Highway slope disease risk zoning method based on three-dimensional positioning and multi-source fusion

Abstract

The present application relates to the technical field of highway slope health monitoring, and provides a highway slope disease risk zoning method based on three-dimensional positioning and multi-source fusion, comprising: acquiring a slope multi-source monitoring data set covering cracks-displacement-deformation three elements with spatial geographic information; constructing an upstream crack three-dimensional positioning module based on YOLO11-MTA and NeRFusion and a downstream slope risk assessment module based on multi-source information fusion; obtaining crack three-dimensional elements with three-dimensional spatial coordinates and geometric attributes based on the upstream crack three-dimensional positioning module; performing spatial co-occurrence and collaborative evolution analysis of the crack-displacement-deformation three elements based on the downstream slope risk assessment module, and performing highway slope disease risk zoning, and outputting three-dimensional risk zoning results. The method can solve the problems of lack of three-dimensional spatial positioning of cracks, insufficient multi-source data fusion, inability to associate displacement anomalies with crack development space, and low risk assessment accuracy in the prior art.

Description

Technical Field

[0001] This invention relates to the field of highway slope health monitoring technology, and in particular to a method for risk zoning of highway slope diseases based on three-dimensional positioning and multi-source fusion. Background Technology

[0002] The safety and stability of highway slopes are directly related to the operational safety of highway traffic. Slope cracks are often accompanied by internal stress redistribution, deformation accumulation, and changes in seepage conditions, and are important apparent precursors in the evolution of slope instability. Therefore, in highway slope operation and maintenance and health monitoring, cracks are usually regarded as a key visual signal requiring attention. However, traditional slope crack monitoring relies heavily on manual inspections or simple two-dimensional image interpretation. These methods are not only time-consuming and labor-intensive, and greatly limited by complex field environments, but their core drawback is that the crack information they output is only at the two-dimensional pixel level, lacking precise three-dimensional spatial positioning. This makes it difficult to uniformly express and directly compare cracks with the slope's three-dimensional topographic structure and internal displacement monitoring data (such as binocular visual displacement and GNSS monitoring points) under the same spatial reference. Crack information can often only be used for qualitative observation and is difficult to deeply support the mechanism analysis and quantitative assessment of the damage.

[0003] To address the problem of slope crack detection and monitoring, the existing technologies mainly include the following implementation schemes: 1. Traditional image processing-based crack detection methods: These methods often employ edge detection operators (such as Sobel and Canny), threshold segmentation, or region growing algorithms to extract cracks. Disadvantages: In natural rock and soil slope scenarios, due to the heterogeneity of the slope material, lighting variations, and severe interference from background textures (such as vegetation and gravel), traditional algorithms are prone to producing numerous artifacts, resulting in extremely low accuracy and poor robustness in extracting thin, low-contrast cracks.

[0004] 2. Deep learning-based 2D crack detection models (such as U-Net and DeepLab): In recent years, deep learning has been widely used for crack segmentation. Although it performs well against clean backgrounds such as concrete, its generalization ability is limited in complex natural slopes. Disadvantages: It can only output a 2D mask and cannot assign precise 3D spatial coordinates to the cracks, resulting in the detection results not being correlated with 3D deformation data.

[0005] 3. Segmentation and modeling techniques based on 3D point clouds (e.g., laser scanning + PointNet): This method acquires slope point clouds using 3D laser scanning (LiDAR) and directly segments cracks in 3D space. Disadvantages: Small surface cracks on natural slopes often lack sufficient geometric feature resolution in point cloud data, and the point clouds lack rich texture information, resulting in low recognition rates for thin, elongated cracks. Furthermore, the direct processing efficiency of massive point clouds is relatively low.

[0006] 4. Isolated Monitoring and Analysis Methods Based on Displacement Monitoring: In highway slope risk assessment, GNSS or binocular vision is typically used independently for point displacement monitoring, UAVs for area deformation observation, and images for crack inspection. Existing solutions based on displacement monitoring have significant drawbacks: the number of monitoring points is limited, reflecting only the temporal changes in displacement at sparse points, and it cannot determine whether abrupt displacement changes have a spatial causal relationship with crack propagation around the monitoring point; when local deformation occurs on the slope, if the displacement monitoring point happens to be located in an undeveloped crack area, it is very easy to miss the true risk. Furthermore, highway slopes are distributed in strips, and displacement monitoring sections are usually set up along the route at intervals of several hundred meters. When a small number of monitoring points show displacement anomalies, existing technology cannot spatially correlate the single-point displacement information with the area crack detection results from UAVs along the route, making it difficult to determine whether the anomaly is a localized isolated phenomenon or represents an overall deterioration trend of a section of the slope, affecting the economy and accuracy of maintenance decisions. Furthermore, the lack of a multi-source information spatial co-occurrence correlation analysis mechanism under a unified spatial benchmark makes it impossible to reveal the co-evolutionary relationship between crack development and slope deformation, which can easily lead to misjudgments or omissions such as "significant deformation but no cracks" or "obvious cracks but no deformation".

[0007] Therefore, it is necessary to provide a method for risk zoning of highway slope diseases based on three-dimensional positioning and multi-source fusion to solve the above-mentioned technical problems. Summary of the Invention

[0008] The main objective of this invention is to provide a risk zoning method for highway slope diseases based on three-dimensional positioning and multi-source fusion, so as to solve the problems in the prior art such as lack of three-dimensional spatial positioning of cracks, insufficient fusion of multi-source data, inability to correlate displacement anomalies with crack development space, and low accuracy of risk assessment.

[0009] The specific technical solution is as follows: The method for risk zoning of highway slope diseases based on three-dimensional positioning and multi-source fusion includes the following steps: Step S1: Collect aerial images and 3D point cloud data of highway slopes by UAV, and obtain the displacement time series data of monitoring points through the slope surface displacement measurement system to obtain a multi-source monitoring dataset of slopes with spatial geographic information and covering the three elements of cracks, displacement and deformation. Step S2: Construct an upstream crack three-dimensional localization module based on YOLO11-MTA and NeRFusion and a downstream slope risk assessment module based on multi-source information fusion; Step S3: Based on the slope multi-source monitoring dataset obtained in Step S1, and combined with the upstream crack three-dimensional localization module based on YOLO11-MTA and NeRFusion, obtain the crack three-dimensional elements with three-dimensional spatial coordinates and geometric attributes. Step S4: Based on multi-source monitoring data of the slope and three-dimensional elements of cracks, combined with the deformation results of the surface area reconstructed by multi-phase UAVs, the spatial co-occurrence and synergistic evolution analysis of the three elements of cracks, displacement and deformation is carried out through the downstream slope risk assessment module based on multi-source information fusion, and then quantitative indicators are extracted; based on the quantitative indicators, the risk zoning of highway slope diseases is carried out, and the three-dimensional risk zoning results are output.

[0010] Furthermore, the YOLO11-MTA model includes an encoder, a decoder, a multi-attention enhancement module, and a detection segmentation head; The encoder uses a CSPDarknet backbone network to extract multi-scale features from the input image and generate feature maps at different levels. The decoder uses a PANet feature fusion structure to upsample and fuse multi-scale features, taking into account both the detailed features of the cracks and the global semantic information. The multi-attention enhancement module is embedded in the feature transfer node of the encoder and decoder, and is used to enhance and weight the crack features; The detection and segmentation head adopts a dual-branch structure, which is used to simultaneously output the bounding box detection results of cracks and the pixel-level segmentation mask.

