An indoor point cloud wall surface segmentation method based on density constraint super voxel

Abstract

The application discloses an indoor point cloud wall surface segmentation method based on density constraint super voxel, and belongs to the technical field of point cloud data processing and three-dimensional space data analysis. The method comprises the following steps: collecting point clouds and eliminating invalid points, dividing structural point clouds by the included angle between normal vectors and the direction of gravity, obtaining a preliminary wall point cloud set through joint constraint of gridding, density constraint and height constraint, adopting an energy function to optimize super voxel segmentation, completing initialization, seed screening and adjacent relationship construction, taking the minimum curvature super voxel as a seed to iteratively grow and traverse and screen to obtain a super voxel cluster, and finally performing binary mask inflation, label updating and point cloud elimination to output complete wall surface point clouds. The method can effectively suppress missegmentation and missing segmentation in a complex indoor scene, and is suitable for point cloud wall segmentation of multi-wall and multi-scale structures.

Description

Technical Field

[0001] This invention relates to the field of point cloud segmentation technology, specifically to a method for indoor point cloud wall segmentation based on density-constrained hypervoxels. Background Technology

[0002] In indoor environments, point cloud data can record the geometric shape and spatial distribution characteristics of building components relatively completely, providing a direct and reliable data foundation for real-scene 3D modeling and indoor structure reconstruction. However, due to the interweaving of various components in 3D space and the influence of light reflection, laser scanning is easily affected by occlusion and noise interference when collecting point clouds in indoor scenes. This results in indoor point clouds exhibiting characteristics such as uneven density, local missing parts, and high noise levels, posing a significant challenge to indoor point cloud segmentation.

[0003] Existing point cloud segmentation methods typically employ model-matching, region growing, feature clustering, and deep learning-based approaches to extract key structural surfaces such as walls, roofs, and floors. Model-fitting methods iteratively fit a pre-defined geometric model to match and segment point clouds. While effective in regular structural scenes, this method is inefficient in real-world indoor environments. Feature clustering segments points with similar features into the same object, but uneven density in indoor environments often hinders stable operation. Deep learning-based methods require large amounts of labeled datasets for training, and their segmentation performance significantly degrades in real-world indoor point cloud tests. In contrast, while region growing offers higher accuracy and is easier to implement, it is susceptible to structural incompleteness in non-Manhattan indoor scenes, making it difficult to maintain wall continuity. Summary of the Invention

[0004] This invention provides a method for indoor point cloud wall segmentation based on density-constrained supervoxels, in order to solve the technical problems in the prior art where indoor point cloud wall segmentation is easily affected by point cloud noise, missing wall point cloud data, and interference from adjacent factors on both sides of the wall.

[0005] A method for indoor point cloud wall segmentation based on density-constrained hypervoxels includes the following steps:

[0006] S1. Collect wall point clouds and remove invalid points, and divide them into vertical structural point clouds and horizontal structural point clouds according to the angle between the normal vector direction and the gravity direction;

[0007] S2. Rasterize the vertical structure point cloud into a horizontal plane to construct a raster density map and a raster height map. Smooth the raster density map and the raster height map respectively, and apply density constraints and raster constraints. Back-project the inner point cloud of the raster that meets the density constraints and height constraints to obtain a preliminary wall point cloud set.

[0008] S3. Using the initial wall point cloud set as the processing object, construct supervoxels by preserving the boundary of the energy function optimization; generate a set of supervoxels with optimization boundary by defining and minimizing the energy function that describes the structural characteristics of the supervoxels.

[0009] S4. Perform initialization, seed selection, and adjacency relationship construction on the supervoxel point cloud of the supervoxel set to obtain the supervoxel adjacency set;

[0010] S5. Initialize all supervoxes in the supervox adjacency set to an unvisited state. Use the unvisited supervoxe with the smallest curvature as the growth seed. Traverse the adjacent supervoxes based on the adjacency relationship. After filtering by normal consistency geometric constraints and curvature similarity geometric constraints, merge them into the current growth cluster. Select the supervoxe with the smallest curvature from the remaining unvisited supervoxes as the new seed. Iterate the seed selection, traversal filtering and merging operations until all supervoxes have been traversed, and obtain the supervoxe cluster with corresponding independent walls and consistent geometric features.

[0011] S6. Based on the wall point cloud corresponding to the supervoxel cluster, recover the supervoxel labels of each 3D point and project them onto the 2D plane. After generating the label raster, construct a binary mask and perform morphological dilation. Update the labels in the dilated binary mask and remove the uncovered point cloud to obtain the complete wall point cloud.

