Building roof and wall point cloud classification and identification method based on UAV technology

Abstract

The invention relates to the technical field of point cloud identification, and discloses a building roof and wall point cloud classification and identification method based on UAV technology, comprising the following steps: step 1, obtaining building point cloud data by adopting a UAV scanning mode; 2, constructing a multi-scale super voxel fusion model according to the point cloud data obtained in the step 1, and fusing the point cloud; and step 3, recognizing building point clouds of different surfaces according to the point clouds fused by the multi-scale super voxel fusion model in the step 2. According to the invention, the method focuses on the utilization of the normal, height, point cloud density and point number to achieve the recognition of the UAV building point cloud wall and roof surface, thereby avoiding the problem that the wall point cloud is too sparse and cannot be recognized, and remarkably improving the recognition efficiency and refinement degree of the UAV building point cloud. The method is suitable for extraction of wall surfaces and roof surfaces of large-scale scenes and a large number of building point clouds acquired by the UAV.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud recognition, and in particular to a method for classifying and recognizing point clouds of building roofs and walls based on UAV technology. Background Art

[0002] With the rapid development of UAV technology and LiDAR technology, the use of UAV (unmanned aerial vehicle, i.e. drone) equipped with LiDAR to collect high-precision building point cloud has become an important means in the fields of building information modeling, urban planning and intelligent buildings. Point cloud data, with its high resolution and accuracy, provides strong support for the reproduction of three-dimensional buildings. However, due to the high density, large scale and complex distribution of point cloud data collected by UAV, and the large difference in the quality of point cloud of building roofs and walls, traditional building point cloud recognition methods are not applicable, resulting in unsatisfactory recognition of building roofs and walls.

[0003] The existing UAV building point cloud roof and wall recognition methods are mainly based on the following technologies: First, the plane recognition method based on regular RANSAC, which is suitable for buildings with simple structures, but is easily affected by point cloud quality and recognition accuracy when dealing with complex building structures, resulting in insufficient model accuracy; second, the method based on region growing. Although this type of method can handle complex structures, it involves more parameters, and different parameters have a greater impact on the recognition results; third, the method based on supervoxels, which has a higher recognition accuracy, but is currently not applicable to the recognition of UAV wall point clouds.

[0004] In response to the above technical problems, there is an urgent need for an efficient, robust and applicable UAV technology-based building roof and wall point cloud classification and recognition method for complex building structures. Summary of the invention

[0005] In order to solve the technical problems raised in the background technology, the present invention provides a method for classifying and identifying building roofs and wall point clouds based on UAV technology. The method should be able to make full use of building height information, roof and wall direction information, and roof point cloud density information to achieve accurate identification of complex building roofs and walls, while having high efficiency and robustness to meet the needs of rapid processing of large-scale point cloud data.

[0006] The present invention is implemented by the following technical solution: A method for classifying and identifying building roof and wall point clouds based on UAV technology includes the following steps: Step 1, using UAV scanning to obtain building point cloud data; Step 2, constructing a multi-scale super-voxel fusion model based on the point cloud data obtained in step 1, and fusing the point cloud; Step 3, identifying building point clouds of different faces according to the point clouds fused by the multi-scale super voxel fusion model in step 2; Step 4, calculating the normal direction of the building point clouds of different faces identified in step 3; and identifying the wall point cloud based on the normal direction information; Step 5, after removing the wall point cloud data in step 4, the point cloud data of the remaining identification surface is obtained, and a probability model of height, density and number of unit points is constructed for the point cloud data of the remaining identification surface, and finally the building roof point cloud is extracted from the remaining identification surface point cloud data.

[0007] Specifically, the construction operation flow of the multi-scale super-voxel fusion model in step 2 is as follows: Step 2.1, constructing supervoxels for the UAV building point cloud data in step 1, and obtaining an initial supervoxel block point cloud, wherein the initial supervoxel block point cloud includes planar supervoxels and non-planar supervoxels; Step 2.2: Consider planar supervoxels and non-planar supervoxels as the basic units of the region growing algorithm, and construct a multi-scale supervoxel fusion algorithm.

