Mountain tunnel excavation face deformation area identification method based on correlation algorithm and data fusion

Abstract

The invention provides a mountain tunnel excavation face deformation area identification method based on a correlation algorithm and data fusion, and the method comprises the steps: 1, synchronously collecting point cloud data and video data of a mountain tunnel excavation face, and carrying out the preprocessing of the data; 2, determining the pose of a video imaging device based on a registration algorithm, and constructing a three-dimensional model fusing geometric and texture information; 3, calculating an initial correlation index of the monitoring sub-region in an initial state, and determining a monitoring reference; 4, calculating correlation indexes of the monitored sub-regions in real time, and identifying a deformed region; and 5, based on an early warning mechanism of a preset threshold value, triggering deformation early warning, and synchronously pushing monitoring early warning information to a construction party. According to the method, the three-dimensional geometric accuracy of the point cloud and the continuous monitoring characteristic of the video are combined, quantitative recognition and unmanned early warning of the deformation area are achieved, manual intervention is reduced, the real-time performance and reliability of tunnel construction monitoring are improved, and key technical guarantee is provided for mountain tunnel construction safety.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering safety monitoring technology, and in particular to a method for identifying deformation areas at the excavation face of a mountain tunnel based on relevant algorithms and data fusion. Background Technology

[0002] With the rapid development of society and economy and technological progress, the number of mountain tunnel projects is constantly increasing. During the excavation and construction of mountain tunnels, the excavation face is affected by factors such as changes in geological conditions and stress redistribution, which can easily lead to deformation problems such as collapse, spalling, and crack expansion. If these problems are not identified and warned in time, they will seriously threaten the safety of construction personnel and equipment.

[0003] Currently, deformation monitoring of mountain tunnel excavation faces mainly relies on two types of technologies: one is total station monitoring technology, which requires the pre-installation of reflecting prisms around the excavation face. The setup and maintenance of measuring points require manual entry into dangerous areas of the tunnel face, posing safety hazards and being easily limited by visibility conditions. If there are obstacles or excessive dust concentration in the tunnel, the measurement will be interrupted. The other type is three-dimensional laser scanning technology, which can accurately acquire three-dimensional point cloud data of the excavation face, including three-dimensional coordinates and reflection intensity information. Deformation can be analyzed by comparing multiple point cloud data. However, this technology has the problems of long monitoring cycles and the need for equipment start-up and shutdown operations for each scan, making it difficult to achieve continuous real-time monitoring. In addition, the simple point cloud data lacks intuitive color and texture information, making it difficult to locate and visualize areas of minute deformation. Summary of the Invention

[0004] To address the aforementioned problems and shortcomings, this invention proposes a method for identifying deformation areas at the excavation face of mountain tunnels based on relevant algorithms and data fusion.

[0005] The specific technical solution of the present invention is as follows: A method for identifying deformation zones at the excavation face of a mountain tunnel based on relevant algorithms and data fusion includes the following steps: Step 1: Synchronously acquire point cloud data and video data of the mountain tunnel excavation face, and preprocess the data. (11): Collect 3D point cloud data and preprocess the data. A 3D laser scanner deployed inside a mountain tunnel is used to scan the tunnel excavation face and obtain high-precision point cloud data. The collected point cloud data is then preprocessed to reduce noise and remove laser points at the edges and those with angles that do not meet the requirements. (12): Acquire video data and preprocess the data. By deploying video imaging devices near the mountain tunnel, two-dimensional color videos of the tunnel excavation face are continuously acquired; the acquired videos are preprocessed by deblurring, white balance correction, and brightness normalization. (13): Synchronously store point cloud data and video data The acquisition timestamps of point cloud data and video data are synchronized using the NTP protocol to ensure data time consistency; after time synchronization and preprocessing, point cloud data and video data with consistent time sequence identifiers are obtained. Step 2: Determine the pose of the video imaging device based on the registration algorithm, and construct a 3D model that integrates geometric and texture information. (21): Determine the feature point extraction Based on the point cloud data and video data obtained in step (13), feature points with stability and representativeness are extracted respectively; then descriptors are used to encode the extracted feature points, and by calculating the similarity between descriptors, a unified descriptor set that can characterize the local geometric and texture features of point cloud and video is generated. (22): Point cloud and video feature point matching According to the descriptor type in step (21), select the feature point matching algorithm to find the initial matching pairs of feature points between the point cloud and the video; use the least squares method to remove the mismatched pairs in the initial matching pairs to ensure the reliability of the correspondence, and finally output the point cloud-video feature points. (23): Initial pose solution Using the point cloud-video feature points output in step 22 as input, and combining the known internal parameters of the video imaging device, the initial pose of the video imaging device relative to the 3D laser scanner is solved by the EPnP algorithm. (24): Optimize the initial pose and construct a 3D model that integrates geometric and texture information. Based on the initial pose obtained in step 23, the pose is iteratively optimized through the projection ICP algorithm, and then the projection error from the point cloud to the video image is minimized to construct a three-dimensional model that integrates geometric and texture information. Step 3: Calculate the initial correlation index of the monitoring sub-region under the initial state and establish the monitoring benchmark. (31): Sub-region division for monitoring the excavation face Based on the three-dimensional model in step (24), the monitoring sub-regions are divided into uniform grids to meet the point cloud quantity requirements and cover the entire excavation face area; the average point cloud reflection intensity and the average RGB color of the corresponding video area of ​​each monitoring sub-region are obtained as the basic features for correlation analysis. (32): Calculation of correlation indicators For each monitoring sub-region, the mean reflectance intensity of the point cloud and the mean RGB color of the corresponding video region are extracted, and the Pearson correlation coefficient between the two is calculated. (33): Construction of Deformation Monitoring Benchmark Database Finally, the Pearson correlation coefficient set of these monitoring sub-regions was used as the initial correlation index to form a deformation monitoring benchmark database for the entire excavation face; Step 4: Calculate the correlation index of the monitored sub-region in real time and identify deformed areas. Calculate the real-time point cloud reflectance mean, video RGB color mean, and Pearson correlation coefficient between the two for each monitoring sub-region; compare the changes in the real-time Pearson correlation coefficient with the initial correlation index. Determine whether the area is a potential deformation area; Step 5: Based on the preset threshold early warning mechanism, trigger the deformation early warning and simultaneously push the monitoring and early warning information to the construction party. The deformation and... were obtained through simulation experiments. The correlation model is used to determine the early warning threshold that meets the engineering safety requirements, and the maximum allowable deformation corresponding to the engineering safety requirements is assigned to it. Set as warning threshold Real-time monitoring of potential deformation areas ,when Exceeding the warning threshold When the potential deformation area is upgraded to a deformation area, an audible and visual alarm is automatically triggered, a monitoring report is pushed to the control terminal, and a report containing the coordinates of the deformation area is simultaneously pushed. The monitoring and early warning signals for information such as estimated deformation are sent to the construction team.

