A method for detecting deformation of tunnel cable supports based on laser radar
Through the lidar point cloud detection technology, the deformation of the tunnel cable bracket is identified and analyzed, and the problems of traditional manual detection are solved, and automated and rapid deformation detection and early warning are realized to ensure the reliable operation of the power grid system.
Patent Information
- Application Number
- CN202111248235.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-26
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2041-10-26
AI Technical Summary
The traditional manual detection of the deformation method of tunnel cable brackets is inefficient, consumes a lot of manpower, increases workers' safety risks, and it is difficult to warn of the cracks of the cable brackets in real time.
Using a point cloud detection method based on lidar, the point cloud data of the tunnel cable bracket is collected and processed, and the deformation area is identified using deep learning object detection algorithm and point cloud segmentation algorithm, the deformation skeleton points are extracted, and the spatial curve is formed, and the cable bracket is analyzed in real time. and provide early warning.
It realizes automated, fast and accurate deformation detection of tunnel cable brackets, and early warning of severe deformation in advance, reducing workers' safety risks and improving the reliability of the power grid system.
Smart Images

Figure CN114092797B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to intelligent tunnel inspection technology, and in particular to a laser radar-based tunnel cable support deformation detection method. Background Art
[0002] In real life, tunnel power grids require the support of cable racks. As the use time increases, the cable racks will gradually become bent, and may also be deformed or even broken due to collisions with other objects. Traditional manual detection methods require a lot of manpower, are time-consuming, inefficient, and increase safety risks for workers. Therefore, a non-manual and accurate detection method is needed to detect whether tunnel cable racks are deformed. Summary of the invention
[0003] In view of the defects of the prior art, the purpose of the present invention is to provide a deformation detection method for a tunnel cable bracket based on lidar, which is used to automatically detect whether the tunnel cable bracket is deformed and to warn of large deformation or even breakage of the cable bracket, and to detect deformation, faults and diseases of the cable bracket, thereby providing guarantee for the reliable operation of the entire power grid system.
[0004] In order to achieve the above object, the present invention adopts the following technical scheme:
[0005] A method for detecting deformation of a tunnel cable support based on laser radar comprises the following steps:
[0006] S1: Collect historical point cloud samples of relevant deformed cable supports in tunnels to make detection labels and build detection models;
[0007] S2: Store several frames of point cloud data in real time, and use the point cloud multi-frame registration and stitching algorithm to stitch the currently measured point cloud set together and add it to the detection candidate area;
[0008] S3: Using the point cloud detection model and the deep learning target detection algorithm, the deformation detection area of the cable bracket is identified and added to the segmentation candidate area;
[0009] S4: Use the point cloud segmentation algorithm to segment the point cloud in the segmentation candidate area and add it to the measurement set;
[0010] S5: Use the point cloud kmeans skeleton extraction technology to extract the deformation skeleton points from the point set of the measurement set to form a space curve;
[0011] S6: Measure the space curve to analyze the deformation of the cable support.
[0012] Furthermore, in step S1, collecting historical point cloud samples of relevant deformed cable supports in the tunnel to make detection labels and construct a detection model includes the following steps:
[0013] S11: Collect point cloud data of the deformed cable bracket in the tunnel to check whether there are deformations in multiple consecutive frames; use the point cloud multi-frame registration and splicing algorithm to splice the point cloud that has deformations in multiple consecutive frames;
[0014] S12: performing data enhancement on the point cloud data of the deformed cable bracket to obtain enhanced point cloud data;
[0015] S13: labeling the deformed area in the enhanced point cloud data as a training data set, and dividing the deformation in the point cloud data set into four labels: small, medium, large, and maintenance, as detection labels;
[0016] S14: Construct a point cloud deep learning detection network pointpillars model for detecting deformed cable brackets, and use the enhanced point cloud data and labels as the input of the network framework to train this deep learning network model to obtain a tunnel cable bracket deformation detection model after deep learning.
