Construction tunnel deformation intelligent detection method and system based on visual perception

Abstract

The invention provides a construction tunnel deformation intelligent detection method and system based on visual perception, and relates to the technical field of tunnel construction detection. Collecting a visible light image and an infrared image of a tunnel monitoring section; based on a visual perception mechanism, after dust interference in the visible light image is guided and suppressed by using infrared image information, illumination adaptive correction is carried out, and a visible light enhanced image is obtained; introducing macroscopic deformation analysis, comparing three-dimensional reconstruction with a dynamically updated reference surface based on an originally acquired image, and calculating the overall deformation quantity of the section; microcosmic crack analysis is introduced, and a crack dynamic risk in the visible light enhanced image is obtained based on a deep learning model; and carrying out fusion decision on the overall deformation quantity and the crack dynamic risk, and generating a comprehensive risk index for triggering graded early warning. Through anti-interference visual perception, multi-scale collaborative analysis and fusion decision making, accurate and real-time monitoring and graded early warning of tunnel overall deformation and local cracks in the severe construction environment are achieved.

Description

Technical Field

[0001] This invention relates to the field of tunnel construction inspection technology, and in particular to an intelligent detection method and system for construction tunnel deformation based on visual perception. Background Technology

[0002] Deformation monitoring during tunnel construction is the core of engineering safety management. It is necessary to accurately capture the overall deformation and local cracks of the surrounding rock and support structure to provide a reliable basis for risk warning. Its monitoring accuracy and anti-interference ability are directly related to construction safety, and are especially critical under complex geological conditions.

[0003] In recent years, tunnel collapse accidents have mostly been caused by monitoring technologies failing to effectively adapt to harsh construction environments and failing to fully perceive safety hazards, resulting in serious losses and highlighting that traditional monitoring methods are no longer sufficient to meet actual needs. Current mainstream detection methods have obvious limitations: manual inspections and contact sensors cannot simultaneously achieve comprehensiveness and real-time performance, while existing visual inspection methods are a focus of the industry, but they are affected by environmental factors such as construction dust and lighting fluctuations, resulting in poor feature extraction stability and an inability to effectively recover clear image data. At the same time, these visual inspection methods mostly focus on single-scale monitoring, or only pay attention to overall deformation or only focus on local defects, failing to achieve collaborative perception of multi-scale information and making it difficult to comprehensively reflect the health status of the tunnel.

[0004] More importantly, existing visual inspection lacks targeted anti-interference visual perception mechanisms and has not built a holistic-local collaborative analysis system, resulting in insufficient data reliability, single monitoring dimensions, and an inability to form a comprehensive security assessment basis, ultimately making it difficult to support accurate and efficient risk warning. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes an intelligent detection method and system for tunnel deformation based on visual perception. Through anti-disturbance visual perception, multi-scale collaborative analysis, and fusion decision-making, it achieves accurate, real-time monitoring and graded early warning of overall tunnel deformation and local cracks under harsh construction environments, significantly improving the level of construction safety management.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for intelligent detection of deformation in construction tunnels based on visual perception, comprising: Collect visible light and infrared images of the tunnel monitoring section; Based on the visual perception mechanism, infrared image information is used to guide and suppress dust interference in visible light images. Then, the interference-free visible light image is subjected to illumination adaptive correction to protect texture details, resulting in a visible light enhanced image. Macroscopic deformation analysis is introduced, which calculates the overall deformation of the cross section by comparing the original acquired images with the dynamically updated reference surface through 3D reconstruction; microscopic crack analysis is introduced, which identifies and quantifies the dynamic risk of cracks in the visible light enhanced images through a deep learning model. The overall deformation and crack dynamic risk are integrated to generate a comprehensive risk index, which is then used to trigger tiered early warnings.