[0011] Furthermore, the upstream crack three-dimensional localization module based on YOLO11-MTA and NeRFusion consists of four parts connected in series: Stage 0, Stage 1, Stage 2 and Stage 3. Stage 0 includes a series of connected image preprocessing modules and a mesh generation and sub-image generation module; Stage 1 includes a series of connected YOLO11-MTA crack segmentation modules, crack mask post-processing modules, and a panoramic stitching module for recognition results; Stage 2 includes a series of connected crack centerline extraction modules, a main direction quantitative calculation module, and a crack orientation automatic classification module; Stage 3 includes a series of connected multi-view projection geometric modeling modules, a 3D coordinate solving module, a NeRFusion geometric constraint optimization module, and a complete crack mask 3D mapping module. The downstream slope risk assessment module based on multi-source information fusion includes two connected parts: Stage 4 and Stage 5. Stage 4 includes a multi-source data spatial benchmark unification module, a displacement-triggered crack correlation sub-module, and a multi-temporal spatial co-occurrence analysis module, which are connected in series. Stage 5 includes a multi-source fusion risk criterion construction module, a risk zoning and assessment implementation module, and an engineering control suggestion output module, which are connected in series.

[0012] Furthermore, step S3 specifically includes: The UAV aerial images are input into the image preprocessing module for preprocessing and output standardized images; the standardized images are then input into the grid division and sub-map generation module to obtain standardized sub-maps and sub-map spatial index information. The YOLO11-MTA crack segmentation module outputs a pixel-level crack segmentation binary mask within the sub-image range based on a standardized sub-image; the crack mask post-processing module preprocesses the pixel-level crack segmentation binary mask within the sub-image range to obtain a continuous crack mask for the sub-image. Based on the continuous crack mask of all subgraphs and the spatial index information of the subgraphs, the panoramic crack segmentation mask within the entire slope area is obtained through the panoramic stitching module of the recognition results. Based on a panoramic crack segmentation mask covering the entire slope area, a crack centerline extraction module outputs a single-pixel-wide, fully connected crack centerline; the main direction quantitative calculation module outputs the crack main direction angle based on the crack centerline. The automatic crack orientation classification module classifies cracks based on their principal direction angle. Output the crack orientation classification results; The camera intrinsic and extrinsic parameters of the UAV aerial imagery are input into the multi-view projection geometry modeling module to achieve the projection transformation from 2D pixel coordinates to 3D world coordinates. The crack pixel coordinates and camera intrinsic and extrinsic parameters are input into the 3D coordinate solving module to obtain the preliminary 3D spatial coordinates of the crack sampling points. Based on multiple crack sampling points within the same crack instance, the preliminary 3D spatial coordinates of the crack centroid are calculated. Based on the preliminary 3D spatial coordinates of the crack sampling points and the crack centroid, and the continuous 3D geometric field of the slope reconstructed by NeRFusion, the high-precision 3D spatial coordinates of the crack sampling points and the crack centroid are obtained through the NeRFusion geometric constraint optimization module. Based on the panoramic crack segmentation mask, the high-precision three-dimensional spatial coordinates of crack sampling points and the high-precision three-dimensional spatial coordinates of crack centroid points, the complete crack mask three-dimensional mapping module outputs crack three-dimensional elements with three-dimensional spatial coordinates and geometric attributes.

[0013] Furthermore, based on the principal direction angle of the crack The classification rules for automated crack classification are as follows: When the principal direction angle of the crack satisfies or At that time, it was determined to be a transverse crack; when At that time, it was determined to be a vertical crack; The remaining cracks are considered diagonal cracks; Calculate the minimum angle between the diagonal crack and the horizontal or vertical direction, when or When, it is determined to be a transverse crack; when At that time, it was determined to be a vertical crack.

[0014] Furthermore, step S4 specifically includes: Based on multi-source monitoring data of slope, three-dimensional elements of cracks with three-dimensional spatial coordinates and geometric attributes obtained in step S3, and deformation results of surface area reconstructed by UAV in multiple phases, the three-dimensional data of cracks, displacement and deformation are fully registered under the same spatial reference through the multi-source data spatial reference unification module, and multi-source registration data and spatial neighborhood index are output. The displacement-triggered crack association submodule enables automatic tracing from displacement anomalies to crack areas of interest. Based on multi-source registration data with the same spatial reference, the spatial co-occurrence relationship between cracks and slope deformation, the multi-temporal collaborative evolution law, and the quantitative results of the engineering indication significance of cracks are output through the multi-temporal spatial co-occurrence analysis module. Based on quantitative indicators, a standardized judgment rule for slope status is constructed through a multi-source fusion risk judgment construction module. The risk zoning and assessment implementation module realizes refined risk zoning of the slope surface according to the standardized judgment rules of slope status, and generates a visualized three-dimensional risk zoning result. The project management recommendation output module outputs differentiated management measures based on the three-dimensional risk zoning results.

[0015] Furthermore, the displacement-triggered crack correlation submodule uses the time-series displacement data of monitoring points as the driving source to achieve automatic tracing from displacement anomalies to the crack area of ​​interest; specifically including: Automatic displacement inflection point detection: A sliding window is used to fit the displacement time-series curve of each monitoring point, and the change in displacement rate within the window is calculated. ;when When the preset threshold is exceeded, the monitoring point is determined to have a displacement acceleration inflection point and is marked as a triggered state; Neighborhood crack reverse query: Centered on the monitoring point in the triggered state, query the crack 3D features generated in step S3 in the spatial neighborhood index, and extract the total crack length within the neighborhood. , percentage of transverse cracks and crack increase rate ; Displacement-crack correlation quantification: Constructing the displacement-crack correlation index (DCI) to quantify the spatial coupling degree between displacement anomalies and crack development, providing supplementary criteria for subsequent risk assessment; The formula for calculating the Displacement-Crack Correlation Index (DCI) is as follows: ; in: This represents the normalized change in displacement rate. This represents the normalized total length of the neighborhood cracks. The percentage of transverse cracks; These are the weighting coefficients, with default values ​​of 0.40, 0.35, and 0.25.

[0016] Furthermore, in the multi-temporal spatial co-occurrence analysis module, to achieve a quantitative characterization of the spatial co-occurrence relationship of multi-source information, three types of quantitative indicators—cracks, displacement, and surface deformation—were constructed: Crack index Let the first Each monitoring point in time phase Within the corresponding local neighborhood, the total length of the transverse cracks is The neighborhood area is ,but: ; Displacement index Let the first The cumulative displacement of each monitoring point is ,but: ; Surface deformation index : Assume time phase The total number of points in the 3D point cloud within the local neighborhood is The number of points whose cumulative deformation exceeds the preset significant deformation threshold is: ,but: ; The crack index is normalized using the range normalization method. Displacement index and surface deformation index A uniform mapping to the [0,1] interval yields a normalized index. , and ; Constructing a comprehensive co-occurrence index Consistency Difference Indicators : ; .

[0017] Furthermore, the multi-source fusion risk judgment construction module uses a comprehensive co-occurrence index. Consistency Difference Indicators Based on this, and combined with the DCI index and dominant information sources, standardized judgment rules for slope condition are output. The specific standardized judgment rules for slope condition are as follows: when ≤0.20 and When the value is less than 0.15, the co-occurrence type is low activity. when ≥0.70 and min( , , When the co-occurrence value is ≥0.50, the co-occurrence type is strongly cooperative. When 0.20 < <0.70 and When the value is less than 0.15, the co-occurrence type is progressively cooperative. when <0.70 and ≥0.15, and satisfy At that time, the co-occurrence type was crack-dominated; when <0.70 and ≥0.15, and satisfy At that time, the co-occurrence type is deformation-dominant; When DCI ≥ 0.65 and At that time, the co-occurrence type was displacement-crack anomaly.