[0008] Specifically, the steps of constructing a multi-scale super-voxel fusion algorithm are as follows: Step 2.21, determine the normal vector of the supervoxel block point cloud; for each point in the supervoxel block point cloud , select its nearest k neighbors, which constitute the point The local neighborhood of , and then calculate the mean of the points in the local neighborhood , as shown in formula (1): (1) According to formula (1), the covariance matrix is ​​constructed C , as shown in formula (2): (2) in, are the coordinates of the points in the neighborhood, is the average coordinate of these points; Covariance matrix C Perform eigenvalue decomposition and obtain the eigenvalue and its corresponding eigenvector , the minimum eigenvalue The corresponding eigenvector That is the estimated normal vector; calculate the normal vectors of all points in the initial supervoxel block The average That is, the normal vector of the supervoxel block point cloud. In order to ensure the consistency of the normal vector of the supervoxel block point cloud, it is necessary to orient the normal vector and select a reference direction. , such as the viewpoint direction or the external normal vector, check the normal vector With reference direction If the angle is greater than 90 degrees, the normal vector is reversed, as shown in formula (3): (3) Then the normal vector is normalized to make it unitary, as shown in formula (4): (4) Step 2.22: Calculate the curvature of the supervoxel point cloud. Curvature is a measure of the curvature of the supervoxel surface. The three eigenvalues ​​of the eigenvalue decomposition of formula (2) are: The curvature of each point in the supervoxel block point cloud is obtained as shown in formula (5); (5) Compute the curvature of all points in a supervoxel point cloud The average value of the final curvature of the supervoxel point cloud is ; Step 2.23, seed supervoxel selection; sort all supervoxel blocks in ascending order according to the curvature value, select the supervoxel block point cloud with the smallest curvature as the seed supervoxel, the supervoxel block point cloud with smaller curvature usually indicates that the supervoxel surface is relatively flat, which is suitable as the starting point of growth, and initialize the recognition result and the visited point set; Step 2.24, growth process: starting from the seed supervoxel, search for its nearest neighbor supervoxel in turn, and grow according to the distance threshold and normal angle threshold; if the growth condition is met, the neighboring supervoxel is added to the seed set and recognition result, and marked as visited, so as to obtain the recognition result of the planar supervoxel ; Step 2.25: For non-planar supervoxels , and the curvature of each non-planar supervoxel is obtained and normal vector , calculate the curvature of non-planar supervoxels Each plane supervoxel curvature is increased with area The difference between (6) Computing non-planar supervoxel normals Grow each plane supervoxel normal vector with the region The angle of (7) Simultaneously calculate the distance of each non-planar supervoxel to the nearest point of each planar supervoxel identification dataset ; According to formula (6), formula (7) and distance , we get the fusion judgment parameters of non-planar supervoxel and regional growth plane recognition supervoxel, as shown in formula (8): (8) in: is the curvature difference between the non-planar supervoxel and the neighboring planar supervoxel, is the difference in normal vector angle between the non-planar supervoxel and the neighboring planar supervoxel, is the shortest distance between a non-planar supervoxel and its neighboring planar supervoxel; Step 2.26, fusion result: when each non-planar supervoxel is transformed into a planar supervoxel recognition dataset When it is the smallest, it is fused into a super voxel block point cloud of the same surface.

[0009] Specifically, the operation of step 2.26 is as follows: when the plane supervoxel is , the plane supervoxel The neighborhood non-planar supervoxel is ; According to formula (8), calculate Neighborhood non-planar supervoxels With planar supervoxels Parameters between ; when When the minimum, non-planar supervoxel With planar supervoxels Blend together.

[0010] Specifically, in step 4, the surface recognition point cloud identifies the wall point cloud based on the normal information of the fused supervoxel block point cloud. When the normal direction is parallel to the xoy plane, it is determined to be a wall point cloud.

