Tunnel face surface reconstruction and joint occurrence automatic extraction method based on radar point cloud

Abstract

The application discloses a tunnel face surface reconstruction and joint occurrence automatic extraction method based on a radar point cloud, and comprises the following steps: step 1, a triangular division is performed on a tunnel face surface model to obtain a plurality of triangular facets; step 2, a normal vector of each triangular facet is calculated, all normal vectors are divided into N groups by using a direction-based clustering method, each group corresponds to a structural plane, and a point cloud area corresponding to a triangular facet with an angle greater than a predetermined value with respect to a tunnel face normal vector is removed in each group of structural planes, then, according to three-dimensional point cloud data of each structural plane obtained by removal, a RANSAC algorithm is used to perform plane fitting to obtain a plurality of groups of structural plane spaces fitting a real posture of a geological body; and step 3, occurrence parameters are calculated based on the normal vectors of the plurality of groups of structural plane spaces and spatial geometric relations; the whole process is automatic, manual intervention is reduced, and the efficiency and objectivity of tunnel geological information collection are significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of tunnel engineering, and in particular relates to a method for automatic extraction of surface reconstruction and joint orientation of tunnel face based on radar point cloud. Background Technology

[0002] The attitude characteristics (including dip, strike, and dip angle) of the surrounding rock structural planes at the tunnel face are crucial for evaluating rock stability and optimizing engineering design. In traditional tunnel geological exploration, structural plane attitude measurements require close proximity to the tunnel face using a geological compass. This not only poses high safety risks to testing personnel but also results in low efficiency. More importantly, the test results are easily influenced by subjective factors such as personnel experience, leading to significant differences in results among different personnel. Currently, traditional structural plane attitude analysis methods are time-consuming, which exacerbates the personnel shortage and can easily delay the construction schedule, thus affecting the smooth progress of subsequent operations. Summary of the Invention

[0003] The purpose of this invention is to provide a method for automatic reconstruction of the tunnel face surface and extraction of joint orientation based on radar point clouds, so as to solve the problem of the influence of subjective factors of personnel on the measurement of the surrounding rock structure in tunnels using the traditional drill-and-blast method.

[0004] This invention adopts the following technical solution: a method for automatic extraction of tunnel face surface reconstruction and joint orientation based on radar point clouds, comprising: Step 1: Obtain full-section laser point cloud data of the tunnel face at different mileages to obtain a high-precision three-dimensional point cloud model, and reshape its surface framework to obtain the tunnel face surface model; then triangulate the tunnel face surface model to obtain several triangular facets. Step 2: Calculate the normal vector of each triangular facet. Use a direction-based clustering method to divide all normal vectors into N groups, each group corresponding to a structural facet. In each group of structural facests, remove the point cloud regions corresponding to triangular facets whose angle with the facet normal vector is greater than a predetermined value. Then, based on the 3D point cloud data of each structural facet obtained after removal, use the RANSAC algorithm to perform plane fitting to obtain multiple sets of structural facet spatial planes that fit the true posture of the geological body. Step 3: Calculate the attitude parameters based on the normal vectors and spatial geometric relationships of multiple sets of structural planes. The attitude parameters are dip angle, strike, and dip direction.

[0005] Furthermore, the predetermined value in step 2 is 10-15 degrees.

[0006] Furthermore, in step 2, after removing the point cloud regions corresponding to the triangular facets whose angle with the facet normal is greater than a predetermined value in each group of structural surfaces, the improved U-Net++ is used to remove narrow crack steep facets. The improvements are as follows: In the encoder part, CSPDarkNet53 is used as the backbone network to replace the convolutional layers of traditional U-Net++, which enhances the feature extraction capability for narrow gaps and steep patches. An SE attention module is added to the decoder section to focus on triangular facets with an angle greater than a predetermined value to the face normal and regions with obvious steep features, thereby suppressing interference from non-steep regions and noise. The loss function is a combination of Dice loss and cross-entropy loss.