[0017] Furthermore, the step S14 constructs a point cloud deep learning detection network pointpillars model for detecting deformed cable brackets, and uses the enhanced point cloud data and labels as network input to train the deep learning network model to obtain a tunnel cable bracket deformation detection model after deep learning, including the following steps:
[0018] S141: performing regional cropping on the collected point cloud data and fixing it within a space with a certain size range of [X, Y, Z];
[0019] S142: The point cloud space is divided into two parts according to the volume size V (w,l,H) The voxels are divided into voxels, where w is the voxel width, l is the voxel length, H is the voxel height, and H = Y is a fixed value; the point cloud data in the training sample are discretized on evenly spaced grids on the XOZ plane, with no restrictions in the Y direction, and a series of cylinders on the XOZ plane are created. The point cloud in each cylinder is sampled, and the voxels with more than N points are uniformly downsampled to N points, and the voxels with less than N points are filled with 0 to N points; the point cloud contained in each cylinder is represented by r, x c ,y c , z c , x p , z p Feature expansion, get the expanded point cloud features (x, y, z, r, x c ,y c , z c , x p , z p ), where x, y, and z represent the initial coordinates of the point cloud; r represents the reflectivity of the point cloud; x c ,y c, z c It represents the coordinate value obtained by taking the arithmetic mean of all the point cloud coordinates in the pillar; p , z p Indicates the deviation of all point clouds in the pillar relative to the coordinate center position in the coordinate system of the current plane; on the XOZ plane, the number of point clouds contained in all non-empty pillars is adjusted to be consistent, and then a dense tensor (D, P, N) is created according to the number of non-empty pillars on the plane, the number of point clouds contained in the non-empty pillars, and the features of the point clouds in the non-empty pillars, that is, the features (D, P, N) of each non-empty pillar in the XOZ plane are obtained, where D represents the feature dimension of the point cloud in the non-empty pillar, P represents the number of non-empty pillars on the XOZ plane, and N represents the number of point clouds contained in the non-empty pillars;
[0020] S143: Use the pointnet network to perform feature learning on the features (D, P, N) of the non-empty pillars on the XOZ plane, and obtain the final features (C, P, N) of the point cloud in each non-empty pillar in the XOZ plane after learning; wherein C represents the new feature dimension of the point cloud after learning by the pointnet network, P represents the number of non-empty pillars on the XOZ plane, and N represents the number of point clouds contained in the non-empty pillars; after the operation, the final features (C, P, N) of the point cloud of each non-empty pillar in the XOZ plane are obtained, and the features (C, P, N) are processed by the maximum pooling operation to obtain a tensor (C, P), and then a sparse pseudo image (C, L, W) is created according to the tensor (C, P), wherein L represents the height of the sparse pseudo image, and W represents the width of the sparse pseudo image;
[0021] S144: Use the Backbone feature extraction network as the backbone network for extracting feature maps to achieve accurate detection, including: using a series of convolutional layers to extract multi-scale features from sparse pseudo images of scale S, which contains a total of three scale feature maps F 11 , F 12 , F 13 , whose sizes are S, S / 2, and S / 4 respectively. The feature maps of these three scales are then pooled and deconvolved to generate three independent pseudo feature maps. They are then fused using the concatenate operation and the 1×1 convolution layer to obtain the final feature map for detection. Finally, three 1×1 convolution layers are used to predict the object category and 3D bounding box respectively.
[0022] Furthermore, in step S2, several frames of point cloud data are stored in real time, and the currently measured point cloud set is added to the detection candidate area using the point cloud multi-frame registration and splicing algorithm, including the following steps:
[0023] S21: real-time storage of point cloud data of the cable support deformation area;
[0024] S22: Using a straight-through filter to crop the point cloud range of the point cloud data to obtain point cloud data within a fixed area;
[0025] S23: using statistical filtering to remove outliers in the point cloud after regional clipping, so as to reduce the point cloud data in the region that interferes with the detection;
[0026] S24: The first frame of data stored is marked as the target point cloud, and the frames stored thereafter are the source point clouds. Each frame of data is filtered, and the source point cloud and the target point cloud are registered using a point cloud multi-frame registration algorithm to determine the rigid body transformation parameters. Then, the transformation is performed according to the rigid body transformation parameters, and this step is repeated over multiple frames to synthesize a target point cloud set, which is added to the point cloud target detection area.