[0011] Specifically, the step 5 is performed as follows: After removing the wall point cloud, the remaining recognition surface point cloud is obtained , using the bounding sphere algorithm to determine the radius The number of points in each bounding sphere of , get the point cloud density of each face recognition , as shown in formula (9).

[0012] (9) in, The number of points for each identified face.

[0013] According to formula (9), the normalized result is obtained as shown in formula (10): (10) At the same time, the average height of each recognition surface point cloud is calculated, as shown in formula (11): (11) According to formula (11), the normalized result is obtained as shown in formula (12): (12) The number of points for each identified face Normalize it, as shown in formula (13): (13) According to equations (10), (12) and (13), the probability of extracting the building roof point cloud is obtained, as shown in equation (14): (14) The probability of formula (14) is used to extract the building roof point cloud from the remaining recognition surface point cloud.

[0014] Specifically, when the probability , it is identified as a building roof point cloud.

[0015] The present invention also proposes an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the above-mentioned UAV technology-based building roof and wall point cloud classification and recognition method.

[0016] A computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the above-mentioned UAV-based building roof and wall point cloud classification and recognition method.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The method in this paper can quickly realize the recognition of building point clouds, and the recognized building surface point clouds are more detailed. The recognition method proposed in this invention is more robust and not easily affected by multiple building surfaces. Therefore, the method proposed in this invention is better than the RANSAC method and the region growing method in the degree of refinement of building point cloud recognition.

[0018] The present invention focuses on using normal, height, point cloud density and number of points to realize the recognition of wall and roof surfaces of building point clouds, thereby avoiding the problem that wall point clouds are too sparse to be recognized. The present invention performs roof and wall point cloud recognition on building point clouds acquired based on UAV technology. It does not involve redundant algorithms and programs, and significantly improves the efficiency and refinement of UAV building point cloud recognition, so it is suitable for the extraction of wall and roof surfaces of a large number of building point clouds in large scenes collected by UAV. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1This is a principle flow chart of the building roof and wall point cloud classification and recognition method proposed by the present invention; Figure 2 Comparison of UAV building roof point clouds extracted by the recognition method, RANSAC method and region growing method proposed in the present invention; Figure 3 A point cloud image of a building roof extracted by the recognition method proposed in the present invention; Figure 4 The wall and roof point cloud images in the building point cloud extracted by the recognition method proposed in the present invention. DETAILED DESCRIPTION

[0020] The present invention is further described below in conjunction with the accompanying drawings and specific implementation methods. It should be noted that, under the premise of no conflict, the various embodiments or technical features described below can be arbitrarily combined to form a new embodiment.

[0021] Example Please combine Figure 1-Figure 4 The method for classifying and identifying building roof and wall point clouds based on UAV technology proposed in this embodiment includes the following steps: Step 1, using UAV scanning to obtain building point cloud data; Step 2, constructing a multi-scale super-voxel fusion model based on the point cloud data obtained in step 1, and fusing the point cloud; Step 3, identifying building point clouds of different faces according to the point clouds fused by the multi-scale super voxel fusion model in step 2; Step 4: Calculate the normal direction of the different surface point clouds identified in step 3; and realize the wall Recognition of surface point clouds; Step 5, removing the wall point cloud in step 4 to obtain the remaining recognition surface point cloud; Step 6: construct a probability model of height, density and number of unit points for the remaining identified surface point cloud in step 5.

[0022] It should be noted that: The UAV building point cloud data in step 1 is constructed with supervoxels to obtain the initial supervoxel block point cloud, which includes planar supervoxels and non-planar supervoxels. Planar supervoxels and non-planar supervoxels are regarded as the basic units of the region growing algorithm. The multi-scale supervoxel fusion algorithm is constructed with the following specific steps: Determine the normal vector of the supervoxel block point cloud: For each point in the supervoxel block point cloud , select its nearest k neighboring points. These points constitute the point Then calculate the mean of the points in the local neighborhood. , as shown in formula (1).

[0023] (1) According to formula (1), the covariance matrix is ​​constructed C , as shown in formula (2): (2) in, are the coordinates of the points in the neighborhood, is the average coordinate of these points.