[0012] Furthermore, the angle between the direction of the normal vector and the direction of gravity includes:

[0013] ;

[0014] In the formula, For the first The angle between the normal vector of a point cloud and the direction of gravity. For the first The normal vector of a point cloud. The direction vector of gravity;

[0015] Set the vertical structure angle threshold Threshold of the angle between the horizontal structure and the horizontal structure , For vertical point clouds, This represents a horizontally structured point cloud.

[0016] Furthermore, the vertical structure point cloud is rasterized in a horizontal plane, including rasterizing the vertical structure point cloud in a horizontal plane, with the grid side length set to... Two-dimensional raster coordinates of vertical structure point cloud :

[0017] ;

[0018] ;

[0019] In the formula, For the first The grid coordinates of a point cloud in the horizontal plane X direction. For the first The coordinates of a point cloud in the horizontal plane X direction. The minimum coordinate in the X direction for all vertical structure point clouds. For the first The grid coordinates of a point cloud in the Y direction of the horizontal plane. For the first The coordinates of a point cloud in the Y direction of the horizontal plane. The minimum value of the Y-axis coordinates of all vertical structure point clouds. This represents the side length of the grid.

[0020] Furthermore, smoothing is applied to the raster density map and raster height map respectively, and density constraints and raster constraints are applied, including based on the two-dimensional raster coordinates of each point cloud. Statistical analysis of two-dimensional grid coordinates The number of inset vertical structure point clouds is used to construct a raster density map. ,right Applying a two-dimensional mean filter to reduce noise yields a smoothed density map. The following wall density constraint formula is used to select grids that meet the density requirements:

[0021] ;

[0022] In the formula, For raster density filtering, In the smoothed raster density map, the coordinates are... The number of point clouds corresponding to the raster. Density threshold;

[0023] Based on the coordinates of each point cloud's two-dimensional raster. Statistical analysis of two-dimensional grid coordinates The number of vertically structured point clouds falling in is used to construct a raster height map. The maximum height value of all point clouds within each grid cell is used as the height value of the corresponding grid cell; for the grid height map Applying maximum value filtering to smooth the image yields the filtered height map. The following wall height constraint formula is used to filter grids that meet the height requirements:

[0024] ;

[0025] In the formula, Filtering identifier for grid height, In the smoothed grid height map, the coordinates are... The point cloud height corresponding to the raster. Ceiling height For use in filtering out low-profile structures.

[0026] Furthermore, the point cloud within the raster that meets the density and height constraints is back-projected to obtain a preliminary wall point cloud set, including wall raster determination through the joint constraint of density and height constraints. The joint constraint is calculated by the following formula:

[0027] ;

[0028] In the formula, For the final filtering identifier of the raster, For raster density filtering, Use grid height as the filter identifier;

[0029] Will satisfy The point cloud within the raster is back-projected into the three-dimensional coordinate space to obtain a preliminary wall point cloud set.

[0030] Furthermore, the energy function includes a difference term and a constraint term, and the energy function is calculated using the following formula:

[0031] ;

[0032] ;

[0033] ;

[0034] In the formula, For point clouds With point clouds Difference value, and The first step in the preliminary wall point cloud The and the first A three-dimensional point, and for and The normal vector, Represents the distance between neighboring points. This represents the statistical value of the number of supervoxels. This is the set of results from supervoxel segmentation. To segment the relation matrix, For indicator functions, This represents the total number of 3D points in the initial wall point cloud that participate in energy function optimization. The optimization objective is to minimize the energy function. The energy function for point cloud segmentation. To preset the number of hypervoxels to be generated, These are the weighting coefficients.

[0035] Furthermore, initialization includes dividing the point cloud into several initial regions based on the color features of the supervoxels, and calculating the centroid for each supervoxel. Average normal vector curvature and the number of point clouds Complete the extraction of core geometric features, set all supervoxels to an unvisited initial state, and aggregate the region growing results. Leave blank;

[0036] Seed selection involves sorting hypervoxels in ascending order of curvature value and selecting the hypervoxel that is currently unvisited and has the smallest curvature. This serves as the initial seed for region growth; if no unvisited supervoxel is found after traversal, the algorithm terminates, and supervoxels that meet the visit criteria are added to the growth queue. And simultaneously mark the status of the hypervoxel as visited;

[0037] Adjacency relationship construction, including using the mass center of the supervoxel. As a spatial search benchmark, a k-nearest neighbor search is performed on each supervoxel to obtain the corresponding local candidate neighborhood. The 3D point coordinates of all supervoxels within the candidate neighborhoods are collected. Based on these point coordinates, a density-based noisy spatial clustering operation is performed to determine the principal local cluster to which the current supervoxel belongs. All supervoxels belonging to the same principal local cluster but not the current supervoxel are considered as the true neighboring voxels of the current supervoxel, generating a supervoxel adjacency set. .