[0007] Secondly, the present invention provides a visual perception-based intelligent detection system for deformation of construction tunnels, comprising: The data acquisition module is used to acquire visible light and infrared images of the tunnel monitoring section; The visual perception module is used to guide and suppress dust interference in visible light images based on the visual perception mechanism and infrared image information. Then, the visible light image after interference removal is subjected to illumination adaptive correction to protect texture details, resulting in a visible light enhanced image. The image processing module is used to introduce macroscopic deformation analysis, which calculates the overall deformation of the cross section by comparing the original acquired image with the dynamically updated reference surface through 3D reconstruction; and to introduce microscopic crack analysis, which identifies and quantifies the dynamic risk of cracks in the visible light enhanced image through a deep learning model. The risk decision module is used to integrate the overall deformation and crack dynamic risk to generate a comprehensive risk index and trigger graded early warning accordingly.

[0008] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the intelligent detection method for deformation of construction tunnels based on visual perception as described in the first aspect.

[0009] Fourthly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the intelligent detection method for deformation of construction tunnels based on visual perception as described in the first aspect.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention overcomes the limitations of visual inspection in harsh construction environments by using infrared-guided dust suppression and adaptive illumination correction, ensuring the reliability of image data. Simultaneously, the dual-channel collaborative analysis of macroscopic deformation and microscopic cracks accurately captures overall structural displacement and quantifies the dynamic risks of local cracks, achieving comprehensive hazard perception. Furthermore, a dynamic decision-making model integrating construction phase and environmental data generates a comprehensive risk index and triggers tiered early warnings, forming a complete closed loop. This significantly improves the accuracy and timeliness of early warnings, providing strong technical support for tunnel construction safety and reducing the risk of collapse accidents.

[0011] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0012] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0013] Figure 1 The main flowchart of a visual perception-based intelligent detection method for construction tunnel deformation is provided in an embodiment of the present invention. Detailed Implementation

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] Example 1 like Figure 1 As shown in the figure, this embodiment discloses an intelligent detection method for deformation of construction tunnels based on visual perception, including the following steps: S1: Acquire visible light and infrared images of the tunnel monitoring section; S2: Based on the visual perception mechanism, infrared image information is used to guide and suppress dust interference in the visible light image, and then the interference-free visible light image is subjected to illumination adaptive correction to protect texture details, resulting in a visible light enhanced image. S3: Introducing macroscopic deformation analysis, based on the original acquired images, the overall deformation of the cross section is calculated by comparing it with the dynamically updated reference surface through 3D reconstruction; introducing microscopic crack analysis, the dynamic risk of cracks in the visible light enhanced images is identified and quantified through a deep learning model. S4: The overall deformation and crack dynamic risk are integrated to generate a comprehensive risk index, and a graded early warning is triggered accordingly.

[0016] Next, combined Figure 1 This embodiment provides a detailed description of a visual perception-based intelligent detection method for construction tunnel deformation.

[0017] S1 collects environmental data related to tunnel construction.

[0018] Visible light data and infrared data were collected separately at typical tunnel sections through hardware collaborative settings.

[0019] Specifically, an array of multiple high-resolution industrial visible light cameras is deployed to cover the cross-section from multiple angles, including the vault and the waist; simultaneously, a long-wave infrared thermal imager is deployed to capture thermal radiation information by utilizing its ability to penetrate dust. In addition, a structured light projector is used to project coded patterns to enhance texture, and a laser contour scanner is used to provide high-precision spatial reference coordinates.

[0020] As one implementation method, a multispectral collaborative imaging architecture is adopted. The main equipment consists of four Sony IMX585 industrial cameras (3840×2160@60fps) arranged in a ring array, deployed at key locations on the arch (15° depression), arch waist (horizontal installation), and rail surface (10° elevation) every 50 meters. An 850nm infrared camera (1920×1080@30fps) is used to penetrate dust interference, and a 635nm structured light projector actively projects coded patterns. A ring-shaped LED supplementary lighting system dynamically adjusts the illuminance (50-500 lux) and stabilizes the color temperature at 5000K±200. A laser profilometer scans at a frequency of 2kHz to generate the cross-sectional baseline, achieving sub-millimeter-level spatial positioning (±0.1mm). All optical equipment is explosion-proof certified (Ex ib IIC T4 Gb), adaptable to the high humidity and high dust environment of the tunnel.

