A highway side slope crack quantification evaluation method and system based on image detection

Abstract

The application discloses a kind of highway side slope crack quantification evaluation method and system based on image detection, it is related to side slope crack evaluation technical field.The present application is carried by unmanned aerial vehicle high-resolution visible light camera, periodically collects highway side slope surface image according to setting route, constructs multi-period time series image database.Adopt instance segmentation network to carry out pixel-level identification to crack target in image, distinguish horizontal crack, longitudinal crack and mesh crack type.Combining morphological processing and skeleton extraction algorithm, the length, average width, extension direction and distribution density of crack such as quantitative parameter are automatically calculated.Based on multi-period image comparative analysis, the displacement vector of crack edge pixel point is tracked using optical flow method, and the crack propagation rate and evolution trend are quantified.The system outputs side slope crack distribution thermodynamic diagram, risk level evaluation and maintenance suggestions, and provides data support for daily inspection and disease control of highway side slope.

Description

Technical Field

[0001] This invention relates to the field of slope crack assessment technology, and in particular to a method and system for quantitative assessment of highway slope cracks based on image detection. Background Technology

[0002] Currently, slope cracks on highways are a crucial precursor to slope instability and geological disasters. The geometric parameters of slope cracks, such as length, width, area, and propagation rate, directly reflect the stress state and deformation evolution trend of the slope's soil and rock mass. Timely and accurate acquisition of crack quantitative parameters and assessment of their hazard level are of great significance for preventing slope collapses, ensuring highway operational safety, and guiding maintenance decisions. Traditional manual inspection methods suffer from drawbacks such as low efficiency, strong subjectivity, and difficulty in accessing hazardous areas. With the development of UAV aerial surveying technology and deep learning image segmentation technology, automatic detection and identification of slope cracks using UAVs equipped with high-resolution cameras has become the mainstream solution.

[0003] Existing image recognition-based methods for quantitative assessment of slope cracks typically employ pixel counting to calculate crack area, skeleton accumulation to calculate crack length, and distance transformation to calculate crack width. Pixel counting is susceptible to segmentation noise, where even a single isolated noise point can lead to an inflated area. Skeleton accumulation suffers from a jagged effect due to pixel-level discretization, resulting in a systematic overestimation of crack length. Distance transformation requires sampling the width along the normal on the skeleton, but local skeleton jitter can cause normal direction shifts, leading to unstable width measurements. Furthermore, multi-period image registration at the pixel level relies on pixel grayscale values, achieved through optical flow or sparse feature point matching. Highway slope images often exhibit significant variations due to factors such as illumination angle, shadows, vegetation growth or decay, and road surface wetness or dryness, resulting in completely different grayscale histogram distributions for the same crack in images from different periods. Traditional registration methods based on mutual information or normalized cross-correlation are prone to failure. Summary of the Invention

[0004] This invention provides a method and system for quantitative assessment of cracks in highway slopes based on image detection, so as to avoid the crack quantification being easily affected by changes in lighting and morphology, and improve the accuracy of slope quantification.

[0005] To address the aforementioned technical problems, this invention provides a method for quantitative assessment of highway slope cracks based on image detection, comprising:

[0006] A time-series image database of highway slope cracks is obtained, and each crack data period in the time-series image database is identified based on a preset instance segmentation network to obtain the binary mask and crack type of each crack in each crack data period. Based on the binary mask, the skeleton lines of each crack are extracted, and the skeleton lines are discretized into several control points along the crack extension direction. At each control point, the left and right boundary points are determined along the normal direction. A triangular mesh is constructed based on the left and right boundary points between two adjacent control points. Using the triangular mesh of the first crack data in the time-series image database as a reference, geometric matching is performed on the remaining crack data in the time-series image database to obtain the global deformation characteristics of each crack in different images. Based on the global deformation features, key deformation points are identified, and the neighborhood range is calculated based on the width and side length of the triangular mesh where the key deformation points are located and the crack type. A spherical neighborhood is constructed with the key deformation points as the center and the neighborhood range as the radius. Based on the global deformation characteristics, the local expansion strength index within each spherical neighborhood is calculated, and an assessment report on highway slope cracks is generated based on the local expansion strength index, crack type, and global deformation characteristics.