[0038] Furthermore, the normal consistency geometric constraint and the curvature similarity geometric constraint include:

[0039] ;

[0040] ;

[0041] In the formula, This represents the absolute value of the angle between the normal vectors of the current supervoxel and its neighboring supervoxels. The average normal vector of the current supervoxel. Let be the average normal vector of the neighboring supervoxels of the current supervoxel. This represents the absolute difference in curvature between the current supervoxel and its neighboring supervoxels. The curvature value of the current supervoxel. The curvature value of the adjacent supervoxels of the current supervoxel;

[0042] Set the threshold for the angle between normal vectors Threshold for difference between curvature value and , will simultaneously satisfy and The adjacent supervoxels are marked as visited and added to the growth queue, and are also incorporated into the current growth cluster.

[0043] Furthermore, based on the wall point cloud corresponding to the supervoxel clusters, the supervoxel labels of each 3D point are recovered and projected onto a 2D plane. After generating a label raster, a binary mask is constructed and morphological dilation is performed, including the 3D points. In the two-dimensional plane coordinates of the projection for:

[0044] ;

[0045] In the formula, Adding clouds to the wall Three-dimensional points Two-dimensional coordinates after projection onto a two-dimensional plane. and The first Three-dimensional points The coordinate values ​​corresponding to the two coordinate axes projected onto a two-dimensional plane, and the combination of coordinate axes satisfies ;

[0046] A grid is constructed based on the two-dimensional planar coordinates of the projection and the grid grid side length is used. Spatially discretize the two-dimensional projection points of each point in the wall point cloud to obtain a two-dimensional label raster. :

[0047] ;

[0048] In the formula, For the first The two-dimensional coordinates of a three-dimensional point projected onto a two-dimensional plane The supervoxel label of the grid location; Adding clouds to the wall A 3D point hypervoxel label;

[0049] Count the number of point clouds contained in the supervoxels and set a threshold for large areas. Construct a set of large faces, which is:

[0050] ;

[0051] In the formula, It is a collection of large pieces of surface. For the first The number of point clouds corresponding to a large area of ​​a supervoxel. For large area thresholds;

[0052] When the number of point clouds corresponding to a large area of ​​a supervoxel exceeds At that time, for large-area sets Each large facet in Constructing in a two-dimensional label grid Corresponding binary mask :

[0053] ;

[0054] In the formula, For the first A large piece of dough The corresponding binary mask, For two-dimensional label grids at grid positions Supervoxel tags at the location, For the first A large, flat hypervoxel label;

[0055] right Perform morphological dilation operation. The expansion radius is determined by the number of iterations. Controlling the process to obtain the expanded binary mask :

[0056] ;

[0057] In the formula, For the first A large piece of dough The corresponding dilated binary mask, For the first A large piece of dough The corresponding binary mask, This represents the number of iterations.

[0058] Furthermore, the labels within the expanded binary mask are updated, and the uncovered point cloud is removed to obtain the complete wall point cloud, including the first... A large piece of dough Corresponding dilated binary mask The covered area is considered as the outer extension of the wall structure in two-dimensional space, and any three-dimensional point in the wall point cloud is considered as such. Get , Falling into A large piece of dough Corresponding dilated binary mask Within the covered area, three-dimensional points The hypervoxel label has been updated to:

[0059] ;

[0060] In the formula, Adding clouds to the wall Three-dimensional points Updated hypervoxel labels;