[0027] Furthermore, in step S3, the point cloud detection model is used to use a deep learning target detection algorithm to identify the deformation detection area of the cable bracket and add it to the segmentation candidate area, including the following steps:
[0028] S31: using the point cloud in the target detection area obtained in step S2 as input of the target detection algorithm;
[0029] S32: Detect the deformation of the target area. If there is no deformation, delete the current point cloud space;
[0030] S33: If there is deformation, mark the deformation levels according to the degree of deformation, small, medium and large, and record the detection results for the next detection;
[0031] S34: For the maintenance level, segmentation candidate regions are added to facilitate subsequent analysis of their specific deformation degrees.
[0032] Furthermore, in step S4, the point cloud in the deformation area is segmented and added to the measurement set using a point cloud segmentation algorithm, including the following steps:
[0033] S41: Voxelize the point cloud data of the detected deformed cable bracket and use the octree algorithm to cover the point cloud:
[0034] S42: Estimate the curvature characteristics of the point cloud using the least squares algorithm;
[0035] S43: Using the region growing segmentation algorithm, all adjacent points with similar curvatures are extracted to establish a deformed point cloud set.
[0036] Furthermore, in step S5, the point cloud kmeans skeleton extraction technology is used to extract deformation skeleton points from the point set of the measurement set to form a space curve, which includes the following steps:
[0037] S51: shrinking the deformation point set extracted in S43 to form skeleton points;
[0038] S52: A closely connected skeleton is formed by adding regularization terms and a local branch curve fitting method to the extracted skeleton points.
[0039] Furthermore, the step S6 of measuring the space curve to analyze the deformation degree of the cable support includes the following steps:
[0040] S61: Establish a spatial straight line equation for the deformation area of the cable bracket with the two end points of the skeleton;
[0041] S62: Calculate the length of the spatial straight line equation;
[0042] S63: Calculate the length of the skeleton space curve extracted in S5;
[0043] S64: Calculate the difference between the space curve and the space straight line, obtain the deformation variable, analyze the influence of the deformation variable and the maintenance and protection measures.
[0044] Compared with the prior art, the present invention has the following outstanding substantive features and remarkable characteristics:
[0045] The method of the present invention first classifies the deformation degree of the tunnel cable support by judging the deformation amount of the support. The cable support with a large degree of deformation may fall off and break, and early warning is given to avoid possible threats to the safety of workers in the cable tunnel. Secondly, Pointpillars is improved to improve the speed and accuracy of cable support deformation detection, and the deformed cable support is segmented by the point cloud segmentation algorithm. The skeleton extraction method is used again to judge the specific deformation of the cable support, and the cable support fault disease detection and identification can be carried out, thereby providing a guarantee for the reliable operation of the entire tunnel power grid system. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 It is a general flow chart of a method for detecting deformation of a tunnel cable support based on laser radar according to the present invention.
[0047] Figure 2 It is the network flow chart of point cloud detection of tunnel cable deformation bracket.
[0048] Figure 3 It is the flow chart of point cloud registration algorithm.
[0049] Figure 4 It is the point cloud segmentation extraction flow chart.
[0050] Figure 5 It is the flowchart of point cloud skeleton point extraction. DETAILED DESCRIPTION
[0051] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments will be described in detail below in conjunction with the accompanying drawings of the present invention.