[0021] S2, preprocesses dual-channel images based on visual perception mechanisms.

[0022] The tunnel construction environment is extremely harsh, presenting two major challenges to visual perception. First, the large amounts of suspended dust and smoke generated during drilling and blasting severely attenuate visible light, resulting in blurred images, reduced contrast, and obscuring key features such as cracks. Second, internal lighting relies on temporary fixtures, leading to uneven illumination, localized overexposure, or insufficient illuminance, causing image shadows, blown-out highlights, and loss of detail. These factors collectively cause conventional visual inspection methods to fail, resulting in unstable feature extraction.

[0023] To address the aforementioned issues, this embodiment proposes a visual perception mechanism that integrates dual-channel information with illumination reconstruction, aiming to recover clear and reliable image data from adverse imaging conditions.

[0024] To recover reliable visual data from harsh imaging environments, this embodiment proposes a tandem visual perception mechanism: first, infrared channel information is used to suppress dust interference and restore the image structure; then, global illumination correction is performed on the interference-free image to optimize the visual effect. The specific steps are as follows: First, a collaborative processing flow for visible light and infrared data is established to suppress dust interference.

[0025] Specifically, the infrared image is preprocessed by performing contrast-limited adaptive histogram equalization (CLAHE) on it, taking advantage of its low susceptibility to dust scattering, to enhance its thermal texture features and obtain enhanced infrared image features.

[0026] To dynamically assess the impact of dust, the enhanced infrared image features are used as a guide to assist the deblurring process of the visible light image. Specifically, by analyzing the difference in texture sharpness between the infrared and visible light images in the same region, the dust concentration distribution map of each region is adaptively estimated. This is represented by calculating the ratio of the average gradient magnitude of the visible light image and the enhanced infrared image within the same local window (e.g., 16x16 pixels). ; Where is the average gradient magnitude of the visible light image, and is the average gradient magnitude of the enhanced infrared image.

[0027] ratio Mapped to a local dust influence factor , ,in This is the attenuation coefficient. The lower the value, the more severe the attenuation of visible light information in that area.

[0028] Subsequently, in this way Factors are used as weighting criteria to apply spatially adaptive guided filtering to visible light images. During the filtering process, in low... In high-value regions (where dust has a significant impact), the filter's guide input and regularization parameters tend to utilize edge information provided by the infrared image to reconstruct the structure and restore obscured details; in high-value regions... In the value region, edge preservation and noise reduction mainly rely on the information of the visible light image itself. This method realizes the conceptual improvement from "fixed weight fusion of the whole image" to "cooperative enhancement based on local features", which can more precisely eliminate the blur caused by dust. It realizes the improvement of dust suppression from qualitative to quantitative and from global to local adaptive, and can more precisely eliminate the blur caused by dust.

[0029] After dust suppression, illumination unevenness correction is performed on the processed visible light image. This is achieved by combining the Retinex theoretical framework with the aforementioned processing results.

[0030] First, the image after dust removal is decomposed into illumination and reflection components.

[0031] Subsequently, a joint optimization model is constructed. When estimating the illumination components, not only pixel brightness is considered, but also enhanced infrared image feature information is introduced as a constraint to avoid misjudging real dark cracks (such as shadow areas) as insufficient illumination and over-brightening them.

[0032] Subsequently, the optimized illumination components are equalized, and an adaptive gamma correction coefficient for each pixel is calculated accordingly, ultimately reconstructing a uniformly illuminated reflection image. This ensures that while improving overall brightness and contrast, the key geological defect textures recovered in the previous step are not distorted or lost.