[0007] The beneficial effects of this invention are: This invention automates the entire process, reducing manual intervention. It automatically completes the surface reconstruction, triangulation, normal vector calculation, structural surface clustering, plane fitting, and attitude parameter calculation of the tunnel face using lidar point cloud data. There is no need to manually select coplanar points or manually measure the attitude, which significantly improves the efficiency and objectivity of tunnel geological information collection. This invention uses a direction-based clustering method to automatically divide the normal vectors into N groups corresponding to different structural surfaces. By removing triangular facets with an angle greater than a predetermined value (10-15 degrees) with the normal vector of the tunnel face, non-structural surface noise is effectively removed. Combined with the RANSAC algorithm for plane fitting, it can resist outlier interference and make the fitted structural surface spatial plane more closely fit the true posture of the rock mass. This invention demonstrates excellent recognition capabilities for narrow, steep structural surfaces. After removing large-angle patches, an improved U-Net++ model is further employed (the encoder uses a CSPDarkNet53 backbone network, the decoder incorporates an SE attention module, and the loss function is a combination of Dice and cross-entropy), enabling accurate segmentation of narrow, steep surfaces. CSPDarkNet53 enhances the extraction of subtle features, the SE attention module focuses on steep regions, and the combined loss function alleviates the sample imbalance problem, thus avoiding missed and false recognitions caused by traditional angle threshold screening.

[0008] The present invention provides accurate and reliable attitude calculation. Based on the fitted structural plane, the dip angle, strike, and dip direction are directly derived using the normal vector and spatial geometric relationship. This avoids the errors caused by manual reading and approximate calculation, and ensures the consistency and accuracy of the extracted results. This invention adapts to complex geological conditions. Starting from the original point cloud, it undergoes robust triangular mesh processing and deep learning for fine screening. It can work stably in tunnel face scenarios with fractured rock, mixed noise, and developed fissures, providing high-precision structural surface information for tunnel advanced geological prediction and construction safety assessment. Attached Figure Description

[0009] Figure 1 Image showing the point cloud separation results at the face of the working face; Figure 2This is a triangular diagram of the tunnel face; Figure 3 This is the result of structural surface attitude information; Figure 4 The distribution of structural planes. Detailed Implementation

[0010] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0011] This invention discloses a method for automatic extraction of surface reconstruction and joint orientation of tunnel face based on radar point cloud, which includes three steps.

[0012] Step 1: Obtain full-section laser point cloud data of the tunnel face at different mileages to obtain a high-precision three-dimensional point cloud model, and reshape its surface framework to obtain the tunnel face surface model; then triangulate the tunnel face surface model to obtain several triangular facets.

[0013] Step 101: Obtain full-section laser point cloud data of the tunnel face at different mileages.

[0014] After the tunnel blasting using the drill-and-blast method is completed, and the tunnel face undergoes slag removal, hazard mitigation, and ventilation for 20 minutes, the operator places the integrated equipment (01) 5-10m in front of the tunnel face at a certain mileage point. The equipment is equipped with a single-echo laser radar with a point cloud output rate ≥240,000 points / second, capable of collecting point cloud data of the tunnel face and surrounding excavation contour lines, with a data acquisition accuracy ≤5mm. The collected point cloud data is transmitted to the data processor via Bluetooth, with the overall acquisition and transmission time controlled within 2 minutes.

[0015] The collected laser point cloud data is synthesized into a high-precision 3D point cloud model by combining the Simultaneous Localization and Mapping (SLAM) algorithm. The implementation process of laser radar synthesis at the tunnel face needs to complete three stages: hardware deployment, data acquisition and processing, and global optimization and model generation. The specific steps are as follows: (1) The hardware deployment stage mainly includes two steps: sensor calibration and ROS environment configuration. The former obtains key system parameters through offline calibration, including the extrinsic parameters of the lidar and IMU (rotation matrix R and translation matrix T) used to transform the lidar point cloud to the IMU coordinate system, the IMU intrinsic parameters (zero bias and scale factor of accelerometer and gyroscope) calibrated by Allan variance analysis, and the lidar time synchronization parameters used to compensate for the time offset between sensors. The latter ensures that the lidar point cloud and IMU data can be correctly received by the algorithm by setting up the ROS Noetic environment, installing the Livox SDK driver and FastLIO algorithm package, and configuring the sensor message publishing node and algorithm subscription node.

