Method for retrieving repetitive building elements based on axial movement and local matching

Abstract

The application discloses a repetitive building element retrieval method based on axial movement and local matching, and specifically comprises the following steps: carrying out preliminary segmentation and optimization on a building facade scene; randomly selecting a building element sample under user driving; determining the X axis of the building facade and other two axes by solving a plane equation; transforming the building element sample to the left lower corner origin of the building facade scene by using an OBB bounding box; extracting key points of the element to be retrieved by using an internal shape descriptor; constructing a local matching constraint by combining FPFH and the OBB bounding box; finally, finding the nearest point of the key points of each element to be retrieved, comparing the difference between the descriptors by using the Euclidean distance, and completing the retrieval of the repetitive building element. The application can directly retrieve the repetitive building element in the building facade scene, and solves the problems of the fuzziness of the retrieval object and the low recall rate in the previous retrieval methods.

Description

Technical Field

[0001] This invention belongs to the field of point cloud semantic segmentation technology, specifically relating to a method for retrieving repetitive building elements based on axial movement and local matching. Background Technology

[0002] 3D point cloud data can be used for on-site inspection and reverse engineering of building models. In recent years, with the rapid development of point cloud data acquisition technology, the demand for 3D point cloud data from existing buildings and civil infrastructure has continued to increase. Among these, building element retrieval, as a major component in building renovation, building analysis, and building conservation, has significant research value and broad application prospects.

[0003] The process of repetitive building element retrieval refers to finding a set of elements in a point cloud dataset that have similar texture, size, density, and other geometric features to all instances of a query object. Point cloud data can come from object categories such as walls, panels, and doors. Previous 3D model retrieval techniques ignored scene factors and directly searched existing 3D model libraries. With the maturation of 3D model retrieval technology, even when considering environmental factors, existing techniques for repetitive building element retrieval still lead to ambiguity in search results for building elements with similar features. Furthermore, deep learning-based methods require large amounts of labeled datasets for training, and creating labeled datasets of building facades requires significant human and material resources. Summary of the Invention

[0004] The purpose of this invention is to provide a repetitive building element retrieval method based on axial movement and local matching, which overcomes the ambiguity problem of existing technologies and does not require a large amount of manual annotation. It can directly use point cloud data to create the element to be retrieved and retrieve matching objects in the point cloud in the building scene.

[0005] The technical solution adopted in this invention is a repetitive building element retrieval method based on axial movement and local matching, which is implemented according to the following steps:

[0006] Step 1, Point Cloud Segmentation: For the original building facade scene, clustering is performed using region growing and color-based region growing methods to obtain the clustered point cloud data of the building facade.

[0007] Step 2, Sample Selection: For the point cloud data of the clustered building facades obtained in Step 1, select a building element to be retrieved, such as a window or door, under the user's guidance.

[0008] Step 3, Axis Determination: Use the Random Sampling Consensus Algorithm (RANSAC) to obtain the normal vector of the building facade and determine the X-axis. Then, use the Principal Component Analysis Algorithm (PCA) to correct the perpendicularity between each principal direction and determine the other axes.

[0009] Step 4, Sample Transformation: Use OBB bounding boxes to transform the building element samples to the bottom left corner of the building facade scene.

[0010] Step 5, Retrieval based on local matching constraints: After determining the axis of movement of the building element sample in the building facade scene in Step 4, local matching constraints are constructed, and the repetitive building element retrieval is completed using the local matching constraints.

[0011] The invention is further characterized in that,

[0012] Step 1 is as follows:

[0013] Step 1.1: Use the region growing method to perform initial clustering on the original building facade scene;

[0014] Step 1.2: Based on the results obtained in Step 1.1, color information is added to perform secondary clustering using a color-based region growing method.

[0015] Step 1.1 specifically involves:

[0016] The normal and curvature are calculated for the original point cloud building facade scene. The point cloud is sorted according to the curvature value of the points, and the point with the smallest curvature is selected as the initial seed point. The region growing algorithm starts growing from the seed point with the smallest curvature. An empty clustering region A and an empty seed point sequence B are defined. The initial seed point is selected and added to the seed point sequence. The neighboring points of the seed point are searched, and the angle between the normal of each neighboring point and the normal of the seed point is calculated. When it is less than the set smoothing threshold, the neighboring point is added to A. At the same time, it is determined whether the curvature value of the neighboring point is less than the curvature threshold. The neighboring points with curvature values ​​less than the curvature threshold are added to the seed point sequence B. After all the neighboring points have been determined, the current seed point is deleted, and a new seed point is selected in B. The above steps are repeated until the sequence in B is empty. At this time, the growth of one region is completed. The above process is repeated until all points in the scene are clustered, and the initial clustering of the building facade scene is completed.

