A tunnel deformation monitoring method based on line structured light

Abstract

The embodiment of the application discloses a tunnel deformation monitoring method based on line structured light, relates to tunnel engineering monitoring technology, and comprises the following steps: acquiring a laser line image projected on the inner wall of a tunnel and performing pretreatment; inputting the pretreated image into a deep learning semantic segmentation model to output a pixel region of the laser line in the image; performing center line extraction processing on the laser line pixel region to obtain a single-pixel-width laser center line; arranging a reference target with a known physical size in a monitoring scene; determining a dynamic scale coefficient between pixel size and physical size by processing an image containing the reference target; calculating physical displacement based on laser center lines at different time points by using the dynamic scale coefficient to determine the tunnel deformation; and the method can effectively overcome the scale uncertainty problem caused by the camera object distance and attitude change, significantly improves the measurement accuracy, environmental adaptability and system robustness, and realizes high-precision real-time automatic monitoring.

Description

Technical Field

[0001] This invention relates to tunnel engineering monitoring technology, and more particularly to a tunnel deformation monitoring method based on line structured light. Background Technology

[0002] As vital transportation and public infrastructure, the structural health and safety of tunnels during operation are of paramount importance. Long-term, real-time, and high-precision monitoring of tunnel structural deformation is a key aspect of ensuring safe operation. Linear structured light visual measurement technology, due to its advantages of being non-contact, highly efficient, and having high data density, is widely used to acquire the contour information of tunnel cross-sections, thereby analyzing their deformation.

[0003] Currently, to accurately extract laser lines from the complex background textures, uneven lighting, and noise interference such as dust and water stains on the tunnel walls, some solutions have begun to employ deep learning methods, such as semantic segmentation models based on convolutional neural networks, to process the acquired laser line images. Compared to traditional image processing algorithms, deep learning models, with their powerful feature learning and representation capabilities, have improved the accuracy and robustness of laser line extraction to a certain extent. However, some inherent defects still exist. On the one hand, measurement accuracy is greatly affected by dynamic factors. In actual monitoring, monitoring devices (especially cameras) may undergo slight changes due to factors such as equipment vibration and temperature variations, introducing significant measurement errors and making it difficult to meet the high precision requirements of millimeters or even sub-millimeter levels required for tunnel safety monitoring. On the other hand, systematic errors are not effectively corrected. During on-site installation, the optical axis of the image acquisition module is often difficult to keep absolutely perpendicular to the laser projection plane. The angle between the two causes projection distortion, resulting in a compression effect on the imaging plane of the actual displacement, leading to inherent deviations in the calculation results of deformation and reducing the reliability of the monitoring data. Summary of the Invention

[0004] The main objective of this invention is to provide a tunnel deformation monitoring method based on line structured light, which aims to solve the technical problems of inaccurate measurement scale and low precision caused by factors such as camera pose, dynamic changes in object distance, and projection distortion in the prior art, so as to achieve high-precision and high-reliability quantitative monitoring of tunnel deformation.

[0005] To achieve the above objectives, the first aspect of this application provides a tunnel deformation monitoring method based on line structured light, the method comprising:

[0006] Acquire an image of the laser line projected onto the inner wall of the tunnel, and preprocess the laser line image;

[0007] The preprocessed laser line image is input into a pre-trained deep learning semantic segmentation model, which identifies and outputs the pixel region of the laser line in the laser line image.

[0008] The centerline of the laser line pixel region is extracted to obtain a laser centerline with a single pixel width;

[0009] A reference target with a known physical size is deployed within the monitoring scene, and a dynamic scaling factor between the pixel size and the physical size is determined by processing an image containing the reference target.

[0010] Based on the laser centerline acquired at least at two different times, the physical displacement between the laser centerlines is calculated using the dynamic scaling factor to determine the amount of tunnel deformation.

[0011] A second aspect of this application provides a tunnel deformation monitoring system based on line structured light, comprising a line structured light projection module, an image acquisition module, and a central processing module, wherein the central processing module is used for:

[0012] The laser line image projected onto the tunnel wall, acquired by the image acquisition module, is preprocessed.

[0013] The preprocessed laser line image is input into a pre-trained deep learning semantic segmentation model, which identifies and outputs the pixel region of the laser line in the laser line image.

