A lane / tunnel pipeline extraction method based on the fusion of point cloud curvature and geometric features

Abstract

A method for extracting lane / tunnel pipelines based on the fusion of point cloud curvature and geometric features. The method includes calculating the curvature value of each point in the lane using the local curvature changes of point cloud data, combined with the point cloud normal vector and spatial distribution characteristics; setting a curvature threshold and combining it with a clustering algorithm to classify the point cloud data and identify point cloud regions that meet pipeline characteristics; optimizing pipeline extraction accuracy by combining the geometric form and curvature characteristics of the pipeline; and fusing the curvature and geometric features to segment and annotate the pipeline using a deep learning-based semantic segmentation and annotation algorithm. This method reduces the impact of noise and interference points, improves pipeline identification accuracy, and achieves accurate pipeline segmentation and annotation, providing more efficient and accurate data support for intelligent inspection, maintenance, and management of mining facilities.

Description

Technical Field

[0001] The present invention relates to a lane / tunnel pipeline extraction method based on the fusion of point cloud curvature and geometric features, belonging to the technical field of pipeline facility three-dimensional point cloud processing. Background Art

[0002] During mining operations, pipelines within tunnels, such as power supply lines, water supply lines, and communication lines, are crucial for supporting safe mine operations. These pipelines are often buried in complex terrain and can be deformed and damaged during mining, necessitating regular inspection and maintenance.

[0003] Traditional pipeline inspection methods mainly rely on manual inspection or simple machine equipment, which cannot meet the needs of high efficiency and accuracy. Manual inspection has problems such as low efficiency, high missed detection rate and susceptibility to human factors. In recent years, with the rapid development of deep learning and point cloud processing technology, three-dimensional point cloud data has become an important means of obtaining spatial information. In complex environments such as underground spaces, mine tunnels, and tunnels, point cloud data can provide rich three-dimensional geometric information, overcoming the limitations of traditional two-dimensional images in processing complex scenes. Point cloud-based deep learning technology, especially semantic segmentation and target detection methods, can extract effective geometric features from point clouds and classify and label them. This provides the possibility for intelligent recognition and accurate detection of pipeline facilities in tunnels. However, the existing point cloud-based pipeline recognition technology still has the following problems: (1) In complex environments such as mine tunnels, point cloud data contains a lot of noise, overlap and irregular geometric shapes, especially the spatial features of pipeline facilities and the surrounding environment are highly overlapped, resulting in unclear edge features of the facilities. Existing point cloud segmentation methods often find it difficult to accurately extract pipelines. Traditional point cloud processing methods mostly rely on global or single features and cannot effectively handle these complexities and noises; (2) Traditional point cloud processing methods usually focus on local features such as normal vectors, geometric shapes, etc., but ignore important geometric information such as curvature and cannot effectively distinguish pipelines from other types of facilities or environments. Summary of the Invention

[0004] The purpose of the present invention is to provide a lane / tunnel pipeline extraction method based on the fusion of point cloud curvature and geometric features. This method can reduce the influence of noise and interference points, improve the recognition accuracy of pipelines, and achieve accurate segmentation and labeling of pipelines, providing more efficient and accurate data support for intelligent detection, maintenance and management of mining facilities.

[0005] To achieve the above objectives, the present invention provides a lane / tunnel pipeline extraction method based on the fusion of point cloud curvature and geometric features, comprising the following steps:

[0006] S1. Calculate the curvature value of each point in the tunnel by using the local curvature change of the point cloud data and combining the point cloud normal vector and spatial distribution characteristics;

[0007] S2. Set the curvature threshold and combine it with the clustering algorithm to classify the point cloud data and identify the point cloud area that meets the pipeline characteristics;

[0008] S3. Optimize pipeline extraction accuracy by combining pipeline geometry and curvature features;

[0009] S4. Fuse curvature features and geometric features, and segment and label the pipeline using a semantic segmentation and labeling algorithm based on deep learning.