[0017] Step 1.2 specifically involves:

[0018] Based on the initial clustering results in step 1.1, define an empty clustering region E and an empty seed point sequence F. Select an initial seed point, add it to the seed point sequence, and search for the neighboring points of the seed point. Calculate the difference between the normal of each neighboring point and the color of the seed point. If the difference is less than the set color threshold, add the neighboring point to E; otherwise, add it to F. Repeat the above steps by selecting a new seed point in F until the sequence in F is empty. At this point, one region has been grown. Repeat the above process, check the average color of adjacent clusters, and if the average color threshold is less than the set threshold, merge the two clusters into one. Continue this process until all color clustering is completed, thus completing the secondary clustering.

[0019] Step 2 is as follows:

[0020] Step 2.1: Using the point cloud file selection function in the PCL point cloud library, input the results of the secondary clustering and enter the selection state;

[0021] Step 2.2: Select an instance of a building element object, remove duplicate selected points during the process, and save the selected point cloud as a .pcd file to obtain a sample of the element to be retrieved.

[0022] Step 3 specifically involves:

[0023] Step 3.1: Use the Random Sampling Consensus Algorithm (RANSAC) to obtain the planar equation of the building facade and determine the main direction, i.e., the X-axis.

[0024] Step 3.2: Based on the X-axis, use the principal component analysis algorithm to correct the X-axis, making each principal direction of the X-axis, Y-axis and Z-axis perpendicular to each other, thereby determining the axial direction of the Y-axis and Z-axis.

[0025] Step 5 specifically involves:

[0026] Step 5.1: In order to speed up the retrieval, key points are extracted from the building element samples and building facades to be retrieved using the Internal Shape Descriptor (ISS). The Fast Point Feature Histogram (FPFH) algorithm is used to calculate the feature descriptors of the building element samples and building facade scenes to be retrieved. Finally, the extracted feature descriptors and the position constraints of the building elements to be retrieved in the scene constitute local matching constraints.

[0027] Step 5.2: Treat the point cloud of the building element samples to be retrieved as a sliding window. Use the sliding window to move axially and partially match the building element samples with the point cloud of the building facade scene to retrieve the target model. Given a threshold, traverse the building element samples to be retrieved from the bottom left corner to the right and from bottom to top, establishing a kd-tree structure to query the point to be retrieved and its nearest neighbors in the building element. Use Euclidean distance to measure similarity, thereby completing the retrieval of repetitive building elements.

[0028] The beneficial effects of this invention are:

[0029] This invention presents a repetitive building element retrieval method based on axial movement and local matching. It selects the building elements to be retrieved, determines their axial direction based on feature analysis, and finally retrieves repetitive elements in the building facade by constructing local matching constraints. During construction, it eliminates the need to pre-establish a computer-aided design model library or building information model library, collect training data, train classifiers, and segment individual objects from scene point clouds. This solves the problems of fuzziness in the retrieved objects and low recall rates inherent in previous retrieval methods. This invention can accurately retrieve building elements of different forms, improving both retrieval accuracy and recall. Attached Figure Description

[0030] Figure 1 This is an example image retrieved in the point cloud scene in step 5 of an embodiment of the repetitive building element retrieval method based on axial movement and local matching of the present invention. Detailed Implementation

[0031] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0032] Example

[0033] This embodiment provides a method for retrieving repetitive building elements based on axial movement and local matching, which is implemented according to the following steps:

[0034] Step 1, Point Cloud Segmentation: The original point cloud data is clustered using region growing and color-based region growing methods to obtain the clustered point cloud data of the building facade.

