Room space segmentation method based on distance constraint
By filtering, voxelizing, and clustering point cloud data based on distance constraints, and combining this with wavefront growing algorithm, the problem of room segmentation in complex 3D indoor environments was solved, achieving effective 3D spatial segmentation.
Patent Information
- Application Number
- CN202410826854.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-25
- Publication Date
- 2024-11-22
AI Technical Summary
Existing two-dimensional methods are difficult to effectively solve the semantic segmentation problem in complex 3D indoor environments, especially in indoor scenes outside the Manhattan world.
A room space segmentation method based on distance constraints is adopted. By performing statistical filtering and voxel downsampling on the original point cloud data, the spatial occupancy probability grid is calculated. The room segmentation is performed by using Euclidean distance and DBSCAN clustering, combined with the wavefront growing algorithm.
It achieves room segmentation of complex interior scenes in three-dimensional space, avoids over-segmentation, effectively segments non-Manhattan scenes, and provides complete room space segmentation results.
Smart Images

Figure CN119006807A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of indoor point cloud scene reconstruction, and relates to a room space segmentation method based on distance constraints. Background Art
[0002] Room segmentation is an important research topic in the field of indoor robotics and is the basis for robot task planning and navigation. The purpose of the room segmentation task is to automatically and robustly segment indoor 3D point clouds. Current methods mainly project a single layer of point cloud onto a 2D occupancy probability image to achieve 2D room segmentation. However, these methods have difficulty solving the semantic segmentation problem in complex 3D indoor environments, and these methods also have poor segmentation effects in non-Manhattan world indoor scenes. Summary of the invention
[0003] The purpose of the present invention is to provide a room space segmentation method based on distance constraints, which is directly implemented in three-dimensional space and solves the room space segmentation in complex indoor scenes.
[0004] The technical solution adopted by the present invention is a room space segmentation method based on distance constraints, which is specifically implemented according to the following steps:
[0005] Step 1: The original point cloud data is subjected to statistical filtering denoising and voxel downsampling, and then the spatial occupancy probability grid is calculated according to the number of point clouds in the voxel, which is divided into free voxels and occupied voxels;
[0006] Step 2: First, calculate the Euclidean distance between the free voxel and its nearest occupied voxel, and then separate the connected rooms according to the distance threshold. Then use the DBSCAN clustering method to cluster the voxels in the same room according to density and mark them with different colors to obtain the room seed region;
[0007] Step 3: Use the wavefront growing algorithm to traverse the 26-neighborhood of the unlabeled voxel. If there is a labeled voxel in the 26-neighborhood of the voxel, the current voxel is labeled according to the labeled voxel, otherwise, other unlabeled voxels are traversed until all unlabeled voxels are assigned a room semantic label, and the room space segmentation is completed.
[0008] The present invention is also characterized in that:
[0009] Step 1 is as follows:
[0010] Step 1.1, by statistically analyzing the points around each point in the point cloud, determine whether the point is an outlier or a noise point, and perform statistical filtering on the points in the point cloud, then downsample the filtered and denoised data and use it as the input data for the next stage;
[0011] Step 1.2, the process of mapping the filtered and denoised point cloud data from a continuous three-dimensional space to a discrete three-dimensional grid or voxel, and then further analyzing the distribution density of points in each voxel, and evaluating whether each voxel is occupied by counting the proportion of the number of point clouds contained in the voxel to the total number of points.
[0012] Step 1.1 is as follows:
[0013] Step 1.1.1, current point P i (x i ,y i ,z i ) and its neighbor point N m (x m ,y m ,z m ) The average distance D between i , the distance between the two points is S i , whose standard deviation is δ, is calculated as (1) and (2).
[0014]
[0015] Among them, P i (x i ,y i ,z i ) any point selected, N m (x m ,y m ,z m ) is P i Neighbor points;
[0016]
[0017] Among them, S i is the distance between two points, D i is the average distance, δ is the standard deviation;
[0018] Step 1.1.2, if D i >D max , then the point is considered to be an outlier and is removed, otherwise it is considered to be a non-outlier and is retained;
[0019] Step 1.1.3, downsample the retained data using the voxel grid method, set the voxel side length, and represent the scene with three-dimensional voxels. Then calculate the center of gravity of each voxel to represent the voxel and obtain the downsampled result.