[0033] Through the aforementioned sequential processing, infrared information is first used as a "guide" to remove dust blur, and then illumination normalization is performed while preserving effective texture, resulting in a high-quality visible light enhanced image after dust suppression and illumination correction. The visual perception mechanism in this embodiment achieves a reliable conversion from dual-source raw data to a high-quality analytical image, laying a solid foundation for subsequent deformation and crack analysis.

[0034] It should be understood that the contrast-limited adaptive histogram equalization (CLAHE) and the calculation of adaptive gamma correction coefficients for each pixel are both achievable by those skilled in the art and will not be elaborated upon here.

[0035] S3 sends the original acquired images and the enhanced visible light images into two analysis channels: the macroscopic deformation channel and the microscopic crack channel. The two channels quantify and perceive the "overall geometric deformation" and "local surface damage" at two scales, respectively. Finally, they are fused through the decision layer to form a complete tunnel health diagnosis report.

[0036] (I) Macroscopic deformation analysis (centimeter-millimeter displacement): This channel is designed to accurately measure large-scale geometric changes such as the overall convergence and settlement of the tunnel cross-section.

[0037] First, a three-dimensional reconstruction is performed and a dynamic reference surface mechanism is introduced.

[0038] Specifically, by utilizing the multi-angle visible light images and structured light coding information synchronously acquired in S1, the two-dimensional image of the current frame is converted into a high-density three-dimensional point cloud model through stereo vision and structured light three-dimensional reconstruction algorithms.

[0039] To eliminate the cumulative errors caused by construction progress and equipment micro-movements, a "sliding reference surface" mechanism is introduced. This mechanism establishes a reference model library that is periodically updated as tunnel excavation progresses. For example, every 10 meters of excavation, the current high-quality point cloud model is saved as a new "reference surface." In subsequent analyses, the most recently acquired point cloud is always compared with the nearest historical "reference surface," thereby controlling long-term drift errors within a single excavation segment.

[0040] Next, high-precision point cloud registration and deformation calculation are performed. Specifically, an Iterative Closest Point (ICP) algorithm accelerated by voxel hash maps is used to quickly and accurately align the current scanned point cloud (denoted as P) with the selected sliding reference plane point cloud (denoted as Q). The registration process is achieved by minimizing the distance between corresponding points in the two point clouds. After registration, the maximum deformation is quantified by calculating the Hausdorff Distance, or the overall convergence trend is quantified by calculating the average distance between corresponding points. The deformation obtained from this calculation is defined as the overall deformation (D) of the macroscopic channel. Specifically:

[0041] in, The Hausdorff distance between the current point cloud and the baseline model is used to characterize the deformation difference between the two. The transformation matrix represents the rigid body transformation during point cloud registration, including operations such as translation and rotation. This represents taking all possible transformations. The minimum value below; This represents the point set of the current point cloud. It is a point in the current point cloud; The point set representing the point cloud of the baseline model. It is a point in the baseline model; This indicates that points a and b have undergone transformation. The Euclidean distance between the points after; The supremum is the maximum value in the set. The infimum represents the minimum value in the set.

[0042] It should be understood that the 3D reconstruction algorithm and the Iterative Closest Point (ICP) algorithm accelerated by voxel hash graphs are both implementable by those skilled in the art.

[0043] Furthermore, to distinguish between actual rock mass displacement and environmental effects such as thermal expansion and contraction of concrete, a temperature-deformation compensation model was also established. This model, based on real-time data from temperature sensors deployed on the cross-section and the material's thermal expansion coefficient, eliminates spurious deformation components caused by temperature from the calculated original deformation, significantly reducing the false alarm rate.

[0044] (II) Microscopic Crack Analysis (Millimeter-to-submillimeter Defects): This channel focuses on automatically identifying and quantifying minute cracks and other apparent damage on the lining surface.