[0007] This invention constructs a triangular mesh, transforming the binary mask of cracks obtained from image segmentation into a continuous set of triangular units composed of control points, left and right boundary points, and topological connections. This eliminates the reliance on independent and noise-sensitive pixel accumulation or skeleton sampling on the geometric properties of cracks. Instead, these properties are unified within the mesh, with the crack length determined by the control point connections, the area accumulated from the triangular units, and the average width naturally derived from the ratio of area to length. These three elements constitute a geometrically self-consistent parameter system, significantly reducing the direct disturbances to quantization parameters caused by segmentation noise, jagged edges, and uneven illumination. Based on this, geometric matching is performed using the first-phase mesh as a reference. The matching process relies solely on vertex coordinates and topological relationships, completely independent of image grayscale values. Therefore, it effectively avoids inconsistencies in grayscale distribution caused by changes in illumination angle, shadow occlusion, vegetation changes, and differences in road surface wetness, ensuring the stability and accuracy of multi-phase deformation tracking. Furthermore, based on global deformation features, key deformation points are identified, and a spherical neighborhood is constructed around these points to calculate local expansion intensity indices. This allows risk assessment to move beyond the overall length or average width of the crack and dynamically focus on the most severely deformed local areas. Simultaneously, the radius of the spherical neighborhood is adaptively determined by the width of the corresponding side and the crack type, ensuring that the scale of the analysis window matches the local crack morphology. This avoids the introduction of background noise in narrow areas or the loss of deformation details in wide areas, which is a problem with fixed windows. The triangular mesh provides a geometric carrier for deformation features and topological anchors for intertemporal matching. The key deformation points and spherical neighborhood mechanism, based on this carrier, enable precise local risk assessment. These two mechanisms are progressively linked and synergistically enhance each other, ultimately enabling the entire quantitative assessment method to effectively resist interference from lighting and morphological changes, significantly improving the accuracy of crack geometric parameter measurement and hazard assessment.

[0008] Furthermore, the step of extracting the skeleton lines of each crack based on the binary mask, discretizing the skeleton lines into several control points along the crack extension direction, and determining the left and right boundary points along the normal direction at each control point includes: Extract a single-pixel-wide skeleton line from the binary mask. The skeleton line is composed of a set of continuously arranged pixels and represents the central axis of the crack. Along the extension direction of the skeleton line, the skeleton line is discretized into several control points arranged in sequence based on the curvature change of the skeleton line, and the line connecting two adjacent control points constitutes the midline broken line of the crack. For each control point, determine the normal direction of the midline broken line at the control point, extend the normal to both sides to the two edges of the binary mask, and determine the intersection of the normal with the two edges as the left boundary point and the right boundary point; wherein, the line connecting the left boundary point and the right boundary point at the same control point constitutes the width edge at that control point.

[0009] This invention extracts single-pixel-wide skeleton lines from a binary mask and discretizes these lines into ordered control points based on curvature changes. This ensures that the distribution of control points adapts to the actual bending shape of the crack, avoiding the loss of detail caused by uniform sampling. At each control point, the left and right boundary points are determined by extending along the normal direction to the mask edge. The intersection of the normal and the edge is accurately locked, forming the width edge. This allows the local width of the crack to be explicitly defined as the length of the line connecting the left and right boundary points. By segmenting the geometric edges of the mask, the stability of width measurement is significantly improved without relying on subsequent complex distance transformations or normal resampling. Especially in complex scenarios such as irregular edges and mesh-like cracks, the width at each control point can be clearly determined, avoiding measurement jumps caused by skeleton jitter or overlapping in traditional normal sampling.

[0010] Furthermore, the construction of a triangular mesh based on the left and right boundary points between two adjacent control points includes: Traverse any two adjacent control points and connect the left boundary point of the previous control point, the right boundary point of the previous control point, and the left boundary point of the next control point in sequence to form the first triangle; Connect the right boundary point of the previous control point, the left boundary point of the next control point, and the right boundary point of the next control point in sequence to form the second triangle; A triangular grid is formed based on the first and second triangles between all adjacent control points.