[0006] Furthermore, in step (11), the three-dimensional laser scanner is a phase laser scanner. The equipment is installed at a preset position around the tunnel excavation face. The three-dimensional laser scanner is installed on a steel support 5-10m behind the mountain tunnel excavation face. The equipment angle, parameters and installation height are adjusted to ensure that the laser scanning range completely covers the excavation face. Statistical filtering and radius filtering are used to remove noise. In step (12), the video imaging device uses an industrial-grade high-definition camera. The video imaging device is installed on the left side of the three-dimensional laser scanner and must meet two field of view requirements at the same time: first, the field of view of the video imaging device covers no less than 80% of the overall area of ​​the excavation face; second, the overlapping area of ​​the field of view must cover the central area of ​​the excavation face and the area must be no less than 20m², so as to avoid monitoring blind spots. The collected video data is deblurred by Wiener filtering, white balance is corrected by gray world algorithm, and brightness is normalized by histogram equalization to eliminate interference from lighting and equipment jitter.

[0007] Furthermore, the more specific process in step (21) is as follows: Based on the point cloud data and video data obtained in step (13), the NARF algorithm is used to extract three-dimensional feature points with strong geometric discriminative power. The algorithm first uses the neighborhood covariance matrix method to search for neighboring points within a radius of 0.1m for each point, constructs a covariance matrix, and obtains the normal vector through eigenvalue decomposition. The direction of the normal vector must be uniformly pointed to the outside of the excavation face. The surface curvature of each point is further calculated, and points with curvature > 0.3 are selected as candidate interest points. Finally, a depth image is constructed along the normal vector direction for the candidate interest points, and the point with the largest radial gradient is extracted as the final feature point. The feature point density needs to be controlled at 50-80 points / m. 2 Ensure coverage of key areas of the excavation face; For the initial frames of the preprocessed video data, the SIFT algorithm is used to extract two-dimensional feature points with scale invariance and rotation invariance. The SIFT algorithm first uses a Gaussian pyramid to scale the video frame to 6 scales, i.e., a scale factor of 1.2. Five layers of Gaussian blur image are constructed at each scale, and the difference in standard deviation between adjacent Gaussian layers is 0.8. Then, difference Gaussian images are calculated between Gaussian images at adjacent scales. Key points are screened by local extremum detection to exclude weak key points with contrast less than 8 bits of gray value and edge points with edge response >10. Finally, for the initially detected key points, sub-pixel level coordinates are accurately calculated by fitting a quadratic function to remove key points with positioning errors greater than 1 pixel. The final number of key points must match the number of feature points in the point cloud. Based on the extracted 3D and 2D feature points, the similarity between the two types of feature points is calculated using BRIEF descriptors: the BRIEF sampling mode is set to 5×5 neighborhood, the binary descriptor length is 256-bit binary vector, Euclidean distance is used to calculate the descriptor similarity, and a Ratio test is introduced to remove mismatched pairs, finally selecting no less than 20 pairs of matching feature points; if the number of matching pairs is insufficient, the contrast threshold can be appropriately reduced and re-extracted to ensure that the registration data has sufficient constraints.

[0008] Furthermore, the more specific process in step (22) is as follows: Preliminary matching and filtering: Based on the point cloud descriptors and video descriptors obtained in step (21), preliminary matching is performed; for each point cloud descriptor, the nearest neighbor and the second nearest neighbor in the video descriptor are searched by the FLANN matcher, and the Euclidean distance between them is calculated; then, if the Euclidean distance of the nearest neighbor ÷ the Euclidean distance of the second nearest neighbor is less than 0.6, the matching pair is determined to be highly unique and is retained as a preliminary matching pair. Matching pair purification: Based on the initial matching pairs, combined with the known internal parameters and imaging model of the video imaging device, the projection error equation is first constructed using RANSAC to initially eliminate obvious error matching points; then, the projection relationship of the remaining matching pairs is optimized using the least squares method, and the reprojection residual of each matching pair is calculated; finally, matching pairs with residuals > 3 times the mean error are eliminated to obtain the final point cloud-video feature point matching pairs. Matching quality verification: If the number of final retained matching pairs is greater than or equal to 70% of the number of initial matching pairs, and the average reprojection error of all retained matching pairs is less than or equal to 2 pixels, the matching is deemed valid; otherwise, the feature extraction and matching process is re-executed.

[0009] Furthermore, the more specific process in step (23) is as follows: Define the point cloud coordinate system as follows: with the center of the excavation face as the origin, the horizontal axis as the X-axis, the vertical axis as the Y-axis, and the tunnel axis as the Z-axis; Define the pixel coordinate system of the video imaging device: the top left corner of the image is the origin, the horizontal rightward axis is the u-axis, and the vertical downward axis is the v-axis; Coordinate transformation: Using the intrinsic parameters of the video imaging device, the video pixel coordinates are transformed into normalized imaging plane coordinates, and the relationship between the point cloud coordinate system and the video imaging device coordinate system is established. Solving the initial pose: Using the purified point cloud-video feature point matching pair as input, and combining the coordinate association relationship, the rotation matrix R and translation vector T are solved by the EPnP algorithm to determine the initial pose of the camera in the point cloud coordinate system.

[0010] Furthermore, the more specific process in step (24) is as follows: Projection ICP algorithm iteration process: ① Based on the current camera pose [R,T] relative to the point cloud coordinate system, convert all points in the point cloud coordinate system to their poses. R 0 and T 0 Transform to the camera coordinate system, and then combine the video device intrinsic parameters to project onto the video image plane, i.e., the pixel coordinate system, to obtain the two-dimensional projection point ( u,v ① Remove invalid projection points that exceed the image boundaries; ② For each valid projection point, in the video frame, use ( u,v Within a 3×3 pixel neighborhood centered on the point cloud, the pixel that best matches the projection direction of the point cloud is searched based on grayscale similarity to determine the correspondence between the point cloud and video pixels; ③ Construct a projection error function and iteratively solve for the optimal transformation matrix using the Gauss-Newton method, i.e., continuously update... R and T ④ Minimize the error function; repeat the first three steps until either of the following conditions is met: the number of iterations reaches 50; the change in error between two consecutive iterations is ≤ ; Initial frame fusion: Based on the optimized pose [R,T], i.e. the final registration matrix, the RGB color values ​​of the video frames are assigned to the corresponding point clouds to generate a 3D model with RGB texture. Fusion quality verification: Randomly select 100 point cloud points, first convert the RGB values ​​of the point cloud texture and the corresponding pixel RGB values ​​of the video to the CIE Lab color space, and then use the CIE76 color difference formula. Requires average color difference ≤10, to ensure consistency between texture and geometry.