[0014] The centerline of the laser line pixel region is extracted to obtain a laser centerline with a single pixel width;

[0015] By processing an image acquired by the image acquisition module that contains a reference target with a known physical size, a dynamic scaling factor between the pixel size and the physical size is determined;

[0016] Based on the laser centerline acquired at at least two different times, and using the dynamic scaling factor, the physical displacement between the laser centerlines is calculated to determine the amount of tunnel deformation.

[0017] A third aspect of this application provides an electronic device including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform steps as described in the first aspect and any possible implementation thereof.

[0018] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the first aspect and any possible implementation thereof.

[0019] This application provides a tunnel deformation monitoring method based on line structured light. The method involves acquiring an image of a laser line projected onto the inner wall of the tunnel and preprocessing the image. The preprocessed image is then input into a pre-trained deep learning semantic segmentation model, which identifies and outputs the pixel region of the laser line in the image. The pixel region is then processed to extract the centerline, resulting in a laser centerline with a single pixel width. Reference targets with known physical dimensions are deployed within the monitoring scene, and an image containing the reference targets is processed to determine a dynamic scaling factor between the pixel size and the physical size. Based on the laser centerline acquired at at least two different times, the dynamic scaling factor is used to calculate the physical displacement between the laser centerlines to determine the amount of tunnel deformation.

[0020] The technical solution provided in this application has the following beneficial effects:

[0021] 1. Significantly improved measurement accuracy. By deploying reference targets with known physical dimensions at the monitoring site and calculating dynamic scaling factors in real time, this application effectively overcomes the uncertainty in measurement scale caused by dynamic factors such as camera distance and attitude changes, achieving online dynamic calibration and greatly improving the accuracy of deformation calculation. 2. Enhanced system environmental adaptability and robustness. This application combines the powerful anti-interference feature extraction capabilities of deep learning models with a dynamic calibration mechanism, enabling the method to not only stably extract laser lines under harsh conditions such as tunnel dust, water seepage, and uneven lighting, but also ensure the accuracy of measurement results, exhibiting strong overall robustness. 3. Achieved high-precision real-time automated monitoring. The technical solution provided in this application has a high degree of automation in its processing flow, enabling rapid analysis of continuously acquired images and real-time output of high-precision deformation data. This meets the dual requirements of timeliness and accuracy for tunnel safety monitoring, providing reliable data for tunnel structure health assessment and early warning decisions. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] in:

[0024] Figure 1 A schematic flowchart illustrating a tunnel deformation monitoring method based on line structured light provided in an embodiment of this application;

[0025] Figure 2A schematic diagram of a model structure based on the Swing Transformer architecture provided for an embodiment of this application;

[0026] Figure 3 A schematic diagram of a laser line skeletonization effect provided in an embodiment of this application;

[0027] Figure 4A A pixel displacement diagram provided for an embodiment of this application;

[0028] Figure 4B A schematic diagram of a standard gauge block provided in an embodiment of this application;

[0029] Figure 5 This is a schematic diagram of the deployment architecture of a tunnel deformation monitoring system based on line structured light, provided in an embodiment of this application. Detailed Implementation

[0030] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0031] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0032] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0033] The embodiments of this application are described below with reference to the accompanying drawings.

[0034] Figure 1 A flowchart illustrating a tunnel deformation monitoring method based on line structured light, as provided in this application embodiment, is shown below. Figure 1 As shown, the method includes:

[0035] 101. Obtain the laser line image projected onto the inner wall of the tunnel, and preprocess the laser line image.

[0036] The execution entity in this application embodiment can be a tunnel deformation monitoring system based on line structured light.

[0037] The core task of step 101 is to provide high signal-to-noise ratio laser line images for subsequent algorithms. Specifically, firstly, raw images of the laser lines on the tunnel inner wall can be continuously captured using a high-resolution industrial camera, and then Gaussian filtering and morphological closing operations can be used to preprocess the images.

[0038] In one optional implementation, the preprocessing described above includes:

[0039] Gaussian filtering was used to denoise the laser line image; and morphological closing operations were used to repair the local breaks in the laser line.