[0016] (2) The data acquisition and processing stage mainly includes two parts: synchronous acquisition of multi-source data and real-time processing using the FastLIO algorithm. The former involves activating the sensor and ROS node, with the LiDAR outputting point cloud data in a Livox-customized Msg format, and the IMU outputting acceleration and angular velocity information at a frequency of 100Hz. The ROS timestamp synchronization mechanism ensures that the LiDAR and IMU data at the same time form a time-aligned dataset. The latter uses IMU pre-integration to predict the carrier's motion state, completes point cloud distortion correction, extracts point cloud features, and matches them with the local map to optimize the carrier's state. A sliding window combined with voxel filtering is used to maintain the local map. The specific steps are as follows: ①IMU pre-integration and state prediction: Based on IMU measurements, the median integration method is used to calculate the carrier motion state (position, velocity, attitude) between adjacent lidar scan frames. The state equation is shown in Equation 1.

[0017]

[0018] in, , , These are position, velocity, and attitude (quaternions). , For IMU acceleration and angular velocity, , For IMU zero bias, For integration time, Exp(·) is the Lie algebra exponent mapping.

[0019] ② Point cloud distortion correction: Based on the motion state predicted by the IMU, the point cloud distortion caused by the carrier motion during the lidar scanning process is compensated. For any point on the lidar scanning line, its distorted coordinates are shown in Formula 2.

[0020]

[0021] in, The scan frame start time, This is the actual scanning time for that point. .

[0022] ③ Feature extraction and state update: The distortion-free point cloud is divided into edge points and planar points, which are then matched with line features and area features in the local map, respectively. A residual function is constructed as shown in Equation 3.

[0023]

[0024] in, The starting point and direction vector of the line feature. Let be the points and normal vectors of the surface features. The Gauss-Newton method is used to minimize the sum of squared residuals to optimize and update the carrier state.

[0025] ④ Local map maintenance: A sliding window mechanism is used to maintain a local map composed of the point clouds of the most recent 50 frames. Voxel filtering is used to reduce the amount of computation and ensure the real-time performance of the algorithm.

[0026] (3) The global optimization and model generation stage mainly includes three steps: segmented point cloud stitching, loop closure detection and optimization, and point cloud post-processing.

[0027] First, the local point cloud output by FastLIO during the moving scan (save a segment every 5 meters) is transformed to the world coordinate system based on the pose information to form a preliminary global point cloud. Second, when the sensor platform returns to the starting area, loop closure is detected and pose correction is calculated using a point cloud matching algorithm. The global pose is then adjusted using a factor graph optimization method to eliminate accumulated errors. Finally, the global point cloud is subjected to statistical filtering for noise reduction, moving least squares smoothing, and meshing processing in sequence to finally generate a 3D model of the tunnel face.

[0028] Step 102: The high-precision 3D point cloud model obtained is reconstructed using tunnel face topology reconstruction technology to obtain the 3D surface model of the tunnel face. The process of synthesizing the tunnel triangular face model requires three stages: data preprocessing, surface reconstruction, and model optimization. The specific steps are as follows: (1) The point cloud preprocessing stage includes two steps: noise removal and downsampling. First, a statistical filtering algorithm is used for each point. The search radius is All points within the neighborhood of , usually Take 1.5-2 times the average spacing of the point cloud and calculate the distance from neighboring points to the target points. average distance and standard deviation ;like The average distance to neighboring points satisfies ( If the threshold coefficient is 3.0 (taken here), the point is retained; otherwise, it is identified as a noise point and discarded. Then, voxel grid filtering is used to divide the point cloud space into a cubic grid with a side length of v. Each grid retains a representative point (usually the center point of the grid or the average value of the points in that grid), thus achieving point cloud downsampling.

