Tunnel lining and pavement disease detection method

Abstract

The invention provides a tunnel lining and pavement disease detection method, and belongs to the technical field of tunnel engineering detection. The method comprises the steps of tunnel detection area differentiation, data acquisition, tunnel three-dimensional model construction based on Gaussian sputtering, and lining and pavement disease intelligent identification. A flight route is planned in a layering-segmenting mode, and images are collected; preprocessing the image, stripping physical information such as position and color of a two-dimensional image to generate a Gaussian kernel, and establishing a three-dimensional model containing information of a whole section of tunnel; based on the constructed three-dimensional model, a YOLOv8 model is adopted for deep learning training, and the purpose of intelligently recognizing diseases is achieved. According to the method, the detection efficiency can be remarkably improved, the detection cost is remarkably reduced, and detection is carried out on the premise that tunnel passing is not affected or safety risk behaviors are not caused.

Description

Technical Field

[0001] This invention belongs to the field of tunnel engineering inspection technology, and relates to a method for detecting tunnel lining and pavement defects. Background Technology

[0002] As a critical component of transportation infrastructure, the long-term health of tunnel lining structures and road surfaces directly impacts operational safety. However, current methods for detecting tunnel lining and road surface defects still face numerous bottlenecks, failing to meet the high standards of efficiency, accuracy, and safety required by modern maintenance management. Existing technologies primarily suffer from the following prominent problems:

[0003] Manual inspection methods are inefficient and highly subjective: Traditional inspection methods rely heavily on technicians using simple tools for close-range visual inspection or manual measurement. This results in extremely low efficiency; conducting a full-coverage inspection of tunnels several kilometers long is time-consuming and cannot meet the frequency requirements of regular inspections. Furthermore, the inspection results are highly dependent on the personal experience and subjective judgment of technicians, lacking unified objective standards, leading to poor data comparability and making it difficult to accurately track and quantitatively analyze the evolution of defects.

[0004] Limited inspection coverage creates safety hazards and blind spots: Manual inspections are limited by visibility and accessibility, often resulting in insufficient or even complete neglect of areas such as tunnel arches and high-altitude sidewalls, creating blind spots and potential safety risks. Furthermore, inspection work typically requires erecting scaffolding inside the tunnel or occupying lanes, which is not only cumbersome but also disrupts traffic, posing both high-altitude work and traffic safety risks to inspection personnel, and reducing tunnel traffic efficiency.

[0005] The limited data dimensions hinder accurate quantification of disease damage: Whether through manual photography or some vehicle-mounted inspection equipment, the acquired data is mostly two-dimensional images or localized information from a single angle, lacking precise three-dimensional spatial coordinates. This makes the measurement of key geometric parameters such as crack width, length, spalling area, pit depth, and volume extremely inaccurate, allowing only qualitative or rough estimations. It fails to provide reliable data support for disease severity classification and repair plan development.

[0006] Existing technologies are either too costly or lack adaptability: While some high-precision vehicle-mounted or mobile measurement systems can improve efficiency, they are expensive, complex to operate, and typically require closed lanes, making them difficult to adopt in small to medium-sized projects or scenarios requiring frequent inspections. Furthermore, these devices are often optimized for a single target such as the lining or pavement, making it difficult to simultaneously and efficiently perform collaborative inspections of the entire tunnel cross-section (including the arch, sidewalls, and pavement), resulting in significant data integration challenges.

[0007] In summary, existing detection methods have significant shortcomings in terms of efficiency, safety, coverage, data accuracy, and economic benefits. There is an urgent need for an intelligent detection technology that can achieve full automation, full coverage, high accuracy, and does not affect the normal operation of tunnels to solve these problems. Summary of the Invention

[0008] The purpose of this invention is to address the aforementioned problems in existing technologies by proposing a method for detecting tunnel lining and pavement defects.

[0009] The objective of this invention can be achieved through the following technical solution: a method for detecting tunnel lining and pavement defects, comprising the following steps:

[0010] S1 Tunnel Inspection Area Division: Based on the geometric characteristics of the tunnel, the inspection area is divided using a combination of segmentation and layering. Segmentation involves dividing the tunnel into multiple inspection segments with a preset length as the unit, and setting positioning marks at the start and end positions of each inspection segment. Layering involves dividing the cross-section of each inspection segment along the height direction into the arch top layer, arch waist layer, sidewall layer, and road surface layer.