[0018] Furthermore, based on the co-occurrence type, the three-dimensional risk zoning results for highway slopes are as follows: When the co-occurrence type is low activity, the risk zoning level is low activity stable zone; When the co-occurrence type is progressive synergistic, the risk zoning level is progressive deformation zone; When the co-occurrence type is strong synergy, the risk zoning level is high-risk activity area; When the co-occurrence type is crack-dominated, the risk zoning level is an area requiring key review; When the co-occurrence type is the deformation-dominant type, the risk zoning level is the area requiring key review / the area with hidden risks; When the co-occurrence type is displacement-crack anomaly, the risk zoning level is high-risk hidden activity zone.

[0019] The beneficial effects achieved by this solution are: (1) This invention integrates the YOLO11-MTA and NeRFusion three-dimensional high-precision crack mapping mechanism. At the front end, by introducing the YOLO11-MTA model with multi-scale channels and spatial attention mechanism, the interference of complex soil and rock background is overcome, and the two-dimensional mask of slender cracks is accurately extracted. At the back end, instead of the traditional ray triangulation method which is susceptible to noise interference, the NeRFusion model is innovatively introduced to reconstruct a continuous three-dimensional geometric field. This geometric field is used to perform line-of-sight back projection and surface intersection point solution, eliminating the ambiguity of two-dimensional projection and realizing high-precision positioning of crack features under a unified three-dimensional spatial reference.

[0020] (2) This invention proposes a quantitative classification mechanism for the engineering significance of cracks. By extracting the centerline of the crack and using principal component analysis (PCA) to calculate its first principal direction vector, the cracks are automatically divided into transverse, vertical and oblique types according to a set threshold, thereby screening out the core disease indicators (such as transverse cracks) that are related to the main deformation height of the slope and eliminating interfering information.

[0021] (3) This invention proposes a crack reverse tracing and correlation quantification mechanism driven by displacement monitoring. Addressing the engineering challenges of large spacing between displacement monitoring points and difficulty in tracing anomalies on highway slopes, this invention proposes a crack reverse tracing method driven by displacement time-series inflection point detection. By constructing the displacement-crack correlation index (DCI), the original unidirectional "crack → risk assessment" path is upgraded to a "displacement → risk assessment" path. The closed-loop analysis architecture with bidirectional correlation of cracks upgrades displacement monitoring data from passive input to core analysis engine, filling the gap in existing technology that cannot trace the development status of cracks from the displacement anomaly space.

[0022] (4) This invention uses displacement monitoring as an anchor point for multi-source collaborative analysis to eliminate assessment blind spots: This invention breaks through the limitation of traditional displacement monitoring, which can only output single-point time-series curves, and unifies slope displacement field monitoring data with three-dimensional crack mapping results and surface deformation results under the same spatial benchmark for co-occurrence analysis. This mechanism enables the temporal anomalies of displacement monitoring points to obtain real-time annotations of the development status of cracks in the vicinity: when the displacement increases suddenly, it can be checked simultaneously whether there is crack expansion or new cracks at the corresponding location; when crack development is significant, it can be verified simultaneously whether the displacement field responds. At the same time, the introduction of the displacement-crack correlation index (DCI) enables the system to automatically identify hidden high-risk areas where "displacement increases but the surface crack response lags behind", effectively avoiding omissions and misjudgments caused by traditional single data sources.

[0023] (5) Constructing a standardized multi-source integrated risk zoning system with strong engineering applicability: Based on the results of multi-source correlation analysis, this invention has formulated a comprehensive risk judgment feature set covering six co-occurrence types, accurately dividing the slope into low-activity stable zone, progressive deformation zone, high-risk activity zone, zone requiring key review, zone requiring key review / hidden risk zone, and high-risk hidden activity zone. This quantitative and refined zoning result directly connects with differentiated engineering management recommendations, providing highly practical decision support for disaster prevention, mitigation, maintenance, and repair of highway slopes. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating the method for risk zoning of highway slope diseases based on three-dimensional positioning and multi-source fusion in an embodiment of the present invention. Figure 2 This is a schematic diagram of the three-dimensional position of each coordinate system relative to the target bounding box in an embodiment of the present invention; Figure 3 This is a schematic diagram of the reconstruction of three-dimensional crack points by the projection light from the multi-view camera in an embodiment of the present invention; Figure 4 This is a flowchart of NeRFusion in an embodiment of the present invention.

[0026] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0028] Example See Figure 1 This invention proposes a method for risk zoning of highway slope diseases based on three-dimensional positioning and multi-source fusion, comprising the following steps: Step S1: Collect aerial images and 3D point cloud data of highway slopes by UAV, and obtain the displacement time series data of monitoring points through a slope surface displacement measurement system represented by a binocular visual displacement monitoring station, to obtain a multi-source monitoring dataset of slopes with spatial geographic information and covering the three elements of cracks, displacement and deformation. Step S2: Construct an upstream crack three-dimensional localization module based on YOLO11-MTA and NeRFusion and a downstream slope risk assessment module based on multi-source information fusion; In this embodiment, the YOLO11-MTA model includes an encoder, a decoder, a multi-attention enhancement module, and a detection segmentation head; The encoder employs a CSPDarknet backbone network to extract multi-scale features from the input image, generating feature maps at different levels. The decoder uses a PANet feature fusion structure to upsample and fuse multi-scale features, taking into account both the detailed features of the cracks and global semantic information. The multi-attention enhancement module is embedded in the feature transfer nodes of the encoder and decoder to enhance and weight the crack features. The multi-attention enhancement module is configured to introduce channel attention and spatial attention during the multi-scale feature transfer process to weight the features of slender and low-contrast cracks and suppress background interference. The detection and segmentation head adopts a dual-branch structure to simultaneously output the crack bounding box detection results and pixel-level segmentation masks.

[0029] Specifically, the pixel-level segmentation mask output by the aforementioned detection and segmentation head is calculated from the crack segmentation probability of the YOLO11-MTA model, and the crack segmentation probability output formula is as follows: ; in: The probability value for each pixel to belong to the crack region; It is a 3×3 convolutional layer; This is the fused feature map output by the decoder; The activation function maps the output value to the 0-1 range, generating a pixel-level crack segmentation mask.

[0030] In this embodiment, NeRFusion consists of a multi-view feature fusion module, a three-dimensional radiation field reconstruction module, and a depth optimization module. See also... Figure 4 The core idea is to introduce multi-view image feature fusion and geometric consistency constraints on the implicit 3D modeling framework of neural radiation fields to achieve high-precision continuous 3D surface reconstruction of complex slope scenes. This provides continuous geometric constraints for the mapping from 2D cracks to 3D space and solves the sensitivity problem of traditional ray triangulation methods under noise and poor viewpoint configuration. The multi-view feature fusion module extracts and matches features from UAV images from different viewpoints to establish cross-viewpoint pixel correspondences; the 3D radiation field reconstruction module constructs a continuous 3D geometric field and radiation field of the slope scene through volume rendering technology; and the depth optimization module optimizes the accuracy of the reconstructed 3D geometric field through multi-view stereo matching constraints to ensure the accuracy of 3D coordinate solution.

[0031] The formulas for volume rendering color and depth output are as follows: ; ; in: For light The rendered color value; For light The corresponding depth value; Cumulative transmittance of light; The density value of the voxel; The distance between adjacent sampling points; The color value of the voxel; The depth value of the sampling point; This represents the number of sampling points for the light rays.