[0010] Furthermore, the specific method of S1 is: according to the different importance of local curvature changes, point cloud normal vectors and spatial distribution features of point cloud data, multi-dimensional feature fusion is performed through a splicing method to obtain the comprehensive feature value of each point cloud data, which is expressed as:

[0011] f i =[n i ,K i ,ρ(p i )];

[0012] Among them, n i is the normal vector, K i is the local curvature value, ρ(p i ) is the spatial distribution characteristic;

[0013] The spatial distribution characteristics are expressed as the distance between point clouds, which is divided into the distance between the pipeline and the inner wall of the lane / tunnel and the distance between the pipeline and the pipeline facilities. The calculation methods are as follows:

[0014] Distance d between pipeline and inner wall of lane / tunnel i Expressed as:

[0015]

[0016] Among them, a point in the point cloud is P i (x i ,y i ,z i ), the equation of the tunnel wall is f(x,y,z)=0;

[0017] The relative distance d between pipelines and pipeline facilities facility Expressed as:

[0018] d facility =‖P i -P facility ‖;

[0019] Among them, ‖P i -Pfacility ‖ is point P i To pipeline facility point P facility The Euclidean distance.

[0020] Furthermore, the method of S2 is as follows: the curvature threshold is selected based on the geometric characteristics of the pipeline structure by calculating the local curvature of the point cloud; similar points are clustered together by a clustering algorithm, and the point cloud area that meets the pipeline characteristics is identified. Based on the clustering results, the cluster that meets the pipeline characteristics is screened out; the local curvature of the point cloud is calculated by PCA. The specific calculation process is as follows:

[0021] S2.1. Use PCA to calculate the covariance matrix of the point cloud within a local neighborhood: For each point p in the point cloud, select a local neighborhood N(p) containing the point and organize the coordinate information of the neighborhood points into a matrix X:

[0022]

[0023] Where n is the number of points in the neighborhood, x i ,y i , z i is the spatial coordinate of the i-th point in the point cloud;

[0024] S2.2. Calculate the covariance matrix C:

[0025]

[0026] in, is the mean vector of the neighborhood points;

[0027] S2.3. Calculate the eigenvalues ​​of the covariance matrix and perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvalues ​​λ1, λ2, and λ3:

[0028] C.v j =λ j ·v j (j=1,2,3);

[0029] Among them, v j is the corresponding eigenvector, representing the main changes of the point cloud in the three main directions; λ1 is the variance along the direction of maximum variance, λ2 is the variance along the direction of second largest variance, and λ3 is the variance along the direction of minimum variance;

[0030] S2.4, the principal curvatures k1 and k2 correspond to the minimum and maximum eigenvalues ​​of the covariance matrix, respectively, which describe the degree of curvature of the local surface of the point cloud in the two main directions:

[0031] k1=λ1,k2=λ2;

[0032] The Gaussian curvature K is the product of two principal curvatures:

[0033] K=k1·k2=λ1·λ2;

[0034] If K = 0, it indicates that the point is on a flat or straight area; if K > 0, it indicates that the point is on a convex surface; if K < 0, the point is on a concave surface;

[0035] S2.5. Clustering similar points into a cluster through a clustering algorithm, focusing on neighborhood calculation, core point definition, cluster expansion, and density reachability. Through the neighborhood radius ∈ and the minimum number of points MinPts, the algorithm can automatically identify and expand clusters and handle complex data with large density variations.

[0036] S2.5.1. The neighborhood calculation method is: for a data point p, its neighborhood N(p) is defined as all points within a given neighborhood radius ε, that is:

[0037] N(p)={q∈D|||pq||≤ε};

[0038] Where D is the set of data points, p and q are two points in the point cloud, ||pq|| is the Euclidean distance between point p and point q, and ε is the neighborhood radius, which represents the maximum distance that p can affect.