[0024] Covariance matrix C Perform eigenvalue decomposition and obtain the eigenvalue and its corresponding eigenvector . Minimum eigenvalue The corresponding eigenvector is the estimated normal vector.

[0025] Compute all normal vectors of points in the initial supervoxel block The average This is the supervoxel block point cloud normal vector. In order to ensure the consistency of the supervoxel block point cloud normal vector, the normal vector needs to be oriented. Select a reference direction , such as the viewpoint direction or the external normal vector. Check the normal vector With reference direction If the angle is greater than 90 degrees, the normal vector is reversed, as shown in formula (3): (3) Then the normal vector is normalized to make it unitary, as shown in formula (4): (4) Curvature calculation of supervoxel point cloud: Curvature is a measure of the curvature of the supervoxel surface. The three eigenvalues ​​of eigenvalue decomposition of formula (2) are The curvature of each point in the supervoxel block point cloud is obtained as shown in formula (5): (5) Compute the curvature of all points in a supervoxel point cloud The average value of the final curvature of the supervoxel point cloud is .

[0026] Seed supervoxel selection: Sort all supervoxel blocks in ascending order according to the curvature value, and select the supervoxel block point cloud with the smallest curvature as the seed supervoxel. The supervoxel block point cloud with smaller curvature usually indicates that the supervoxel surface is relatively flat, which is suitable as the starting point for growth and initializing the recognition results and the set of visited points.

[0027] Growth process: Starting from the seed supervoxel, search for its nearest neighbor supervoxel in turn, and grow according to the distance threshold and normal angle threshold. If the growth conditions are met, the neighboring supervoxel is added to the seed set and recognition result, and marked as visited. Thus, the recognition result of the planar supervoxel is obtained. .

[0028] For non-planar supervoxels , and obtain the curvature of each non-planar supervoxel and normal vector . Compute the curvature of non-planar supervoxels Each plane supervoxel curvature is increased with area The difference: (6) Computing non-planar supervoxel normals Grow each plane supervoxel normal vector with the region Angle: (7) Simultaneously calculate the distance of each non-planar supervoxel to the nearest point of each planar supervoxel identification dataset .

[0029] According to formula (6) (7) and distance , we get the fusion judgment parameters of non-planar supervoxel and regional growth plane recognition supervoxel, as shown in formula (8): (8) in: is the curvature difference between the non-planar supervoxel and the neighboring planar supervoxel, is the difference in normal vector angle between the non-planar supervoxel and the neighboring planar supervoxel, is the shortest distance between a non-planar supervoxel and its neighboring planar supervoxel.

[0030] Fusion results: When each non-planar supervoxel is transformed into a planar supervoxel, the dataset is identified. When it is the smallest, they are fused into a super voxel block point cloud of the same surface.

[0031] For example, if the planar supervoxel is , the plane supervoxel The neighborhood non-planar supervoxel is According to formula (8), calculate Neighborhood non-planar supervoxels With planar supervoxels Parameters between .when When the minimum, non-planar supervoxel With planar supervoxels Blend together.

[0032] Therefore, the surface recognition of the building point cloud is achieved based on the above fusion results.

[0033] The building surface segmentation results obtained by this scheme are compared with the RANSAC method and the region growing method. Figure 2 As shown. Figure 2 The RANSAC method in part (a) is the same as Figure 2 By comparing the proposed solution with the solution in part (c), it can be seen that the wall of the building point cloud is successfully segmented into three small walls by the recognition method proposed in this solution, while the RANSAC method fails to segment the three small walls.

[0034] Depend on Figure 2 The region growing method in part (b) is similar to Figure 2 By comparing the scheme in part (c) with the present scheme, it can be seen that the region growing method not only fails to segment the right part of the roof of the building, but also produces a rather chaotic segmentation result.

[0035] And by Figure 2 It can be clearly seen from part (c) that the recognition method proposed in this scheme successfully segments the right part of the roof.