[0040] Tunnel environments typically present challenges such as strong ambient light interference, suspended dust, water seepage reflections, and speckle noise. These factors cause uneven brightness, broken lines, and background clutter in the acquired laser lines, severely impacting subsequent accurate laser line extraction. To improve the image signal-to-noise ratio and enhance laser line continuity, this application proposes an image preprocessing workflow based on Gaussian filtering and morphological closing operations. This system eliminates environmental noise and interference, laying the foundation for subsequent accurate detection.

[0041] Specifically, firstly, the original acquired images I ( x , y Perform Gaussian filtering. G σ ( x , y The definition of ) is:

[0042]

[0043] in, σ The standard deviation of the Gaussian kernel determines the smoothness. This invention, based on the tunneling laser linewidth and noise level, preferably... σ The value is between 0.8 and 2.0 to smooth high-frequency noise while preserving the edge details of the laser line. Gaussian filtering results. I G ( x , y This can be obtained through convolution calculation:

[0044]

[0045] Gaussian filtering not only suppresses random noise generated by dust particles but also homogenizes the background grayscale, enhancing the grayscale contrast between the laser line and the background. After Gaussian filtering, to further repair local breaks in the laser line, this invention employs morphological closing operations. I C ( x , y ) is defined as:

[0046]

[0047] Where ⊕ represents the expansion operation and ⊖ represents the corrosion operation. B These are morphological structural elements. Dilation operations fill small gaps in laser lines, while erosion is used to restore the actual width of the lines. Structural elements B A 3×3 or 5×5 square or elliptical core is typically used, and its shape and size are determined experimentally to accommodate laser linewidths under different tunnel conditions. The expansion and corrosion operation formulas are as follows:

[0048]

[0049]

[0050] Morphological closing operations, through a combination of dilation and erosion, can fill small gaps in laser lines, connect broken segments, and smooth line edges. This process is particularly crucial because tunnels contain numerous micro-cracks, dust obstructions, and other factors that can easily cause laser line interruptions. Directly using broken laser lines for subsequent analysis would inevitably lead to discontinuous cross-section fitting results, affecting measurement accuracy. Closing operations not only restore the continuity of the lines but also suppress some isolated noise points, making the laser lines smoother and more continuous.

[0051] By employing a combined preprocessing method of Gaussian filtering and morphological closing operations, this embodiment of the application can preserve the main morphological features of the laser line while maximally suppressing environmental noise and optical interference, laying a solid foundation for subsequent sub-pixel precision extraction. This process is particularly suitable for harsh working conditions such as dusty and highly reflective environments, significantly improving the system's environmental adaptability.

[0052] 102. Input the preprocessed laser line image into a pre-trained deep learning semantic segmentation model, and the model identifies and outputs the pixel region of the laser line in the laser line image.

[0053] The preprocessed image is input into a pre-trained deep learning semantic segmentation model for recognition. Optionally, a model based on the Swing Transformer architecture or a model based on the U-Net architecture can be used.

[0054] In tunnel monitoring scenarios, laser lines typically appear as long, continuous, bright areas in images, with an extremely low pixel ratio, representing a typical problem of sparse distribution of small targets. Traditional image processing algorithms struggle to accurately extract such sparse linear targets in heavily interfering backgrounds and are highly susceptible to factors such as lighting changes, structural textures, and reflections, leading to false detections, missed detections, or positional shifts. To achieve high-precision identification and extraction of laser lines, this application's embodiments introduce a semantic segmentation network based on a visual Transformer architecture, employing the Swing Transformer as the backbone feature extraction network. This fully leverages its global modeling capabilities and local window attention mechanism, effectively enhancing the perception of fine-grained linear targets while maintaining computational efficiency.

[0055] The Swin Transformer features a hierarchical feature extraction structure and a sliding window self-attention mechanism, enabling it to capture local and global features in images at different spatial scales, making it particularly suitable for non-uniformly distributed target detection tasks. In this embodiment, the Swin Transformer can fully model the contextual relationship between laser lines and their surrounding environment in an image, significantly improving the ability to recognize fine lines while maintaining spatial structure. Furthermore, the inter-layer patch merging operation can enhance the network's expressive power while maintaining image resolution, effectively mitigating the feature vanishing problem caused by excessively thin laser lines, thus ensuring the model maintains good robustness and generalization ability in complex tunnel environments.