[0011] Furthermore, the more specific process in step (31) is as follows: Mesh division parameters are determined as follows: Based on the size of the excavation face and the density of the point cloud, the side length of the mesh is calculated. Squares are preferred, and special irregular excavation faces can be locally adjusted to rectangles. The three-dimensional model after fusion in step (24) is divided into several monitoring sub-regions according to a uniform mesh. The size of the mesh is adjusted to a 50cm×50cm square area according to the monitoring accuracy requirements to ensure that each monitoring sub-region contains no less than 30 point cloud data points and covers all areas of the excavation face. Verification of grid division results: Count the number of point clouds in all monitored sub-regions to ensure that more than 95% of the monitored sub-regions meet the requirement of ≥50 points; visualize the grid division results, i.e., overlay them on the 3D model of the excavation face to check whether they cover the entire area of ​​the excavation face and whether the edge merging areas are reasonable. Monitoring sub-region generation and edge processing: For each monitoring sub-region, a uniform grid is generated along the side length in both the extension and width directions, with the lower left corner of the 3D model of the excavation face as the origin, to obtain the set of each monitoring sub-region. For incomplete monitoring sub-regions at the edge of the excavation face, a strategy of merging adjacent regions is adopted for merging.

[0012] Furthermore, the more specific process in step (32) is as follows: Point cloud reflectance mean extraction: for each monitoring sub-region Filter out all point cloud data within its three-dimensional range; extract the reflection intensity value of each point. Remove outliers and calculate the arithmetic mean of the remaining intensity values, denoted as . , for The mean reflection intensity of the point cloud; RGB color mean extraction of video region: This is achieved through spatial projection relationships, i.e., based on the extrinsic parameters of the video imaging device, to extract the RGB color mean of the monitored sub-region. The three-dimensional range is projected onto the initial frame of the video to obtain the corresponding two-dimensional pixel region. ;extract RGB values ​​of all pixels ,inp To count the number of pixels, remove outliers; calculate the arithmetic mean of the R, G, and B channels, denoted as . , , ; Pearson correlation coefficient calculation: for each monitoring sub-region Two sets of variables are constructed, which are point cloud feature variables. and video feature variables , , If the grayscale mean is used as the video feature, calculate and Pearson correlation coefficient: The correlation with the RGB channels was calculated separately to obtain... , , Take the largest absolute value as Representative correlation coefficient ; The more specific process in step (33) is as follows: Integrate the characteristics and correlation indicators of all monitored sub-regions in step (32), store them in a relational database or file format and ensure they are indexable, generate a correlation heatmap to visually display the distribution, and output a benchmark report containing data statistics and anomaly descriptions; verify the consistency of correlation coefficients by randomly selecting 10% of sub-regions with an error ≤5%, and archive the database to the project data management system after version marking, as a comparison benchmark for subsequent deformation monitoring.

[0013] Furthermore, the more specific process in step 4 is as follows: Real-time data processing: According to the acquisition cycle of step 1, continuously acquire point cloud data and corresponding video frames at subsequent time nodes, repeat the preprocessing process of step 1 and the registration process of step 2, and establish the mapping relationship between point clouds and corresponding video frames at each time node. Dynamic correlation calculation: For the excavation face at each time point, calculate the average real-time reflection intensity of each monitoring sub-region according to the regional division method in step 3. Compared with real-time RGB color average , , And calculate the real-time Pearson correlation coefficient. ; Deformation region identification: Calculate the change in real-time correlation index of each monitoring sub-region compared to the initial baseline index. If a certain monitoring sub-region If the value increases over time, it indicates that the matching relationship between the point cloud geometric information and the video texture information in this region has changed, and the region is determined to be a potential deformation region.

[0014] Furthermore, the more specific process in step 5 is as follows: Based on the "Technical Specifications for Highway Tunnel Construction," a model similar to the actual tunnel excavation face was built indoors. Known deformations were artificially applied, and point cloud data and video data under different deformations were collected to calculate the corresponding values. Establish deformation and The correlation model will determine the maximum allowable deformation that meets engineering safety requirements. Set as warning threshold ; Monitor potential deformation areas in real time according to steps 1-4. ,when Exceeding the warning threshold When the system automatically initiates a multi-dimensional response, the on-site terminal simultaneously triggers an audible and visual alarm and sends a warning text message to on-site management personnel, along with a message containing the coordinates of the deformed area. The monitoring and early warning signals for information such as estimated deformation are sent to the construction team.

[0015] The beneficial effects of this invention are as follows: This invention combines the three-dimensional geometric accuracy of point clouds with the continuous monitoring characteristics of video to achieve quantitative identification and unmanned early warning of deformed areas, reduce manual intervention, improve the real-time performance and reliability of tunnel construction monitoring, and provide key technical support for the safety of mountain tunnel construction. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a step diagram of the present invention; Figure 2 The Leica RTC360LT is used in this invention; Figure 3 This refers to the AI ​​video surveillance measuring instrument in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 A method for identifying deformation areas at the excavation face of a mountain tunnel based on relevant algorithms and data fusion includes the following steps: Step 1: Synchronously acquire point cloud data and video data of the mountain tunnel excavation face, and preprocess the data. (11): Collect 3D point cloud data and preprocess the data. A 3D laser scanner deployed inside a mountain tunnel is used to scan the tunnel excavation face and obtain high-precision point cloud data. The collected point cloud data is then preprocessed to reduce noise and remove laser points at the edges and those with angles that do not meet the requirements.