[0045] First, inference is performed using a lightweight two-branch deep learning model. Specifically, the visible light-enhanced image obtained from S2 is input into a dedicated convolutional neural network deployed on an edge computing device. This network employs a two-branch design: 1. Global Structure Branch: Employing lightweight structures such as the Inception module, this branch downsamples and extracts features from the image, focusing on understanding the overall composition of the tunnel cross-section and signs of large-scale instability (such as large-scale peeling and bulging). This branch outputs a feature map representing the overall structural information. .

[0046] 2. Local Detail Branch: Employing a dilated convolution layer, this branch expands the receptive field without excessively reducing resolution, specifically designed to capture pixel-level fine details such as cracks, water stains, and other high-frequency details. This branch outputs a feature map rich in high-frequency detail information. .

[0047] Subsequently, by channel splicing or weighted summation, and The features are fused to obtain composite features that contain both global semantics and local details, which are then used for subsequent precise crack segmentation and quantization.

[0048] Furthermore, the model not only outputs the pixel-level location of the crack, but also performs precise quantization.

[0049] First, the composite features are input into a segmentation head. Through convolution and upsampling operations, a pixel-level segmentation map of the cracks with the same resolution as the input image is generated, where each pixel is classified as either a crack or background. Based on this segmentation map, the centerline of the crack is extracted using a morphological skeletonization algorithm. The cumulative length of the centerline pixels is converted to the crack length (L) using a scale bar. Simultaneously, the average width (W) can be approximated by dividing the area of ​​the segmented crack region by its skeleton length. The morphological skeletonization algorithm is implementable by those skilled in the art.

[0050] Next, for each identified crack, calculate the risk factor C for each crack (considering crack width, length, and direction): ; in, and These represent the length and average width of the crack, respectively. and This is a preset normalized reference value; The rate at which the crack length expands per unit time can be obtained using an inter-frame tracking algorithm; , , These are the weighting coefficients for each indicator, which can be adjusted according to engineering priorities. The crack inter-frame tracking algorithm establishes cross-frame correlations between cracks by comparing the shape, location, and feature descriptions of cracks in multiple consecutive frames, thereby calculating their propagation rate. Duration A higher risk weight is assigned to newly formed cracks that are continuously and rapidly developing.

[0051] The macroscopic deformation channel and the microscopic crack channel process image data from the same time and cross-section in parallel, complementing each other in terms of physical scale and target of interest. The system's backend decision module synchronously receives the outputs from both channels: the overall deformation (D) from the macroscopic channel and the crack risk coefficient set (C) and its distribution from the microscopic channel. These data will be fused to generate a comprehensive report containing an overall stability assessment and details of local damage, providing a quantitative basis for subsequent risk warning and decision-making.

[0052] S4 integrates macroscopic deformation variables (D) with microscopic crack risk coefficients (C) at the decision-making level and triggers tiered early warnings.

[0053] First, a dynamic weighted risk index model is constructed. This model integrates macro and micro indicators, environmental data, and construction phase information for unified calculation, generating a comprehensive risk index R. Its calculation formula is as follows: ; in, , , is a dynamic weighting coefficient that can be automatically adjusted according to different construction stages of the cross-section (such as after excavation or after support); D is the macroscopic deformation. The crack risk is a function of its coefficient C and duration t; It is a correction factor that takes into account environmental factors such as temperature and humidity.

[0054] Subsequently, the system automatically triggers four levels of warnings—blue, yellow, orange, and red—based on the risk index R. When a higher-level warning (such as orange or red) is triggered, the system will simultaneously push alarm information to the mobile terminal of the person in charge at the site and activate the on-site audible and visual alarm devices.

[0055] All data, including raw data collection, processing results at all levels, risk index calculation process and early warning records, are stored in real time in a security audit chain based on blockchain technology to ensure that the data throughout the process is tamper-proof and traceable, meeting the requirements for security audit and liability determination.