[0056] Figure 2 This is a schematic diagram of a model structure based on the Swing Transformer architecture provided for an embodiment of this application.

[0057] like Figure 2As shown, the model specifically employs an encoder-decoder structure. Specifically, the network model in this application starts with the input image, forms initial features through image patch segmentation and linear embedding, and then enters the hierarchical extraction process of the Swin-T backbone. At each stage, multi-head self-attention and shifted-window self-attention work alternately, in conjunction with a feedforward network, to finely model local patterns and establish global connections across windows under the shifting mechanism. Downsampling and channel representation are improved through image patch merging between stages, ensuring that the thin and sparse laser lines remain clearly discernible at the feature level. The multi-scale features output from the backbone are fed into the UPerNet decoder. First, the pyramid pooling module obtains contextual information under different receptive fields. Then, it aligns and fuses with the backbone features along a top-down fusion path, gradually restoring spatial details through skip connections and upsampling. Finally, it projects the laser line segmentation result onto the category space via pointwise convolution. Ultimately, a binary segmentation mask with the same size as the input image is output. In this mask, regions with a pixel value of 1 (or 255) precisely correspond to the location of the laser line, while regions with a pixel value of 0 represent the background.

[0058] During the training phase, cross-entropy with class weights is employed to alleviate class imbalance, and online hard sample mining is used to highlight easily confused regions. Dice Loss is also introduced to enhance supervision of fine boundaries and reduce breakage and offset. In the feature extraction process, SE-Unet-style channel attention is superimposed as a feature modulation module. Through compression and excitation, channel importance is adaptively recalibrated, improving the response of weak-contrast laser lines and enhancing robustness under complex lighting and strong reflection conditions, thereby achieving stable and high-precision extraction of laser lines in tunnel scenes.

[0059] The Dice Loss mentioned in this embodiment is a loss function commonly used in image segmentation tasks. It is based on the Dice coefficient and is used to measure the similarity between the predicted segmentation result and the ground truth annotation. The SE-Unet mentioned in this embodiment is an improved U-Net model that incorporates the Squeeze-and-Excitation (SE) attention mechanism and is widely used in medical image segmentation and other image segmentation tasks.

[0060] Additionally, optionally, to address the training sample imbalance problem caused by the extremely low proportion of positive laser line targets, this application proposes an image cropping and sample selection strategy based on semantic mask content. By performing a fixed-window sliding cropping of the original image, only image patches containing laser line targets are retained for training, significantly increasing the proportion of positive samples in the training data and mitigating the adverse effects of class imbalance on model training. Simultaneously, this cropping strategy maintains the diversity of the original image's spatial distribution, avoiding the offset caused by the laser line targets being concentrated in a certain image region.

[0061] In addition, to further improve the model's recognition ability under various laser line postures, strategies such as random rotation, mirror flipping, and illumination perturbation can be introduced in the data augmentation stage to simulate the distribution characteristics of laser lines in different directions and under different brightness conditions in the tunnel, thereby improving the model's adaptability in unstructured scenarios.

[0062] 103. Perform centerline extraction processing on the above laser line pixel area to obtain a laser centerline with a single pixel width.

[0063] The laser line region output by the deep learning model is a bright band with a width of several pixels. For accurate geometric measurement, it needs to be refined into a single-pixel center line.

[0064] In one alternative implementation, the skeletonization algorithm is used to process the recognition results of deep learning, and the laser line with a certain thickness is refined into a single-pixel laser center line. This can solve the problem that directly using the recognition results of deep learning will lead to unstable feature points in subsequent displacement calculations.

[0065] Although deep learning methods can achieve sub-pixel accuracy in locating laser lines, in real-world images, laser lines still retain a certain width, especially in environments with high brightness and strong scattering, where they appear as bright bands several pixels wide. Directly using such wide lines for cross-sectional analysis results in significant data redundancy and is prone to matching errors. Therefore, this embodiment further employs a skeletonization algorithm to perform single-pixel processing on the extracted laser lines.