[0061] 3D points in wall point cloud of Not by any If covered, then three-dimensional points are considered. 3D points not belonging to the wall structure are directly removed. After updating the labels and removing the uncovered point clouds from the wall point cloud, a complete wall point cloud is obtained.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] This invention employs hypervoxels as the core processing unit, combining density constraints and morphological dilation optimization to effectively suppress point cloud noise interference. Simultaneously, it fills in missing areas of wall data and distinguishes adjacent wall boundaries, avoiding segmentation interruptions and region mismerging issues caused by local anomalies or structural proximity in existing technologies. This makes it suitable for segmentation needs of complex interior structures. Through energy function optimization, this invention achieves precise definition of hypervoxel boundaries. Combined with a region growth strategy constrained by normal consistency and curvature similarity, it can accurately divide independent wall regions, reducing segmentation fragments and ensuring clear wall boundaries and complete morphology. Compared to existing technologies, it significantly reduces segmentation errors and improves the accuracy and completeness of wall segmentation. This invention forms a complete closed-loop process of "point cloud screening - hypervoxel construction - cluster growth - refined repair," requiring minimal manual parameter adjustment. It can automatically complete invalid point removal, hypervoxel clustering, and wall repair, adapting to wall point cloud data in different interior scenarios. It is convenient to operate, highly efficient, and easier to apply to real-world engineering scenarios. It has been specifically optimized to address typical indoor point cloud problems such as adjacent walls on both sides, missing data, and noise interference. It can not only handle conventional indoor wall segmentation, but also adapt to indoor scenes with complex structures and many interference factors, making it more applicable to a wider range of scenarios compared to traditional methods. Attached Figure Description

[0064] Figure 1 This is a flowchart illustrating the technical process of the present invention.

[0065] Figure 2 This is a flowchart of a raster filtering process based on raster density and raster height constraints.

[0066] Figure 3 This is a flowchart of the region growth process based on boundary-preserving supervoxels. Detailed Implementation

[0067] To enable those skilled in the art to better understand the technical solutions in this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this disclosure.

[0068] like Figure 1 As shown, an indoor point cloud wall segmentation method based on density-constrained hypervoxels includes the following steps:

[0069] S1. Collect wall point clouds. For the effective wall point clouds obtained after removing invalid points, calculate the normal vector direction of each effective point using the principal component analysis method. Divide the effective point clouds into vertical structural point clouds and horizontal structural point clouds according to the angle between the normal vector direction and the gravity direction.

[0070] The angle between the direction of the normal vector and the direction of gravity is calculated using the following formula:

[0071] ;

[0072] In the formula, For the first The angle between the normal vector of a point cloud and the direction of gravity. For the first The normal vector of a point cloud. The direction vector of gravity;

[0073] Set the vertical structure angle threshold Threshold of the angle between the horizontal structure and the horizontal structure , This is a vertical point cloud. This results in a horizontally structured point cloud.

[0074] S2. For example Figure 2 As shown, the filtered vertical post-node point cloud is rasterized on a horizontal plane, with the grid side length set to 1. Coordinates of a 2D raster of a vertically structured point cloud :

[0075] ;

[0076] ;

[0077] In the formula, For the first The grid coordinates of a point cloud in the horizontal plane X direction. For the first The coordinates of a point cloud in the horizontal plane X direction. The minimum coordinate in the X direction for all vertical structure point clouds. For the first The grid coordinates of a point cloud in the Y direction of the horizontal plane. For the first The coordinates of a point cloud in the Y direction of the horizontal plane. The minimum value of the Y-axis coordinates of all vertical structure point clouds. This represents the side length of the grid.

[0078] Based on the two-dimensional raster coordinates of each point cloud Statistical analysis of two-dimensional grid coordinates The number of inset vertical structure point clouds is used to construct a raster density map. ,right Applying a two-dimensional mean filter to reduce noise yields a smoothed density map. The following wall density constraint formula is used to select grids that meet the density requirements:

[0079] ;

[0080] In the formula, For raster density filtering, In the smoothed raster density map, the coordinates are... The number of point clouds corresponding to the raster. Density threshold;

[0081] Based on the two-dimensional raster coordinates of each point cloud Statistical analysis of two-dimensional grid coordinates The number of vertically structured point clouds falling in is used to construct a raster height map. The maximum height value of all point clouds within each grid cell is used as the height value of the corresponding grid cell; for the grid height map Applying maximum value filtering to smooth the image yields the filtered height map. The following wall height constraint formula is used to filter grids that meet the height requirements:

[0082] ;

[0083] In the formula, Filtering identifier for grid height, In the smoothed grid height map, the coordinates are... The point cloud height corresponding to the raster. Ceiling height Used to filter out low-profile structures;

[0084] The wall grid is then determined by the combined constraints of density and height, calculated using the following formula:

[0085] ;

[0086] In the formula, For the final filtering identifier of the raster, For raster density filtering, Use grid height as the filter identifier;

[0087] Will satisfy The point cloud within the raster is back-projected into the three-dimensional coordinate space to obtain a preliminary wall point cloud set.