[0052] like Figure 1 As shown, a method for detecting deformation of a tunnel cable support based on laser radar comprises the following steps:
[0053] S1: Collect historical point cloud samples of relevant deformed cable supports in tunnels to make detection labels and build detection models; including the following steps, such as Figure 2 As shown:
[0054] S11: Install a 3D laser radar on the unmanned vehicle, use the 3D laser radar to scan the deformed cable bracket in the tunnel, collect relevant point cloud data, and pre-process the point cloud data; check the collected point cloud to see if there is deformation in consecutive frames. If so, synthesize multiple frames of point cloud data into one frame of point cloud. The specific process is as follows: Figure 3 As shown;
[0055] S12: Perform data enhancement on point cloud data;
[0056] S13: Label the point cloud after data enhancement, mark the real deformation area, and divide the real deformation area into four labels: small, medium and large repair, as input data for the detection network model. Among them, 20% of the verification data is reserved to verify whether the trained network model has overfitting.
[0057] S14: Build a point cloud deep learning detection network framework for detecting deformed cable brackets, and use the enhanced point cloud data and labels as input to the network framework, train the framework of this deep learning network model, and obtain a tunnel cable bracket deformation detection model after deep learning.
[0058] The step S14 comprises the following steps:
[0059] S141: performing regional cropping on the collected point cloud data and fixing it within a space with a certain size range of [X, Y, Z];
[0060] S142: The point cloud space is divided into two parts according to the volume size V (w,l,H) The voxels are divided into voxels, where w is the voxel width, l is the voxel length, H is the voxel height, and H = Y is a fixed value; the point cloud data in the training sample are discretized on evenly spaced grids on the XOZ plane, with no restrictions in the Y direction, and a series of cylinders on the XOZ plane are created. The point cloud in each cylinder is sampled, and the voxels with more than N points are uniformly downsampled to N points, and the voxels with less than N points are filled with 0 to N points; the point cloud contained in each cylinder is represented by r, x c ,y c , z c , x p , z p Feature expansion, get the expanded point cloud features (x, y, z, r, xc ,y c , z c , x p , z p ), where x, y, and z represent the initial coordinates of the point cloud; r represents the reflectivity of the point cloud; x c ,y c , z c It represents the coordinate value obtained by taking the arithmetic mean of all the point cloud coordinates in the pillar; p , z p Indicates the deviation of all point clouds in the pillar relative to the coordinate center position in the coordinate system of the current plane; on the XOZ plane, the number of point clouds contained in all non-empty pillars is adjusted to be consistent, and then a dense tensor (D, P, N) is created according to the number of non-empty pillars on the plane, the number of point clouds contained in the non-empty pillars, and the features of the point clouds in the non-empty pillars, that is, the features (D, P, N) of each non-empty pillar in the XOZ plane are obtained, where D represents the feature dimension of the point cloud in the non-empty pillar, P represents the number of non-empty pillars on the XOZ plane, and N represents the number of point clouds contained in the non-empty pillars;
[0061] S143: Use the pointnet network to perform feature learning on the features (D, P, N) of the non-empty pillars on the XOZ plane, and obtain the final features (C, P, N) of the point cloud in each non-empty pillar in the XOZ plane after learning; wherein C represents the new feature dimension of the point cloud after learning by the pointnet network, P represents the number of non-empty pillars on the XOZ plane, and N represents the number of point clouds contained in the non-empty pillars; after the operation, the final features (C, P, N) of the point cloud of each non-empty pillar in the XOZ plane are obtained, and the features (C, P, N) are processed by the maximum pooling operation to obtain a tensor (C, P), and then a sparse pseudo image (C, L, W) is created according to the tensor (C, P), wherein L represents the height of the sparse pseudo image, and W represents the width of the sparse pseudo image;
[0062] S144: Use the Backbone feature extraction network as the backbone network for extracting feature maps to achieve accurate detection, including: using a series of convolutional layers to extract multi-scale features from sparse pseudo images of scale S, which contains a total of three scale feature maps F 11 , F 12 , F 13 , whose sizes are S, S / 2, and S / 4 respectively. The feature maps of these three scales are then pooled and deconvolved to generate three independent pseudo feature maps. They are then fused using the concatenate operation and the 1×1 convolution layer to obtain the final feature map for detection. Finally, three 1×1 convolution layers are used to predict the object category and 3D bounding box respectively.