[0039] The Euclidean distance calculation formula is:

[0040]

[0041] Where p=(x p ,y p ,z p ) and q=(x q ,y q ,z q ) are the coordinates of the three-dimensional point;

[0042] S2.5.2. A core point is defined as follows: If a point p is a core point, it is only if its neighborhood N(p) contains at least MinPts points:

[0043] |N(p)|≥MinPts;

[0044] Where |N(p)| represents the number of points contained in the neighborhood of point p, and MinPts is a preset parameter that represents the minimum number of neighborhood points required for a point to be considered a core point.

[0045] S2.5.3. After finding the core point, the cluster is expanded based on the density-based clustering method, and points in the neighborhood are iteratively added to the cluster. The expansion process is performed according to the following formula:

[0046] E = {p|p∈D, p is a core point or p is density-reachable};

[0047] Here, E represents the cluster being expanded. E is a set that contains all the points currently known to belong to a cluster. The expansion of the core point continues to explore its neighborhood and add new points until there are no more points that can be added.

[0048] S2.5.4 A point p is density-reachable if and only if it is a core point or is in the neighborhood of a core point and is connected from the core point through a series of density-reachable points.

[0049] Points p and q must be density-connected if and only if there exists a core point r such that both p and q are within the neighborhood of the core point r and are connected by r.

[0050] Furthermore, the method of S3 is: fitting the straight line portion of the pipeline using the least squares method, and obtaining the direction, length and position information of the pipeline by fitting several points in the point cloud;

[0051] S3.1. The relevant formula for straight line fitting is as follows:

[0052] For a set of points P1, P2, ..., P n The goal of the least squares method is to find a straight line y = mx + t such that:

[0053]

[0054] Among them, x i ,y i It's point P i The coordinates of , m and t are the parameters of the fitted line;

[0055] For the curved pipe section, the least squares method is used to obtain:

[0056]

[0057] Among them, a, b, h are fitting curve parameters;

[0058] S3.2. For a portion of the pipeline corresponding to the identified point cloud cluster C, the curvature is calculated as follows:

[0059]

[0060] Among them, k i is the curvature of point i, r i is the radius of curvature at that point, and are the tangent and normal of the point respectively;

[0061] S3.3. Set a curvature threshold based on the calculated curvature value. During the optimization process, use geometric information to determine the pipeline boundary. The curvature information is used to determine the curved portion of the pipeline. For the curved portion, the curvature trend is combined to determine whether it is a natural bend of the pipeline.

[0062] Furthermore, the specific process of S4 is as follows:

[0063] S4.1. Input the comprehensive feature value into the improved Pointnet++ network and extract the multi-scale local features of the point cloud through the Set Abstraction layer. Combine the local geometric information with the global information and extract rich feature information by sampling and aggregating features at different scales of the point cloud data. In this layer, the features of each point in the point cloud are processed by the MLP multi-layer perceptron to generate local features. Then the global feature representation of each sampling point is obtained through pooling operation;

[0064] S4.2, Feature Propagation layer is used to fuse local features at different levels with global features and restore the high-resolution structure of point cloud data;

[0065] S4.3. After feature extraction in the Set Abstraction and Feature Propagation layers, the improved PointNet++ classifies or segments the point cloud through the fully connected layer (FC). The FC layer is used to map the fused multimodal features to the final label space, thereby performing pipeline segmentation.

[0066]

[0067] Among them, y i It's point p i The classification or segmentation label, Softmax function is used to calculate the probability that each point belongs to each category;

[0068] S4.4. For each point p i , whose category probability is p i =[p1,p2,…,p k ], the loss function evaluates the segmentation effect by calculating the cross entropy of each point. The specific formula is:

[0069]

[0070] Where s is the total number of points in the point cloud, k is the number of categories, and y i,g It's point p i The true label of category g, p i,g is the probability predicted by the network that the point belongs to category g.