[0020] Step 1.2 is as follows:
[0021] Step 1.2.1, according to the ray casting method, multiple beams are emitted from the laser scanning point, the scanning point is the viewpoint, and when a beam encounters an obstacle, it stops. The voxel at the time of collision is the occupied voxel. The space between the viewpoint and the obstacle is called a free voxel;
[0022] Step 1.2.2, using the Bresenham line algorithm, determine which voxels constitute the line between the two points, and mark them according to the collision situation. The marked voxels are divided into three three-dimensional occupancy probability grids: "occupied voxels", "free voxels" and "unknown". The assignment rules are:
[0023]
[0024] Step 2 is as follows:
[0025] Step 2.1, using a transformation based on Euclidean distance, the distance value from the free voxel to its nearest occupied voxel is mapped to the voxel gray value, where the gray value of each grid is equal to its distance to the nearest target point;
[0026] Step 2.2, after the Euclidean distance transformation is completed, a distance threshold is set to determine which points belong to the same room, and then a density-based clustering method is used to cluster the point cloud after threshold segmentation.
[0027] Step 2.1 is as follows:
[0028] Step 2.1.1, given a grid map M, let p and q be the coordinates of any two grids in the grid M, and the Euclidean distance between the two grids is as shown in formula (4);
[0029]
[0030] Among them, the point (x i ,y i ,z i ) is the free grid coordinate, point (x j ,y j ,z j ) is the occupied grid coordinate, and the free grid and the occupied grid constitute the grid map M.
[0031] Step 2.1.2, define the Euclidean distance transformation voxel execution state queue Q, and determine the state of some voxels through the ray casting algorithm. When a cell pops out of the queue, it first needs to check whether it has been processed before. If it has not been processed, it needs to confirm the type of wave it belongs to;
[0032] Step 2.1.3, determine whether it is an ascending wave or a descending wave based on the current voxel state and the voxel occupancy probability value. If the current voxel is a free space voxel and the voxel state is F, change the voxel state to T. If the current voxel state is T and the occupancy probability value has stored "occupied", add the current voxel to the queue Q and update the probability value of the voxel to "free". At this time, the two processes are ascending waves;
[0033] Step 2.1.4, if the current voxel state is F, change the voxel state to T, and add the current voxel to queue Q. If the current voxel state is T, and the occupancy probability value has stored "free", add the current voxel to queue Q, and update the probability value of the voxel to "occupied". At this time, the two processes are descending waves;
[0034] Step 2.1.5, update the circular queue Q, and repeat steps 2.1.1-2.1.4 until the three-dimensional Euclidean distance transformation of all voxels is completed.
[0035] Step 2.2 is as follows:
[0036] Step 2.2.1, set a distance threshold to determine whether the distance from the voxel to the nearest occupied voxel is less than or equal to the set distance threshold. If it is less than or equal to the distance threshold, the point is marked as inactive, indicating that the voxel belongs to the same room; if it is greater than the distance threshold, the point is marked as active, and the state of the part of the distance value minus the distance threshold is set to pending, and the remaining state is set to inactive;
[0037] Step 2.2.2, use the DBSCAN algorithm to cluster the point cloud after threshold segmentation. First, randomly select an unvisited data point p and mark it as visited;
[0038] Step 2.2.3, with p as the center, check whether there are at least MinPts data points in the ε neighborhood of p. If so, mark all these points as visited and form a new cluster;
[0039] Step 2.2.4, if p is not a core object, mark p as a noise point;
[0040] Step 2.2.5, repeat steps 2.2.2-2.2.4 until all data points have been visited.
[0041] Step 3 is as follows:
[0042] Step 3.1, assign the initial room semantic tag ID to the room seed region after density analysis, and mark the voxels contained in each occupied space in the seed region as the corresponding initial room tag ID, and obtain the voxel grid S containing the room seed region information i_lableThen, all the initial voxels are subtracted from the marked voxels to obtain the voxel grid that needs to be marked, denoted as S c_lable ;
[0043] Step 3.2, randomly select a voxel unit s(s∈S c_lable ), by traversing the voxel in the initial room semantic label S i_lable If there is a room semantic label value S1 in the 26-neighborhood, the current voxel is also labeled as S1;
[0044] Step 3.3, using S c_lable Update S i_lable ;
[0045] Step 3.4, repeat steps 3.1-3.3 until all unlabeled voxels are assigned a room semantic label, and then obtain the final room space segmentation result.