[0066] The core purpose of skeletonization is to simplify a multi-pixel wide laser line into a single-pixel wide center line while preserving its geometric topology. Compared to directly using a multi-pixel wide laser line, skeletonization results in a thinner and more continuous center line, significantly reducing computational redundancy and positioning errors in subsequent point cloud fitting and matching. Furthermore, the skeletonized center line exhibits extremely high noise resistance, further suppressing stray line segments caused by environmental interference, preserving the true physical path of the laser line, thereby effectively compressing data volume and enhancing noise resistance. The skeletonized center line also possesses extremely high noise resistance, suppressing pseudo-line segments caused by environmental interference such as reflections and crack shadows, thus preserving the true physical path of the laser line and ensuring the accuracy and stability of subsequent analysis. This application's embodiments can use the Zhang-Suen skeletonization algorithm, performing iterative pruning based on pixel neighborhood conditional judgments, which can remove redundant pixels while preserving the image topology.

[0067] Please see Figure 3 The image visually illustrates the effect before and after skeletonization. This algorithm is an iterative thinning algorithm that, in two sub-iteration steps, checks the eight-neighborhood connectivity of each target pixel in parallel and, if specific deletion conditions are met, progressively deletes pixels located at the region edges that do not affect the overall topological connectivity. This process is repeated until no pixels can be deleted, and the final set of remaining pixels constitutes the single-pixel-width skeleton of the original region, which is the laser centerline. Specifically, the laser centerline acquired at the initial time (e.g., after the system deployment is complete and stable) can be denoted as the baseline centerline C(t0), and the laser centerline acquired at the current time t can be denoted as C(t). Optionally, the above centerline extraction process uses a skeletonization algorithm.

[0068] 104. Deploy reference targets with known physical dimensions within the monitoring scene, and determine the dynamic scaling factor between pixel size and physical size by processing images containing the aforementioned reference targets.

[0069] To achieve high-precision quantitative monitoring of tunnel cross-sectional deformation, this application proposes a deformation analysis method based on the ratio method. This method converts the two-dimensional pixel displacement in optical measurements into actual physical displacement by arranging standard gauge blocks of known dimensions on the tunnel inner wall, and proposes a dynamic monitoring model, effectively improving measurement accuracy.

[0070] Figure 4A This is a schematic diagram of pixel displacement provided for an embodiment of this application. For example... Figure 4A As shown, a1, b1, and c1 in the right image are image feature points, corresponding to points a, b, and c in the standard block in the left original image. The distance between points a1 and b1 in the image is the pixel displacement.

[0071] Figure 4B This is a schematic diagram of a standard gauge block provided for an embodiment of this application.

[0072] Specifically, reference targets with known physical dimensions (such as standard gauge blocks) can be deployed within the monitoring scene. By detecting their pixel length in the image, the "pixel-physical" ratio coefficient, i.e., the aforementioned dynamic ratio coefficient, is calculated and updated in real time. This coefficient dynamically adjusts with changes in shooting distance and angle to ensure that subsequent displacement results always correspond to the actual physical quantity. In one implementation, the calculation method and its derivation include:

[0073] Ideally, the light plane and the image plane are strictly parallel, and the object lies in the camera coordinate system. x Actual displacement Δ in the direction X With pixel displacement Δ u The following linear relationship exists:

[0074]

[0075] in, p This represents the physical size of a single pixel in the camera. Z The object distance is the spatial distance between the object being measured and the camera. f Δ is the lens focal length; u This represents the pixel displacement. However, in tunnel monitoring scenarios, there is often an angle between the camera and the laser projection plane, causing the actual displacement to vary slightly from the camera's position. x The direction produces a projection compression effect. To correct this projection distortion, a spatial rotation matrix can be used. R Describe the attitude difference between the light plane and the image plane. Assume the laser plane is rotated about the y-axis by an angle of... θ Then the projection matrix is:

[0076]

[0077] In this geometric model, the object is along the radial direction of the tunnel, i.e., the direction within the laser plane. X The true displacement Δ′ X ′, it is in the camera x The projection components of the direction are:

[0078]

[0079] This indicates that when the actual displacement is projected onto the image plane, a compression effect will occur, with a compression ratio of cos... θ Ignoring this projection correction will lead to systematic errors in deformation calculations, especially when the angle between the camera and the tunnel wall is large. Therefore, this invention proposes a dynamic correction scaling factor. k dyn Its definition is:

[0080]