[0063] S2: Store several frames of point cloud data in real time, and use the point cloud multi-frame registration and stitching algorithm to stitch the currently measured point cloud set together and add it to the detection candidate area;
[0064] S21: Use an unmanned vehicle equipped with a 3D laser radar to store point cloud data of the cable bracket deformation area in real time;
[0065] S22: Using a straight-through filter to crop the point cloud range of the point cloud data to obtain point cloud data of a fixed area;
[0066] S23: Use statistical filtering to remove outliers in the point cloud;
[0067] S24: The first frame of data stored is marked as the target point cloud, and the frames stored later are the source point cloud. Each frame of data is filtered. The source point cloud and the target point cloud are registered using the point cloud multi-frame registration algorithm to determine the rigid body transformation parameters. Then, the rigid body transformation parameters are converted, and this step is repeated over multiple frames to synthesize a target point cloud set and add it to the point cloud target detection area.
[0068] S3: Using the point cloud detection model and the deep learning target detection algorithm, the deformation detection area of the cable bracket is identified and added to the segmentation candidate area; Figure 4 As shown, the specific steps are as follows:
[0069] S31: using the point cloud in the target detection area obtained in S2 as input of the target detection algorithm;
[0070] S32: Detect the deformation of the target area. If there is no deformation, delete the current point cloud space;
[0071] S33: if there is deformation, mark the deformations of small, medium and large levels according to the degree of deformation for the next inspection;
[0072] S34: For the maintenance level, a segmentation candidate region is added to facilitate the subsequent analysis of its specific deformation degree;
[0073] S4: Use the point cloud segmentation algorithm to segment the point cloud in the candidate segmentation area and add it to the measurement set; the specific steps are as follows:
[0074] S41: extracting the point cloud data detected to be deformed, voxelizing the point cloud in the segmentation candidate area, and covering the input point cloud using the octree algorithm;
[0075] S42: sort the entire point cloud according to the curvature of each point, search for the point with the smallest curvature, and add it as a seed point;
[0076] S43: Search for points in the neighborhood of the seed point. If the difference between the normal vector of a point in the neighborhood and the normal vector of the seed point is less than a certain range, calculate the curvature of the point. If the curvature of the point is less than the set threshold, the point belongs to the current part. Remove the points that have passed the second step test from the original point cloud. Set the minimum point cloud cluster (the minimum number of segmentations) and the maximum point cloud cluster (the maximum number of segmentations) according to the number of points in the point cloud. Repeat the above steps, the original point cloud will be segmented into parts that meet the above minimum and maximum point cloud clusters, and these point clouds will be recorded. When the number of remaining points is less than the set point clusters, stop segmenting. Extract all recorded points, which is the segmented point cloud prediction set.
[0077] S5: Use the point cloud kmeans skeleton extraction technology to extract the deformation skeleton points from the point set of the measurement set to form a spatial curve; for the predicted measurement set generated by the segmentation in S4, such as Figure 5 As shown, the specific steps are as follows:
[0078] S51: Extract skeleton points from the points extracted by octree segmentation using a skeleton extraction algorithm.
[0079] Give segmentation point cloud And the sampling point set A series of skeleton points can be obtained using the optimization formula:
[0080]
[0081]
[0082] θ(r)=exp[-4(r / h) 2 ] (3)
[0083]
[0084] The first term in formula (1) determines the local L1-median contraction skeleton point, which moves toward the local median based on the gravitational effect of the neighborhood points of the sampling point in the input point set; the second term is the regularization term, which is based on the neighborhood points x in the sampling point set. i ′ forms a repulsive force to avoid shrinking too close, which is expressed by formula (2) according to the point distribution, where γ i is the parameter that balances attraction and repulsion. θ(r) is the Gaussian weight function. From formula (3), we can see that under a fixed neighborhood radius r, the farther away from the sampling point, the smaller the contribution weight h, thus playing the role of median filtering. By introducing point x in the regularization term i The neighborhood direction angle is used to determine the distribution of the points, and the local repulsion β is automatically calculated and adjusted. From formula (4), we know that, i is x i Corresponding to the three eigenvalues of the covariance matrix, λ i0 ≤λ i 1 ≤λ i 2 When σ i The closer it is to 1, the The sum is less than Compared to time point x i and its neighborhood are basically distributed in a linear manner, and the closer they are to the skeleton branches. Substituting formula (3) into the L1-median solution term yields:
[0085]
[0086] in
[0087] In order to reduce the number of iterations, an adaptive radius skeleton extraction algorithm based on region division is used. Different initial neighborhood radii are set for different regions to avoid an increase in the number of iterations due to a too small initial radius. The radius is expanded at the same growth rate, new branches are found and skeleton connections are made, and branches that have shrunk into place are fixed.