[0035] Step 1.1: Calculate the normal and curvature of the original point cloud building facade scene. Sort the point cloud according to the curvature values ​​and select the point with the smallest curvature as the initial seed point. The region growing algorithm starts growing from the seed point with the smallest curvature. Define an empty clustering region A and an empty seed point sequence B. Select the initial seed point and add it to the seed point sequence. Search for the neighboring points (the 50 nearest neighbors) of the seed point. Calculate the angle between the normal of each neighboring point and the normal of the seed point. If it is less than the set smoothing threshold, add the neighboring point to A. At the same time, determine if the curvature value of the neighboring point is less than the curvature threshold. Add the neighboring points with curvature values ​​less than the curvature threshold to the seed point sequence B. After all neighboring points have been determined, delete the current seed point and reselect a new seed point in B, repeating the above steps until the sequence in B is empty. At this point, one region growth is complete. Repeat the above process until all points in the scene are clustered, completing the initial clustering and initial segmentation of the point cloud data.

[0036] Step 1.2: Based on the initial clustering results, define an empty clustering region E and an empty seed point sequence F. Select an initial seed point, add it to the seed point sequence, and search for the seed point's neighboring points (the 10 nearest neighbors). Calculate the difference between the normal of each neighboring point and the color of the seed point. If the difference is less than a set color threshold, add the neighboring point to E; otherwise, add it to F. Repeat the above steps, selecting a new seed point in F, until the sequence in F is empty. At this point, one region has been grown. Repeat the above process, checking the average color of adjacent clusters. If the average color threshold is less than a set threshold, merge the two clusters into one. Continue this process until all color clustering is completed, thus completing the secondary clustering. Information optimization is then performed using the labeled colors.

[0037] Step 2, Sample Selection: For the point cloud data of the clustered building facades obtained in Step 1, select a building element to be retrieved, such as a window or door, under the user's guidance.

[0038] Step 2.1: Use the point cloud file selection function registerAreaPickingCallback() in the point cloud library PCL, input the result of the secondary clustering, and press the X key to enter the selection state;

[0039] Step 2.2: Select an instance of a building element object, remove duplicate selected points during the process, and save the selected point cloud as a .pcd format. This gives you a sample of the element to be retrieved.

[0040] Step 3, Axis Determination: Use the Random Sampling Consensus Algorithm (RANSAC) to obtain the normal vector of the building facade and determine the X-axis. Then, use the Principal Component Analysis Algorithm (PCA) to correct the perpendicularity between each principal direction and determine the other axes.

[0041] Step 3.1: Use the Random Sampling Consensus Algorithm (RANSAC) to obtain the planar equation of the building facade and determine the main direction, i.e., the X-axis.

[0042] Step 3.1 specifically involves:

[0043] Step 3.1.1: According to the plane equation (1), three points are needed to fit the plane. Randomly select three points on the plane and calculate the plane model parameters A, B, C, and D.

[0044] A·x+B·y+C·z=D (1)

[0045] Step 3.1.2: Use the remaining data points to test the model and calculate the error. If the error is less than the given threshold, then the store is identified as an inside point. Count and record the number of inside points under the model parameters.

[0046] Step 3.1.3: Repeat steps 3.1.1 and 3.1.2. If the number of inliers in the current model is greater than the maximum number of inliers that have been saved, then update the model parameters. The model parameters that are saved are always the model parameters with the most inliers.

[0047] Step 3.1.4: Repeat steps 3.1.1, 3.1.2 and 3.1.3, iterating continuously until the iteration threshold is reached, find the parameter model with the most interior points, and finally estimate the model parameters again using the interior points to obtain the final model parameters A, B, C and D.

[0048] Step 3.1.5: After obtaining the fitted plane parameters as described above, save them in the predefined parameter matrix E in the order A, B, C, D, and define a vector. The normalized x, y, z coordinates of the plane normal are x = E[0], y = E[1], and z = E[2], respectively. This determines the axis of x.

[0049] Step 3.2 Use principal component analysis algorithm to correct the X-axis, making each principal direction of the X-axis, Y-axis and Z-axis perpendicular to each other, thereby determining the axis of the Y-axis and Z-axis;

[0050] Step 3.2 specifically involves:

[0051] Step 3.2.1, let the centroid coordinates P c =(P x ,P y ,P z ), calculate the centroid coordinates of the building facade scene according to formula (2);

[0052]

[0053] In equation (2): n represents the total number of point cloud scenes of building facades; x i The x-coordinate and y-coordinate of each point in the point cloud scene representing the building facade are... i The z-coordinates of each point in the point cloud scene representing the building facade are... i The z-coordinate of each point in the point cloud scene representing the building facade, i∈[1,n];

[0054] Step 3.2.2, for each point P in the point cloud scene of the building facade... i According to formula (3), calculate the normalized covariance matrix C corresponding to each point in the building facade scene; and obtain the result.