[0071] This invention accurately identifies and extracts pipelines within complex tunnels by fusing point cloud curvature features with geometric structure information. It further precisely segments and optimizes pipeline regions through point cloud region partitioning based on the DBSCAN clustering algorithm. Finally, an improved deep learning network model fuses curvature and geometric features, improving the accuracy of pipeline extraction and annotation, enabling high-quality pipeline classification in complex environments. Curvature information effectively distinguishes the geometric differences between pipelines and other facilities, while geometric features such as normal vectors and PCA enhance understanding of the spatial distribution and morphology of pipelines. Compared to traditional point cloud segmentation methods, this invention enables pipeline identification in dense, overlapping environments, improves the accuracy and robustness of pipeline extraction, and reduces mis-segmentation and missed segmentations. It consistently outputs accurate segmentation results, particularly in complex and dynamic mining environments. By combining curvature and geometric features, this deep learning-based point cloud segmentation method significantly improves efficiency in automated pipeline identification and segmentation tasks. This invention automatically identifies and annotates pipelines within tunnels, reducing manual intervention and improving work efficiency. The present invention is suitable for the rapid processing of large-scale point cloud data, meeting the needs for real-time and accurate pipeline monitoring and protection in modern mine and urban infrastructure management. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is a workflow diagram of the present invention;

[0073] Figure 2 This is a visualization diagram of the local point cloud normal vector of the present invention, wherein Figure (a) is a schematic diagram of the local pipeline normal vector, and Figure (b) is a schematic diagram of the local tunnel inner wall normal vector;

[0074] Figure 3 is a schematic diagram of point cloud curvature information of the present invention;

[0075] Figure 4 It is a curvature threshold and clustering algorithm workflow diagram of the present invention;

[0076] Figure 5 This is the result diagram of the curvature information clustering algorithm based on the present invention, wherein Figure (a) is the original point cloud information diagram of the pipeline, Figure (b) is the result diagram of the pipeline point cloud clustering processing, Figure (c) is the original point cloud information diagram of the working transmission line facility, and Figure (d) is the result diagram of the working transmission line facility point cloud clustering processing;

[0077] Figure 6 Schematic diagram of the mine tunnel point cloud dataset used to train the network model in the present invention, where (a) shows the labeled pipeline, (b) shows the transportation track facilities, (c) shows the lighting equipment line facilities, and (d) and (e) show the working power transmission line facilities;

[0078] Figure 7 This is a workflow diagram of the deep learning network based on the fusion of curvature features and geometric features of the present invention. DETAILED DESCRIPTION

[0079] The present invention will be further described below with reference to the accompanying drawings.

[0080] like Figure 1 As shown in FIG, a lane / tunnel pipeline extraction method based on the fusion of point cloud curvature and geometric features includes the following steps:

[0081] S1. Calculate the curvature value of each point in the tunnel by using the local curvature change of the point cloud data and combining the point cloud normal vector and spatial distribution characteristics;

[0082] S2. Set the curvature threshold and combine it with the clustering algorithm to classify the point cloud data and identify the point cloud area that meets the pipeline characteristics;

[0083] S3. Optimize pipeline extraction accuracy by combining pipeline geometry and curvature features;

[0084] S4. Fuse curvature features and geometric features, and segment and label the pipeline using a semantic segmentation and labeling algorithm based on deep learning.

[0085] As a preferred embodiment, in step S1, representative geometric features in the point cloud are extracted by normal vector estimation and local curvature calculation. After loading the point cloud data, the point cloud is analyzed using PCA principal component analysis to obtain the local features of each point. In order to improve the accuracy of normal vector estimation, the least squares method based on neighborhood points is used to calculate the normal vector. On this basis, the local neighborhood of each point is analyzed by PCA to extract the main direction and curvature features of the point cloud. During the PCA calculation, the eigenvalues ​​of the covariance matrix of the neighborhood points are taken, and the minimum eigenvalue reflects the curvature information of the point in the neighborhood. Through this process, the key features in the point cloud are efficiently extracted, providing basic data for subsequent pipeline extraction and optimization. The calculated normal vector and curvature information also provide valuable input for the deep learning model, further improving the accuracy of pipeline segmentation and annotation. The specific method of S1 is: in view of the different importance of local curvature changes, point cloud normal vectors and spatial distribution features of point cloud data, multi-dimensional feature fusion is performed through splicing method to obtain the comprehensive eigenvalues ​​of each point cloud data, expressed as:

[0086] f i =[n i ,K i ,ρ(p i )];

[0087] Among them, n i is the normal vector, Ki is the local curvature value, ρ(p i ) is the spatial distribution characteristic;

[0088] The spatial distribution characteristics are expressed as the distance between point clouds, which is divided into the distance between the pipeline and the inner wall of the lane / tunnel and the distance between the pipeline and the pipeline facilities. The calculation methods are as follows:

[0089] Distance d between pipeline and inner wall of lane / tunnel i Expressed as:

[0090]

[0091] Among them, a point in the point cloud is P i (x i ,y i ,z i ), the equation of the tunnel wall is f(x,y,z)=0;

[0092] The relative distance d between pipelines and pipeline facilities facility Expressed as:

[0093] d facility =‖P i -P facility ‖;

[0094] Among them, ‖P i -P facility ‖ is point P i To pipeline facility point P facility The Euclidean distance.

[0095] like Figure 2 As shown in (a) and (b), the normal vector of the point cloud is calculated and visualized through the geometric distribution information of the point cloud data; the curvature difference of the point cloud data is visualized by calculating the local curvature, as shown in Figure 3 shown.

[0096] As a preferred embodiment, the method of S2 is as follows: the curvature threshold is selected based on the geometric characteristics of the pipeline structure by calculating the local curvature of the point cloud; similar points are clustered together by a clustering algorithm, and the point cloud area that meets the pipeline characteristics is identified. The cluster that meets the pipeline characteristics is screened out based on the clustering results; the local curvature of the point cloud is calculated by PCA. The specific calculation process is as follows:

[0097] S2.1. Use PCA to calculate the covariance matrix of the point cloud within a local neighborhood: For each point p in the point cloud, select a local neighborhood N(p) containing the point and organize the coordinate information of the neighborhood points into a matrix X:

[0098]

[0099] Where n is the number of points in the neighborhood, x i ,y i , z i is the spatial coordinate of the i-th point in the point cloud;

[0100] S2.2. Calculate the covariance matrix C:

[0101]

[0102] in, is the mean vector of the neighborhood points;

[0103] S2.3. Calculate the eigenvalues ​​of the covariance matrix and perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvalues ​​λ1, λ2, and λ3:

[0104] C.v j =λ j ·v j (j=1,2,3);

[0105] Among them, v j is the corresponding eigenvector, representing the main changes of the point cloud in the three main directions; λ1 is the variance along the direction of maximum variance, λ2 is the variance along the direction of second largest variance, and λ3 is the variance along the direction of minimum variance;

[0106] S2.4, the principal curvatures k1 and k2 correspond to the minimum and maximum eigenvalues ​​of the covariance matrix, respectively, which describe the degree of curvature of the local surface of the point cloud in the two main directions:

[0107] k1=λ1,k2=λ2;

[0108] The Gaussian curvature K is the product of two principal curvatures:

[0109] K=k1·k2=λ1·λ2;

[0110] If K = 0, it indicates that the point is on a flat or straight area; if K > 0, it indicates that the point is on a convex surface; if K < 0, the point is on a concave surface;

[0111] S2.5. Clustering similar points into a cluster through a clustering algorithm. This is done around neighborhood calculation, core point definition, cluster expansion, and density reachability. Using the neighborhood radius ε and the minimum number of points MinPts, the algorithm can automatically identify and expand clusters and handle complex data with large density variations.

[0112] S2.5.1. The neighborhood calculation method is: for a data point p, its neighborhood N(p) is defined as all points within a given neighborhood radius ε, that is:

[0113] N(p)={q∈D|||pq||≤ε};

[0114] Where D is the set of data points, p and q are two points in the point cloud, ||pq|| is the Euclidean distance between point p and point q, and ε is the neighborhood radius, which represents the maximum distance that p can affect.