[0020] The specific operation is as follows: a phase-type laser scanner can be selected for the 3D laser scanner, such as... Figure 2 The phase-detection laser scanner can be the Leica RTC360 LT, which has a scanning speed of 1 million points per second, a point accuracy of up to 1.9 mm, and a ranging accuracy of 1 mm + 10 ppm. The equipment is installed at a pre-set location around the tunnel excavation face, mounted on a steel support 5–10 m behind the excavation face in the mountain tunnel. The equipment angle, parameters, and installation height are adjusted to ensure the laser scanning range completely covers the excavation face. The Leica RTC360 LT 3D laser scanner continuously acquires point cloud data of the mountain tunnel excavation face at a frequency of 1 million points per second, stored in .las format. The acquired 3D point cloud data can be uniformly filtered using statistical filtering (K-nearest neighbors = 50, standard deviation threshold = 1.0) and radius filtering (radius = 0.01 m, minimum number of points = 5) to remove noise and outliers.

[0021] (12): Acquire video data and preprocess the data. By deploying video imaging devices near mountain tunnels, two-dimensional color videos of the tunnel excavation face are continuously acquired; the acquired videos are then preprocessed by deblurring, white balance correction, and brightness normalization.

[0022] The specific operation is as follows: The video imaging device can use an industrial-grade high-definition camera, such as... Figure 3Industrial-grade high-definition cameras can be equipped with AI video monitoring and measuring instruments. The core parameters of the AI ​​video monitoring and measuring instrument are a resolution of 1920×1080, a frame rate of 25fps, a horizontal field of view of ≥160° and a vertical field of view of ≥80° for panoramic images. The AI ​​video monitoring and measuring instrument is installed on the left side of the 3D laser scanner and must meet two field of view requirements at the same time: first, the monitoring field of view of the video imaging device must cover at least 80% of the overall area of ​​the excavation face; second, the overlapping area of ​​the field of view must cover the central area of ​​the excavation face and the area must be at least 20m², so as to avoid monitoring blind spots. The acquired video data is then subjected to deblurring (Wiener filtering can be used), white balance correction (grayscale world algorithm can be used), and brightness normalization processing (histogram equalization can be used) to eliminate interference from lighting and equipment jitter.

[0023] (13): Synchronously store point cloud data and video data The acquisition timestamps of point cloud data and video data are synchronized using the NTP protocol to ensure data time consistency (ensuring that the time error between the two is less than 10ms). In the case of unstable tunnel network environment, a local NTP server or local time drift compensation module can be used to obtain point cloud data and video data with consistent time sequence identifiers after time synchronization and preprocessing.

[0024] Step 2: Determine the pose of the video imaging device based on the registration algorithm, and construct a 3D model that integrates geometric and texture information. (21): Determine the feature point extraction Based on the point cloud data and video data obtained in step (13), feature points with stability and representativeness are extracted respectively; then, the extracted feature points are encoded using BRIEF descriptors, and a unified descriptor set that can characterize the local geometric and texture features of point cloud and video is generated by calculating the similarity between descriptors.

[0025] The specific operation is as follows: Based on the point cloud data and video data obtained in step (13), the NARF algorithm is used to extract three-dimensional feature points with strong geometric discrimination. The algorithm first uses the neighborhood covariance matrix method to search for neighboring points within a radius of 0.1m for each point, constructs a covariance matrix, and obtains the normal vector through eigenvalue decomposition. The direction of the normal vector must be uniformly pointed to the outside of the excavation face. Further, the surface curvature of each point is calculated, and points with curvature > 0.3 are selected as candidate interest points. Finally, for the candidate interest points, a depth image is constructed along the direction of the normal vector, and the point with the largest radial gradient is extracted as the final feature point. The feature point density needs to be controlled at 50-80 points / m. 2 Ensure that key areas of the excavation face are covered.

[0026] For the initial frames of the preprocessed video data, the SIFT algorithm is used to extract two-dimensional feature points with scale invariance and rotation invariance. The SIFT algorithm first uses a Gaussian pyramid to scale the video frame to 6 scales (i.e., scale factor 1.2), and constructs 5 layers of Gaussian blurred images at each scale, with the difference in standard deviation between adjacent Gaussian layers being 0.8. Further, difference Gaussian images are calculated between Gaussian images at adjacent scales, and key points are screened by local extremum detection, excluding weak key points with contrast less than 8 bits of grayscale value and edge points with edge response >10. Finally, for the initially detected key points, sub-pixel level coordinates are accurately calculated by fitting a quadratic function, and key points with positioning errors greater than 1 pixel are removed. The final number of key points must match the number of feature points in the point cloud.

[0027] Based on the extracted 3D and 2D feature points, the similarity between the two types of feature points is calculated using BRIEF descriptors: the BRIEF sampling mode is set to 5×5 neighborhood, the binary descriptor length is 256-bit binary vector, Euclidean distance is used to calculate the descriptor similarity, and a Ratio test is introduced to remove mismatched pairs, finally selecting no less than 20 pairs of matching feature points; if the number of matching pairs is insufficient, the contrast threshold can be appropriately reduced and re-extracted to ensure that the registration data has sufficient constraints.

[0028] (22): Point cloud and video feature point matching According to the descriptor type selection algorithm in step (21), the initial matching pairs of feature points of point cloud and video are found; the least squares method is used to eliminate the mismatched pairs in the initial matching pairs to ensure the reliability of the correspondence, and finally the point cloud-video feature points are output.

[0029] More specific instructions are as follows: Preliminary matching and filtering: Based on the point cloud descriptors and video descriptors obtained in step (21), preliminary matching is performed. For each point cloud descriptor, the nearest neighbor and the second nearest neighbor in the video descriptor are searched by the FLANN matcher, and the Euclidean distance between them is calculated. Then, if the Euclidean distance of the nearest neighbor ÷ the Euclidean distance of the second nearest neighbor is less than 0.6, the matching pair is determined to be highly unique and is retained as a preliminary matching pair.

[0030] Matching pair purification: Based on the initial matching pairs, combined with the known internal parameters and imaging model of the video imaging device, the projection error equation is first constructed using RANSAC to initially eliminate obvious error matching points; then, the projection relationship of the remaining matching pairs is optimized using the least squares method, and the reprojection residual of each matching pair is calculated; finally, matching pairs with residuals > 3 times the mean error are eliminated to obtain the final point cloud-video feature point matching pairs.

[0031] Matching quality verification: If the number of final retained matching pairs is greater than or equal to 70% of the number of initial matching pairs, and the average reprojection error of all retained matching pairs is less than or equal to 2 pixels, the matching is deemed valid; otherwise, the feature extraction and matching process is re-executed.