[0011] S2 Data Acquisition: Using a drone equipped with a lidar and a multispectral camera, image data and point cloud data are acquired based on flight path rules. During the acquisition process, the drone's position is calibrated by both positioning markers and real-time SLAM positioning by lidar, and missing data areas are automatically captured. Flight path rules: A serpentine flight pattern parallel to the tunnel axis is used in straight sections, and a circular flight pattern around the tunnel cross-section is used in curved sections.

[0012] S3 Gaussian sputtering 3D modeling: The collected data is preprocessed sequentially, followed by tunnel global coordinate transformation, Gaussian sputtering technology processing, iterative algorithm optimization of Gaussian kernel parameters, and fusion of image texture information to obtain a tunnel 3D model;

[0013] In the Gaussian sputtering process, a Gaussian kernel is generated based on the Python framework. The Gaussian kernel is obtained through a joint image feature-laser method. An isotropic Gaussian kernel is initialized at each point.

[0014] S4 Defect Identification: The defect detection model is used to identify defects in the 3D tunnel model. Then, the quantitative parameters of the defects are calculated based on the spatial coordinates of the 3D tunnel model, and a detection report is generated. The types of defects include cracks, spalling, and water leakage in the lining, and potholes, settlement, and cracks in the road surface. The defect detection model is the YOLOv8 model, which incorporates a 3D coordinate attention mechanism and a multispectral feature fusion module.

[0015] Preferably, in step S1, the geometric features include length, cross-sectional dimensions, and structural features; the positioning identifier is an invisible fluorescent QR code; the arch top layer is within 1.5 meters below the top of the tunnel, and the road surface layer is within 0.5 meters above the ground.

[0016] Preferably, the serpentine flight is a three-dimensional serpentine flight that reciprocates within a layer and ascends and descends between layers: within a single layer, it maintains a sinusoidal waveform parallel to the tunnel axis; after completing the current layer, the UAV moves to the next layer along a Z-shaped turning track and immediately reverses the serpentine flight.

[0017] Preferably, for tunnels longer than 5000 meters, multiple drones are used to collect data collaboratively, and the trajectory coordination and data synchronization between drones are achieved through ground station software; for underwater or high-humidity tunnels, drones are equipped with IP67 waterproof protective shells, and anti-fog lenses are added to the lidar and multispectral cameras. When performing Gaussian sputtering 3D modeling, a humidity correction module based on multispectral data is added; during the acquisition process, if the image overlap rate is insufficient or the point cloud is missing due to occlusion, the drone is set to automatically start a local detour reshoot mode.

[0018] Preferably, in the serpentine flight mode of step S2, the distance between the flight trajectory and the tunnel wall is maintained at 1.5-2 meters; in the circular flight mode, the image overlap rate of each layer area is ≥75%.

[0019] Preferably, the multispectral camera is set to RGB mode to simultaneously record ambient light intensity data; the lidar is set to high-density scanning mode with a point cloud density of ≥800 points / square centimeter to simultaneously collect spatial coordinate information.

[0020] Preferably, in the preprocessing operation of step S3, radial-tangential distortion correction and brightness equalization processing are performed on the image data. Among them, the ambient light-compensated gamma correction method is used for brightness equalization processing; and a filtering algorithm based on point cloud intensity and spatial distribution statistics is used to remove noise points and isolated points in the point cloud data.

[0021] Preferably, in the iterative algorithm optimization of Gaussian kernel parameters, the rendering loss L = Σ‖I_k-C_k‖ is used. 2 +λΣα_i 2 To achieve the objective, Adam is used to simultaneously differentiate with respect to μ, Σ, c, and α. The rendering loss function comprises two parts: the first part L = Σ‖I_k-C_k‖ 2 The photometric consistency loss is used to constrain the tunnel surface texture generated by the 3D Gaussian model to maintain consistency with the real acquired images, ensuring accurate restoration of defect features (such as cracks and seepage). Specifically: I_k is the observed value, which is the original image taken by the tunnel inspection robot, and C_k is the predicted value, which is the virtual tunnel surface reconstructed by the rendering algorithm; the second part λΣα_i 2The opacity regularization term is used to constrain the sparsity of the Gaussian sphere distribution and remove suspension noise caused by the dim tunnel environment. Specifically: λ is a weighting coefficient and a balancing factor that adjusts the denoising intensity; the opacity parameter α distinguishes between the diseased entity and the background space. The optimizer uses gradient descent to update the following Gaussian properties: position parameter μ: accurately locates the 3D coordinates of the disease in the tunnel space; covariance parameter Σ: adaptively fits the geometric shape of strip-shaped or point-shaped diseases; color parameter c: restores the visual characteristics of the disease; and the opacity parameter α is the same as above.