[0036] In this solution, for the face recognition point cloud, the wall recognition point cloud can be determined based on the normal direction of the fused supervoxel block point cloud, that is, when the normal direction is parallel to the xoy plane, it is judged to be a wall point cloud.

[0037] After removing the wall point cloud, the remaining recognition surface point cloud is obtained .

[0038] For the remaining identified surface point clouds , using the bounding sphere algorithm to determine the radius The number of points in each bounding sphere of .

[0039] Get the point cloud density of each face recognition , as shown in formula (9).

[0040] (9) in, The number of points for each identified face.

[0041] According to formula (9), the normalized result is obtained as shown in formula (10): (10) At the same time, the average height of each recognition surface point cloud is calculated, as shown in formula (11): (11) According to formula (11), the normalized result is obtained as shown in formula (12): (12) The number of points for each identified face Normalize it, as shown in formula (13): (13) According to equations (10), (12) and (13), the probability of extracting the building roof point cloud is obtained, as shown in equation (14): (14) The probability of formula (14) is used to extract the building roof point cloud from the remaining recognition surface point cloud. Specifically, in this embodiment, when the probability When , it is the point cloud of the building roof.

[0042] During specific implementation, the above process can be automatically operated using computer software technology.

[0043] The above-mentioned embodiments are only preferred embodiments of the present invention and cannot be used to limit the scope of protection of the present invention. Any non-substantial changes and substitutions made by technicians in this field on the basis of the present invention shall fall within the scope of protection required by the present invention.

Claims

1. A method for classifying and identifying building roof and wall point clouds based on UAV technology, characterized in that: The steps include: Step 1, using UAV scanning to obtain building point cloud data; Step 2, constructing a multi-scale super-voxel fusion model based on the point cloud data obtained in step 1, and fusing the point cloud; Step 3, identifying building point clouds of different faces according to the point clouds fused by the multi-scale super voxel fusion model in step 2; Step 4, calculating the normal direction of the building point clouds of different faces identified in step 3, and identifying the wall point clouds based on the normal direction information; Step 5, after removing the wall point cloud data in step 4, the point cloud data of the remaining identification surface is obtained, and a probability model of height, density and number of unit points is constructed for the point cloud data of the remaining identification surface, and finally the building roof point cloud is extracted from the remaining identification surface point cloud data.

2. The method for classifying and identifying building roof and wall point clouds based on UAV technology as claimed in claim 1, characterized in that: The construction operation flow of the multi-scale super-voxel fusion model in step 2 is as follows: Step 2.1, constructing supervoxels on the building point cloud data in step 1 to obtain an initial supervoxel block point cloud, wherein the initial supervoxel block point cloud includes planar supervoxels and non-planar supervoxels; Step 2.2: Consider planar supervoxels and non-planar supervoxels as the basic units of the region growing algorithm, and construct a multi-scale supervoxel fusion algorithm.