[0046] The beneficial effects of the present invention are:
[0047] The room space segmentation method based on distance constraints of the present invention first simplifies the original point cloud, and distinguishes between free voxels and occupied voxels by calculating the spatial occupancy probability grid; secondly, the three-dimensional Euclidean distance between the free voxel and its nearest occupied voxel is calculated; on this basis, the room seed area is obtained by distance threshold constraint and DBSCAN clustering method; then, the unlabeled accessible area is expanded by adopting the wavefront growth algorithm based on 26 neighborhoods, and it is merged into the room seed area, which solves the problem that some voxels that belong to the room are not labeled, and finally obtains a complete room space segmentation result. The method of the present invention makes full use of the rich geometric and spatial structure information of the three-dimensional space, solves the room space segmentation in complex indoor scenes, and can not only segment non-Manhattan scenes, but also will not excessively divide the room into multiple parts, solving the problem of over-segmentation. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 It is a flow chart of the room space segmentation method based on distance constraint of the present invention;
[0049] FIG2 is a schematic diagram of a scene of Scene 1 of the present invention, wherein FIG2(a) is a schematic diagram of a scene of original scene data, FIG2(b) is a schematic diagram of a scene of a denoising result, FIG2(c) is a schematic diagram of a scene of a downsampling result, and FIG2(d) is a schematic diagram of a scene after an occupancy probability grid is generated;
[0050] FIG3 is a schematic diagram of the process results of Scene 1 of the present invention, wherein FIG3(a) is a schematic diagram of the three-dimensional distance transformation result, FIG3(b) is a schematic diagram of the threshold constraint result, FIG3(c) is a schematic diagram of the DBSCAN clustering result, and FIG3(d) is a schematic diagram of the wavefront growth result;
[0051] 4 is a diagram showing the segmentation results of Scene 1 of the present invention when different distance thresholds are taken, wherein FIG4(a) is a diagram showing the segmentation result when the distance value is 0.3, FIG4(b) is a diagram showing the segmentation result when the distance value is 0.42, and FIG4(c) is a diagram showing the segmentation result when the distance value is 0.5. DETAILED DESCRIPTION
[0052] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0053] Example 1
[0054] This embodiment provides a room space segmentation method based on distance constraints, such as Figure 1 As shown, the specific steps are as follows:
[0055] Step 1: The original point cloud data is subjected to statistical filtering denoising and voxel downsampling, and then the spatial occupancy probability grid is calculated according to the number of point clouds in the voxel, which is divided into free voxels and occupied voxels;
[0056] Step 1.1, by performing statistical analysis on the points around each point in the point cloud, determine whether the point is an outlier or a noise point, and perform statistical filtering on the points in the point cloud, then downsample the filtered and denoised data and use it as the input data for the next stage. The original data is shown in Figure 2(a), and the voxel filtered downsampled data is shown in Figure 2(b) and Figure 2(c);
[0057] Step 1.2, the process of mapping the filtered and denoised point cloud data from a continuous three-dimensional space to a discrete three-dimensional grid or voxel, and then further analyzing the distribution density of points in each voxel, and evaluating whether each voxel is occupied by counting the proportion of the number of point clouds contained in the voxel to the total number of points, as shown in Figure 2(d).
[0058] Step 2: First, calculate the Euclidean distance between the free voxel and its nearest occupied voxel, and then separate the connected rooms according to the distance threshold. Then use the DBSCAN clustering method to cluster the voxels in the same room according to density and mark them with different colors to obtain the room seed region;
[0059] Step 2.1, using the Euclidean distance-based transformation, as shown in Figure 3(a), maps the distance value from the free voxel to its nearest occupied voxel to the voxel grayscale value, where the grayscale value of each grid is equal to its distance to the nearest target point;
[0060] Step 2.2, after the Euclidean distance transformation is completed, set a distance threshold to determine which points belong to the same room, such as Figure 4(a)-Figure 4(c) As shown in Figure 3(b), the point cloud after threshold segmentation is clustered using a density-based clustering method, as shown in Figure 3(c).