[0029] (2) The normal calculation stage includes three steps: neighborhood point search, covariance matrix construction, and normal estimation and orientation adjustment. First, for the preprocessed point cloud, for each point... Search for the k nearest neighbors to form a neighborhood point set. Then, the covariance matrix of the neighborhood point set is calculated. Finally, regarding the covariance matrix... Perform eigenvalue decomposition to obtain eigenvalues. and the corresponding feature vectors Minimum eigenvalue corresponding feature vector That is, a point The normal vector; using the region growing method, with the normal direction of the seed point as the reference, the normal directions of adjacent points are adjusted to ensure... ,So Let be the normal vector of the adjacent points, as shown in Formula 4.

[0030]

[0031] in It is the center point of the neighborhood point set.

[0032] (3) The Poisson surface reconstruction stage includes four steps: multi-level octree construction, indicator function gradient estimation, Poisson equation solution, and triangular face generation. First, to ensure the capture of tunnel surface details, the point cloud space is divided into multi-level octrees, and a minimum high-density point cloud count of 20 or more is set as the stopping iteration condition for octree depth search. The segmentation accuracy is adaptively adjusted according to the point cloud density, and the maximum depth is set to 12 levels. Then, using the point cloud and its normal information, the gradient value of each node in the octree is calculated. ,in For indicator functions, inside the object ,external Next, solve the Poisson equation. This yields a continuous indicator function. By setting the isosurface threshold The tunnel surface is extracted. Finally, the moving cube algorithm is used to process the octree nodes. Based on the distribution of indicator function values ​​within the nodes, triangular patches are generated and connected to form isosurfaces, thus forming the initial triangular network model.

[0033] The center point and side length of the octree are determined based on the maximum and minimum values ​​of the high-density point cloud, as shown in Formula 5: Among them is The coordinates of the center point of the octree, These are the maximum and minimum x-coordinate values ​​of the high-density point cloud. These are the maximum and minimum ordinate values ​​of the high-density point cloud. yes The length of the side in the direction, yes The length of the side in the direction, It is the side length of the octree.

[0034] The gradient estimation function for the indicator function is calculated as shown in Equation 6: in The weighting function is Gaussian. For point The normal vector.

[0035] (4) The model optimization stage includes three steps: mesh simplification, hole filling, and smoothing optimization. First, an edge folding algorithm is used to calculate the folding error of each edge (the degree of influence on the model surface after folding), and the edges are folded in order of increasing error until the number of triangles is reduced to the target value (usually retaining 30%-50% of the original number). Then, boundary edges in the mesh (edges belonging to only one triangle) are detected, and hole boundary loops are constructed; the least squares fitting method is used to estimate the surface parameters of the hole region, and new triangles are generated to fill the holes, ensuring the closure of the model. Finally, Laplacian smoothing is performed on the mesh to adjust each vertex. The position is shown in Formula 7.

[0036]

[0037] in This is the Laplace operator (the average positional difference between a vertex and its adjacent vertices). The smoothing coefficient is set to (0.1-0.3), and grid smoothing is achieved through multiple iterations, while avoiding over-smoothing that could lead to feature loss. Step 103: Triangulate the separated face model to obtain several triangular facets, so that the face model perfectly matches the actual shape on site.

[0038] Step 2: Calculate the normal vector of each triangular facet. Use a direction-based clustering method to divide all normal vectors into N groups, each corresponding to a structural plane. Within each structural plane group, remove the point cloud regions corresponding to triangular facets whose angle with the facet normal is greater than a predetermined value. Then, based on the 3D point cloud data of each structural plane obtained after removal, use the RANSAC algorithm to perform plane fitting to obtain multiple sets of spatial planes that conform to the true orientation of the geological body. The predetermined value in Step 2 is 10-15 degrees.