[0115] The Euclidean distance calculation formula is:

[0116]

[0117] Where p=(x p ,y p ,z p ) and q=(x q ,y q ,z q ) are the coordinates of the three-dimensional point;

[0118] S2.5.2. A core point is defined as follows: If a point p is a core point, it is only if its neighborhood N(p) contains at least MinPts points:

[0119] |N(p)|≥MinPts;

[0120] Where |N(p)| represents the number of points contained in the neighborhood of point p, and MinPts is a preset parameter that represents the minimum number of neighborhood points required for a point to be considered a core point.

[0121] S2.5.3. After finding the core point, the cluster is expanded based on the density-based clustering method, and points in the neighborhood are iteratively added to the cluster. The expansion process is performed according to the following formula:

[0122] E = {p|p∈D, p is a core point or p is density-reachable};

[0123] Here, E represents the cluster being expanded. E is a set that contains all the points currently known to belong to a cluster. The expansion of the core point continues to explore its neighborhood and add new points until there are no more points that can be added.

[0124] S2.5.4 A point p is density-reachable if and only if it is a core point or is in the neighborhood of a core point and is connected from the core point through a series of density-reachable points.

[0125] Points p and q must be density-connected if and only if there exists a core point r such that both p and q are within the neighborhood of the core point r and are connected by r.

[0126] Through this step, the area that meets the pipeline characteristics can be effectively separated from the complex point cloud data. Figure 4The figure shows the flow chart of curvature threshold and clustering algorithm module; Figure 5 As shown in the figure, the original point cloud data obtained by the clustering algorithm and the point cloud data after clustering are displayed, where: Figure 5 (a) is the original point cloud information map of the pipeline, (b) is the result map of the pipeline point cloud clustering processing, (c) is the original point cloud information map of the working transmission line facilities, and (d) is the result map of the working transmission line facility point cloud clustering processing; Figure 6 The figure shows a mine tunnel point cloud data set trained by the method of the present invention, wherein: Figure 6 (a) is a marked pipeline diagram, (b) is a transportation track facility diagram, (c) is a lighting equipment line facility diagram, (d) and (e) are working power transmission line facility diagrams;

[0127] As a preferred embodiment, the method of S3 is: using the least squares method to fit the straight line portion of the pipeline, and obtaining the direction, length and position information of the pipeline by fitting several points in the point cloud;

[0128] S3.1. The relevant formula for straight line fitting is as follows:

[0129] For a set of points P1, P2, ..., P n The goal of the least squares method is to find a straight line y = mx + t such that:

[0130]

[0131] Among them, x i ,y i It's point P i The coordinates of , m and t are the parameters of the fitted line;

[0132] For the curved pipe section, the least squares method is used to obtain:

[0133]

[0134] Among them, a, b, and h are fitting curve parameters;

[0135] S3.2. For a portion of the pipeline corresponding to the identified point cloud cluster C, the curvature is calculated as follows:

[0136]

[0137] Among them, k i is the curvature of point i, r i is the radius of curvature at that point, and are the tangent and normal of the point respectively;

[0138] S3.3. Set a curvature threshold based on the calculated curvature value. During the optimization process, use geometric information to determine the pipeline boundary. The curvature information is used to determine the curved portion of the pipeline. For the curved portion, the curvature trend is combined to determine whether it is a natural bend of the pipeline.

[0139] This step can refine and optimize the pipeline extraction results, improve the pipeline segmentation accuracy, and provide a more accurate basis for subsequent pipeline labeling and identification.