[0061] Step 3: Use the wavefront growing algorithm to traverse the 26-neighborhood of the unlabeled voxel. If there is a labeled voxel in the 26-neighborhood of the voxel, the current voxel is labeled according to the labeled voxel. Otherwise, traverse other unlabeled voxels until all unlabeled voxels are assigned a room semantic label, and the room space segmentation is completed.
[0062] Step 3.1, assign the initial room semantic tag ID to the room seed region after density analysis, and mark the voxels contained in each occupied space in the seed region as the corresponding initial room tag ID, and obtain the voxel grid S containing the room seed region information i_lable Then, all the initial voxels are subtracted from the marked voxels to obtain the voxel grid that needs to be marked, denoted as S c_lable ;
[0063] Step 3.2, randomly select a voxel unit s(s∈S c_lable ), by traversing the voxel in the initial room semantic label S i_lable If there is a room semantic label value S1 in the 26-neighborhood, the current voxel is also labeled as S1;
[0064] Step 3.3, using S c_lable Update S i_lable ;
[0065] Step 3.4: Repeat steps 3.1-3.3 until all unlabeled voxels are assigned a room semantic label, and then obtain the final room space segmentation result.
[0066] Example 2
[0067] This embodiment provides a room space segmentation method based on distance constraints. Based on Embodiment 1, step 1.1 is specifically as follows:
[0068] Step 1.1.1, current point P i (x i ,y i ,z i ) and its neighbor point N m (x m ,y m ,z m ) The average distance D between i , the distance between the two points is Si , whose standard deviation is δ, is calculated as (1) and (2).
[0069]
[0070] In formula (1), P i (x i ,y i ,z i ) any point selected, N m (x m ,y m ,z m ) is P i Neighbor points;
[0071]
[0072] In formula (2), S i is the distance between two points, D i is the average distance, δ is the standard deviation;
[0073] Assume that there are n points in the initial point cloud and the distance threshold D max is the average distance D i It is obtained by adding α times the standard deviation δ (α is a constant);
[0074] Step 1.1.2, if D i >D max , then the point is considered an outlier and is removed. Otherwise, the point is considered a non-outlier and is retained;
[0075] Step 1.1.3, downsample the retained data using the voxel grid method, set the voxel side length, and represent the scene with three-dimensional voxels. Then calculate the center of gravity of each voxel to represent the voxel, and get the downsampled result, as shown in Figure 2(c).
[0076] Step 1.2 is as follows:
[0077] Step 1.2.1, according to the ray casting method, multiple beams are emitted from the laser scanning point, the scanning point is the viewpoint, when a beam encounters an obstacle, it stops, the voxel at the time of collision is the occupied voxel, and the space between the viewpoint and the obstacle is called the free voxel;
[0078] Step 1.2.2, using the Bresenham line algorithm, determine which voxels constitute the straight line between the two points, and mark them according to the collision situation. The marked voxels are divided into three three-dimensional occupancy probability grids: "occupied voxels", "free voxels" and "unknown". The assignment rule is:
[0079]
[0080] Example 3
[0081] This embodiment provides a room space segmentation method based on distance constraints. Based on Embodiment 1-2, step 2.1 is specifically as follows:
[0082] Step 2.1.1, given a grid map M (representing the distance from each grid to the nearest obstacle), let p and q be the coordinates of any two grids in the grid M, and the Euclidean distance between the two grids is as shown in formula (4);
[0083]
[0084] In formula (4), point (x i ,y i ,z i ) is the free grid coordinate, point (x j ,y j ,z j ) is the occupied grid coordinate, and the free grid and the occupied grid constitute the grid map M;
[0085] Step 2.1.2, define the Euclidean distance transformation voxel execution state queue Q. The state of some voxels has been determined by the ray casting algorithm. When a cell is popped out of the queue, it is first checked whether it has been processed before. If it has not been processed, it is necessary to confirm the type of wave it belongs to.
[0086] Step 2.1.3, determine whether it is an ascending wave or a descending wave according to the current voxel state and the voxel occupancy probability value. If the current voxel is a free space voxel and the voxel state is F, change the voxel state to T. If the current voxel state is T and the occupancy probability value has stored "occupied", add the current voxel to the queue Q and update the probability value of the voxel to "free". The two processes at this time are ascending waves;
[0087] Step 2.1.4, if the current voxel state is F, then change the voxel state to T, and add the current voxel to the queue Q. If the current voxel state is T, and the occupancy probability value has stored "free", then add the current voxel to the queue Q, and update the probability value of the voxel to "occupied". The two processes at this time are descending waves;
[0088] Step 2.1.5, update the circular queue Q, and repeat steps 2.1.1-2.1.4 until the three-dimensional Euclidean distance transformation of all voxels is completed.