[0088] S3. Taking the initial wall point cloud set as the processing object, the structural characteristics of the super voxels are described by an explicitly defined energy function, and the energy function is minimized by selecting a subset of the point cloud, thereby generating super voxels with optimized boundary definitions.

[0089] The energy function includes a difference term and a constraint term, and is calculated using the following formula:

[0090] ;

[0091] ;

[0092] ;

[0093] In the formula, For point clouds With point clouds Difference value, and The first step in the preliminary wall point cloud The and the first A three-dimensional point, and for and The normal vector, Represents the distance between neighboring points. This represents the statistical value of the number of supervoxels. This is the set of results from supervoxel segmentation. To segment the relation matrix, For indicator functions, This represents the total number of 3D points in the initial wall point cloud that participate in energy function optimization. The optimization objective is to minimize the energy function. The energy function for point cloud segmentation. To preset the number of hypervoxels to be generated, These are the weighting coefficients.

[0094] S4. For example Figure 3As shown, the input is a supervoxel point cloud segmented using the Boundary Preserving Supervoxel Segmentation (BPSS) method. First, the point cloud is divided into several regions based on the supervoxel color. Then, features, including the centroid, are calculated for each supervoxel. Average normal vector curvature and the number of point clouds Set all supervoxels to an unvisited initial state and set the region growth results together. Set to empty; then perform seed filtering, which involves sorting hypervoxels in ascending order of curvature value and selecting the hypervoxel that is currently unvisited and has the smallest curvature. This serves as the initial seed for region growth; if no unvisited supervoxel is found after traversal, the algorithm terminates, and supervoxels that meet the visit criteria are added to the growth queue. Simultaneously, the state of the supervoxel is marked as visited, and then adjacency relationships are constructed. The adjacency relationship construction includes the centroid of the supervoxel. As a spatial search benchmark, k-nearest neighbor search is performed on each supervoxel to obtain the corresponding local candidate neighborhood. The 3D point coordinates of all supervoxels within the candidate neighborhoods are collected. Based on these point coordinates, density-based noisy spatial clustering (DBSCAN) is performed to determine the principal local cluster to which the current supervoxel belongs. All supervoxels belonging to the same principal local cluster but not the current supervoxel are considered as the true neighboring voxels of the current supervoxel, generating a supervoxel adjacency set. Obtain the adjacency set It can effectively avoid the erroneous connectivity between parallel walls caused by the traditional k-nearest neighbor method.

[0095] S5. Initialize all supervoxels in the supervoxel adjacency set to an unvisited state. Select the supervoxel with the smallest curvature as the growth seed and add it to the growth queue. Sequentially retrieve the current supervoxel from the growth queue, traverse the adjacent supervoxels, and make judgments based on the geometric constraints of normal consistency and curvature similarity between the current supervoxel and its adjacent supervoxels.

[0096] Normal consistency geometric constraints and curvature similarity geometric constraints include:

[0097] ;

[0098] ;

[0099] In the formula, This represents the absolute value of the angle between the normal vectors of the current supervoxel and its neighboring supervoxels. The average normal vector of the current supervoxel. Let be the average normal vector of the neighboring supervoxels of the current supervoxel. This represents the absolute difference in curvature between the current supervoxel and its neighboring supervoxels. The curvature value of the current supervoxel. The curvature value of the adjacent supervoxels of the current supervoxel;

[0100] Set the threshold for the angle between normal vectors Threshold for difference between curvature value and , will simultaneously satisfy and The adjacent supervoxels are marked as visited and added to the growth queue, and simultaneously merged into the current growth cluster. The current growth cluster is added to the global result set. From the remaining unvisited supervoxels, the unvisited supervoxel with the smallest curvature is selected as the new growth seed. The seed selection, traversal filtering and merging operations are iteratively executed until all supervoxels have been traversed, and finally a wall structure region with consistent geometric features is obtained.

[0101] S6. After the supervoxel region growth and clustering in S5, wall structure regions with consistent geometric features are obtained. However, the wall edges and corner structures are prone to fluctuations due to local geometric features. Affected by the dual geometric constraints of normal consistency and curvature similarity, the structural protrusions are prone to forming fragmented small voxels that cannot be merged, which destroys the overall continuity of the wall and affects the integrity of the wall. Therefore, the supervoxel point cloud after growth and clustering is subjected to fine repair and integrity optimization processing: the supervoxel labels of each three-dimensional point are restored according to the RGB encoding of the input point cloud, and the point cloud is projected onto a two-dimensional plane.