[0140] As a preferred embodiment, Figure 7 As shown, the specific process of S4 is:

[0141] S4.1. Input the comprehensive feature value into the improved Pointnet++ network and extract the multi-scale local features of the point cloud through the Set Abstraction layer. Combine the local geometric information with the global information and extract rich feature information by sampling and aggregating features at different scales of the point cloud data. In this layer, the features of each point in the point cloud are processed by the MLP multi-layer perceptron to generate local features. Then the global feature representation of each sampling point is obtained through pooling operation;

[0142] S4.2, Feature Propagation layer is used to fuse local features at different levels with global features and restore the high-resolution structure of point cloud data;

[0143] S4.3. After feature extraction in the Set Abstraction and Feature Propagation layers, the improved PointNet++ classifies or segments the point cloud through the fully connected layer (FC). The FC layer is used to map the fused multimodal features to the final label space, thereby performing pipeline segmentation.

[0144]

[0145] Among them, y i It's point p i The classification or segmentation label, Softmax function is used to calculate the probability that each point belongs to each category;

[0146] S4.4. For each point p i , whose category probability is p i =[p1,p2,…,p k ], the loss function evaluates the segmentation effect by calculating the cross entropy of each point. The specific formula is:

[0147]

[0148] Where s is the total number of points in the point cloud, k is the number of categories, and y i,g It's point p i The true label of category g, p i,g is the probability predicted by the network that the point belongs to category g.

Claims

1. A lane / tunnel pipeline extraction method based on the fusion of point cloud curvature and geometric features, characterized in that: The steps include: S1. Calculate the curvature value of each point in the tunnel by using the local curvature change of the point cloud data and combining the point cloud normal vector and spatial distribution characteristics; S2. Set the curvature threshold and combine it with the clustering algorithm to classify the point cloud data and identify the point cloud area that meets the pipeline characteristics; S3. Optimize pipeline extraction accuracy by combining pipeline geometry and curvature features; S4. Fusing curvature features and geometric features, the pipeline is segmented and labeled using a semantic segmentation and labeling algorithm based on deep learning. The specific method of S1 is: according to the different importance of local curvature changes, point cloud normal vectors and spatial distribution features of point cloud data, multi-dimensional feature fusion is performed through the splicing method to obtain the comprehensive feature value of each point cloud data, which is expressed as: ； in, is the normal vector, is the local curvature value, is the spatial distribution feature; The spatial distribution characteristics are expressed as the distance between point clouds, which is divided into the distance between the pipeline and the inner wall of the lane / tunnel and the distance between the pipeline and the pipeline facilities. The calculation methods are as follows: Distance between pipeline and roadway / tunnel inner wall Expressed as: ； Among them, a point in the point cloud is , the equation of the tunnel wall is ; Relative distance between pipelines and pipeline facilities Expressed as: ； in, for point To pipeline facility The Euclidean distance of The method of S2 is as follows: the curvature threshold is selected based on the geometric characteristics of the pipeline structure by calculating the local curvature of the point cloud; similar points are clustered together by a clustering algorithm, and the point cloud area that meets the pipeline characteristics is identified. The cluster that meets the pipeline characteristics is screened out based on the clustering results; the local curvature of the point cloud is calculated by PCA. The specific calculation process is as follows: S2.

1. Use PCA to calculate the covariance matrix of the point cloud in a local neighborhood: For each point p in the point cloud, select a local neighborhood N(p) containing the point and organize the coordinate information of the neighborhood points into a matrix X : ； in, is the number of points in the neighborhood, It is the first The spatial coordinates of a point; S2.

2. Calculate the covariance matrix C: ； in, is the mean vector of the neighborhood points; S2.

3. Calculate the eigenvalues ​​of the covariance matrix and perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvalues. and : ； in, Represents a point in the point cloud or a sample in the point cloud, is the corresponding eigenvector, representing the main changes of the point cloud in the three main directions; is the variance along the direction of maximum variance, is the variance along the direction of the second largest variance, is the variance along the direction of minimum variance; S2.4, principal curvature and They correspond to the minimum and maximum eigenvalues ​​of the covariance matrix, respectively, which describe the curvature of the local surface of the point cloud in the two main directions: ； Gaussian curvature is the product of the two principal curvatures: ； like K =0, indicating that the point is in a flat or straight area; if K >0, indicating that the point is located on a convex surface; if K <0, the point is located on a concave surface; S2.5, cluster similar points into a cluster through clustering algorithm, focusing on neighborhood calculation, core point definition, cluster expansion and density reachability, and using neighborhood radius and minimum points ,The algorithm can automatically identify and expand clusters and handle complex data with large density variations; S2.5.