[0032] In this embodiment, the upstream crack three-dimensional localization module based on YOLO11-MTA and NeRFusion includes four parts connected in series: Stage 0, Stage 1, Stage 2 and Stage 3. Stage 0 includes a series of connected image preprocessing modules and a mesh generation and sub-image generation module; Stage 1 includes a series of connected YOLO11-MTA crack segmentation modules, crack mask post-processing modules, and a panoramic stitching module for recognition results; Stage 2 includes a series of connected crack centerline extraction modules, a main direction quantitative calculation module, and a crack orientation automatic classification module; Stage 3 includes a series of connected multi-view projection geometric modeling modules, a 3D coordinate solving module, a NeRFusion geometric constraint optimization module, and a complete crack mask 3D mapping module. The downstream slope risk assessment module based on multi-source information fusion includes two connected parts: Stage 4 and Stage 5. Stage 4 includes a multi-source data spatial benchmark unification module, a displacement-triggered crack correlation sub-module, and a multi-temporal spatial co-occurrence analysis module, which are connected in series. Stage 5 includes a multi-source fusion risk criterion construction module, a risk zoning and assessment implementation module, and an engineering control suggestion output module, which are connected in series.

[0033] Step S3: Based on the slope multi-source monitoring dataset obtained in Step S1, and combined with the upstream crack 3D localization module based on YOLO11-MTA and NeRFusion, obtain the 3D features of the crack with 3D spatial coordinates and geometric attributes; specifically: The UAV aerial images are input into the image preprocessing module of Stage 0 for preprocessing. Image standardization processing algorithms such as distortion correction, dehazing, and contrast enhancement are used to eliminate problems such as lens distortion, uneven lighting, and fog interference in the UAV aerial images, and output standardized images to provide a high-quality image foundation for subsequent crack identification. Preferably, the dehazing adopts the dark channel prior dehazing algorithm. The specific processing process is as follows: the atmospheric light value and transmittance of the original image are estimated using the dark channel prior algorithm to eliminate fog, histogram equalization is used to eliminate uneven lighting, and distortion correction is performed in combination with the camera intrinsic parameter matrix.

[0034] The standardized image is input into the Stage 0 gridding and sub-image generation module. Through proportional grid cropping and sub-image normalization, standardized sub-images and sub-image spatial index information are obtained. This solves the problems that ultra-high resolution images cannot be directly input into deep learning models and that scaling the entire image can easily cause loss of crack details. The specific processing procedure is as follows: a fixed sub-image size is set for the high-resolution image, and sliding window cropping is performed according to proportional gridding. During cropping, the pixel offset of the upper left corner of each sub-image in the original image is recorded as spatial index information.

[0035] Furthermore, the formula for the subgraph space index record is as follows: ; in: For the first Line 1 Spatial index information for subgraphs; This represents the pixel offset of the top-left corner of the sub-image within the original image. The width of the subgraph; The height of the subgraph; the spatial index information of the subgraph provides a spatial reference for subsequent identification result stitching.

[0036] Stage 1's YOLO11-MTA crack segmentation module takes the standardized sub-image obtained in Stage 0 as input and the pixel-level crack segmentation binary mask within the sub-image range as output. Through pre-training and fine-tuning of the YOLO11-MTA model inference, it achieves end-to-end accurate crack identification in complex soil and rock slope scenarios, providing high-quality two-dimensional basic data for subsequent three-dimensional mapping. The pre-training and fine-tuning process here belongs to the well-known deep learning supervised training technique in this field. The specific process is as follows: First, the YOLO11-MTA model is pre-trained using a general image dataset to obtain basic image feature extraction capabilities. Then, a publicly available crack dataset (such as the GeoCrack dataset) containing natural rock and soil surface textures and complex lighting conditions is introduced for fine-tuning the network weight parameters. The fine-tuned model must effectively suppress the interference of slope material heterogeneity and background texture, and possess the generalization ability to segment low-contrast, thin cracks at the pixel level. During the inference phase, the standardized sub-image is input into the fine-tuned model for forward propagation calculation, directly outputting a pixel-level crack segmentation binary mask within the sub-image.

[0037] The crack mask post-processing module preprocesses the pixel-level crack segmentation binary mask within the sub-image range to obtain a continuous crack mask for the sub-image. The preprocessing includes morphological opening and closing operations. Specifically, the morphological opening operation is first applied to remove isolated noise pixels, and then the closing operation is applied to fill in small holes and breaks in the crack area to ensure the continuity of the crack area.

[0038] The panoramic stitching module takes the continuous crack mask of all sub-images and the spatial index information of the sub-images recorded in Stage 0 as input. Through inverse spatial index transformation and edge stitching fusion, it outputs a panoramic crack segmentation mask within the entire slope area. The specific processing is as follows: read the spatial index information of the sub-images, inversely transform the coordinates of the sub-image masks to restore them to the original image coordinate system, and use the maximum value fusion strategy at the overlapping edges to eliminate the stitching gaps at the edges of the sub-images, forming a continuous crack segmentation mask within the entire slope area, which fully preserves the detailed features and spatial continuity of the cracks.

[0039] The crack centerline extraction module in Stage 2 takes the panoramic crack segmentation mask obtained in Stage 1 as input and processes it through the Zhang-Suen morphological thinning algorithm to output a crack centerline with a single pixel width and complete connectivity. The specific processing is as follows: by iteratively judging and eliminating the boundary pixels of the crack connected domain until the crack skeleton with a single pixel width is retained, the influence of crack width difference on orientation calculation is eliminated, and standardized input data is provided for subsequent orientation calculation.

[0040] The main direction quantitative calculation module takes the single-pixel wide crack centerline as input and the crack main direction angle as input. For output, specifically, the covariance matrix of the pixel set of the centerline of the crack is calculated by principal component analysis (PCA) and eigenvalue decomposition is performed. The eigenvector corresponding to the largest eigenvalue is obtained as the first principal direction vector of the crack, and then the angle between this direction vector and the horizontal axis of the image is calculated. This enables the quantitative characterization of crack orientation.

[0041] The automatic crack orientation classification module classifies cracks based on their principal direction angle. Output the crack orientation classification results; use the calculated included angle Logical judgments are performed to automatically classify cracks and distinguish between crack types with different engineering indications. The specific classification rules are as follows: When the principal direction angle of the crack satisfies or When the cracks are identified as transverse cracks, they are often distributed along contour lines or approximately perpendicular to the main sliding direction of the slope. They are highly correlated with the overall tensile or shear deformation process of the slope and are the core surface response index of slope displacement evolution, which is given higher weight in subsequent analysis. when When the cracks are identified as vertical cracks, they are often developed along the main slope direction. Their formation is often related to long-term surface runoff, drainage erosion, or local weathering, and their direct correlation with the overall slope displacement is relatively weak. The remaining cracks are considered diagonal cracks; calculate the minimum angle between the diagonal crack and the horizontal or vertical direction, when... or When, it is determined to be a transverse crack; when At that time, it was determined to be a vertical crack.

[0042] In this embodiment, the single-pixel wide crack centerline serves as the reference input for subsequent 3D coordinate solving, the skeleton constraint for the complete crack mask 3D mapping, and the geometric basis for outlier filtering; the crack principal direction angle The crack orientation classification results are used for adaptive optimization of subsequent three-dimensional mapping sampling density and analysis of slope stress field evolution trends; the crack orientation classification results are used for subsequent engineering attribute labeling of three-dimensional crack elements, calculation of crack risk indicators, and formulation of differentiated engineering control measures.