[0088] S52: The skeleton obtained by calculating the shrinkage still has problems of continuity and smoothness. It is necessary to connect the extracted skeleton points and add a regular term through formula (6) for curve fitting of local branches to form a closely connected skeleton and form a spatial curve:
[0089]
[0090] N(x i ) is the neighborhood point set of the point, υ(x i ) is the number of points in the neighborhood. The smoothing function P(x) optimizes the position of the skeleton points by minimizing the distance between the neighborhood points to obtain a locally fitted curve skeleton.
[0091] S6: Measure the space curve to analyze the deformation of the cable bracket; the specific steps are as follows:
[0092] S61: For the two cable support endpoints (x0, y0, z0) and (x m ,y m z m ), calculate the length L0 of the space line equation.
[0093]
[0094] S62: Calculate the length of the skeleton space curve extracted in S5. t (x t ,y t , zt )(t=1,2,...,m), forming a straight line segment connected end to end, thus forming a spatial discount that is infinitely close to the actual trajectory to represent the three-dimensional curve. In this way, the length calculation of the space curve can be simplified to m-1 straight line segments P connected end to end formed by m discrete points. t P t+1 Calculation of the length of the composed spatial polyline L.
[0095]
[0096] S63: Calculate the difference between the space curve and the space straight line to obtain the deformation variable d.
[0097] d=L-L0 (9)
[0098] S64: Analyze the specific deformation cause based on the specific deformation amount and take appropriate maintenance measures.
[0099] The above embodiment is based on the laser radar deformation detection method of the tunnel cable support, collects the laser radar point cloud information of the cable support in the tunnel, uses the point cloud multi-frame registration and splicing algorithm to splice the current point cloud set and add it to the detection candidate area; identifies the deformed area in the deformation detection candidate area, adds the segmentation candidate area; then performs point cloud segmentation on the segmentation candidate area; then uses the point cloud skeleton extraction technology to extract the skeleton points of the deformed cable support, forms a spatial curve, and analyzes the deformation degree of the curve. This method uses traditional methods combined with deep learning methods, uses the deformation characteristics of the cable support to detect its deformation, measures and analyzes its deformation degree, thereby providing a guarantee for the reliable operation of the entire power grid system.
[0100] The above describes the embodiments of the present invention in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Various changes can be made according to the purpose of the invention. Any changes, modifications, substitutions, combinations or simplifications made according to the spirit and principle of the technical solution of the present invention should be equivalent replacement methods. As long as they meet the purpose of the invention and do not deviate from the technical principles and inventive concepts of the present invention, they belong to the protection scope of the present invention.
Claims
1. A method for detecting deformation of a tunnel cable support based on laser radar, characterized in that: The steps include: S1: Collect historical point cloud samples of relevant deformed cable supports in tunnels to make detection labels and build detection models; S2: Store several frames of point cloud data in real time, and use the point cloud multi-frame registration and stitching algorithm to stitch the currently measured point cloud set together and add it to the detection candidate area; S3: Using the point cloud detection model and the deep learning target detection algorithm, the deformation detection area of the cable bracket is identified and added to the segmentation candidate area; S4: Use the point cloud segmentation algorithm to segment the point cloud in the segmentation candidate area and add it to the measurement set; S5: Use the point cloud kmeans skeleton extraction technology to extract the deformation skeleton points from the point set of the measurement set to form a space curve; S6: Measure the space curve to analyze the deformation of the cable support.