[0081] Using this scaling factor, the pixel displacement measured by the camera can be accurately converted into the actual displacement Δ of the object in the laser plane. X ′. Further Δ X proj With pixel displacement Δ u Connecting them, we can

[0082]

[0083] Therefore, the actual displacement can be expressed as:

[0084]

[0085] The scaling factor at any pixel position u can be written as:

[0086]

[0087]

[0088] in, k eff ( u ; Z , t () is the equivalent proportionality coefficient, representing the distance between objects. Z and time t Below, the pixel x-coordinate is u The conversion relationship between the pixel displacement and the actual displacement. This indicates that it is based on a reference scaling factor. k dyn Based on this, a spatial adaptive scale factor is obtained after projection geometry correction. In practical applications, the angle... θ Usually unknown, it can be solved by experimentally calibrating data. Let the x-axis be... u i Collected from M i Group of observation pairs {(Δ u ij , Δ X ’ ij Then we have the observation equation.

[0089]

[0090] in, ε ij This is to account for observation noise. When there is limited calibration data, a weighted least squares fit can be used to estimate the optimal value for all observation points. θ Then, the effective proportional coefficient is calculated using the geometric correction term; when calibration data is abundant, there is no need to explicitly estimate the angle parameters. θ It can be directly applied to multiple pixels. u i The empirical scaling factor is obtained based on location. This takes into account the pixel x-coordinate in tunnel monitoring scenarios. u The relationship between the actual displacement and the actual displacement is nonlinear. At the same time, it is necessary to balance computational stability and fitting accuracy. This invention selects a fifth-order polynomial as the basis function, which can balance fitting flexibility and numerical stability.

[0091] To balance the quality of different experimental data, a diagonal weight matrix is ​​defined. The coefficient vector can be obtained through the least squares solution, forming a continuous mapping function. k eff This allows pixel displacement to be accurately converted into real displacement.

[0092] It should be noted that Δ in the above formula X′ This mainly corresponds to the radial displacement in the cross-sectional direction of the tunnel, i.e., convergent deformation. Similarly, the settlement deformation Δ of the tunnel surrounding rock... Y′ The same derivation process as described above can be used.

[0093] 105. Based on the laser centerlines obtained at at least two different times, the physical displacement between the laser centerlines is calculated using the dynamic scaling factor to determine the amount of tunnel deformation.

[0094] The system registers the laser centerline acquired at least two different times, and uses the obtained dynamic scaling coefficient to convert the pixel-level displacement into the real physical displacement, thereby obtaining the convergence, settlement or horizontal displacement deformation of the tunnel structure during that time period, which can realize real-time quantitative assessment of the tunnel's safety status.

[0095] In an optional implementation, step 105 includes:

[0096] By matching the laser centerline at the current moment with that at the previous moment point by point, the displacement of each point in the image coordinate system can be obtained.

[0097] Based on the aforementioned dynamic proportional coefficient, the displacement is converted into actual physical displacement to obtain the deformation of the tunnel.

[0098] When an object deforms, the pixel displacement Δ is measured in real time. u Multiply k dyn This allows us to obtain the actual deformation of the object:

[0099]

[0100] Dynamically updated under different object distances Z n pxThis method enables online correction of the proportionality coefficient, overcoming the limitation of traditional ratio methods which are only applicable to parallel installations, and significantly improves the measurement accuracy and reliability of tunnel deformation monitoring. Furthermore, this method is not only applicable to single-point displacement measurement but also to multi-point cross-sectional deformation quantification analysis based on continuous laser centerlines. The embodiments of this application utilize a dynamic calibration mechanism to organically integrate projection compression compensation with the ratio method, retaining the advantages of the ratio method's simplicity and efficiency while significantly improving measurement accuracy under complex spatial postures, making it particularly suitable for real-time deformation monitoring in narrow spatial environments such as tunnels.

[0101] The tunnel deformation monitoring method based on line structured light in this application employs a Swin Transformer network for laser line extraction, solving the challenging problem of laser line extraction in tunnel environments. Traditional methods often rely on threshold segmentation or edge detection, which are susceptible to background noise and lack accuracy. In contrast, the Swin Transformer leverages its advantages in processing local and global image features to achieve high-precision laser line extraction.