[0102] Let three-dimensional points In the two-dimensional plane coordinates of the projection for:

[0103] ;

[0104] In the formula, Adding clouds to the wall Three-dimensional points Two-dimensional coordinates after projection onto a two-dimensional plane. and The first Three-dimensional points The coordinate values ​​corresponding to the two coordinate axes projected onto a two-dimensional plane, and the combination of coordinate axes satisfies ;

[0105] A grid is constructed based on the two-dimensional planar coordinates of the projection and the grid grid side length is used. Spatially discretize the two-dimensional projection points of each point in the wall point cloud to obtain a two-dimensional label raster. :

[0106] ;

[0107] In the formula, For the first The two-dimensional coordinates of a three-dimensional point projected onto a two-dimensional plane The supervoxel label of the grid location; Adding clouds to the wall A 3D point hypervoxel label;

[0108] To avoid fragmented hypervoxels dominating subsequent processing, the number of points for each hypervoxel is counted, and a set of large patches is constructed:

[0109] ;

[0110] In the formula, It is a collection of large pieces of surface. For the first The number of point clouds corresponding to a large area of ​​a supervoxel. For large area thresholds;

[0111] When the number of point clouds corresponding to a large area of ​​a supervoxel exceeds At that time, it was assumed that the region corresponding to the supervoxel had sufficient geometric stability in the wall structure. Since the wall point cloud would exhibit local fractures in three-dimensional space, connectivity constraints needed to be applied to large patches in two-dimensional grid space for each large patch. Constructing in a two-dimensional label grid Corresponding binary mask :

[0112] ;

[0113] In the formula, For the first A large piece of dough The corresponding binary mask, For two-dimensional label grids at grid positions Supervoxel tags at the location, For the first A large, flat hypervoxel label;

[0114] right Perform morphological dilation operation. The expansion radius is determined by the number of iterations. Controlling the process to obtain the expanded binary mask The dilated binary mask :

[0115] ;

[0116] In the formula, For the first A large piece of dough The corresponding dilated binary mask, For the first A large piece of dough The corresponding binary mask, This represents the number of iterations.

[0117] The first A large piece of dough Corresponding dilated binary mask The covered area is considered as the outer extension of the wall structure in two-dimensional space, and any three-dimensional point in the wall point cloud is considered as such. Get , Falling into A large piece of dough Corresponding dilated binary mask Within the covered area, three-dimensional points The hypervoxel label has been updated to:

[0118] ;

[0119] In the formula, Adding clouds to the wall Three-dimensional points Updated hypervoxel labels;

[0120] 3D points in wall point cloud of Not by any If covered, then three-dimensional points are considered. 3D points not belonging to the wall structure are directly removed. After updating the labels and removing the uncovered point clouds from the wall point cloud, a complete wall point cloud is obtained.

[0121] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for indoor point cloud wall segmentation based on density-constrained hypervoxels, characterized in that, Includes the following steps: S1. Collect wall point clouds and remove invalid points, and divide them into vertical structural point clouds and horizontal structural point clouds according to the angle between the normal vector direction and the gravity direction; S2. Rasterize the vertical structure point cloud into a horizontal plane, construct a raster density map and a raster height map, smooth the raster density map and the raster height map respectively, and apply density constraints and raster constraints. Back-project the point cloud in the raster that meets the density constraints and height constraints to obtain a preliminary wall point cloud set. S3. Using the initial wall point cloud set as the processing object, the boundary is optimized by energy function, and the supervoxel segmentation method is used to construct supervoxels; By defining and minimizing an energy function that describes the structural properties of a supervoxel, a set of supervoxels with an optimization boundary is generated. S4. Perform initialization, seed selection, and adjacency relationship construction on the supervoxel point cloud of the supervoxel set to obtain the supervoxel adjacency set; S5. Initialize all supervoxes in the supervox adjacency set to an unvisited state. Use the unvisited supervoxe with the smallest curvature as the growth seed. Traverse the adjacent supervoxes based on the adjacency relationship. After filtering by normal consistency geometric constraints and curvature similarity geometric constraints, merge them into the current growth cluster. Select the supervoxe with the smallest curvature from the remaining unvisited supervoxes as the new seed. Iterate the seed selection, traversal filtering and merging operations until all supervoxes have been traversed, and obtain the supervoxe cluster with corresponding independent walls and consistent geometric features. S6. Based on the wall point cloud corresponding to the supervoxel cluster, recover the supervoxel labels of each 3D point and project them onto the 2D plane. After generating the label raster, construct a binary mask and perform morphological dilation. Update the labels in the dilated binary mask and remove the uncovered point cloud to obtain the complete wall point cloud.