2. The method for detecting deformation of a tunnel cable support based on laser radar according to claim 1 is characterized in that: In step S1, historical point cloud samples of relevant deformed cable supports in the tunnel are collected to make detection labels and construct a detection model, including the following steps: S11: Collect point cloud data of the deformed cable bracket in the tunnel to check whether there are deformations in multiple consecutive frames; use the point cloud multi-frame registration and splicing algorithm to splice the point cloud that has deformations in multiple consecutive frames; S12: performing data enhancement on the point cloud data of the deformed cable bracket to obtain enhanced point cloud data; S13: labeling the deformed area in the enhanced point cloud data as a training data set, and dividing the deformation in the point cloud data set into four labels: small, medium, large, and maintenance, as detection labels; S14: Construct a point cloud deep learning detection network pointpillars model for detecting deformed cable brackets, and use the enhanced point cloud data and labels as the input of the network framework to train this deep learning network model to obtain a tunnel cable bracket deformation detection model after deep learning.
3. The method for detecting deformation of a tunnel cable support based on laser radar according to claim 2 is characterized in that: The step S14 constructs a point cloud deep learning detection network pointpillars model for detecting deformed cable supports, and uses the enhanced point cloud data and labels as network input to train the deep learning network model to obtain a tunnel cable support deformation detection model after deep learning, including the following steps: S141: performing regional cropping on the collected point cloud data and fixing it within a space with a certain size range of [X, Y, Z]; S142: The point cloud space is divided into two parts according to the volume size V (w,l,H) The voxels are divided into voxels, where w is the voxel width, l is the voxel length, H is the voxel height, and H = Y is a fixed value; the point cloud data in the training sample are discretized on evenly spaced grids on the XOZ plane, with no restrictions in the Y direction, and a series of cylinders on the XOZ plane are created. The point cloud in each cylinder is sampled, and the voxels with more than N points are uniformly downsampled to N points, and the voxels with less than N points are filled with 0 to N points; the point cloud contained in each cylinder is represented by r, x c ,y c , z c , x p , z p Feature expansion, get the expanded point cloud features (x, y, z, r, x c ,y c , z c , x p , z p ), where x, y, and z represent the initial coordinates of the point cloud; r represents the reflectivity of the point cloud; x c ,y c , z c It represents the coordinate value obtained by taking the arithmetic mean of all the point cloud coordinates in the pillar; p , z p Indicates the deviation of all point clouds in the pillar relative to the coordinate center position in the coordinate system of the current plane; on the XOZ plane, the number of point clouds contained in all non-empty pillars is adjusted to be consistent, and then a dense tensor (D, P, N) is created according to the number of non-empty pillars on the plane, the number of point clouds contained in the non-empty pillars, and the features of the point clouds in the non-empty pillars, that is, the features (D, P, N) of each non-empty pillar in the XOZ plane are obtained, where D represents the feature dimension of the point cloud in the non-empty pillar, P represents the number of non-empty pillars on the XOZ plane, and N represents the number of point clouds contained in the non-empty pillars; S143: Use the pointnet network to perform feature learning on the features (D, P, N) of the non-empty pillars on the XOZ plane, and obtain the final features (C, P, N) of the point cloud in each non-empty pillar in the XOZ plane after learning; wherein C represents the new feature dimension of the point cloud after learning by the pointnet network, P represents the number of non-empty pillars on the XOZ plane, and N represents the number of point clouds contained in the non-empty pillars; after the operation, the final features (C, P, N) of the point cloud of each non-empty pillar in the XOZ plane are obtained, and the features (C, P, N) are processed by the maximum pooling operation to obtain a tensor (C, P), and then a sparse pseudo image (C, L, W) is created according to the tensor (C, P), wherein L represents the height of the sparse pseudo image, and W represents the width of the sparse pseudo image; S144: Use the Backbone feature extraction network as the backbone network for extracting feature maps to achieve accurate detection, including: using a series of convolutional layers to extract multi-scale features from sparse pseudo images of scale S, which contains a total of three scale feature maps F 11 , F 12 , F 13 , whose sizes are S, S / 2, and S / 4 respectively. The feature maps of these three scales are then pooled and deconvolved to generate three independent pseudo feature maps. They are then fused using the concatenate operation and the 1×1 convolution layer to obtain the final feature map for detection. Finally, three 1×1 convolution layers are used to predict the object category and 3D bounding box respectively.