[0022] For the disease detection model, a 3D-coordinate attention module is introduced at the end of the backbone network to concatenate the Gaussian-rendered depth map Z with RGB as a 4-channel input to generate a weight map M to strengthen the crack edges; the Neck layer adds a 1×1 convolution to unify the RGB and near-infrared channels and then average the elements to achieve multispectral feature fusion.

[0023] Preferably, in step S4, regarding the definition of the recognition result, the crack constraint calculation is as follows: length L = ||P_start - P_end||, width W = max(projection frame width × GSD), depth D = Z_max - Z_min; the pit constraint calculation is as follows: area A = pixel count × GSD 2 The volume V = A × average depth; Peeling limitations: area and maximum thickness are calculated using the same method, but the location differs from pits and craters. Specifically: for cracks, the length L is calculated using the starting point coordinates P_start and the ending point coordinates P_end; the width W is calculated by combining the projection frame width and the ground sampling distance GSD; the depth D is determined by the difference between the maximum and minimum Z-axis values ​​Z_max and Z_min. For volumetric defects such as pits and craters, the area A is calculated based on pixel count and GSD, and then the volume V is calculated by combining the average depth.

[0024] Compared with existing technologies, the intelligent detection method for tunnel lining and pavement defects based on UAV and Gaussian sputtering technology provided by this invention achieves multi-dimensional and significant beneficial effects through systematic innovation and integration of technical processes, as detailed below:

[0025] 1. Achieving full automation of the detection process, significantly improving efficiency and objectivity: This invention constructs a fully automated technology system encompassing "region division → image acquisition → 3D modeling → disease identification." The drone autonomously flies and collects data based on a pre-defined "segmentation-layering" scheme. The backend algorithm automatically completes data processing, model building, and intelligent identification, completely replacing the traditional, highly manual detection mode. This technology improves detection efficiency by 15-25 times compared to manual inspection and completely avoids subjective judgment errors caused by factors such as human experience and fatigue, ensuring the consistency and objectivity of the detection results.

[0026] 2. Achieving full-section, blind-spot-free tunnel coverage, reaching a new level of detection completeness: Through an innovative "segmented-layered" area division strategy, this invention ensures comprehensive coverage of the detection range. As shown in the figure, the tunnel cross-section is precisely divided into four layers: the arch, the arch waist, the sidewalls, and the road surface. Optimal flight paths are designed for straight and curved sections respectively, thereby achieving comprehensive coverage of all tunnel surfaces (including the arch area, which is difficult for humans to reach), fundamentally eliminating detection blind spots and providing an unprecedentedly complete data foundation for tunnel structural health assessment.

[0027] 3. Generate high-fidelity 3D models, achieving breakthroughs in disease identification and quantification: Utilizing Gaussian sputtering technology, 2D images are fused with laser point clouds to generate a tunnel 3D model that combines "high-resolution texture" and "precise spatial coordinates." Subsequently, disease detection is performed using an improved detection model. After identifying the disease, the quantitative parameters of the disease (such as crack length, depth, and pit volume) are calculated based on the spatial coordinates of the 3D model, and a detection report is generated. This report can be directly used for subsequent maintenance plan formulation, providing precise data support for disease severity assessment (such as classifying minor, moderate, and severe according to the JTGH12-2015 standard) that is unattainable by traditional methods.

[0028] In summary, through technological integration and innovation, this invention has produced significant beneficial effects in multiple key dimensions such as detection efficiency, coverage integrity, and quantification accuracy, providing a comprehensive and reliable technical solution for achieving long-term, intelligent, and refined management and maintenance of tunnel infrastructure. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the technical route of the present invention.