[0102] Furthermore, considering the extremely small proportion of laser lines in tunnel images and the severe shortage of positive sample data, this application employs data augmentation techniques, such as cropping, rotation, and mirroring, to synthesize more positive sample data, thereby improving the network training effect and the robustness of laser line extraction. This application ensures accurate calculation of deformation under different monitoring conditions by real-time calibration of the scaling factor, effectively improving the accuracy of tunnel deformation monitoring, and is particularly suitable for real-time monitoring of millimeter-level deformation.

[0103] Based on the aforementioned method embodiments, this application also provides a tunnel deformation monitoring system based on line structured light.

[0104] Please see Figure 5 , Figure 5 This is a schematic diagram of the deployment architecture of a tunnel deformation monitoring system based on line structured light, provided in an embodiment of this application.

[0105] like Figure 5As shown, a monitoring device is deployed at the tunnel cross-section to be monitored. This device mainly includes a line structured light projector 2, an industrial camera 3, and a central processing module 5. The line structured light projector 2 can be a 5-watt line laser, which is firmly mounted on one side of the tunnel to project a clear and uniform red and blue laser line 6 onto the opposite tunnel wall 1. The industrial camera 3 can be an industrial-grade camera, installed alongside or at a certain angle to the line structured light projector 2, with its lens field of view completely covering the projection area of ​​the laser line 6. It should be noted that within the field of view of the industrial camera 3, one or more standard gauge blocks 4 with known physical dimensions are firmly fixed to the tunnel wall 1. In this embodiment, the reference target 4 is specifically a precision-machined high-speed steel standard gauge block with a length L. real To ensure accurate known values, the central processing module 5 can be an industrial control computer or an embedded computing device, internally equipped with a high-performance processor (such as a multi-core central processing unit and graphics processing unit), memory, storage, and communication interfaces. The central processing module 5 is connected to the industrial camera 3 via a data cable to receive image data, and is connected to the line structured light projector 2 via a control line to control its switching. Accordingly, the central processing module 5 serves as the core of the system, on which software programs implementing the method of this application can run.

[0106] like Figure 1 The method shown can be executed periodically (e.g., once per minute) by a software program deployed on the central processing module 5 to achieve continuous monitoring of tunnel deformation.

[0107] Optionally, an inertial measurement unit (IMU) 3a can be added to the aforementioned embodiments. This IMU 3a is securely integrated or attached to the housing of the industrial camera 3. Its coordinate system is pre-calibrated with the camera's imaging coordinate system, enabling it to measure and output the attitude angles of the industrial camera 3 in three-dimensional space, particularly its pitch and yaw angles, in real time and with high precision. The IMU 3a is connected to the central processing module 5 via a serial interface or bus, allowing its attitude data to be read in real time. The remaining hardware configuration, including the line structured light projector 2, the industrial camera 3, the reference target 4 (standard gauge block), and the central processing module 5, is the same as in the aforementioned embodiments.

[0108] Regarding the methodology, this embodiment still follows the following... Figure 1 The method flow is shown, but a step to correct projection distortion is added to the deformation calculation step. In actual field installation, the optical axis of the industrial camera 3 is usually difficult to be completely perpendicular to the laser projection plane where the laser line is located; there is often a non-negligible angle between them, which is called the projection angle. θ This included angle will cause a projective compression effect: when the tunnel inner wall 1 undergoes a real physical displacement Δ XAt that time, its projected displacement Δ on the camera's imaging plane X proj It will be less than the actual displacement. The relationship between them approximately satisfies: Δ X proj If this effect is not corrected, the measurement results will be systematically underestimated, thus affecting the accuracy of monitoring.

[0109] In this embodiment, the system error can be corrected online in the following way:

[0110] During each deformation calculation, the central processing module 5 first reads real-time attitude angle data from the inertial measurement unit 3a associated with the industrial camera 3 via an interface. Using a pre-calibrated coordinate transformation relationship, the projection angle between the current camera imaging plane and the laser projection plane can be calculated. θ ;

[0111] Similar to the aforementioned method embodiments, the program identifies reference target 4 (standard block) in the image and measures its pixel length. n px Calculate the dynamic scaling factor. k dyn When, the projection angle is introduced. θ Perform correction. The corrected dynamic proportional coefficient. k dyn_corr The calculation formula is:

[0112]

[0113] The physical meaning of this formula is that, due to projection compression, the pixel length of the reference target itself in the image... n px This is also the result after compression, therefore it needs to be divided by cos( θ This is used to restore the correct pixel length under vertical projection, thus obtaining a more accurate scaling factor.