[0043] The single-pixel wide crack centerline and crack principal direction angle output by Stage 2 The crack orientation classification results are input into Stage 3; among them, the centerline of a single-pixel wide crack serves as the sampling reference for solving the 3D coordinates and the skeleton constraint for the 3D mapping of the complete crack mask, and the crack principal orientation angle is used. Adaptive adjustment of 3D mapping sampling density is used, and crack orientation classification results serve as engineering attribute labels for crack 3D elements. The camera intrinsic and extrinsic parameters of UAV aerial photography are input into Stage3's multi-view projection geometry modeling module. This module takes the camera intrinsic and extrinsic parameters of the UAV aerial photography as input and outputs a mathematical model of the projection transformation from 2D pixel coordinates to 3D world coordinates. Through pinhole camera modeling and coordinate system transformation relationship construction, it establishes the geometric relationships of multi-view imaging, providing a theoretical basis for subsequent 3D coordinate solutions. Specifically, see Figure 2 and Figure 3 This embodiment establishes the UAV body coordinate system. Camera coordinate system Image pixel coordinate system With world coordinate system The geometric mapping relationship between them. During spatial positioning, starting from the camera's optical center and passing through specific crack pixels on the image plane... A line-of-sight ray is formed; this ray is represented in the world coordinate system as a vector direction originating from the camera's optical center. The physical intersection of this line-of-sight ray and the slope surface is solved by introducing a continuous 3D surface reconstructed by NeRFusion as a geometric constraint. This allows for accurate positioning of the crack in the three-dimensional world coordinate system. This mechanism effectively eliminates the ambiguity of projection depth uncertainty under a single viewpoint by checking the consistency of reprojection under multiple views.

[0044] A 3D point in the camera coordinate system is rotated, translated, and projected onto the camera intrinsic parameter matrix, ultimately mapping to pixel coordinates on the 2D image plane. The transformation formula between pixel coordinates and world coordinates is as follows: ; ; in: , These are the depth values ​​of the three-dimensional crack point from the perspectives of camera I and camera II, respectively, which are the vertical distances from the optical center of the camera to the three-dimensional spatial point, in mm; , These are the homogeneous pixel coordinates of the crack pixel in the corresponding images of camera I and camera II, respectively. , , , This represents the row and column number of a pixel in the image plane, expressed in pixels. , These are the internal parameter matrices for camera I and camera II, with a dimension of 3×3, containing intrinsic parameters such as camera focal length and principal point coordinates, which were obtained in advance through Zhang Zhengyou's calibration method; , These are the rotation matrices of camera I and camera II relative to the world coordinate system, with a dimension of 3×3, describing the camera's attitude in three-dimensional space. They are obtained by solving the UAV aerial POS data and the bundle adjustment method. , These are the translation matrices of camera I and camera II relative to the world coordinate system, with a dimension of 3×1, describing the position of the camera optical center in three-dimensional space. They are obtained by solving the UAV aerial POS data and the bundle adjustment method. The extrinsic parameter matrix of the camera is derived from the rotation matrix. Translation matrix It is pieced together, with dimensions of 3×4; Three-dimensional crack point Homogeneous coordinates in the world coordinate system , , These are the coordinates of a 3D point along the X, Y, and Z axes in the world coordinate system, in mm. In this embodiment, a 3D crack point (i.e., a crack sampling point) refers to a 3D spatial sampling point obtained by solving multi-view geometric constraints from crack pixel points or crack centerline sampling points; a crack centroid point refers to the central representative point of a single crack connected domain or crack instance, obtained by statistical calculation from multiple 3D crack points corresponding to that crack. The crack centroid point is used to characterize the overall spatial position of a single crack, while the 3D crack points are used to characterize the complete crack geometry.

[0045] The crack pixel coordinates and camera intrinsic and extrinsic parameters are input into the 3D coordinate solving module. A coefficient matrix is ​​constructed based on the camera intrinsic and extrinsic parameters from multiple perspectives. The linear equation system is solved by least squares optimization algorithms such as singular value decomposition (SVD). The specific process is as follows: the least squares fitting solution is performed on the ray intersection equations from multiple perspectives using SVD to obtain the approximate values ​​of the crack sampling points at the spatial intersection points. The preliminary 3D spatial coordinates of the crack centroid are calculated based on multiple crack sampling points within the same crack instance.

[0046] The NeRFusion geometric constraint optimization module takes the preliminary 3D spatial coordinates of the crack sampling points and crack centroids, as well as the continuous 3D geometric field of the slope reconstructed by NeRFusion, as input. Through line-of-sight back projection and surface intersection solving, it obtains the high-precision 3D spatial coordinates of the crack sampling points and crack centroids. Specifically, it extracts the surface isosurface of the NeRFusion continuous geometric field, back-projects the line-of-sight ray onto the isosurface to solve for the precise intersection, eliminates coordinate deviations caused by reconstruction noise and lens distortion, and finally obtains the high-precision 3D spatial coordinates of the crack sampling points and crack centroids.

[0047] Based on the high-precision 3D spatial coordinates of the panoramic crack segmentation mask, crack sampling points, and crack centroid, a complete crack mask 3D mapping module outputs 3D crack elements with 3D spatial coordinates and geometric attributes. Specifically, the complete crack mask 3D mapping module takes the high-precision 3D spatial coordinates of the panoramic crack segmentation mask, crack sampling points, and crack centroid as input. Through pixel-by-pixel 3D coordinate solving, point cloud registration and index construction, and outlier filtering, it outputs 3D crack elements with 3D spatial coordinates and geometric attributes. The specific processing is as follows: the solved 3D crack coordinates are spatially registered with the slope 3D point cloud model using a nearest neighbor search algorithm (such as KD-Tree), and outliers that are significantly deviated from the slope surface are removed. This achieves a complete conversion from 2D semantic information to 3D spatial objects, establishing a spatial correspondence between the 3D crack elements and the slope 3D point cloud model, providing a unified geometric carrier for subsequent multi-source correlation analysis.

[0048] Step S4: Based on multi-source monitoring data of the slope and three-dimensional features of cracks, combined with the deformation results of the surface area reconstructed by multi-phase UAV three-dimensional reconstruction, the spatial co-occurrence and co-evolution analysis of the three elements of cracks, displacement, and deformation is carried out through the downstream slope risk assessment module based on multi-source information fusion, and then quantitative indicators (i.e., comprehensive co-occurrence index) are extracted. Consistency difference index The Displacement-Crack Correlation Index (DCI) is used to zon the risk of highway slope diseases based on quantitative indicators, and the three-dimensional risk zoning results are output. Specifically, this includes: Based on multi-source monitoring data of slope, three-dimensional elements of cracks with three-dimensional spatial coordinates and geometric attributes obtained in step S3, and deformation results of surface area reconstructed by UAV in multiple phases, the three heterogeneous data of cracks, displacement and deformation are fully registered under the same spatial reference through the multi-source data spatial reference unification module, and multi-source registration data and spatial neighborhood index are output.

[0049] The multi-source data spatial benchmark unification module takes the crack three-dimensional elements obtained from Stage 3, the displacement time series data of the slope binocular monitoring points, the surface deformation results of the UAV multi-stage three-dimensional reconstruction, and the slope three-dimensional point cloud model as inputs, and the multi-source registration data and spatial neighborhood index under the unified spatial benchmark as outputs to realize the spatial registration of multi-source heterogeneous data and establish a unified spatial analysis benchmark. Specifically, the coordinate system sub-process uses the world coordinate system of UAV 3D reconstruction as the sole reference. It prioritizes a seven-parameter coordinate system transformation of the monitoring point coordinates of slope surface displacement measurement devices, represented by binocular vision monitoring stations, ensuring that the displacement monitoring data, crack 3D elements, and slope 3D point cloud model are all under a completely unified spatial reference and time cross-section, with coordinate transformation errors controlled within 5mm. Simultaneously, the surface deformation results obtained from multi-phase UAV 3D reconstruction are registered point-by-point with the slope 3D point cloud, ensuring that each terrain point in the slope 3D model carries corresponding cumulative deformation and deformation rate attributes. Ultimately, this achieves complete registration of the three heterogeneous data types—cracks, displacement, and deformation—under the same spatial reference. The spatial neighborhood index construction sub-process, based on an octree spatial index structure, matches a corresponding spatial neighborhood for each crack 3D element, with a default neighborhood radius of 10m. Simultaneously, it establishes a corresponding spatial influence range for each binocular monitoring point, extracting crack development characteristics and surface deformation characteristics within the range. This ensures that multi-source data are analyzed within the same spatial unit, eliminating analytical errors caused by spatial misalignment.