[0011] This invention traverses adjacent control points, connecting the left and right boundary points of the previous control point with the left boundary point of the next control point to form a first triangle, and then connecting the right boundary points of the previous control point with the left and right boundary points of the next control point to form a second triangle. This creates a continuous triangular mesh covering the entire crack region, ensuring that the width edge of each control point is explicitly preserved as an edge of a triangle. Furthermore, adjacent triangles share boundary vertices, guaranteeing the topological continuity of the mesh. This triangular mesh structure not only represents cracks as a set of triangular elements but also provides natural adjacency constraints for vertex matching between subsequent mesh phases. During cross-phase matching, the adjacency relationship of triangles can be used to verify the rationality of vertex correspondences, effectively preventing erroneous matching caused by local occlusion or crack expansion, thereby improving the reliability of multi-phase deformation tracking.

[0012] Furthermore, the time-series image database includes crack data from various periods ordered chronologically; the step of using the triangular mesh of the first-period crack data in the time-series image database as a reference to perform geometric matching on the remaining crack data in the time-series image database to obtain the global deformation features of each crack across different periods of images includes: Obtain the triangular mesh of each crack in the first phase crack data of the time-series image database, take all vertices of the triangular mesh as geometric primitives to be matched, the vertices include the left boundary point and the right boundary point corresponding to each control point, and obtain the topological connection relationship between each vertex; Based on the topological connection relationship, the crack data of each period are matched to obtain the matching vertices in the crack data of each period that correspond to the geometric primitives; For each phase of crack data, calculate the displacement vector between the geometric primitive and the matching vertex, and construct a displacement field based on the displacement vectors of all matching vertices; The area change rate and side length change rate of each triangle in the triangular mesh are calculated based on the displacement field, and a global deformation feature is formed based on the displacement vector of each matching vertex, the area change rate and the side length change rate.

[0013] This invention uses a triangular mesh of the first phase of crack data as a reference, treating all vertices as geometric primitives and recording the topological connections between them. Subsequent phases of crack data are matched while maintaining the same topological structure to obtain matching vertices corresponding to the geometric primitives. Utilizing the inherent topological constraints of the mesh, it does not rely on the grayscale distribution of the image, thus completely avoiding grayscale inconsistencies caused by lighting, shadows, and seasonal changes. The displacement vector is calculated from the matching results, and a displacement field is constructed to obtain the area change rate and side length change rate of each triangle, forming global deformation features. Since the displacement field is directly established on the mesh vertices, the deformation features correspond one-to-one with the physical expansion of the crack, avoiding the drift caused by insufficient texture or large deformation at the pixel level in traditional optical flow methods, significantly improving the quantitative accuracy of crack deformation parameters in complex environments.

[0014] Furthermore, the step of identifying deformation key points based on the global deformation features, and constructing a spherical neighborhood with the deformation key points as the center and the width side length of the triangular mesh containing the deformation key points as the radius, includes: A cumulative deformation field is constructed based on the global deformation characteristics corresponding to crack data from each period; the cumulative deformation field is calculated by accumulating the displacement vectors between adjacent crack data periods period by period. The displacement modulus, displacement gradient modulus, and triangle change rate of the geometric primitive are calculated based on the cumulative deformation field. Vertices whose displacement modulus is higher than a preset first threshold, or whose displacement gradient modulus is higher than a second threshold, or whose triangle change rate is higher than a preset third threshold are identified as deformation key points. The triangle change rate is calculated based on the area change rate and the side length change rate. For each deformation key point, determine the triangular cell to which the deformation key point belongs in the triangular mesh of the first phase crack data, and take the control point in the triangular cell that is closest to the deformation key point as the target control point; Calculate the width length between the left and right boundary points corresponding to the target control point; The radius coefficient is determined based on the crack type, and the neighborhood range is calculated based on the radius coefficient and the width side length.