4. The method for detecting deformation of a tunnel cable support based on laser radar according to claim 1, characterized in that: In step S2, several frames of point cloud data are stored in real time, and the currently measured point cloud set is added to the detection candidate area using the point cloud multi-frame registration and splicing algorithm, which includes the following steps: S21: real-time storage of point cloud data of the cable support deformation area; S22: Using a straight-through filter to crop the point cloud range of the point cloud data to obtain point cloud data within a fixed area; S23: using statistical filtering to remove outliers in the point cloud after regional clipping, so as to reduce the point cloud data in the region that interferes with the detection; S24: The first frame of stored data is marked as the target point cloud, and the frames stored thereafter are the source point clouds. Each frame of data is filtered, and the source point cloud and the target point cloud are registered using a point cloud multi-frame registration algorithm to determine the rigid body transformation parameters. Then, the point cloud is converted according to the rigid body transformation parameters, and this step is repeated over multiple frames to synthesize a target point cloud set, which is added to the point cloud target detection area.
5. The method for detecting deformation of a tunnel cable support based on laser radar according to claim 1, characterized in that: In step S3, the point cloud detection model is used to use a deep learning target detection algorithm to identify the deformation detection area of the cable bracket and add it to the segmentation candidate area, including the following steps: S31: using the point cloud in the target detection area obtained in step S2 as input of the target detection algorithm; S32: Detect the deformation of the target area. If there is no deformation, delete the current point cloud space; S33: If there is deformation, mark the deformation levels according to the degree of deformation, small, medium and large, and record the detection results for the next detection; S34: For the maintenance level, segmentation candidate regions are added to facilitate subsequent analysis of their specific deformation degrees.
6. The method for detecting deformation of a tunnel cable support based on laser radar according to claim 1, characterized in that: The step S4 uses a point cloud segmentation algorithm to segment the point cloud in the deformation area and add it to the measurement set, including the following steps: S41: Voxelize the point cloud data of the detected deformed cable bracket and use the octree algorithm to cover the point cloud: S42: Estimate the curvature characteristics of the point cloud using the least squares algorithm; S43: Using the region growing segmentation algorithm, all adjacent points with similar curvatures are extracted to establish a deformed point cloud set.
7. The method for detecting deformation of a tunnel cable support based on laser radar according to claim 1, characterized in that: In step S5, the point cloud kmeans skeleton extraction technology is used to extract the deformation skeleton points from the point set of the measurement set to form a space curve, including the following steps: S51: shrinking the deformation point set extracted in S43 to form skeleton points; S52: A closely connected skeleton is formed by adding regularization terms and a local branch curve fitting method to the extracted skeleton points.
8. The method for detecting deformation of a tunnel cable support based on laser radar according to claim 1, characterized in that: The step S6 of measuring the space curve and analyzing the deformation degree of the cable support comprises the following steps: S61: Establish a spatial straight line equation for the deformation area of the cable bracket with the two end points of the skeleton; S62: Calculate the length of the spatial straight line equation; S63: Calculate the length of the skeleton space curve extracted in S5; S64: Calculate the difference between the space curve and the space straight line, obtain the deformation variable, analyze the influence of the deformation variable and the maintenance and protection measures.
Citation Information
Patent Citations
Tunnel structure detection method and device and electronic device
CN111102935A
Omnibearing obstacle detection method based on multi-sensor fusion
CN111583337A
Lane detection method and system based on vision and lidar multi-level fusion
US10929694B1
Method, system and computer readable storage media for registering intraoral measurements
US20210090272A1