[0055]

[0056] In equation (3): n represents the total number of point cloud scenes of building facades; P c The centroid of the point cloud scene representing the building facade, i∈[1,n].

[0057] Step 3.2.3: Decompose the covariance matrix C obtained after step 3.2.2 into eigenvalues ​​and eigenvectors according to formula (4). It is necessary to convert the covariance matrix into a diagonal matrix, according to the content of linear algebra.

[0058]

[0059] In equation (4): λ j Let j be the j-th eigenvalue of the covariance matrix; Let j be the j-th eigenvector;

[0060] Step 3.2.4 involves generating a 3x3 matrix F from the eigenvectors obtained in step 3.2.3. Each column of matrix F contains an eigenvector (col). The X-axis has already been determined in step 3.2.1. Let the axis of Y be The axis of Z is Then correct the perpendicularity between each principal direction, first taking... Based on formula (5), the axes of X, Y, and Z are obtained respectively. The main direction of the point cloud scene of the building facade is then aligned with the determined X, Y, and Z axes.

[0061]

[0062] Step 4, Sample Transformation: Use OBB bounding boxes to transform the building element samples to the bottom left corner of the building facade scene, thus completing the transformation of the sample point cloud.

[0063] Step 4 specifically involves:

[0064] After determining the axes above, the aligned element samples to be retrieved and the building facade scene are obtained. The bounding boxes in the Y and Z directions are calculated using OBB respectively, and the minimum values ​​among all points in the transformed element samples to be retrieved and the building facade scene are obtained respectively. Let min_p and min_c be the minimum main axis values ​​among all points in the element samples to be retrieved and the building facade scene respectively. The translation distances of the Y and Z axes are calculated using formula (6) so that the point cloud of the transformed sample to be retrieved is translated to the position of the origin in the lower left corner of the scene.

[0065]

[0066] In equation (6): D y D represents the translation distance of the element to be retrieved along the Y-axis. zmin_p.y represents the translation distance of the element to be retrieved along the Z-axis, min_p.z represents the minimum Y value among all points of the element to be retrieved, min_c.y represents the minimum Y value among all points of the building facade scene, and min_c.z represents the minimum Z value among all points of the building facade scene.

[0067] Step 5, Retrieval based on local matching constraints: After determining the axis of movement of the building element samples in the building facade scene in Step 4, local matching constraints are constructed, and the retrieval of repetitive building elements is completed using the local matching constraints, such as... Figure 1 The image shows the completion of the retrieval of repetitive building elements;

[0068] Step 5.1: In order to speed up the retrieval, key points are extracted from the building element samples and building facades to be retrieved using the Internal Shape Descriptor (ISS). The Fast Point Feature Histogram (FPFH) algorithm is used to calculate the feature descriptors of the building element samples and building facade scenes to be retrieved. Finally, the extracted feature descriptors and the position constraints of the building elements to be retrieved in the scene constitute local matching constraints.

[0069] Step 5.1 specifically involves:

[0070] Step 5.1.1, for each point k of the element to be retrieved i Establish a local coordinate system and set a search radius r for each point;

[0071] Step 5.1.2: Construct a kd-tree with a search radius r, and calculate the weight W of all points within the range according to formula (7). ij ;

[0072]

[0073] In equation (7): |k i -k j | Set the search radius r, k for each point. j Represents sample points of building elements;

[0074] Step 5.1.3, calculate k for each point according to formula (8). i The covariance matrix cov(k) i ), eigenvalues ​​are obtained from matrix decomposition. And arrange them in descending order;

[0075]

[0076] Step 5.1.4: Set threshold ε1 and threshold ε2. If the following equation (9) is satisfied, then the point is considered to be a key point to be extracted.