[0030] Figure 2 This is a schematic diagram of the segmentation and layering of the tunnel detection area and the serpentine flight pattern of the present invention.

[0031] Figure 3 This is a schematic diagram of the layered annular flight cross-section of the UAV tunnel curve according to the present invention.

[0032] Figure 4 This is a diagram showing the results of intelligent disease identification according to the present invention. Detailed Implementation

[0033] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0034] like Figure 1 As shown, taking a highway tunnel with a length of 2000 meters and a cross-sectional dimension of "10 meters wide × 7 meters high" as an example, the method of this invention is used for lining and pavement defect detection. The specific steps are as follows:

[0035] S1 Tunnel Inspection Area Division: The 2000-meter tunnel is divided into 20 inspection sections, each 100 meters long. At the start and end points of each section, invisible fluorescent QR codes measuring 30cm x 30cm are sprayed onto the walls. Each QR code contains the section number and coordinate information. The global coordinates of the QR codes are calibrated using a total station.

[0036] The tunnel cross-section will be divided into four layers along its height: the arch top layer (7m-5.5m), the arch waist layer (5.5m-3m), the sidewall layer (3m-0.5m), and the road surface layer (0.5m-0m). This layering method ensures that the angle between the UAV camera's optical axis and the surface being inspected is ≤30°, guaranteeing consistent texture resolution across the entire tunnel. Furthermore, it provides clear layer identifiers for subsequent Gaussian sputtering modeling, allowing for independent kernel density weighting of the lining and road surface areas. Simultaneously, it completely eliminates the overlapping blind spots created by single-layer flight paths between the arch top and the road surface, achieving comprehensive data acquisition across the entire tunnel cross-section.

[0037] S2 Data Acquisition: The DJI Matrice 400 drone is equipped with a TB60 smart flight battery, Zenmuse S1 LiDAR and Zenmuse H30 multispectral camera, and is paired with DJI GSPro ground station software and data storage server.

[0038] like Figure 2 and Figure 3 As shown, the straight section adopts a serpentine flight pattern, with a flight height of 6.25 meters for the top arch layer, 4.25 meters for the arch waist layer, 1.75 meters for the sidewall layer, and 0.25 meters for the road surface layer, at a flight speed of 2.2 meters per second. The flight path is designed as a three-dimensional serpentine pattern of "reciprocating within layers and ascending and descending between layers": maintaining a sinusoidal waveform parallel to the tunnel axis within a single layer, with a wavelength of 20 meters and a peak distance of 1.5-2 meters from the wall; after completing the current layer, the UAV climbs or descends along a Z-shaped turning track to the next layer and immediately reverses the serpentine pattern, thus forming a "four-layer, four-directional" coverage within the 100-meter detection section; the curved section adopts a circular pattern with a flight radius of 5 meters and a flight speed of 1.8 meters per second. This method can provide a constant ground sampling distance (GSD≈0.45mm / pixel) and ≥75% heading / lateral overlap in the straight section, while reducing large yaws, reducing heading errors caused by tunnel magnetic interference, and meeting the feature redundancy requirements of SfM and Gaussian sputtering. During the data acquisition process, the drone's position is calibrated using both positioning markers and real-time SLAM positioning with LiDAR, and the missing data areas are automatically re-captured.

[0039] Each drone is responsible for data collection for two detection sections per flight, and a total of 10 flights are required to complete the data collection in the entire tunnel; a total of 9,800 images and approximately 210 million LiDAR point cloud data points are collected, with a total collection time of approximately 2.5 hours. During this period, the tunnel is open to traffic and there are no traffic closures.