[0032] (23): Initial pose solution Using the point cloud-video feature points output in step 22 as input, and combining the known internal parameters of the video imaging device, the initial pose of the video imaging device relative to the 3D laser scanner is solved using the EPnP algorithm.

[0033] The specific steps are as follows: Define the point cloud coordinate system as follows: with the center of the excavation face as the origin, the horizontal axis as the X-axis, the vertical axis as the Y-axis, and the tunnel axis as the Z-axis.

[0034] Define the pixel coordinate system of the video imaging device: the top left corner of the image is the origin, the horizontal axis to the right is the u-axis, and the vertical axis downwards is the v-axis.

[0035] Coordinate transformation: The video pixel coordinates are transformed into normalized imaging plane coordinates using the intrinsic parameters of the video imaging device, and the relationship between the point cloud coordinate system and the video imaging device coordinate system is established.

[0036] Solving the initial pose: Using the purified point cloud-video feature point matching pair as input, and combining the coordinate association relationship, the rotation matrix R and translation vector T are solved by the EPnP algorithm to determine the initial pose of the camera in the point cloud coordinate system.

[0037] (24): Optimize the initial pose and construct a 3D model that integrates geometric and texture information. Based on the initial pose obtained in step 23, the pose is iteratively optimized through the projection ICP algorithm, and then the projection error from the point cloud to the video image is minimized to construct a 3D model that integrates geometric and texture information.

[0038] The specific steps are as follows: Projection ICP algorithm iteration process: ① Based on the current camera pose [R,T] relative to the point cloud coordinate system, convert all points in the point cloud coordinate system to their poses ( R 0 and T 0 The coordinates are transformed to the camera coordinate system, and then combined with the intrinsic parameters of the video device and projected onto the video image plane (pixel coordinate system) to obtain the two-dimensional projection points. u,v ① Remove invalid projection points that exceed the image boundaries; ② For each valid projection point, in the video frame, use ( u,vWithin a 3×3 pixel neighborhood centered on the point cloud, the pixel that best matches the projection direction of the point cloud is searched based on grayscale similarity to determine the correspondence between the point cloud and video pixels; ③ Construct a projection error function and iteratively solve for the optimal transformation matrix using the Gauss-Newton method (i.e., continuously update) R and T ), minimize the error function; ④ Repeat the first three steps until either of the following conditions is met: the number of iterations reaches 50; the change in error between two consecutive iterations is ≤ .

[0039] Initial frame fusion: Based on the optimized pose [R,T] (i.e. the final registration matrix), the RGB color values ​​of the video frames are assigned to the corresponding point clouds to generate a 3D model with RGB texture.

[0040] Fusion quality verification: Randomly select 100 point cloud points, first convert the RGB values ​​of the point cloud texture and the corresponding pixel RGB values ​​of the video to the CIE Lab color space, and then use the CIE76 color difference formula. Requires average color difference ≤10, to ensure consistency between texture and geometry.

[0041] Step 3: Calculate the initial correlation index of the monitoring sub-region under the initial state and establish the monitoring benchmark. (31): Sub-region division for monitoring the excavation face Based on the three-dimensional model in step (24), the monitoring sub-regions are divided into uniform grids to meet the point cloud quantity requirements and cover the entire excavation face area; the average point cloud reflection intensity and the average RGB color of the corresponding video area of ​​each monitoring sub-region are obtained as the basic features for correlation analysis.

[0042] The specific steps are as follows: Mesh division parameters are determined as follows: Based on the size of the excavation face and the density of the point cloud, the side length of the mesh is calculated. Squares are preferred, and special irregular excavation faces can be locally adjusted to rectangles. The three-dimensional model after fusion in step (24) is divided into several monitoring sub-regions according to the uniform mesh. The size of the mesh is adjusted to a 50cm×50cm square area according to the monitoring accuracy requirements to ensure that each monitoring sub-region contains no less than 30 point cloud data points and covers all areas of the excavation face.

[0043] Verification of the division results: Count the number of point clouds in all monitored sub-regions to ensure that more than 95% of the monitored sub-regions meet the requirement of ≥50 points; visualize the grid division results (overlaid on the 3D model of the excavation face) to check whether it covers the entire area of ​​the excavation face and whether the edge merging areas are reasonable.

[0044] Monitoring sub-region generation and edge processing: For each monitoring sub-region, a uniform grid is generated along the side length in both the extension and width directions, with the lower left corner of the 3D model of the excavation face as the origin, to obtain the set of each monitoring sub-region. For incomplete monitoring sub-regions at the edge of the excavation face, a strategy of merging adjacent regions is adopted for merging.

[0045] (32): Calculation of correlation indicators For each monitoring sub-region, the mean point cloud reflectance intensity and the mean RGB color of the corresponding video region are extracted, and the Pearson correlation coefficient between the two is calculated.

[0046] The specific steps are as follows: Point cloud reflectance mean extraction: for each monitoring sub-region Filter out all point cloud data within its three-dimensional range; extract the reflection intensity value of each point. Remove outliers and calculate the arithmetic mean of the remaining intensity values, denoted as . (Right now (mean value of point cloud reflection intensity).

[0047] RGB color mean extraction of video region: By using spatial projection relationships (i.e., based on the extrinsic parameters of the video imaging device), the monitored sub-region is... The three-dimensional range is projected onto the initial frame of the video to obtain the corresponding two-dimensional pixel region. ;extract RGB values ​​of all pixels ,in p To count the number of pixels, remove outlier pixels (e.g., RGB values ​​outside the range [0, 255] are invalid); calculate the arithmetic mean of the R, G, and B channels respectively, and denot it as... , , (Right now (The average RGB color value of the corresponding video area).

[0048] Pearson correlation coefficient calculation: for each monitoring sub-region Construct two sets of variables (point cloud feature variables) and video feature variables , , If the grayscale mean is used as the video feature, calculate... and Pearson correlation coefficient: The correlation with the RGB channels was calculated separately to obtain... , , Take the largest absolute value as Representative correlation coefficient .

[0049] (33): Construction of Deformation Monitoring Benchmark Database Finally, the Pearson correlation coefficient set of these monitored sub-regions was used as the initial correlation index to form a deformation monitoring benchmark database for the entire excavation face.

[0050] The specific steps are as follows: Integrate the characteristics and correlation indicators of all monitored sub-regions in step (32), store them in a relational database or file format and ensure they are indexable, generate a correlation heatmap to visually display the distribution, and output a benchmark report containing data statistics and anomaly descriptions; verify the consistency of correlation coefficients (error ≤ 5%) by randomly selecting 10% of sub-regions, and archive the database to the project data management system after version marking, as a comparison benchmark for subsequent deformation monitoring.

