterating the target point based on the neighborhood of the cylinder to obtain the fused point cloud includes: Determine redundant reference points; The target point is optimized and iterated based on the redundant reference points to obtain the fusion point position; The fusion point cloud is obtained based on the location of the fusion point.

7. The method for real-time destacking and fusion of point clouds in multi-line structured light 3D reconstruction frames according to claim 6, characterized in that, The expression for the redundant reference point is: ； in, For the target point, For neighboring points, It is a direction vector. and They are respectively exist and The forward projection length along the normal direction. As a redundant reference point, for and The direction vector.

Images