[0040] S3 Gaussian sputtering 3D modeling: The DJI SDK tools were used to correct image distortion. The Zenmuse H30 factory calibration tables k1, k2, p1, and p2 were loaded, and radial-tangential distortion correction was performed on the 5472×3648 original image. Simultaneously, the ambient light sensor illuminance E was read, and the brightness was corrected using E-adaptive gamma I_out = (I_in / 255)^(1+0.2·(E-Ref) / Ref)×255, reducing the brightness deviation at the tunnel entrance and exit from ±18% to ±5%, ensuring consistent texture colors in subsequent steps. CloudCompare software was used to remove point cloud intensities ≤15 from the point cloud data. Interference points are then removed using a "statistical outlier" model (k=20, σ=1.0), achieving a retention rate of ≥98%. This invention makes two improvements: First, it introduces ambient light-compensated gamma correction, simultaneously recording the illuminance E in the 0.3-2.5μm band, and uses the E-adaptive gamma function to compress the brightness difference between the tunnel entrance and exit from ±18% to ±5%, ensuring consistent texture color in the subsequent process. Second, it proposes a "strength-distance" joint filtering method, first removing noise points based on reflection intensity ≤15, and then removing distant flying points based on a distance >1.5 times the mode distance, thus maintaining a point cloud retention rate of >98% while avoiding excessive corrosion of the lining microtexture.

[0041] The local ENU coordinates of the UAV are converted into the global coordinates of the tunnel based on the National Geodetic Coordinate System 2000 (CGCS2000); a Gaussian kernel is generated by joint mapping, and SIFT features are extracted from the corrected image and sparsely reconstructed to obtain a color point cloud P_color with RGB information; it is then ICP registered with the lidar P_laser under CGCS2000 to obtain a joint point cloud P_fused. Using P_fused as input, the gaussian-splatting-py framework is invoked to initialize an isotropic Gaussian kernel for each point: g_i=(μ_i,Σ_i,c_i,α_i), where μ_i=(x_i,y_i,z_i), Σ_i=σ²I, σ is taken as the average nearest neighbor distance × 0.3, c_i=(r_i,g_i,b_i), and α_i=1; density weighting is applied, based on the layer label, if μ_i is located in the lining or pavement layer, then the kernel density weight ρ_struct=2.5, and the covariance is scaled to Σ_i′=Σ_i / ρ_struct; the air region maintains ρ=1. This density value is verified by kernel counting in a 0.5m×0.5m sliding window, with a target value of 1200 kernels / cm² for the lining region. 2 In unstructured areas, the density is 500 cores / cm², and encryption continues until this target is met. Here, "Gaussian kernel density" refers to the number of 3D Gaussians per unit area: 1200 cores / cm² for lining / pavement layers. 2 Used to preserve 0.15mm crack details; air zone set to 500 cores / cm 2 This can reduce redundant calculations and suppress floating artifacts.

[0042] In the specific implementation of building the 3D model, the image is first distorted using the camera's factory calibration parameters k1, k2 (radial distortion) and p1, p2 (tangential distortion). Simultaneously, by combining the ambient light sensor illuminance E with a set reference value Ref, the original input brightness I_in is adjusted to the output brightness I_out to achieve brightness equalization. Subsequently, a colored point cloud P_color is generated through sparse reconstruction and registered and fused with the LiDAR point cloud P_laser to obtain a joint point cloud P_fused. Based on this, a Gaussian kernel g_i is initialized, and the kernel density weight ρ_struct is set according to the layering (e.g., 2.5 for the lining layer), thereby scaling the covariance Σ_i′ to achieve differentiated encryption in different regions.

[0043] Iterative optimization, with rendering loss L = Σ‖I_k-C_k‖ 2 +λΣα_i 2 To achieve the objective, Adam is used to simultaneously differentiate with respect to μ, Σ, c, and α, with a learning rate of 1×10⁻⁶. -3 →1×10 -5 The model underwent 1200 iterations, achieving a PSNR ≥ 32dB and a model-to-real-image pixel error of 2.8%. The generated 3D model clearly shows a 0.15 mm wide crack on the lining surface and a 3 mm deep pothole on the road surface.

[0044] By adjusting the parameters of the Gaussian kernel through an iterative optimization algorithm, the error between the rendering effect of the model and the actual image is ≤3%. At the same time, the high-resolution texture information of Zenmuse H30 is integrated to generate a tunnel 3D model that combines high-resolution texture and accurate spatial coordinates.

[0045] S4 Defect Identification: Collect typical tunnel defect samples, such as lining cracks, spalling, and water leakage; road surface potholes, settlement, and cracks. Extract the three-dimensional features (such as the length, width, and depth of cracks, and the area and volume of potholes) and texture features (such as color differences in spalling areas and water stain textures from water leakage) of each defect. Establish a defect feature library containing 5,000+ labeled samples, covering defect morphology annotations under different lighting and humidity environments.