[0039] A plane can be defined by three non-collinear points. In the generated triangular mesh tunnel face rock mass surface model, each triangle can be considered a small structural plane. In the engineering geological decomposition standard, a structural plane is defined as a surface with a certain scale, direction, and characteristics within the rock mass. A rock mass structural plane can be considered as a cluster of triangular facets with approximately the same normal vector (same direction characteristic), a certain scale, satisfying certain undulation constraints (planar characteristics), and interconnected.

[0040] The general equation of the fitted structural surface space plane is as follows: At least three non-collinear coplanar points (from the trace and the exposed surface) are required to fit the plane equation using the least squares method: Assume there is Points Construct the error function Find the derivative and minimize it to solve for the coefficient. The optimal solution (usually by normalizing the equation, letting) (to simplify subsequent calculations).

[0041] This invention first filters target regions based on the normal vector characteristics of triangular mesh patches. By calculating the angle between the normal vector of each patch and a preset reference direction (such as the X-axis), patch clusters with angles meeting threshold requirements are selected. Then, noise and edge interference regions are eliminated by combining conditions such as region area and boundary distance, resulting in a highly reliable target geological body region. Subsequently, the RANSAC algorithm is used to perform optimal plane fitting on the vertex data of this region. This algorithm constructs a plane through iterative sampling and filters interior points, effectively eliminating the influence of outliers and outputting a plane equation (including plane normal vector) that fits the true attitude of the geological body. Finally, the attitude parameters are calculated based on the normal vector of the fitted plane and the spatial geometric relationship. The dip angle is derived from the angle between the normal vector and the vertical direction, the dip direction is determined by the projection direction of the normal vector on the horizontal plane, and the strike direction is the direction of the dip rotated 90° clockwise. Ultimately, the quantitative extraction of the attitude of the target geological body is achieved.

[0042] In step 2, after removing the point cloud regions corresponding to the triangular facets whose angle with the face normal is greater than a predetermined value in each group of structural surfaces, the improved U-Net++ is used to remove narrow crack steep facets. The improvements are as follows: In the encoder part, CSPDarkNet53 is used as the backbone network to replace the convolutional layers of traditional U-Net++, which enhances the feature extraction capability for narrow gaps and steep patches. In the decoder section, an SE attention module is added to focus on triangular facets with an angle greater than a predetermined value to the facet normal vector and regions with obvious steep features, while suppressing interference from non-steep regions and noise. The loss function is a combination of Dice loss and cross-entropy loss.

[0043] Traditional methods obtain 3D point cloud regions of steep structural surfaces by filtering triangular facets whose normal vectors have an angle greater than a predetermined value with respect to the X-axis. Relying solely on a single angle threshold, this approach is susceptible to normal vector deviations and surface undulations, leading to missed or incorrect identification of steep structural surfaces (e.g., misclassifying non-steep facets with large undulations as steep facets). Introducing deep learning technology to construct a semantic segmentation model enables accurate identification and filtering of steep structural surfaces. Combining the spatial features of triangular meshes with image features enhances the intelligence of the filtering process.

[0044] Step 201: Dataset Construction and Preprocessing A triangular mesh model dataset of tunnel face was constructed, covering structural face samples with different geological conditions (such as weathered rock, intact rock, and fractured rock) and varying steepness. Each sample was labeled, including steep structural face patches, non-steep structural face patches, and noisy face patches. The triangular mesh model was converted into a format suitable for deep learning model input. Specifically, the multi-dimensional features (geometric features, texture features) of the triangular face patches were mapped into feature maps, and combined with rock surface images (if available), a dual input data of "feature map + image" was constructed. The data was augmented (e.g., random rotation, translation, scaling, and noise addition) to expand the dataset size, avoid model overfitting, and improve the model's generalization ability. This approach referenced practical experience in data augmentation for deep learning-based borehole image rock structure face recognition, which can effectively improve the model's adaptability to complex geological scenarios.