1. The neighborhood calculation method is: for a data point , its neighborhood Defined as the radius of the neighborhood All points within, that is: ； in, is a collection of data points, are two points in the point cloud, Yes and point The Euclidean distance between is the neighborhood radius, indicating The maximum distance that can be affected; The Euclidean distance calculation formula is: ； in, and are the coordinates of a three-dimensional point; S2.5.2, the definition of core point is: if the point is a core point if and only if its neighborhood Contains at least Points: ； in, Indicates a point The number of points contained in the neighborhood of It is a preset parameter, indicating the minimum number of neighborhood points required for a point to be considered a core point; S2.5.

3. After finding the core point, the cluster is expanded based on the density-based clustering method, and points in the neighborhood are iteratively added to the cluster. The expansion process is performed according to the following formula: ； in, Represents the cluster being expanded, which is a set containing all the points currently known to belong to a cluster; the expansion of the core point continues to explore its neighborhood and add new points until there are no more points that can be added; S2.5.4, Point Is density-reachable if and only if the point Is the core point or point In the neighborhood of a core point, and connected from the core point through a series of density-reachable points; point and point It is a density connection that must be satisfied if and only if there is a core point , so that the point and point All are core points points in the neighborhood, and both pass through Connect them.

2. The lane / tunnel pipeline extraction method based on the fusion of point cloud curvature and geometric features according to claim 1 is characterized in that: The method of S3 is: using the least squares method to fit the straight part of the pipeline, and obtaining the direction, length and position information of the pipeline by fitting several points in the point cloud; S3.

1. The relevant formula for straight line fitting is as follows: For a set of points , the goal of the least squares method is to find a straight line , such that: ； in, Yes The coordinates of are the parameters of the fitted straight line; For the curved pipe section, the least squares method is used to obtain: ； in, are the fitting curve parameters; S3.

2. For the identified point cloud clusters For a certain part of the pipeline, the curvature is calculated as: ； in, for point The curvature of for point The radius of curvature, Points The tangent and normal of ; S3.

3. Set a curvature threshold based on the calculated curvature value. During the optimization process, use geometric information to determine the pipeline boundary. The curvature information is used to determine the curved portion of the pipeline. For the curved portion, the curvature trend is combined to determine whether it is a natural bend of the pipeline.

3. The lane / tunnel pipeline extraction method based on the fusion of point cloud curvature and geometric features according to claim 2 is characterized in that: The specific process of S4 is: S4.

1. Input the comprehensive feature value into the improved Pointnet++ network and extract the multi-scale local features of the point cloud through the Set Abstraction layer; Combine local geometric information with global information, and extract rich feature information by sampling and feature aggregation at different scales of point cloud data. In this layer, the features of each point in the point cloud are processed by the MLP multi-layer perceptron to generate local features. , and then the global feature representation of each sampling point is obtained through pooling operation; S4.2, Feature Propagation layer is used to fuse local features at different levels with global features and restore the high-resolution structure of point cloud data; S4.

3. After feature extraction in the Set Abstraction and Feature Propagation layers, the improved PointNet++ classifies or segments the point cloud through the fully connected layer (FC). The FC layer is used to map the fused multimodal features to the final label space, thereby performing pipeline segmentation. ； in, Yes The classification or segmentation label, The function is used to calculate the probability that each point belongs to each category; S4.

4. For each point , and its category probability is , the loss function evaluates the segmentation effect by calculating the cross entropy of each point. The specific formula is: ； in, is the total number of points in the point cloud, is the number of categories, Yes Belong to category The true label, is the point predicted by the network Belong to category probability.

Images