2. The method for indoor point cloud wall segmentation based on density-constrained hypervoxels according to claim 1, characterized in that, The angle between the direction of the normal vector and the direction of gravity includes: ； In the formula, For the first The angle between the normal vector of a point cloud and the direction of gravity. For the first The normal vector of a point cloud. The direction vector of gravity; Set the vertical structure angle threshold Threshold of the angle between the horizontal structure and the horizontal structure , For vertical point clouds, This represents a horizontally structured point cloud.

3. The method for indoor point cloud wall segmentation based on density-constrained hypervoxels according to claim 1, characterized in that, The vertical structure point cloud is rasterized into a horizontal plane, which includes rasterizing the vertical structure point cloud onto a horizontal plane, with the grid side length set to 1. Two-dimensional raster coordinates of vertical structure point cloud : ； ； In the formula, For the first The grid coordinates of a point cloud in the horizontal plane X direction. For the first The coordinates of a point cloud in the horizontal plane X direction. The minimum coordinate in the X direction for all vertical structure point clouds. For the first The grid coordinates of a point cloud in the Y direction of the horizontal plane. For the first The coordinates of a point cloud in the Y direction of the horizontal plane. The minimum value of the Y-axis coordinates of all vertical structure point clouds. This represents the side length of the grid.

4. The method for indoor point cloud wall segmentation based on density-constrained hypervoxels according to claim 1, characterized in that, Smoothing is applied to both the raster density map and the raster height map, and density constraints and raster constraints are applied, including based on the two-dimensional raster coordinates of each point cloud. Statistical analysis of two-dimensional grid coordinates The number of inset vertical structure point clouds is used to construct a raster density map. ,right Applying a two-dimensional mean filter to reduce noise yields a smoothed density map. The following wall density constraint formula is used to select grids that meet the density requirements: ； In the formula, For raster density filtering, In the smoothed raster density map, the coordinates are... The number of point clouds corresponding to the raster. Density threshold; Based on the coordinates of each point cloud's two-dimensional raster. Statistical analysis of two-dimensional grid coordinates The number of vertically structured point clouds falling in is used to construct a raster height map. The maximum height value of all point clouds within each grid cell is used as the height value of the corresponding grid cell; for the grid height map Applying maximum value filtering to smooth the image yields the filtered height map. The following wall height constraint formula is used to filter grids that meet the height requirements: ； In the formula, Filtering identifier for grid height, In the smoothed grid height map, the coordinates are... The point cloud height corresponding to the raster. Ceiling height For use in filtering out low-profile structures.

5. The method for indoor point cloud wall segmentation based on density-constrained hypervoxels according to claim 1, characterized in that, Back-projecting the point cloud within the raster that meets both density and height constraints yields a preliminary wall point cloud set. This includes determining the wall raster based on the joint constraints of density and height constraints, calculated using the following formula: ； In the formula, For the final filtering identifier of the raster, For raster density filtering, Use grid height as the filter identifier; Will satisfy The point cloud within the raster is back-projected into the three-dimensional coordinate space to obtain a preliminary wall point cloud set.

6. The method for indoor point cloud wall segmentation based on density-constrained hypervoxels according to claim 1, characterized in that, The energy function includes a difference term and a constraint term, and is calculated using the following formula: ； ； ； In the formula, For point clouds With point clouds Difference value, and The first step in the preliminary wall point cloud The and the first A three-dimensional point, and for and The normal vector, Represents the distance between neighboring points. Z represents the statistical value of the number of supervoxels, and Z is the result set of supervoxel segmentation. To segment the relation matrix, For indicator functions, This represents the total number of 3D points in the initial wall point cloud that participate in energy function optimization. The optimization objective is to minimize the energy function. The energy function for point cloud segmentation. To preset the number of hypervoxels to be generated, These are the weighting coefficients.