[0050] The displacement-triggered crack correlation submodule enables automatic tracing from displacement anomalies to crack areas of interest. Addressing the challenges of large spacing between highway slope displacement monitoring points and difficulties in anomaly tracing, this submodule uses time-series data from displacement monitoring points as the driving source to achieve automatic tracing from displacement anomalies to crack areas of interest. Specifically, it includes: (1) Automatic detection of displacement inflection points: The displacement time-series curve of each monitoring point is fitted with a sliding window, and the change in displacement rate within the window is calculated. ;when When the preset threshold is exceeded, the monitoring point is determined to have a displacement acceleration inflection point and is marked as a triggered state; (2) Neighborhood crack reverse query: Taking the monitoring point in the triggered state as the center, query the crack three-dimensional features generated in step S3 in the spatial neighborhood index, and extract the total length of cracks in the neighborhood. , percentage of transverse cracks and crack increase rate ; (3) Quantification of displacement-crack correlation: Construct the Displacement-Crack Index (DCI) to quantify the spatial coupling between displacement anomalies and crack development, providing supplementary criteria for subsequent risk assessment; DCI is an event-driven quantitative indicator that is only calculated when an acceleration inflection point appears at the displacement monitoring point. Its core purpose is to verify whether the displacement anomaly is caused by the overall instability of the slope and to identify the hidden slip risk that traditional methods easily miss, namely, "deep deformation has not yet been transmitted to the surface". The formula for calculating the Displacement-Crack Correlation Index (DCI) is as follows: ; in: This represents the normalized change in displacement rate. This represents the normalized total length of the neighborhood cracks. The percentage of transverse cracks; These are the weighting coefficients, with default values ​​of 0.40, 0.35, and 0.25.

[0051] Based on multi-source registration data with the same spatial reference, the spatial co-occurrence relationship between cracks and slope deformation, the multi-temporal collaborative evolution law, and the quantitative results of the engineering indication significance of cracks are output through the multi-temporal spatial co-occurrence analysis module. The multi-temporal spatial co-occurrence analysis module takes multi-source registration data with the same spatial benchmark as input. Through typical temporal phase selection, core index extraction, and co-evolutionary pattern analysis, it outputs the spatial co-occurrence relationship between cracks and slope deformation, the multi-temporal co-evolution law, and the quantitative results of the engineering indication significance of cracks. The specific processing is as follows: within a unified spatial neighborhood, the total length density of transverse cracks, the cumulative displacement of points, and the proportion of significant deformation points in the area are extracted according to a preset algorithm and the range is normalized to reveal the intrinsic relationship between crack development and slope deformation.

[0052] Spatial co-occurrence refers to the phenomenon that, under a unified spatial benchmark, three types of heterogeneous data—cracks, displacement, and surface deformation—occur simultaneously within the same local spatial unit and exhibit an inherent mechanical relationship. Its function is to break through the limitations of traditional isolated analysis of multi-source data, integrate scattered point, line, and surface monitoring data in the spatial dimension, reveal the inherent causal relationship between crack development and slope deformation, and avoid misjudgments or omissions such as "significant deformation but no cracks" or "obvious cracks but no deformation." The spatial co-occurrence relationship, the multi-temporal co-evolution law, and the quantitative results of the engineering indicative significance of cracks are a progressive causal relationship: the spatial co-occurrence relationship is a static spatial correlation description of three types of indicators under the same time section; the multi-temporal co-evolution law is a dynamic trend of spatial co-occurrence relationship under multiple continuous time sections; and the quantitative results of the engineering indicative significance of cracks are quantitative conclusions on the contribution of cracks with different orientations and different development stages to the overall risk of the slope, based on the former two.

[0053] To achieve a quantitative characterization of the spatial co-occurrence relationship of multi-source information, three types of quantitative indicators were constructed: cracks, displacements, and surface deformations. Crack index Let the first Each monitoring point in time phase Within the corresponding local neighborhood, the total length of the transverse cracks is The neighborhood area is ,but: ; Crack indices are used to characterize the intensity of crack development within a local area; among them, the total length of transverse cracks... Based on the crack orientation classification results output by Stage 2, transverse cracks were selected from the crack 3D features and statistically analyzed. Displacement index , used to characterize the Each monitoring point in time phase The local motion response intensity; let the first... The cumulative displacement of each monitoring point is ,but: ; Surface deformation index , used to characterize the Each monitoring point in time phase The degree of deformation activity of the surface region within the neighborhood; let the time phase be... The total number of points in the 3D point cloud within the local neighborhood is The number of points whose cumulative deformation exceeds the preset significant deformation threshold is: ,but: ; Considering the different dimensions of the various indicators, the range normalization method is used to normalize the crack indicators. Displacement index and surface deformation index A uniform mapping to the [0,1] interval yields a normalized index. , and ; Based on this, a comprehensive co-occurrence index is constructed. Consistency Difference Indicators : ; ; in: Used to characterize the overall level of coordination between cracks, displacements, and surface deformations within the same local neighborhood; This value is used to characterize the consistency of the three types of information. The smaller the value, the higher the consistency; the larger the value, the clearer the dominant information source.

[0054] It should be noted that the comprehensive co-occurrence index Consistency Difference Indicators The calculation is not limited to the arithmetic mean mentioned above; it can also be based on expert experience in actual engineering scenarios or machine learning models to normalize the index. , and Different weighting factors are assigned to each component for weighted calculation.

[0055] Based on quantitative indicators, a standardized judgment rule for slope status is constructed through a multi-source fusion risk judgment construction module. The multi-source fusion risk judgment construction module uses a comprehensive co-occurrence index. Consistency Difference Indicators Based on this, combined with the DCI index and the dominant information source (the dominant information source refers to the crack index after range normalization within the same spatial assessment unit), Displacement index Surface deformation index Of the three, the index with the largest value represents the dominant driving factor in the evolution of slope diseases in that area; the determination rule is: in Under the premise of ≥0.15 (poor consistency among the three types of indicators, with significant differences), through Judgment: If The maximum value indicates that the dominant information source is crack development, corresponding to a crack-dominant type; if or The maximum value indicates that the dominant information source is displacement / area deformation, corresponding to the deformation-dominant type. The output slope state classification rules are as follows: the extracted quantitative index values ​​are input into the multi-source information co-occurrence type comprehensive judgment rule base for logical threshold comparison, and the established standardized judgment rules (multi-source information co-occurrence type comprehensive judgment rules) are shown in Table 1. Table 1 Comprehensive Judgment Rules for Co-occurrence Types of Multi-Source Information

[0056] The risk zoning and assessment implementation module replaces the traditional single threshold alarm with the above six co-occurrence types to realize the refined risk zoning of the slope surface. The specific processing process is as follows: the slope is divided into multiple spatial assessment units and the above rule matching is performed unit by unit. The obtained classification results are assigned to the corresponding geometric region of the three-dimensional point cloud model to generate a visualized three-dimensional risk zoning result.