[0114] Calculate the average pixel displacement Δ between the current centerline and the reference centerline. u This step is the same as in the previous embodiments.

[0115] Finally, the dynamic scaling factor corrected for projection distortion is used. k dyn_corr To calculate the final, more accurate physical deformation. The result The systematic projection error caused by the camera installation angle has been compensated.

[0116] By introducing an inertial measurement unit 3a and performing real-time correction on the dynamic scaling factor, this embodiment effectively eliminates systematic errors introduced by the camera not being installed directly facing the tunnel wall, making the monitoring results closer to the physical truth. This has extremely high practical value in many practical engineering application scenarios where installation conditions are limited and it is impossible to guarantee that the camera is directly facing the tunnel wall, further improving the accuracy and reliability of the entire monitoring system.

[0117] In one embodiment, a computer-readable storage medium is also provided, which stores a computer program that, when executed by a processor, causes the processor to perform any of the steps in the above method embodiments.

[0118] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0119] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0120] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for monitoring tunnel deformation based on line structured light, characterized in that, The method includes: Acquire an image of a laser line projected onto the inner wall of a tunnel, and preprocess the laser line image, the preprocessing including: denoising the laser line image using Gaussian filtering; and repairing local breaks in the laser line using morphological closing operations; The preprocessed laser line image is input into a pre-trained deep learning semantic segmentation model, which identifies and outputs the pixel region of the laser line in the laser line image. The laser line is processed by centerline extraction in the pixel region of the laser line image to obtain a laser centerline with a single pixel width. Specifically, this includes: using a skeletonization algorithm to process the recognition results of deep learning and refining the laser line into a laser centerline with the single pixel width. A reference target with a known physical size is deployed within the monitoring scene, and a dynamic scaling factor between the pixel size and the physical size is determined by processing an image containing the reference target. Based on the laser centerline acquired at least two different times, the physical displacement between the laser centerlines is calculated using the dynamic scaling factor to determine the deformation of the tunnel, specifically including: By matching the laser centerline at the current moment with that at the previous moment point by point, the displacement of each point in the image coordinate system is obtained; based on the dynamic scaling factor, the displacement is converted into actual physical displacement to obtain the deformation of the tunnel.

2. The tunnel deformation monitoring method based on line structured light according to claim 1, characterized in that, The deep learning semantic segmentation model is a model based on the Swing Transformer architecture; When the deep learning semantic segmentation model is a model based on the Swing Transformer architecture, it uses at least one of the following for model optimization: cross-entropy loss function with class weights, online hard sample mining strategy, and DiceLoss.

3. The tunnel deformation monitoring method based on line structured light according to claim 2, characterized in that, The deep learning semantic segmentation model is trained through a training process that includes data augmentation. The data augmentation includes cropping, randomly rotating, and mirroring image blocks containing laser line targets.

4. The tunnel deformation monitoring method based on line structured light according to claim 1, characterized in that, The reference target is a standard gauge block.

5. A tunnel deformation monitoring system based on line structured light, characterized in that, The system is used to perform the steps of the method as described in any one of claims 1-4; the system includes a line structured light projection module, an image acquisition module, and a central processing module, wherein the central processing module is used for: The laser line image projected onto the tunnel wall, acquired by the image acquisition module, is preprocessed. The preprocessed laser line image is input into a pre-trained deep learning semantic segmentation model, which identifies and outputs the pixel region of the laser line in the laser line image. The centerline of the laser line in the pixel region of the laser line image is extracted to obtain a laser centerline with a single pixel width. By processing an image acquired by the image acquisition module that contains a reference target with a known physical size, a dynamic scaling factor between the pixel size and the physical size is determined; Based on the laser centerline acquired at at least two different times, and using the dynamic scaling factor, the physical displacement between the laser centerlines is calculated to determine the amount of tunnel deformation.

6. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1-4.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it causes the processor to perform the steps of the method as described in any one of claims 1-4.

Images