[0089] Step 2.2 is as follows:
[0090] Step 2.2.1, set a distance threshold to determine whether the distance from the voxel to the nearest occupied voxel is less than or equal to the set distance threshold. If it is less than or equal to the distance threshold, the point is marked as inactive, indicating that the voxel belongs to the same room; if it is greater than the distance threshold, the point is marked as active, and the state of the part of the distance value minus the distance threshold is set to pending, and the remaining state is set to inactive. As shown in Figure 3 and Figure 4(b);
[0091] Step 2.2.2, use the DBSCAN algorithm to cluster the point cloud after threshold segmentation. First, randomly select an unvisited data point p and mark it as visited;
[0092] Step 2.2.3, with p as the center, check whether there are at least MinPts data points in the ε neighborhood of p. If so, mark all these points as visited and form a new cluster;
[0093] Step 2.2.4, if p is not a core object, mark p as a noise point;
[0094] Step 2.2.5, repeat steps 2.2.2-2.2.4 until all data points have been visited.
Claims
1. A room space segmentation method based on distance constraints, characterized in that: The specific steps are as follows: Step 1: The original point cloud data is subjected to statistical filtering denoising and voxel downsampling, and then the spatial occupancy probability grid is calculated according to the number of point clouds in the voxel, which is divided into free voxels and occupied voxels; Step 2: First, calculate the Euclidean distance between the free voxel and its nearest occupied voxel, and then separate the connected rooms according to the distance threshold. Then, use the DBSCAN clustering method to cluster the voxels in the same room according to the density and mark them with different colors to obtain the room seed region. Step 3: Use the wavefront growing algorithm to traverse the 26-neighborhood of the unlabeled voxel. If there is a labeled voxel in the 26-neighborhood of the voxel, the current voxel is labeled according to the labeled voxel. Otherwise, traverse other unlabeled voxels until all unlabeled voxels are assigned a room semantic label and the room space segmentation is completed.
2. The room space segmentation method based on distance constraint according to claim 1, characterized in that: The step 1 is specifically as follows: Step 1.1, by statistically analyzing the points around each point in the point cloud, determine whether the point is an outlier or a noise point, and perform statistical filtering on the points in the point cloud, then downsample the filtered and denoised data and use it as the input data for the next stage; Step 1.2, the process of mapping the filtered and denoised point cloud data from a continuous three-dimensional space to a discrete three-dimensional grid or voxel, and then further analyzing the distribution density of points in each voxel, and evaluating whether each voxel is occupied by counting the proportion of the number of point clouds contained in the voxel to the total number of points.
3. The room space segmentation method based on distance constraint according to claim 2, characterized in that: The step 1.1 is specifically as follows: Step 1.1.1, current point P i (x i ,y i ,z i ) and its neighbor point N m (x m ,y m ,z m ) The average distance D between i , the distance between the two points is S i , whose standard deviation is δ, is calculated as (1) and (2). Among them, P i (x i ,y i ,z i ) any point selected, N m (x m ,y m ,z m ) is P i Neighbor points of Among them, S i is the distance between two points, D i is the average distance, δ is the standard deviation; Assume that there are n points in the initial point cloud and the distance threshold D max is the average distance D i It is obtained by adding α times the standard deviation δ (α is a constant); Step 1.1.2, if D i >D max , then the point is considered to be an outlier and is removed, otherwise it is considered to be a non-outlier and is retained; Step 1.1.3, use the voxel grid method to downsample the denoised data, set the voxel side length, represent the scene with three-dimensional voxels, and then calculate the center of gravity of each voxel to represent the voxel, and obtain the downsampling result.