[0057] The project management recommendation output module outputs differentiated management measures based on the 3D risk zoning results. See Table 2 for details. Table 2 Risk Zoning Results and Differentiated Control Measures

[0058] The method of this invention was used in a long-term monitoring experiment and verification for 26 weeks in a highway slope scenario to prove the feasibility of the technical solution of this invention, as detailed below: A highway slope with typical soil and rock characteristics was selected as the test area, and six binocular visual displacement monitoring points were set up on the slope. The experiment used UAV aerial photography and a binocular monitoring system as data sources, focusing on comparing data from three typical time phases: week 2, week 14, and week 26. The specific process included: First, for ultra-high resolution UAV images with a resolution of up to 6K, a strategy of "large image cropping - small image recognition - result stitching" was used for preprocessing. Then, the images were input into a finely tuned YOLO11-MTA model based on a joint public fracture dataset for inference, outputting a pixel-level segmentation mask. Next, the crack centerline was extracted, and the principal direction was calculated using PCA to complete the orientation classification. In the 3D mapping stage, combining the UAV camera intrinsic and extrinsic parameters and the continuous 3D geometric field reconstructed by NeRFusion, the intersection points were solved using singular value decomposition (SVD), projecting the crack centroid and the complete mask onto the 3D point cloud. Finally, under a unified spatial benchmark, the crack distribution, point displacement, and surface deformation in the neighborhood of each monitoring point are extracted, spatiotemporal co-evolution analysis is carried out, and the displacement-crack correlation index (DCI) is calculated simultaneously.

[0059] Experimental results: (1) Stable and reliable three-dimensional mapping: The experiment successfully extracted about 36 crack centroids and simultaneously obtained the three-dimensional crack sampling point set corresponding to each crack instance; through the line-of-sight back projection mechanism based on NeRFusion, the crack centroids and crack sampling point set were accurately mapped to the surface of the three-dimensional slope model, effectively avoiding projection distortion and completely preserving the geometric shape and topological relationship of the crack in three-dimensional space.

[0060] (2) Spatial co-occurrence analysis reveals multiple evolutionary patterns: Through tracking 6 monitoring points for 26 weeks, based on the proposed method of this invention... Index and The indicator judgment rules clearly reveal the complex co-evolution path of slope deformation. Specific experimental data are shown in Table 3 below: Table 3. Statistical results of normalized multi-source indicators and co-occurrence types at typical monitoring points

[0061] (3) Verification of displacement-crack correlation index: Taking points 2 and 6 as examples, the displacement-crack correlation index (DCI) is further calculated.

[0062] Table 4. Calculation results of DCI index for typical monitoring points

[0063] The DCI index at point 2 rapidly climbed to 0.871, indicating a high degree of coupling between displacement anomaly and crack development, verifying its "strong synergistic" determination. The DCI index at point 6 remained below 0.40 (although the displacement normalized value reached 0.73), confirming its "deformation-dominated" characteristic—displacement was increasing, but the surface crack response was severely lagging. If only crack inspection is relied upon in this type of area, it will lead to missed detections, fully demonstrating the necessity of displacement-crack bidirectional correlation analysis.

[0064] Based on the data in Tables 3 and 4, the following conclusions can be drawn: Point 2 (rapidly evolving into a strongly collaborative state): In the early stage, deformation dominated; in the middle and later stages, transverse cracks increased significantly, point displacement increased substantially, and surface deformation continued to intensify, rapidly transitioning to a strongly collaborative state. (This area) was accurately identified as the region with the highest activity.

[0065] Point 6 (typical deformation-dominated): Displacement and surface deformation increased significantly during the monitoring period, but the response of the transverse surface cracks lagged severely. The model accurately identified the point as deformation-dominated (with a DCI as high as 0.460). Further verification using the DCI index showed that the DCI at point 6 remained consistently below 0.40, further confirming its deep-seated, latent deformation characteristics from a displacement-crack correlation perspective. This verifies the engineering pain point of traditional "crack inspection" methods, which are prone to overlooking deep-seated, latent deformation risks.

[0066] Points 1, 4, and 5 (gradually co-evolutionary): They show a development from low activity to gradually enhanced co-evolution. Among them, point 4 is dominated by displacement in the later stage, and the model identifies it as deformation-dominated based on feature recognition.

[0067] Location 3 (Evolution from low activity to later stage): It remained low activity in the early stage, but by week 26 the system detected an increase in displacement response, and it was objectively classified as deformation-dominated.

[0068] Experimental results show that this invention can not only stably and accurately achieve three-dimensional spatial positioning of cracks in complex slopes, but also utilizes a "six-type" spatial co-occurrence determination model driven by displacement monitoring data and constrained by three-dimensional crack mapping. index, The indicators (including the DCI index) are fully capable of performing high-dimensional deconstruction of the "crack-displacement-area" structure. The resulting risk evolution assessment is highly consistent with the actual mechanical mechanism of the slope, indicating that this technology is feasible for large-scale application in real highway slope scenarios.

[0069] The above description is only a preferred embodiment of the present invention and does not limit the scope of protection of the present invention. All equivalent structural transformations made under the inventive concept of the present invention using the contents of the present invention specification and drawings, or direct / indirect applications in other related technical fields, are included within the scope of protection of the present invention.

Claims

1. A method for risk zoning of highway slope diseases based on three-dimensional positioning and multi-source fusion, characterized in that, The steps include the following: Step S1: Collect aerial images and 3D point cloud data of highway slopes by UAV, and obtain the displacement time series data of monitoring points through the slope surface displacement measurement system to obtain a multi-source monitoring dataset of slopes with spatial geographic information and covering the three elements of cracks, displacement and deformation. Step S2: Construct an upstream crack three-dimensional localization module based on YOLO11-MTA and NeRFusion and a downstream slope risk assessment module based on multi-source information fusion; Step S3: Based on the slope multi-source monitoring dataset obtained in Step S1, and combined with the upstream crack three-dimensional localization module based on YOLO11-MTA and NeRFusion, obtain the crack three-dimensional elements with three-dimensional spatial coordinates and geometric attributes. Step S4: Based on multi-source monitoring data of the slope and three-dimensional elements of cracks, combined with the deformation results of the surface area reconstructed by UAVs in multiple phases, the spatial co-occurrence and co-evolution analysis of the three elements of cracks, displacement and deformation is carried out through the downstream slope risk assessment module based on multi-source information fusion, and then quantitative indicators are extracted. Based on quantitative indicators, highway slope disease risk zoning is performed, and three-dimensional risk zoning results are output; specifically including: Based on multi-source monitoring data of slope, three-dimensional elements of cracks with three-dimensional spatial coordinates and geometric attributes obtained in step S3, and deformation results of surface area reconstructed by UAV in multiple phases, the three-dimensional data of cracks, displacement and deformation are fully registered under the same spatial reference through the multi-source data spatial reference unification module, and multi-source registration data and spatial neighborhood index are output. The displacement-triggered crack association submodule enables automatic tracing from displacement anomalies to crack areas of interest. Based on multi-source registration data with the same spatial reference, the spatial co-occurrence relationship between cracks and slope deformation, the multi-temporal collaborative evolution law, and the quantitative results of the engineering indication significance of cracks are output through the multi-temporal spatial co-occurrence analysis module. Based on quantitative indicators, a standardized judgment rule for slope status is constructed through a multi-source fusion risk judgment construction module. The risk zoning and assessment implementation module realizes refined risk zoning of the slope surface according to the standardized judgment rules of slope status, and generates a visualized three-dimensional risk zoning result. The project management and control suggestion output module outputs differentiated management and control measures based on the three-dimensional risk zoning results; The displacement-triggered crack correlation submodule uses the time-series displacement data of monitoring points as the driving source to achieve automatic tracing from displacement anomalies to the crack area of ​​interest; specifically, it includes: Automatic displacement inflection point detection: A sliding window is used to fit the displacement time-series curve of each monitoring point, and the change in displacement rate within the window is calculated. ;when When the preset threshold is exceeded, the monitoring point is determined to have a displacement acceleration inflection point and is marked as a triggered state; Neighborhood crack reverse query: Centered on the monitoring point in the triggered state, query the crack 3D features generated in step S3 in the spatial neighborhood index, and extract the total crack length within the neighborhood. , percentage of transverse cracks and crack increase rate ; Displacement-crack correlation quantification: Constructing the displacement-crack correlation index (DCI) to quantify the spatial coupling degree between displacement anomalies and crack development, providing supplementary criteria for subsequent risk assessment; The formula for calculating the Displacement-Crack Correlation Index (DCI) is as follows: ； in: This represents the normalized change in displacement rate. This represents the normalized total length of the neighborhood cracks. The percentage of transverse cracks; These are the weighting coefficients, with default values ​​of 0.40, 0.35, and 0.