[0051] Step 4: Calculate the correlation index of the monitored sub-region in real time and identify deformed areas. Calculate the real-time point cloud reflectance mean, video RGB color mean, and Pearson correlation coefficient between the two for each monitoring sub-region; compare the changes in the real-time Pearson correlation coefficient with the initial correlation index. Determine whether the area is a potential deformation area.

[0052] The specific steps are as follows: Real-time data processing: Following the acquisition cycle of step 1, continuously acquire point cloud data and corresponding video frames for subsequent time nodes, repeat the preprocessing process of step 1 and the registration process of step 2, and establish the mapping relationship between point clouds and corresponding video frames at each time node.

[0053] Dynamic correlation calculation: For the excavation face at each time point, calculate the average real-time reflection intensity of each monitoring sub-region according to the regional division method in step 3. Compared with real-time RGB color average , , And calculate the real-time Pearson correlation coefficient. .

[0054] Deformation region identification: Calculate the change in real-time correlation index of each monitoring sub-region compared to the initial baseline index. If a certain monitoring sub-region If the value increases over time, it indicates that the matching relationship between the point cloud geometric information (reflection intensity related to three-dimensional shape) and the video texture information (color) in this area has changed, and the area is determined to be a potential deformation area.

[0055] Step 5: Based on the preset threshold early warning mechanism, trigger the deformation early warning and simultaneously push the monitoring and early warning information to the construction party. The deformation and... were obtained through simulation experiments. The correlation model is used to determine the early warning threshold that meets the engineering safety requirements, and the maximum allowable deformation corresponding to the engineering safety requirements is assigned to it. Set as warning threshold Real-time monitoring of potential deformation areas ,when Exceeding the warning threshold When the potential deformation area is upgraded to a deformation area, an audible and visual alarm is automatically triggered, a monitoring report is pushed to the control terminal, and a report containing the coordinates of the deformation area is simultaneously pushed. The monitoring and early warning signals for information such as estimated deformation are sent to the construction team.

[0056] The specific steps are as follows: Based on the "Technical Specification for Highway Tunnel Construction" (JTG / T 3660-2020), a model similar to the actual tunnel excavation face was built indoors. Known deformations were artificially applied, and point cloud data and video data under different deformations were collected to calculate the corresponding values. Establish deformation and The correlation model will determine the maximum allowable deformation that meets engineering safety requirements. Set as warning threshold .

[0057] Monitor potential deformation areas in real time according to steps 1-4. ,when Exceeding the warning threshold When the system automatically initiates a multi-dimensional response, the on-site terminal simultaneously triggers audible and visual alarms (blue warning light for "Attention Level", yellow warning light with low-frequency buzzer for "Early Warning Level", and red warning light with high-frequency buzzer for "Emergency Level"), and sends warning text messages to on-site management personnel, while simultaneously sending messages containing the coordinates of the deformed area, The monitoring and early warning signals for information such as estimated deformation are sent to the construction team.

[0058] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying deformation zones at the excavation face of a mountain tunnel based on correlation algorithms and data fusion, characterized in that: Includes the following steps: Step 1: Synchronously acquire point cloud data and video data of the mountain tunnel excavation face, and preprocess the data. (11): Collect 3D point cloud data and preprocess the data. A 3D laser scanner deployed inside a mountain tunnel is used to scan the tunnel excavation face and obtain high-precision point cloud data. The collected point cloud data is then preprocessed to reduce noise and remove laser points at the edges and those with angles that do not meet the requirements. (12): Acquire video data and preprocess the data. By deploying video imaging devices near the mountain tunnel, two-dimensional color videos of the tunnel excavation face are continuously acquired; the acquired videos are preprocessed by deblurring, white balance correction, and brightness normalization. (13): Synchronously store point cloud data and video data The point cloud data and video data are synchronized using the NTP protocol to ensure data time consistency. After time synchronization and preprocessing, point cloud data and video data with consistent temporal identifiers are obtained; Step 2: Determine the pose of the video imaging device based on the registration algorithm, and construct a 3D model that integrates geometric and texture information. (21): Determine the feature point extraction Based on the point cloud data and video data obtained in step (13), feature points with stability and representativeness are extracted respectively; then descriptors are used to encode the extracted feature points, and by calculating the similarity between descriptors, a unified descriptor set that can characterize the local geometric and texture features of point cloud and video is generated. (22): Point cloud and video feature point matching According to the descriptor type in step (21), select the feature point matching algorithm to find the initial matching pairs of feature points between the point cloud and the video; use the least squares method to remove the mismatched pairs in the initial matching pairs to ensure the reliability of the correspondence, and finally output the point cloud-video feature points. (23): Initial pose solution Using the point cloud-video feature points output in step 22 as input, and combining the known internal parameters of the video imaging device, the initial pose of the video imaging device relative to the 3D laser scanner is solved by the EPnP algorithm. (24): Optimize the initial pose and construct a 3D model that integrates geometric and texture information. Based on the initial pose obtained in step 23, the pose is iteratively optimized through the projection ICP algorithm, and then the projection error from the point cloud to the video image is minimized to construct a three-dimensional model that integrates geometric and texture information. Step 3: Calculate the initial correlation index of the monitoring sub-region under the initial state and establish the monitoring benchmark. (31): Sub-region division for monitoring the excavation face Based on the three-dimensional model in step (24), the monitoring sub-regions are divided into uniform grids to meet the point cloud quantity requirements and cover the entire excavation face area; the average point cloud reflection intensity and the average RGB color of the corresponding video area of ​​each monitoring sub-region are obtained as the basic features for correlation analysis. (32): Calculation of correlation indicators For each monitoring sub-region, the mean reflectance intensity of the point cloud and the mean RGB color of the corresponding video region are extracted, and the Pearson correlation coefficient between the two is calculated. (33): Construction of Deformation Monitoring Benchmark Database Finally, the Pearson correlation coefficient set of these monitoring sub-regions was used as the initial correlation index to form a deformation monitoring benchmark database for the entire excavation face; Step 4: Calculate the correlation index of the monitored sub-region in real time and identify deformed areas. Calculate the real-time point cloud reflectance mean, video RGB color mean, and Pearson correlation coefficient between the two for each monitoring sub-region; compare the changes in the real-time Pearson correlation coefficient with the initial correlation index. Determine whether the area is a potential deformation area; Step 5: Based on the preset threshold early warning mechanism, trigger the deformation early warning and simultaneously push the monitoring and early warning information to the construction party. The deformation and... were obtained through simulation experiments. The correlation model is used to determine the early warning threshold that meets the engineering safety requirements, and the maximum allowable deformation corresponding to the engineering safety requirements is assigned to it. Set as warning threshold Real-time monitoring of potential deformation areas ,when Exceeding the warning threshold When the potential deformation area is upgraded to a deformation area, an audible and visual alarm is automatically triggered, a monitoring report is pushed to the control terminal, and a report containing the coordinates of the deformation area is simultaneously pushed. The monitoring and early warning signals for information such as estimated deformation are sent to the construction team.