[0046] An improved YOLOv8 model was trained on a GPU server. Specifically, this improved YOLOv8 model incorporates a 3D coordinate attention mechanism and a multispectral feature fusion module. The model was trained using texture images and point cloud feature maps of the 3D model as input and a disease feature database as training samples. During training, a public tunnel disease dataset (5000 images) and labeled images (1500 images) were fused, augmented to 26000 images through HSV perturbation, affine transformation, and multispectral channel shuffle. A 3D-coordinate attention module was introduced at the end of the backbone network, concatenating the Gaussian-rendered depth map Z with RGB as a 4-channel input to generate a weight map M that strengthens crack edges. A 1×1 convolution was added to the Neck layer to unify the RGB and near-infrared channels and then element-wise averaged to achieve multispectral feature fusion. Training hyperparameters: input 1536×1024, batch size 16, initial lr 1×10 -3 Cosine annealing to 1×10 -5 The loss function includes Cls+Bbox+DFL+CA-loss (weights 1:1:1:0.5), spanning 100 epochs. The output mAP@0.5 for lining cracks reaches 97.2%, and mAP@0.5 for road potholes reaches 99.1%, with a 3D localization error ≤0.5m. During the training phase, the training / validation / test sets are divided in an 8:1:1 ratio. The model outputs defect category IDs, 2D bounding boxes, and confidence scores. Using a 3D Gaussian Z-buffer, the center pixel of the bounding box is back-projected to global coordinates. Crack constraints are defined as follows: length L = ||P_start - P_end||, width W = max(projected box width × GSD), depth D = Z_max - Z_min. Pothole constraints are defined as: area A = pixel count × GSD. 2 Volume V = A × average depth; Exfoliation limits: area and maximum thickness are calculated using the same method, but location differs from pits. Classify into three levels—minor, moderate, and severe—according to JTGH12-2015, and record in a quantitative data table. Use the trained model to identify diseases and calculate quantitative parameters, generating a detection report including a 3D model diagram, a disease location annotation map, and a quantitative data table. For example... Figure 4 As shown, a total of 15 lining cracks were identified, with the longest crack measuring 4.2 meters and the widest crack measuring 0.9 millimeters; 3 instances of lining spalling were identified, with a total area of ​​approximately 0.7 square meters; and 10 potholes were identified on the road surface, with the largest pothole measuring approximately 0.03 cubic meters. The identification accuracy rate was 97.8%, and the consistency with the results of manual verification reached 99%.

[0047] The specific embodiments described herein are merely illustrative examples illustrating the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or substitute them with similar methods, without departing from the spirit of the invention or exceeding its defined scope. Although the invention has been detailed and described in the accompanying drawings and foregoing description, such descriptions are considered illustrative or exemplary rather than restrictive. It should be understood that changes and modifications can be made by those skilled in the art within the scope of the following claims.

Claims

1. A method for detecting tunnel lining and pavement defects, characterized in that, Includes the following steps: S1 Tunnel Inspection Area Division: Based on the geometric characteristics of the tunnel, the inspection area is divided using a combination of segmentation and layering. Segmentation involves dividing the tunnel into multiple inspection segments with a preset length as the unit, and setting positioning marks at the start and end positions of each inspection segment. Layering involves dividing the cross-section of each inspection segment along the height direction into the arch top layer, arch waist layer, sidewall layer, and road surface layer. S2 Data Acquisition: Using a drone equipped with a lidar and a multispectral camera, image data and point cloud data are acquired based on flight path rules. During the acquisition process, the drone's position is calibrated by both positioning markers and real-time SLAM positioning by lidar, and missing data areas are automatically captured. Flight path rules: A serpentine flight pattern parallel to the tunnel axis is used in straight sections, and a circular flight pattern around the tunnel cross-section is used in curved sections. S3 Gaussian sputtering 3D modeling: The collected data is preprocessed sequentially, followed by tunnel global coordinate transformation, Gaussian sputtering technology processing, iterative algorithm optimization of Gaussian kernel parameters, and fusion of image texture information to obtain a tunnel 3D model; In the Gaussian sputtering process, a Gaussian kernel is generated based on the Python framework. The Gaussian kernel is obtained through a joint image feature-laser method. An isotropic Gaussian kernel is initialized at each point. S4 Defect Identification: The defect detection model is used to identify defects in the 3D tunnel model. Then, the quantitative parameters of the defects are calculated based on the spatial coordinates of the 3D tunnel model, and a detection report is generated. The types of defects include cracks, spalling, and water leakage in the lining, and potholes, settlement, and cracks in the road surface. The defect detection model is the YOLOv8 model, which incorporates a 3D coordinate attention mechanism and a multispectral feature fusion module.