7. The method for indoor point cloud wall segmentation based on density-constrained hypervoxels according to claim 1, characterized in that, Initialization includes dividing the point cloud into several initial regions based on the color features of the supervoxels, and calculating the centroid for each supervoxel. Average normal vector curvature and the number of point clouds Complete the extraction of core geometric features, set all supervoxels to an unvisited initial state, and aggregate the region growing results. Leave blank; Seed selection involves sorting hypervoxels in ascending order of curvature value and selecting the hypervoxel that is currently unvisited and has the smallest curvature. This serves as the initial seed for region growth; if no unvisited supervoxel is found after traversal, the algorithm terminates, and supervoxels that meet the visit criteria are added to the growth queue. And simultaneously mark the status of the hypervoxel as visited; Adjacency relationship construction, including using the mass center of the supervoxel. As a spatial search benchmark, a k-nearest neighbor search is performed on each supervoxel to obtain the corresponding local candidate neighborhood. The 3D point coordinates of all supervoxels within the candidate neighborhoods are collected. Based on these point coordinates, a density-based noisy spatial clustering operation is performed to determine the principal local cluster to which the current supervoxel belongs. All supervoxels belonging to the same principal local cluster but not the current supervoxel are considered as the true neighboring voxels of the current supervoxel, generating a supervoxel adjacency set. .

8. The method for indoor point cloud wall segmentation based on density-constrained hypervoxels according to claim 1, characterized in that, Normal consistency geometric constraints and curvature similarity geometric constraints include: ； ； In the formula, This represents the absolute value of the angle between the normal vectors of the current supervoxel and its neighboring supervoxels. The average normal vector of the current supervoxel. Let be the average normal vector of the neighboring supervoxels of the current supervoxel. This represents the absolute difference in curvature between the current supervoxel and its neighboring supervoxels. The curvature value of the current supervoxel. The curvature value of the adjacent supervoxels of the current supervoxel; Set the threshold for the angle between normal vectors Threshold for difference between curvature value and , will simultaneously satisfy and The adjacent supervoxels are marked as visited and added to the growth queue, and are also incorporated into the current growth cluster.

9. The method for indoor point cloud wall segmentation based on density-constrained hypervoxels according to claim 1, characterized in that, Based on the wall point cloud corresponding to the supervoxel clusters, the supervoxel labels of each 3D point are recovered and projected onto a 2D plane. After generating a label raster, a binary mask is constructed and morphological dilation is performed, including the 3D points. In the two-dimensional plane coordinates of the projection for: ； In the formula, Adding clouds to the wall Three-dimensional points Two-dimensional coordinates after projection onto a two-dimensional plane. and The first Three-dimensional points The coordinate values ​​corresponding to the two coordinate axes projected onto a two-dimensional plane, and the combination of coordinate axes satisfies ; A grid is constructed based on the two-dimensional planar coordinates of the projection and the grid grid side length is used. Spatially discretize the two-dimensional projection points of each point in the wall point cloud to obtain a two-dimensional label raster. : ； In the formula, For the first The two-dimensional coordinates of a three-dimensional point projected onto a two-dimensional plane The supervoxel label of the grid location; Adding clouds to the wall A 3D point hypervoxel label; Count the number of point clouds contained in the supervoxels and set a threshold for large areas. Construct a set of large faces, which is: ； In the formula, It is a collection of large pieces of surface. For the first The number of point clouds corresponding to a large area of ​​a supervoxel. For large area thresholds; When the number of point clouds corresponding to a large area of ​​a supervoxel exceeds At that time, for large-area sets Each large facet in Constructing in a two-dimensional label grid Corresponding binary mask : ； In the formula, For the first A large piece of dough The corresponding binary mask, For two-dimensional label grids at grid positions Supervoxel tags at the location, For the first A large, flat hypervoxel label; right Perform morphological dilation operation. The expansion radius is determined by the number of iterations. Controlling the process to obtain the expanded binary mask : ； In the formula, For the first A large piece of dough The corresponding dilated binary mask, For the first A large piece of dough The corresponding binary mask, This represents the number of iterations.

10. The method for indoor point cloud wall segmentation based on density-constrained hypervoxels according to claim 1, characterized in that, The labels within the expanded binary mask are updated, and the uncovered point cloud is removed to obtain the complete wall point cloud, including the first label. A large piece of dough Corresponding dilated binary mask The covered area is considered as the outer extension of the wall structure in two-dimensional space, and any three-dimensional point in the wall point cloud is considered as such. Get , Falling into A large piece of dough Corresponding dilated binary mask Within the covered area, three-dimensional points The hypervoxel label has been updated to: ； In the formula, Adding clouds to the wall Three-dimensional points Updated hypervoxel labels; 3D points in wall point cloud of Not by any If covered, then three-dimensional points are considered. 3D points not belonging to the wall structure are directly removed. After updating the labels and removing the uncovered point clouds from the wall point cloud, a complete wall point cloud is obtained.

Images