2. The method for identifying deformation areas at the excavation face of a mountain tunnel based on relevant algorithms and data fusion as described in claim 1, characterized in that: In step (11), the three-dimensional laser scanner is a phase laser scanner. The equipment is installed at a preset position around the tunnel excavation face. The three-dimensional laser scanner is installed on a steel support 5-10m behind the mountain tunnel excavation face. The equipment angle, parameters and installation height are adjusted to ensure that the laser scanning range completely covers the excavation face. Statistical filtering and radius filtering are used to remove noise. In step (12), the video imaging device uses an industrial-grade high-definition camera; the video imaging device is installed on the left side of the three-dimensional laser scanner and must meet two field of view requirements at the same time: first, the field of view of the video imaging device covers no less than 80% of the overall area of ​​the excavation face; second, the overlapping area of ​​the field of view must cover the central area of ​​the excavation face and the area must be no less than 20m², so as to avoid monitoring blind spots; the collected video data is deblurred by Wiener filtering, white balance is corrected by gray world algorithm, and brightness is normalized by histogram equalization to eliminate interference from lighting and equipment jitter.

3. The method for identifying deformation areas at the excavation face of a mountain tunnel based on relevant algorithms and data fusion as described in claim 1, characterized in that: The more specific process in step (21) is as follows: Based on the point cloud data and video data obtained in step (13), the NARF algorithm is used to extract three-dimensional feature points with strong geometric discriminative power. The algorithm first uses the neighborhood covariance matrix method to search for neighboring points within a radius of 0.1m for each point, constructs a covariance matrix, and obtains the normal vector through eigenvalue decomposition. The direction of the normal vector must be uniformly pointed to the outside of the excavation face. The surface curvature of each point is further calculated, and points with curvature > 0.3 are selected as candidate interest points. Finally, a depth image is constructed along the normal vector direction for the candidate interest points, and the point with the largest radial gradient is extracted as the final feature point. The feature point density needs to be controlled at 50-80 points / m. 2 Ensure coverage of key areas of the excavation face; For the initial frames of the preprocessed video data, the SIFT algorithm is used to extract two-dimensional feature points with scale invariance and rotation invariance. The SIFT algorithm first uses a Gaussian pyramid to scale the video frame to 6 scales, i.e., a scale factor of 1.

2. Five layers of Gaussian blur image are constructed at each scale, and the difference in standard deviation between adjacent Gaussian layers is 0.

8. Then, difference Gaussian images are calculated between Gaussian images at adjacent scales. Key points are screened by local extremum detection to exclude weak key points with contrast less than 8 bits of gray value and edge points with edge response >10. Finally, for the initially detected key points, sub-pixel level coordinates are accurately calculated by fitting a quadratic function to remove key points with positioning errors greater than 1 pixel. The final number of key points must match the number of feature points in the point cloud. Based on the extracted 3D and 2D feature points, the similarity between the two types of feature points is calculated using BRIEF descriptors: the BRIEF sampling mode is set to 5×5 neighborhood, the binary descriptor length is 256-bit binary vector, Euclidean distance is used to calculate the descriptor similarity, and a Ratio test is introduced to remove mismatched pairs, finally selecting no less than 20 pairs of matching feature points; if the number of matching pairs is insufficient, the contrast threshold can be appropriately reduced and re-extracted to ensure that the registration data has sufficient constraints.

4. The method for identifying deformation areas at the excavation face of a mountain tunnel based on relevant algorithms and data fusion as described in claim 1, characterized in that: The more specific process in step (22) is as follows: Preliminary matching and filtering: Based on the point cloud descriptors and video descriptors obtained in step (21), preliminary matching is performed; for each point cloud descriptor, the nearest neighbor and the second nearest neighbor in the video descriptor are searched by the FLANN matcher, and the Euclidean distance between them is calculated; then, if the Euclidean distance of the nearest neighbor ÷ the Euclidean distance of the second nearest neighbor is less than 0.6, the matching pair is determined to be highly unique and is retained as a preliminary matching pair. Matching pair purification: Based on the initial matching pairs, combined with the known internal parameters and imaging model of the video imaging device, the projection error equation is first constructed using RANSAC to initially eliminate obvious error matching points; then, the projection relationship of the remaining matching pairs is optimized using the least squares method, and the reprojection residual of each matching pair is calculated; finally, matching pairs with residuals > 3 times the mean error are eliminated to obtain the final point cloud-video feature point matching pairs. Matching quality verification: If the number of final retained matching pairs is greater than or equal to 70% of the number of initial matching pairs, and the average reprojection error of all retained matching pairs is less than or equal to 2 pixels, the matching is deemed valid; otherwise, the feature extraction and matching process is re-executed.

5. The method for identifying deformation areas at the excavation face of a mountain tunnel based on relevant algorithms and data fusion as described in claim 1, characterized in that: The more specific process in step (23) is as follows: Define the point cloud coordinate system as follows: with the center of the excavation face as the origin, the horizontal axis as the X-axis, the vertical axis as the Y-axis, and the tunnel axis as the Z-axis; Define the pixel coordinate system of the video imaging device: the top left corner of the image is the origin, the horizontal rightward axis is the u-axis, and the vertical downward axis is the v-axis; Coordinate transformation: Using the intrinsic parameters of the video imaging device, the video pixel coordinates are transformed into normalized imaging plane coordinates, and the relationship between the point cloud coordinate system and the video imaging device coordinate system is established. Solving the initial pose: Using the purified point cloud-video feature point matching pair as input, and combining the coordinate association relationship, the rotation matrix R and translation vector T are solved by the EPnP algorithm to determine the initial pose of the camera in the point cloud coordinate system.