4. The room space segmentation method based on distance constraint according to claim 3, characterized in that: The step 1.2 is specifically as follows: Step 1.2.1, according to the ray casting method, multiple beams are emitted from the laser scanning point. When a beam encounters an obstacle, it stops. The voxel at the time of collision is the occupied voxel, and the space between the viewpoint and the obstacle is called the free voxel; Step 1.2.2, using the Bresenham line algorithm, determine which voxels constitute the straight line between the two points, and mark them according to the collision situation. The marked voxels are divided into three three-dimensional occupancy probability grids: "occupied voxels", "free voxels" and "unknown", and the assignment rules are as follows:
5. The room space segmentation method based on distance constraint according to claim 4, characterized in that: The step 2 is specifically as follows: Step 2.1, using a transformation based on Euclidean distance, the distance value from the free voxel to its nearest occupied voxel is mapped to the voxel gray value, where the gray value of each grid is equal to its distance to the nearest target point; Step 2.2, after the Euclidean distance transformation is completed, a distance threshold is set to determine which points belong to the same room, and then a density-based clustering method is used to cluster the point cloud after threshold segmentation.
6. The room space segmentation method based on distance constraint according to claim 5, characterized in that: The step 2.1 is specifically as follows: Step 2.1.1, given a grid map M, let p and q be the coordinates of any two grids in the grid M, and the Euclidean distance between the two grids is as shown in formula (4); Among them, the point (x i ,y i ,z i ) is the free grid coordinate, point (x j ,y j ,z j ) is the occupied grid coordinate, and the free grid and the occupied grid constitute the grid map M; Step 2.1.2, define the Euclidean distance transformation voxel execution state queue Q. The state of some voxels has been determined by the ray casting algorithm. When a cell is popped out of the queue, it is first checked whether it has been processed before. If it has not been processed, it is necessary to confirm the type of wave it belongs to. Step 2.1.3, determine whether it is an ascending wave or a descending wave according to the current voxel state and the voxel occupancy probability value. If the current voxel is a free space voxel and the voxel state is F, change the voxel state to T. If the current voxel state is T and the occupancy probability value has stored "occupied", add the current voxel to the queue Q and update the probability value of the voxel to "free". The two processes at this time are ascending waves; Step 2.1.4, if the current voxel state is F, then change the voxel state to T, and add the current voxel to the queue Q. If the current voxel state is T, and the occupancy probability value has stored "free", then add the current voxel to the queue Q, and update the probability value of the voxel to "occupied". The two processes at this time are descending waves; Step 2.1.5, update the circular queue Q, and repeat steps 2.1.1-2.1.4 until the three-dimensional Euclidean distance transformation of all voxels is completed.
7. The room space segmentation method based on distance constraint according to claim 6, characterized in that: The step 2.2 is specifically as follows: Step 2.2.1, set a distance threshold to determine whether the distance from the voxel to the nearest occupied voxel is less than or equal to the set distance threshold. If it is less than or equal to the distance threshold, the point is marked as inactive, indicating that the voxel belongs to the same room; if it is greater than the distance threshold, the point is marked as active, and the state of the part of the distance value minus the distance threshold is set to pending, and the remaining state is set to inactive; Step 2.2.2, use the DBSCAN algorithm to cluster the point cloud after threshold segmentation. First, randomly select an unvisited data point p and mark it as visited; Step 2.2.3, with p as the center, check whether there are at least MinPts data points in the ε neighborhood of p. If so, mark all these points as visited and form a new cluster; Step 2.2.4, if p is not a core object, mark p as a noise point; Step 2.2.5, repeat steps 2.2.2-2.2.4 until all data points have been visited.
8. The room space segmentation method based on distance constraint according to claim 7, characterized in that: The step 3 is specifically as follows: Step 3.1, assign the initial room semantic tag ID to the room seed region after density analysis, and mark the voxels contained in each occupied space in the seed region as the corresponding initial room tag ID, and obtain the voxel grid S containing the room seed region information i_lable Then, all the initial voxels are subtracted from the marked voxels to obtain the voxel grid that needs to be marked, denoted as S c_lable ; Step 3.2, randomly select a voxel unit s(s∈S c_lable ), by traversing the voxel in the initial room semantic label S i_lable If there is a room semantic label value S1 in the 26-neighborhood, the current voxel is also labeled as S1; Step 3.3, using S c_lable Update S i_lable ; Step 3.4, repeat steps 3.1-3.3 until all unlabeled voxels are assigned a room semantic label, and then obtain the final room space segmentation result.