[0077]

[0078] Step 5.1.5, for each query point P s Calculate a series of α values ​​between this point and its neighboring points. θ, value, denoted as SP, according to formula (10) at any point P t Define a local coordinate system u,v,w on it;

[0079]

[0080] In the formula: P t P s It refers to any two points in the point cloud; n s It is point P s The corresponding normal; n t It is point P t The corresponding normal;

[0081] Step 5.1.6, using three angle values ​​α, θ, and estimate the normal n according to formula (11) based on the local coordinate system obtained in step 5.1.5. s and n t The relative positional deviation between normals;

[0082]

[0083] Step 5.1.7, repeat step 5.1.5, calculate SP in the k-neighborhood of other points, and calculate P according to formula (12). d Feature descriptor;

[0084]

[0085] In equation (12): W k It is P d The distance between it and each of its neighboring points;

[0086] Step 5.1.8: FPFH divides each angle into 11 sub-intervals, thus obtaining a 33-dimensional local feature descriptor. OBB is used to calculate the maximum and minimum values ​​of the element to be retrieved in the building facade scene. According to formula (13), the local position constraints of the building element to be retrieved in the building facade scene are constructed.

[0087]

[0088] In equation (13), V represents the local position constraint of the building element to be retrieved in the building facade scene. x V represents the cell length with the search element in the X direction. yV represents the cell length with searchable elements in the Y direction. z This indicates the cell length of the element to be searched in the Z direction. max_p.x, max_py, and max_pz represent the maximum X, Y, and Z values ​​of the element to be searched in the building facade scene, respectively.

[0089] Step 5.1.9: The local feature descriptors extracted by FPFH and the position constraints of the building element to be retrieved in the building facade scene constitute a local matching constraint S = (F, V), where F represents the difference between the descriptors and V represents the local position constraint of the building element to be retrieved in the building facade scene.

[0090] Step 5.2: Treat the point cloud of the building element samples to be retrieved as a sliding window. Use the sliding window to move axially to partially match the building element samples with the point cloud of the building facade scene to retrieve the target model. Given a threshold, traverse the building element samples to be retrieved from the bottom left corner to the right and from bottom to top, establishing a kd-tree structure. Compare the building element samples to be retrieved to perform local matching constraints on the building facade scene, and use Euclidean distance for similarity measurement to complete the retrieval of repetitive building elements.

[0091] Step 5.2 specifically involves:

[0092] Step 5.2.1: Based on the global coordinate system of the point cloud data of the building facade scene, establish a cube bounding box containing all point clouds. For each cube containing more than 1 point, construct a segmentation plane. The two segmented subspaces and the points on the segmentation plane constitute branches and connection points. If the number of internal points of the segmented subspace exceeds 1, continue to segment to obtain the established kd tree.

[0093] Step 5.2.2: Since the target is a building facade scene, the scene is divided according to the threshold. The element to be retrieved is regarded as a sliding window. Given the threshold, the sliding window will start to traverse the transformed point cloud from left to right and from top to bottom.

[0094] Step 5.2.3: During the traversal, use nearest neighbor search to find the k nearest neighbors of the key point of the element to be retrieved obtained after step 5.1 in the kd-tree built after step 5.2.1, and save the index of each k nearest neighbor.

[0095] Step 5.2.4: Compare the feature descriptors of the key points of the element to be retrieved obtained after step 5.1.5 with the feature descriptors of the nearest neighbor points to determine whether the neighboring points and the detection points can be classified as the same building element instance.

[0096] Step 5.2.5: Use the Euclidean distance formula (14) to compare the similarity between two descriptors. If distance(A,B) is less than a threshold δ, it is considered that the key point is matched with its nearest neighbor. If distance(A,B) is greater than a threshold δ, it is considered that the key point is not matched with its nearest neighbor. Finally, store the matching point pairs in the vector.

[0097]

[0098] In the formula: distance(A,B) is the distance between two corresponding points;

[0099] Step 5.2.6: After each translation following a certain threshold, repeat steps 5.2.4 and 5.2.5 to retrieve the corresponding instances, determine the range of building element instances to be detected, the number of stored instances, and the range of each instance;

[0100] Step 5.2.7: Display instances in the scene that are identical to the element to be retrieved using different colors.

[0101] As can be seen from the above, the present invention is a repetitive building element retrieval method based on axial movement and local matching. For building facades, it adopts the idea of ​​axial movement based on samples to construct local matching constraints and complete the retrieval of repetitive building elements in building facades. The present invention can directly retrieve repetitive building elements in building facade scenes, solving the problems of fuzziness of retrieval objects and low recall rate in previous retrieval methods.