[0045] Step 202: Deep Learning Model Selection and Training The U-Net++ semantic segmentation model was selected as the core recognition model. This model has an encoder-decoder structure, which can effectively capture structural surface features at different scales and is suitable for the recognition of small targets (scattered steep patches) and irregular regions (clusters of steep structural surfaces).

[0046] The model is improved as follows: (1) In the encoder part, CSPDarkNet53 is used as the backbone network to replace the convolutional layer of the traditional U-Net, which enhances the feature extraction capability, especially the feature capture of fine structural surfaces (such as narrow cracks and steep surfaces). CSPDarkNet53 can improve the stability and effectiveness of feature extraction by introducing batch normalization layers and Mish activation function.

[0047] (2) In the decoder part, an attention mechanism (SE attention module) is added to focus on the region with large normal vector angle and obvious steep features, and suppress the interference of non-steep regions and noise; (3) In the loss function, a combination loss function of Dice loss + cross-entropy loss is adopted to solve the imbalance problem between steep structural surface samples and non-steep samples in the dataset, and improve the model's sensitivity to the recognition of steep structural surfaces.

[0048] Step 203: Screening and Optimization of Steep Structural Surfaces The preprocessed triangular mesh feature data is input into the trained U-Net++ model. The model outputs the class probability (steep / non-steep) of each triangular facet. A probability threshold (e.g., 0.8) is set, and faces with a probability greater than or equal to the threshold are selected to form a 3D point cloud region of steep structural surfaces. Compared with traditional angle selection methods, this method combines multi-dimensional features and the feature learning capabilities of deep learning. It can automatically identify facet clusters that are characterized by "large angle between the normal vector and the X-axis + undulations that meet the requirements of steep structural surfaces + good connectivity," avoiding the limitations of a single angle threshold. At the same time, through the model's semantic segmentation capabilities, noisy faces can be removed, improving the purity of the steep structural surface region. In addition, a fine boundary delineation (CED) method can be introduced to optimize the boundaries of the identified steep structural surfaces, enhance the capture of local boundaries, improve the integrity of the structural surface region, and adapt to the structural surface recognition needs in complex construction scenarios such as TBM tunnels.

[0049] Traditional methods employ the least squares approach to fit the plane to the screened structural regions. This requires manual selection of at least three non-collinear coplanar points (from the trace line and the exposed surface), which is susceptible to human selection bias. Furthermore, the least squares method is sensitive to outliers (such as protrusions or depressions on the structural surface), leading to deviations in the plane equation fitting and consequently affecting the accuracy of attitude calculation. Introducing robust regression and regression prediction models from machine learning can automate and improve the accuracy of plane fitting while reducing human intervention.

[0050] A coplanar point identification model is constructed using the random forest algorithm in machine learning. This model automatically selects coplanar points (at least three non-collinear points) that meet the requirements from the 3D point cloud region of steep structural surfaces. The spatial coordinates (x, y, z), normal vectors, curvature, and other features of the point cloud are input into the random forest model. The model learns the feature differences between "coplanar points" and "non-coplanar points" in the training set, automatically identifying coplanar points on the structural surface traces and exposed surfaces. Simultaneously, a voting mechanism within the model filters out highly reliable coplanar points and eliminates outliers, ensuring that the selected coplanar points meet the requirements of plane fitting (non-collinearity and strong coplanarity).

[0051] Robust regression algorithms from machine learning (such as Iterative Weighted Least Squares, IWLS) are used to replace traditional least squares methods for plane fitting. The IWLS algorithm assigns different weights to each coplanar point, giving lower weights to outliers (such as noise points or convex points on the structural surface) to reduce their impact on the fitting results, and higher weights to points with strong coplanarity and high reliability to improve fitting accuracy. The general equation of a plane is Ax + By + Cz = D. By fitting the coordinates of coplanar points using the IWLS algorithm, the optimal coefficients A, B, C, and D are solved, ensuring that the fitted plane more closely matches the true shape of steep structural surfaces. Furthermore, it can be combined with point cloud deep learning models such as PointNet to directly regress the plane equation coefficients from point cloud features, further improving fitting efficiency and accuracy, especially suitable for plane fitting of complex structural surfaces.