[0056] This specific embodiment constructs an infrared-visible light dual-channel serial visual perception mechanism. First, infrared images guide adaptive suppression of dust interference, then Retinex theory is used for illumination correction, efficiently restoring clear and usable enhanced visible light images and solving the problem of unstable feature extraction under harsh environments. Simultaneously, a dual-channel collaborative analysis system for macroscopic deformation and microscopic cracks is designed. The macroscopic channel relies on dynamic reference surfaces and point cloud registration to accurately quantify overall deformation, while the microscopic channel uses a dual-branch deep learning model to accurately identify cracks and dynamically quantify risks, overcoming the limitations of single-scale monitoring. Finally, a dynamic weighted decision-making model integrating construction stage and environmental data is constructed to generate a comprehensive risk index and trigger tiered early warnings, forming a closed loop of data acquisition, intelligent analysis, and decision response. This significantly improves the reliability of monitoring data and the accuracy of early warnings, providing comprehensive technical support for tunnel construction safety.

[0057] Example 2 This embodiment provides a vision-based intelligent detection system for tunnel deformation during construction, comprising: The data acquisition module is used to acquire visible light and infrared images of the tunnel monitoring section; The visual perception module is used to guide and suppress dust interference in visible light images based on the visual perception mechanism and infrared image information. Then, the visible light image after interference removal is subjected to illumination adaptive correction to protect texture details, resulting in a visible light enhanced image. The image processing module is used to introduce macroscopic deformation analysis, which calculates the overall deformation of the cross section by comparing the original acquired image with the dynamically updated reference surface through 3D reconstruction; and to introduce microscopic crack analysis, which identifies and quantifies the dynamic risk of cracks in the visible light enhanced image through a deep learning model. The risk decision module is used to integrate the overall deformation and crack dynamic risk to generate a comprehensive risk index and trigger graded early warning accordingly.

[0058] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the intelligent detection method for deformation of construction tunnels based on visual perception as described in Embodiment 1 above.

[0059] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the intelligent detection method for deformation of construction tunnels based on visual perception as described in Embodiment 1 above.

[0060] The steps or modules involved in Embodiments 2 to 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent detection of deformation in construction tunnels based on visual perception, characterized in that, include: Collect visible light and infrared images of the tunnel monitoring section; Based on the visual perception mechanism, infrared image information is used to guide and suppress dust interference in visible light images. Then, the interference-free visible light image is subjected to illumination adaptive correction to protect texture details, resulting in a visible light enhanced image. Macroscopic deformation analysis is introduced, which calculates the overall deformation of the cross section by comparing the original acquired images with the dynamically updated reference surface through 3D reconstruction; microscopic crack analysis is introduced, which identifies and quantifies the dynamic risk of cracks in the visible light enhanced images through a deep learning model. The overall deformation and crack dynamic risk are integrated to generate a comprehensive risk index, which is then used to trigger tiered early warnings.

2. The intelligent detection method for construction tunnel deformation based on visual perception as described in claim 1, characterized in that, The method of using infrared image information to guide and suppress dust interference in visible light images specifically includes: The acquired infrared images are enhanced, and reliable texture regions are extracted. Calculate the difference in feature sharpness between the infrared image and the visible light image in the same local area, and map an influence factor map reflecting the spatial distribution of dust based on the difference; Using the aforementioned influence factor map as a weighting basis, spatially adaptive guided filtering is performed on the visible light image. In areas with significant dust impact, infrared image features are preferentially used to reconstruct edges and details, suppressing blurring caused by dust.

3. The intelligent detection method for construction tunnel deformation based on visual perception as described in claim 2, characterized in that, The step of performing adaptive illumination correction on the distorted visible light image to preserve texture details, resulting in an enhanced visible light image, specifically includes: The visible light image after dust removal is decomposed into illumination and reflection components; When estimating the illumination components, constraints are constructed by combining the reliable texture region information to prevent over-brightening of real dark cracks; The optimized illumination components are equalized, and an adaptive correction coefficient is calculated for each pixel to finally synthesize a visible light enhanced image.