6. The method for identifying deformation areas at the excavation face of a mountain tunnel based on correlation algorithms and data fusion as described in claim 1, characterized in that: The more specific process in step (24) is as follows: Projection ICP algorithm iteration process: ① Based on the current camera pose [R,T] relative to the point cloud coordinate system, convert all points in the point cloud coordinate system to their poses. R 0 and T 0 Transform to the camera coordinate system, and then combine the video device intrinsic parameters to project onto the video image plane, i.e., the pixel coordinate system, to obtain the two-dimensional projection point ( u,v ① Remove invalid projection points that exceed the image boundaries; ② For each valid projection point, in the video frame, use ( u,v Within a 3×3 pixel neighborhood centered on the point cloud, the pixel that best matches the projection direction of the point cloud is searched based on grayscale similarity to determine the correspondence between the point cloud and the video pixel. ③ Construct the projection error function and iteratively solve for the optimal transformation matrix using the Gauss-Newton method, i.e., continuously update... R and T ④ Minimize the error function; repeat the first three steps until either of the following conditions is met: the number of iterations reaches 50; the change in error between two consecutive iterations is ≤ ; Initial frame fusion: Based on the optimized pose [R,T], i.e. the final registration matrix, the RGB color values ​​of the video frames are assigned to the corresponding point clouds to generate a 3D model with RGB texture. Fusion quality verification: Randomly select 100 point cloud points, first convert the RGB values ​​of the point cloud texture and the corresponding pixel RGB values ​​of the video to the CIE Lab color space, and then use the CIE76 color difference formula. Requires average color difference ≤10, to ensure consistency between texture and geometry.

7. The method for identifying deformation areas at the excavation face of a mountain tunnel based on correlation algorithms and data fusion as described in claim 1, characterized in that: The more specific process in step (31) is as follows: Mesh division parameters are determined as follows: Based on the size of the excavation face and the density of the point cloud, the side length of the mesh is calculated. Squares are preferred, and special irregular excavation faces can be locally adjusted to rectangles. The three-dimensional model after fusion in step (24) is divided into several monitoring sub-regions according to a uniform mesh. The size of the mesh is adjusted to a 50cm×50cm square area according to the monitoring accuracy requirements to ensure that each monitoring sub-region contains no less than 30 point cloud data points and covers all areas of the excavation face. Verification of grid division results: Count the number of point clouds in all monitored sub-regions to ensure that more than 95% of the monitored sub-regions meet the requirement of ≥50 points; visualize the grid division results, i.e., overlay them on the 3D model of the excavation face to check whether they cover the entire area of ​​the excavation face and whether the edge merging areas are reasonable. Monitoring sub-region generation and edge processing: For each monitoring sub-region, a uniform grid is generated along the side length in both the extension and width directions, with the lower left corner of the 3D model of the excavation face as the origin, to obtain the set of each monitoring sub-region. For incomplete monitoring sub-regions at the edge of the excavation face, a strategy of merging adjacent regions is adopted for merging.

8. The method for identifying deformation areas at the excavation face of a mountain tunnel based on correlation algorithms and data fusion as described in claim 1, characterized in that: The more specific process in step (32) is as follows: Point cloud reflectance mean extraction: for each monitoring sub-region Filter out all point cloud data within its three-dimensional range; extract the reflection intensity value of each point. Remove outliers and calculate the arithmetic mean of the remaining intensity values, denoted as . , for The mean reflection intensity of the point cloud; RGB color mean extraction of video region: This is achieved through spatial projection relationships, i.e., based on the extrinsic parameters of the video imaging device, to extract the RGB color mean of the monitored sub-region. The three-dimensional range is projected onto the initial frame of the video to obtain the corresponding two-dimensional pixel region. ;extract RGB values ​​of all pixels ,in p To count the number of pixels, remove outliers; calculate the arithmetic mean of the R, G, and B channels, denoted as . , , ; Pearson correlation coefficient calculation: for each monitoring sub-region Two sets of variables are constructed, which are point cloud feature variables. and video feature variables , , If the grayscale mean is used as the video feature, calculate and Pearson correlation coefficient: The correlation with the RGB channels was calculated separately to obtain... , , Take the largest absolute value as Representative correlation coefficient ; The more specific process in step (33) is as follows: Integrate the characteristics and correlation indicators of all monitored sub-regions in step (32), store them in a relational database or file format and ensure they are indexable, generate a correlation heatmap to visually display the distribution, and output a benchmark report containing data statistics and anomaly descriptions; verify the consistency of correlation coefficients by randomly selecting 10% of sub-regions with an error ≤5%, and archive the database to the project data management system after version marking, as a comparison benchmark for subsequent deformation monitoring.

9. The method for identifying deformation areas at the excavation face of a mountain tunnel based on correlation algorithms and data fusion as described in claim 1, characterized in that: The more detailed process in step 4 is as follows: Real-time data processing: According to the acquisition cycle of step 1, continuously acquire point cloud data and corresponding video frames at subsequent time nodes, repeat the preprocessing process of step 1 and the registration process of step 2, and establish the mapping relationship between point clouds and corresponding video frames at each time node. Dynamic correlation calculation: For the excavation face at each time point, calculate the average real-time reflection intensity of each monitoring sub-region according to the regional division method in step 3. Compared with real-time RGB color average , , And calculate the real-time Pearson correlation coefficient. ; Deformation region identification: Calculate the change in real-time correlation index of each monitoring sub-region compared to the initial baseline index. ; If a certain monitoring sub-area If the value increases over time, it indicates that the matching relationship between the point cloud geometric information and the video texture information in this region has changed, and the region is determined to be a potential deformation region.

10. The method for identifying deformation areas at the excavation face of a mountain tunnel based on correlation algorithms and data fusion as described in claim 1, characterized in that: The more detailed process in step 5 is as follows: Based on the "Technical Specifications for Highway Tunnel Construction," a model similar to the actual tunnel excavation face was built indoors. Known deformations were artificially applied, and point cloud data and video data under different deformations were collected to calculate the corresponding values. Establish deformation and The correlation model will determine the maximum allowable deformation that meets engineering safety requirements. Set as warning threshold ; Monitor potential deformation areas in real time according to steps 1-4. ,when Exceeding the warning threshold When the system automatically initiates a multi-dimensional response, the on-site terminal simultaneously triggers an audible and visual alarm and sends a warning text message to on-site management personnel, along with a message containing the coordinates of the deformed area. The monitoring and early warning signals for information such as estimated deformation are sent to the construction team.

Images