25.

2. The method for risk zoning of highway slope diseases based on three-dimensional positioning and multi-source fusion according to claim 1, characterized in that, The YOLO11-MTA model includes an encoder, a decoder, a multi-attention enhancement module, and a detection segmentation head; The encoder uses a CSPDarknet backbone network to extract multi-scale features from the input image and generate feature maps at different levels. The decoder uses a PANet feature fusion structure to upsample and fuse multi-scale features, taking into account both the detailed features of the cracks and the global semantic information. The multi-attention enhancement module is embedded in the feature transfer node of the encoder and decoder, and is used to enhance and weight the crack features; The detection and segmentation head adopts a dual-branch structure, which is used to simultaneously output the bounding box detection results of the crack and the pixel-level segmentation mask.

3. The method for risk zoning of highway slope diseases based on three-dimensional positioning and multi-source fusion according to claim 2, characterized in that, The upstream crack three-dimensional localization module based on YOLO11-MTA and NeRFusion consists of four parts connected in series: Stage 0, Stage 1, Stage 2 and Stage 3. Stage 0 includes a series of connected image preprocessing modules and a mesh generation and sub-image generation module; Stage 1 includes a series of connected YOLO11-MTA crack segmentation modules, crack mask post-processing modules, and a panoramic stitching module for recognition results; Stage 2 includes a series of connected crack centerline extraction modules, a main direction quantitative calculation module, and a crack orientation automatic classification module; Stage 3 includes a series of connected multi-view projection geometric modeling modules, a 3D coordinate solving module, a NeRFusion geometric constraint optimization module, and a complete crack mask 3D mapping module. The downstream slope risk assessment module based on multi-source information fusion includes two connected parts: Stage 4 and Stage 5. Stage 4 includes a multi-source data spatial benchmark unification module, a displacement-triggered crack correlation sub-module, and a multi-temporal spatial co-occurrence analysis module, which are connected in series. Stage 5 includes a multi-source fusion risk criterion construction module, a risk zoning and assessment implementation module, and an engineering control suggestion output module, which are connected in series.

4. The method for risk zoning of highway slope diseases based on three-dimensional positioning and multi-source fusion according to claim 3, characterized in that, Step S3 specifically includes: The UAV aerial images are input into the image preprocessing module for preprocessing and output standardized images; the standardized images are then input into the grid division and sub-map generation module to obtain standardized sub-maps and sub-map spatial index information. The YOLO11-MTA crack segmentation module outputs a pixel-level crack segmentation binary mask within the sub-image range based on a standardized sub-image; the crack mask post-processing module preprocesses the pixel-level crack segmentation binary mask within the sub-image range to obtain a continuous crack mask for the sub-image. Based on the continuous crack mask of all subgraphs and the spatial index information of the subgraphs, the panoramic crack segmentation mask within the entire slope area is obtained through the panoramic stitching module of the recognition results. Based on a panoramic crack segmentation mask covering the entire slope area, a crack centerline extraction module outputs a single-pixel-wide, fully connected crack centerline; the main direction quantitative calculation module outputs the crack main direction angle based on the crack centerline. The automatic crack orientation classification module classifies cracks based on their principal direction angle. Output the crack orientation classification results; The camera intrinsic and extrinsic parameters of the UAV aerial imagery are input into the multi-view projection geometry modeling module to achieve the projection transformation from 2D pixel coordinates to 3D world coordinates. The crack pixel coordinates and camera intrinsic and extrinsic parameters are input into the 3D coordinate solving module to obtain the preliminary 3D spatial coordinates of the crack sampling points. Based on multiple crack sampling points within the same crack instance, the preliminary 3D spatial coordinates of the crack centroid are calculated. Based on the preliminary 3D spatial coordinates of the crack sampling points and the crack centroid, and the continuous 3D geometric field of the slope reconstructed by NeRFusion, the high-precision 3D spatial coordinates of the crack sampling points and the crack centroid are obtained through the NeRFusion geometric constraint optimization module. Based on the panoramic crack segmentation mask, the high-precision three-dimensional spatial coordinates of crack sampling points and the high-precision three-dimensional spatial coordinates of crack centroid points, the complete crack mask three-dimensional mapping module outputs crack three-dimensional elements with three-dimensional spatial coordinates and geometric attributes.

5. The method for risk zoning of highway slope diseases based on three-dimensional positioning and multi-source fusion according to claim 4, characterized in that, Based on the principal direction angle of the crack The classification rules for automated crack classification are as follows: When the principal direction angle of the crack satisfies or At that time, it was determined to be a transverse crack; when At that time, it was determined to be a vertical crack; The remaining cracks are considered diagonal cracks; Calculate the minimum angle between the diagonal crack and the horizontal or vertical direction, when or When, it is determined to be a transverse crack; when At that time, it was determined to be a vertical crack.

6. The method for risk zoning of highway slope diseases based on three-dimensional positioning and multi-source fusion according to claim 5, characterized in that, In the multi-temporal spatial co-occurrence analysis module, three types of quantitative indicators—cracks, displacement, and surface deformation—were constructed to achieve a quantitative representation of the spatial co-occurrence relationship of multi-source information. Crack index Let the first Each monitoring point in time phase Within the corresponding local neighborhood, the total length of the transverse crack is The neighborhood area is ,but: ； Displacement index Let the first The cumulative displacement of each monitoring point is ,but: ； Surface deformation index : Assume time phase The total number of points in the 3D point cloud within the local neighborhood is The number of points whose cumulative deformation exceeds the preset significant deformation threshold is: ,but: ； The crack index is normalized using the range normalization method. Displacement index and surface deformation index A uniform mapping to the [0,1] interval yields a normalized index. , and ; Constructing a comprehensive co-occurrence index Consistency Difference Indicators : ； 。 7. The method for risk zoning of highway slope diseases based on three-dimensional positioning and multi-source fusion according to claim 6, characterized in that, The multi-source fusion risk criterion construction module uses a comprehensive co-occurrence index. Consistency Difference Indicators Based on this, and combined with the DCI index and dominant information sources, standardized judgment rules for slope condition are output. The specific standardized judgment rules for slope condition are as follows: when ≤0.20 and When the value is less than 0.15, the co-occurrence type is low activity. when ≥0.70 and min( , , When the co-occurrence value is ≥0.50, the co-occurrence type is strongly cooperative. When 0.20 < <0.70 and When the value is less than 0.15, the co-occurrence type is progressively cooperative. when <0.70 and ≥0.15, and satisfy At that time, the co-occurrence type was crack-dominated; when <0.70 and ≥0.15, and satisfy At that time, the co-occurrence type is deformation-dominant; When DCI ≥ 0.65 and At that time, the co-occurrence type was displacement-crack anomaly.

8. The method for risk zoning of highway slope diseases based on three-dimensional positioning and multi-source fusion according to claim 7, characterized in that, Based on the co-occurrence type, the three-dimensional risk zoning results for highway slopes are as follows: When the co-occurrence type is low activity, the risk zoning level is low activity stable zone; When the co-occurrence type is progressive synergistic, the risk zoning level is progressive deformation zone; When the co-occurrence type is strong synergy, the risk zoning level is high-risk activity area; When the co-occurrence type is crack-dominated, the risk zoning level is an area requiring key review; When the co-occurrence type is the deformation-dominant type, the risk zoning level is the area requiring key review / the area with hidden risks; When the co-occurrence type is displacement-crack anomaly, the risk zoning level is high-risk hidden activity zone.

Images