2. The method for detecting tunnel lining and pavement defects as described in claim 1, characterized in that, In step S1, the geometric features include length, cross-sectional dimensions, and structural features; the positioning identifier is an invisible fluorescent QR code; the arch top layer is within 1.5 meters below the top of the tunnel, and the road surface layer is within 0.5 meters above the ground.

3. The method for detecting tunnel lining and pavement defects as described in claim 1, characterized in that, The serpentine flight is a three-dimensional serpentine flight that reciprocates within a layer and ascends and descends between layers: within a single layer, it maintains a sinusoidal waveform parallel to the tunnel axis; after completing the current layer, the UAV moves to the next layer along a Z-shaped turning track and immediately reverses the serpentine flight.

4. The method for detecting tunnel lining and pavement defects as described in claim 1, characterized in that, For tunnels longer than 5000 meters, multiple drones are used to collect data collaboratively, and the trajectory coordination and data synchronization between drones are achieved through ground station software. For underwater or high-humidity tunnels, drones are equipped with IP67 waterproof protective shells, and anti-fog lenses are added to the lidar and multispectral cameras. A humidity correction module based on multispectral data is added when performing Gaussian sputtering 3D modeling. During the data collection process, if the image overlap rate is insufficient or the point cloud is missing due to occlusion, the drones are set to automatically start a local detour reshoot mode.

5. The method for detecting tunnel lining and pavement defects as described in claim 1, characterized in that, In the serpentine flight mode of step S2, the distance between the flight trajectory and the tunnel wall is maintained at 1.5-2 meters; in the circular flight mode, the image overlap rate of each layer area is ≥75%.

6. The method for detecting tunnel lining and pavement defects as described in claim 1, characterized in that, The multispectral camera is set to RGB mode to simultaneously record ambient light intensity data; the lidar is set to high-density scanning mode with a point cloud density of ≥800 points / square centimeter to simultaneously collect spatial coordinate information.

7. The method for detecting tunnel lining and pavement defects as described in claim 1, characterized in that, In the preprocessing operation of step S3, radial-tangential distortion correction and brightness equalization are performed on the image data. Among them, the ambient light-compensated gamma correction method is used for brightness equalization. A filtering algorithm based on point cloud intensity and spatial distribution statistics is used to remove noise points and isolated points in the point cloud data.

8. The method for detecting tunnel lining and pavement defects as described in claim 1, characterized in that, In the iterative algorithm optimization of Gaussian kernel parameters, the rendering loss is L = Σ‖I_k-C_k‖. 2 +λΣα_i 2 With the objective function, use Adam to differentiate with respect to μ, Σ, c, and α simultaneously.

9. The method for detecting tunnel lining and pavement defects as described in claim 1, characterized in that, For the disease detection model, a 3D-coordinate attention module is introduced at the end of the backbone network to concatenate the Gaussian-rendered depth map Z with RGB as a 4-channel input and generate a weight map M to strengthen the crack edges. Adding a 1×1 convolution to the Neck layer unifies the RGB and near-infrared channels and then averaging the elements to achieve multispectral feature fusion.

10. The method for detecting tunnel lining and pavement defects as described in claim 1, characterized in that, In step S4, regarding the definition of the recognition results, the crack constraint calculation is as follows: length L = ||P_start - P_end||, width W = max(projection frame width × GSD), depth D = Z_max - Z_min; the pit constraint calculation is as follows: area A = pixel count × GSD. 2 Volume V = A × average depth; Peeling limit: area and maximum thickness are calculated in the same way, but the location is different from that of pits and trenches.

Images