3. The method for classifying and identifying building roof and wall point clouds based on UAV technology as claimed in claim 1, characterized in that: The steps of constructing the multi-scale super-voxel fusion algorithm are as follows: Step 2.21, determine the normal vector of the supervoxel block point cloud; for each point in the supervoxel block point cloud , select its nearest k neighbors, which constitute the point The local neighborhood of , and then calculate the mean of the points in the local neighborhood , as shown in formula (1): ；（1） According to formula (1), the covariance matrix is ​​constructed C , as shown in formula (2): ；（2） in, are the coordinates of the points in the neighborhood, is the average coordinate of these points; Covariance matrix C Perform eigenvalue decomposition and obtain the eigenvalue and its corresponding eigenvector , the minimum eigenvalue The corresponding eigenvector , which is the estimated normal vector; Compute all normal vectors of points in the initial supervoxel block The average , which is the normal vector of the supervoxel point cloud; Orient the normal vector and choose a reference direction , such as the viewpoint direction or the external normal vector, check the normal vector With reference direction If the angle is greater than 90 degrees, the normal vector is reversed, as shown in formula (3): ；（3） Then the normal vector is normalized to make it unitary, as shown in formula (4): ；（4） Step 2.22: Calculate the curvature of the supervoxel point cloud. Curvature is a measure of the curvature of the supervoxel surface. The three eigenvalues ​​of the eigenvalue decomposition of formula (2) are: The curvature of each point in the supervoxel block point cloud is obtained as shown in formula (5); ；（5） Then calculate the curvature of all points in the supervoxel point cloud The average value of the final curvature of the supervoxel point cloud is ; Step 2.23, seed supervoxel selection: sort all supervoxel block point clouds in ascending order according to the curvature value, select the supervoxel block point cloud with the smallest curvature as the seed supervoxel, as the starting point of growth, and initialize the recognition result and the visited point set; Step 2.24, growth process: starting from the seed supervoxel, search for its nearest neighbor supervoxel in turn, and grow according to the distance threshold and normal angle threshold; if the growth condition is met, the neighboring supervoxel is added to the seed set and the recognition result, and marked as visited, so as to obtain the recognition result of the planar supervoxel; ； Step 2.25: For non-planar supervoxels , and the curvature of each non-planar supervoxel is obtained and normal vector , calculate the curvature of non-planar supervoxels Each plane supervoxel curvature is increased with area The difference between ；（6） Computing non-planar supervoxel normals Grow each plane supervoxel normal vector with the region The angle of ；（7） Simultaneously calculate the distance of each non-planar supervoxel to the nearest point of each planar supervoxel identification dataset ; According to formula (6), formula (7) and distance , we get the fusion judgment parameters of non-planar supervoxel and regional growth plane recognition supervoxel, as shown in formula (8): ；（8） in: is the curvature difference between the non-planar supervoxel and the neighboring planar supervoxel, is the difference in normal vector angle between the non-planar supervoxel and the neighboring planar supervoxel, is the shortest distance between a non-planar supervoxel and its neighboring planar supervoxel; Step 2.26, fusion result: when each non-planar supervoxel is transformed into a planar supervoxel recognition dataset When it is the smallest, it is fused into a super voxel block point cloud of the same surface.

4. The method for classifying and identifying building roof and wall point clouds based on UAV technology as claimed in claim 3, characterized in that: The operation of step 2.26 is as follows: when the plane supervoxel is , the plane supervoxel The neighborhood non-planar supervoxel is ; According to formula (8), calculate Neighborhood non-planar supervoxels With planar supervoxels Parameters between ; when When the minimum, non-planar supervoxel With planar supervoxels Blend together.

5. The method for classifying and identifying building roof and wall point clouds based on UAV technology as claimed in claim 1, characterized in that: In step 4, the wall point cloud is identified according to the normal information of the fused supervoxel block point cloud. Specifically, when the normal direction is parallel to the xoy plane, it is determined to be a wall point cloud.

6. The method for classifying and identifying building roof and wall point clouds based on UAV technology as claimed in claim 1, characterized in that: The step 5 is performed as follows: After removing the wall point cloud, the remaining recognition surface point cloud is obtained , using the bounding sphere algorithm to determine the radius The number of points in each bounding sphere of , get the point cloud density of each face recognition , as shown in formula (9); ； （9） in, is the number of points for each identified face; According to formula (9), the normalized result is obtained as shown in formula (10): ；（10） At the same time, the average height of each recognition surface point cloud is calculated, as shown in formula (11): ；（11） According to formula (11), the normalized result is obtained as shown in formula (12): ；（12） The number of points for each identified face Normalize it, as shown in formula (13): ；（13） According to equations (10), (12) and (13), the probability of extracting the building roof point cloud is obtained, as shown in equation (14): ；（14） The probability of formula (14) is used to extract the building roof point cloud from the remaining recognition surface point cloud.

7. The method for classifying and identifying building roof and wall point clouds based on UAV technology as claimed in claim 6, characterized in that: When the probability , it is identified as a building roof point cloud.

8. An electronic device, characterized in that: It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the building roof and wall point cloud classification and recognition method based on UAV technology as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the building roof and wall point cloud classification and recognition method based on UAV technology as described in any one of claims 1 to 7.

Images