Claims

1. A method for repetitive building element retrieval based on axial movement and local matching, characterized in that, The specific steps are as follows: Step 1, Point Cloud Segmentation: For the original building facade scene, clustering is performed using region growing and color-based region growing methods to obtain the clustered point cloud data of the building facade. Step 2, Sample Selection: Based on the point cloud data of the clustered building facades obtained in Step 1, select a building element to be retrieved under the user's guidance; Step 3, Axis Determination: Use the random sampling consensus algorithm to obtain the normal vector of the building facade and determine the X-axis. Then use the principal component analysis algorithm to correct the perpendicularity between each principal direction and determine the other axes. Step 4, Sample Transformation: Use OBB bounding boxes to transform the building element samples to the bottom left corner of the building facade scene. Step 5, Retrieval based on local matching constraints: After determining the axis of movement of the building element sample in the building facade scene in Step 4, local matching constraints are constructed, and the repetitive building element retrieval is completed using the local matching constraints. Step 1 specifically involves: Step 1.1: Use the region growing method to perform initial clustering on the original building facade scene; Step 1.2: Based on the results obtained in Step 1.1, color information is added to perform secondary clustering using a color-based region growing method; Step 1.1 specifically includes: The normal and curvature are calculated for the original point cloud building facade scene. The point cloud is sorted according to the curvature value of the points, and the point with the smallest curvature is selected as the initial seed point. The region growing algorithm starts growing from the seed point with the smallest curvature. An empty clustering region A and an empty seed point sequence B are defined. The initial seed point is selected and added to the seed point sequence. The neighboring points of the seed point are searched, and the angle between the normal of each neighboring point and the normal of the seed point is calculated. When it is less than the set smoothing threshold, the neighboring point is added to A. At the same time, it is determined whether the curvature value of the neighboring point is less than the curvature threshold. Neighboring points with curvature values ​​less than the curvature threshold are added to the seed point sequence B. After all neighboring points have been determined, the current seed point is deleted, and a new seed point is selected in B. The above steps are repeated until the sequence in B is empty. At this time, the growth of one region is completed. The above process is repeated until all points in the scene are clustered, and the initial clustering of the building facade scene is completed. Step 1.2 specifically includes: Based on the initial clustering results in step 1.1, define an empty clustering region E and an empty seed point sequence F. Select an initial seed point, add it to the seed point sequence, and search for the neighboring points of the seed point. Calculate the difference between the normal of each neighboring point and the color of the seed point. If the difference is less than the set color threshold, add the neighboring point to E; otherwise, add it to F. Select a new seed point in F and repeat the above steps until the sequence in F is empty. At this point, one region has been grown. Repeat the above process, check the average color of adjacent clusters, and if the average color threshold is less than the set threshold, merge the two clusters into one cluster. Continue until all color clustering is completed, thus completing the secondary clustering. Step 5 specifically involves: Step 5.1: In order to speed up the retrieval, key points are extracted from the building element samples and building facades to be retrieved by using internal shape descriptors. The fast point feature histogram algorithm is used to calculate the feature descriptors of the building element samples and building facade scenes to be retrieved. Finally, the extracted feature descriptors and the position constraints of the building elements to be retrieved in the scene constitute local matching constraints. Step 5.2: Treat the point cloud of the building element sample to be retrieved as a sliding window. Use the sliding window to move axially and partially match the building element sample with the point cloud of the building facade scene to retrieve the target model. Given a threshold, start traversing the building element sample to be retrieved from the lower left corner to the right and from bottom to top. Establish a kd-tree structure to query the point to be retrieved and its nearest neighbor in the building element. Use Euclidean distance to measure similarity, thereby completing the retrieval of repetitive building elements.

2. The method for repetitive building element retrieval based on axial movement and local matching according to claim 1, characterized in that, Step 2 specifically involves: Step 2.1: Using the point cloud file selection function in the PCL point cloud library, input the results of the secondary clustering and enter the selection state; Step 2.2: Select an instance of a building element object, remove duplicate selected points during the process, and save the selected point cloud as a .pcd file to obtain a sample of the element to be retrieved.

3. The method for repetitive building element retrieval based on axial movement and local matching according to claim 1, characterized in that, Step 3 specifically involves: Step 3.1: Use the random sampling consensus algorithm to obtain the planar equation of the building facade and determine the main direction, i.e., the X-axis. Step 3.2: Based on the X-axis, use the principal component analysis algorithm to correct the X-axis, making each principal direction of the X-axis, Y-axis and Z-axis perpendicular to each other, thereby determining the axial direction of the Y-axis and Z-axis.

Images