[0052] Step 3: Calculate the attitude parameters based on the normal vectors and spatial geometric relationships of multiple sets of structural planes. The attitude parameters are dip angle, strike, and dip direction.

[0053] Structural plane space equation The coefficient has the following relationship with the dip direction and dip angle, as shown in Formula 8: From this equation, after obtaining the coefficients of the plane equation, the attitude can be calculated.

[0054] Example: Figure 1 The image shows the result of point cloud separation at the tunnel face. Delaunay triangulation was performed on the separated point cloud, generating approximately 1.2 million triangular facets. The triangular facet image of the tunnel face is shown below. Figure 2 As shown, the average side length of the facets is 0.05m. The triangular mesh model closely matches the actual shape of the tunnel face, with no obvious distortion or gaps. The normal vector of each triangular facet was calculated, and K-Means++ clustering was used to divide all normal vectors into three groups, each corresponding to a structural facet. Point cloud regions corresponding to triangular facets with an angle greater than 15 degrees to the tunnel facet normal vector were removed from each group of structural faces. Then, based on the 3D point cloud data of each structural facet obtained after the removal, the RANSAC algorithm was used to perform planar fitting on the three structural faces. The number of iterations was set to 1000, and the interior point threshold was 0.04m. The interior point rate of the fitting results all exceeded 90%, indicating that the fitted plane can effectively reflect the true shape of the structural facet. The obtained structural facet attitude information is shown below. Figure 3 As shown, the obtained structural surface region distribution is as follows: Figure 4 As shown.

Claims

1. A method for automatic surface reconstruction and joint orientation extraction of tunnel face based on radar point clouds, characterized in that, include: Step 1: Obtain full-section laser point cloud data of the tunnel face at different mileages to obtain a high-precision three-dimensional point cloud model, and reshape its surface framework to obtain the tunnel face surface model; then triangulate the tunnel face surface model to obtain several triangular facets. Step 2: Calculate the normal vector of each triangular facet. Use a direction-based clustering method to divide all normal vectors into N groups, each group corresponding to a structural facet. In each group of structural facests, remove the point cloud regions corresponding to triangular facets whose angle with the facet normal vector is greater than a predetermined value. Then, based on the 3D point cloud data of each structural facet obtained after removal, use the RANSAC algorithm to perform plane fitting to obtain multiple sets of structural facet spatial planes that fit the true posture of the geological body. Step 3: Calculate the attitude parameters based on the normal vectors and spatial geometric relationships of multiple sets of structural planes. The attitude parameters are dip angle, strike, and dip direction.

2. The method for automatic surface reconstruction and joint orientation extraction of tunnel face based on radar point cloud as described in claim 1, characterized in that, The predetermined value in step 2 is 10-15 degrees.

3. The method for automatic surface reconstruction and joint orientation extraction of tunnel face based on radar point cloud as described in claim 1, characterized in that, In step 2, after removing the point cloud regions corresponding to the triangular facets whose angle with the face normal is greater than a predetermined value in each group of structural surfaces, the improved U-Net++ is used to remove narrow crack steep facets. The improvements are as follows: In the encoder part, CSPDarkNet53 is used as the backbone network to replace the convolutional layers of traditional U-Net++, which enhances the feature extraction capability for narrow gaps and steep patches. An SE attention module is added to the decoder section to focus on triangular facets with an angle greater than a predetermined value to the face normal and regions with obvious steep features, thereby suppressing interference from non-steep regions and noise. The loss function is a combination of Dice loss and cross-entropy loss.

Images