4. The intelligent detection method for deformation of construction tunnels based on visual perception as described in claim 1, characterized in that, The introduction of macroscopic deformation analysis involves comparing the original acquired images with a dynamically updated reference surface through 3D reconstruction to calculate the overall deformation of the cross-section. Specifically, this includes: Using the acquired raw visible light images and structured light information of the tunnel monitoring section from multiple angles, a high-density three-dimensional point cloud of the current section is generated through stereo vision and three-dimensional reconstruction algorithms. A sliding reference surface mechanism is adopted, using a historical point cloud model that is periodically updated with tunnel excavation as a comparison benchmark; The current point cloud is registered with the selected reference model using the iterative nearest point algorithm; The spatial deviation between the two cloud points after registration is calculated, and thermal expansion compensation is performed in combination with real-time temperature data. Finally, the overall macroscopic deformation of the current cross section is output.

5. The intelligent detection method for construction tunnel deformation based on visual perception as described in claim 1, characterized in that, The introduction of micro-crack analysis, which uses a deep learning model to identify and quantify the dynamic risk of cracks in visible light enhanced images, specifically includes: The enhanced visible light image is input into a lightweight convolutional neural network; The network encodes and extracts features from the image, and outputs pixel-level crack segmentation maps through a decoding segmentation head; Based on the segmentation map, the geometric parameters of each crack, including length and average width, are extracted through morphological processing. The same crack in consecutive images is matched using an inter-frame tracking algorithm, and its expansion rate and duration are calculated. Based on the combined geometric parameters and dynamic expansion information, the dynamic risk coefficient of each crack is calculated.

6. The intelligent detection method for deformation of construction tunnels based on visual perception as described in claim 5, characterized in that, The lightweight convolutional neural network employs a parallel dual-branch architecture for feature extraction. The global branch extracts the overall structural semantic features of the image, while the local detail branch uses dilated convolution to expand the receptive field and capture subtle crack texture features. The feature maps output from the two branches are fused, and the fused features are then input into the segmentation head to complete crack segmentation and quantization.

7. The intelligent detection method for deformation of construction tunnels based on visual perception as described in claim 1, characterized in that, The process of fusing the overall deformation with the dynamic risk of cracks to generate a comprehensive risk index, and triggering tiered early warnings accordingly, specifically includes: The overall deformation obtained from macroscopic analysis, the crack risk coefficient and its duration obtained from microscopic analysis, the current construction stage information, and environmental sensor data are all input into a dynamic weighted model. The model adaptively adjusts the weights of each input item according to the construction stage to calculate the comprehensive risk index; The system presets four warning thresholds: blue, yellow, orange, and red, corresponding to different risk index ranges. When the risk index falls into a certain range, the system automatically triggers the corresponding warning level and releases the warning information through preset communication and alarm channels.

8. A visual perception-based intelligent detection system for deformation of construction tunnels, characterized in that, include: The data acquisition module is used to acquire visible light and infrared images of the tunnel monitoring section; The visual perception module is used to guide and suppress dust interference in visible light images based on the visual perception mechanism and infrared image information. Then, the visible light image after interference removal is subjected to illumination adaptive correction to protect texture details, resulting in a visible light enhanced image. The image processing module is used to introduce macroscopic deformation analysis, which calculates the overall deformation of the cross section by comparing the original acquired image with the dynamically updated reference surface through 3D reconstruction; and to introduce microscopic crack analysis, which identifies and quantifies the dynamic risk of cracks in the visible light enhanced image through a deep learning model. The risk decision module is used to integrate the overall deformation and crack dynamic risk to generate a comprehensive risk index and trigger graded early warning accordingly.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the intelligent detection method for deformation of construction tunnels based on visual perception as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the intelligent detection method for deformation of construction tunnels based on visual perception as described in any one of claims 1-7.

Images