[0015] This invention constructs a cumulative deformation field by progressively accumulating the displacement vectors between adjacent periods. This comprehensively reflects the overall deformation trend of cracks throughout the monitoring period, avoiding interference from local noise in a single period on the identification of key deformation points. Based on the cumulative deformation field, the displacement modulus, displacement gradient modulus, and triangle change rate obtained by combining the area change rate and the side length change rate for each grid vertex are calculated. Vertices exceeding a preset threshold are identified as key deformation points. This identification strategy does not rely on the inherent geometric features of the crack, such as endpoints or bifurcation points, but dynamically selects the locations with the most concentrated stress or the most severe deformation based on the actual degree of deformation, making the location of key points more accurate. Subsequently, with the key deformation point as the center, the neighborhood radius is determined based on the width and side length of its location and the crack type. The width and side length reflect the local structural scale, and the crack type provides morphological priors. The combination of these two factors allows the spatial range of the spherical neighborhood to adaptively match the local features of the crack, laying a spatial foundation for the subsequent refined calculation of local expansion intensity.

[0016] Furthermore, the calculation of the local expansion intensity index within each spherical neighborhood based on the global deformation features includes: For each spherical neighborhood, the triangular units that intersect with the spherical neighborhood are determined as target triangular units, and the intersection area of ​​each target triangle with the spherical neighborhood is calculated. The weighting coefficients of each target triangle are determined based on the intersection area, and the weighted change rate is calculated based on the weighting coefficients and the area change rate of each target triangle. The total intersection area of ​​the spherical neighborhood is calculated based on the intersection areas of all target triangles, and the average area change rate is calculated based on the weighted change rate and the total intersection area. The average area change rate is used as an index of local expansion intensity within the spherical neighborhood.

[0017] This invention selects intersecting triangular units as target triangles for each spherical neighborhood, and calculates the intersection area between each target triangle and the spherical neighborhood. This intersection area is used as a weighting factor to calculate a weighted sum of the area change rates, which is then divided by the total intersection area to obtain the average area change rate, which is used as an indicator of local expansion intensity. This calculation method directly utilizes the area change rate of each triangle already obtained from the global deformation features, eliminating the need for re-interpolation or constructing a continuous displacement field. It is computationally efficient and has clear physical meaning. Furthermore, using the actual intersection area as the weight ensures that triangles with higher overlap with the spherical neighborhood contribute more to the local expansion intensity, avoiding the bias caused by simple averaging. Since the average area change rate integrates the overall expansion or contraction information of multiple triangles, it has a smoothing and suppressing effect on noise at the edges of individual triangles. This improves the stability and reliability of the local expansion intensity indicator while maintaining high sensitivity, accurately reflecting the net expansion degree of cracks in the region.

[0018] Furthermore, the generation of the highway slope crack assessment report based on the local expansion strength index, crack type, and global deformation characteristics includes: The average expansion strength index is calculated based on all local expansion strength indices in the crack. Calculate the overall length propagation rate and average width change rate of each crack based on the global deformation characteristics; Based on the crack type, a risk weight coefficient is determined, and a comprehensive score is calculated based on the risk weight coefficient, the average expansion strength index, the overall length expansion rate, and the average width change rate. The crack level is then determined based on the comprehensive score. A heat map of slope crack distribution is generated based on the triangular mesh of the cracks. The heat map is then labeled based on the crack level, crack type, overall length expansion rate, average width change rate, spatial location of each spherical neighborhood and its local expansion intensity index, and an assessment report of highway slope cracks is generated.

[0019] This invention calculates the average expansion intensity index based on the local expansion intensity index of all spherical neighborhoods within a crack. Simultaneously, it extracts the overall length expansion rate and average width change rate of each crack from global deformation characteristics, achieving multi-scale deformation quantification from local to global levels. Risk weighting coefficients are determined according to crack type, and the average expansion intensity, overall length expansion rate, and average width change rate are fused to obtain a comprehensive score, which is used to classify crack levels. This assessment method not only considers the overall development trend of cracks but also highlights the comprehensive impact of local deformation through the average expansion intensity index, avoiding the potential for a single global parameter to mask high-risk local areas. Finally, a heat map of slope crack distribution is generated based on a triangular mesh, and the heat map is annotated with the comprehensive level, type, expansion rate, location of each spherical neighborhood, and expansion intensity, making the assessment report comprehensive and clearly structured, providing intuitive and accurate data support for maintenance decisions.

[0020] Furthermore, the acquisition of the time-series image database of highway slope cracks includes: Based on the visible light camera mounted on the drone, and according to the preset sampling period and fixed route, the highway slope is regularly inspected and mapped to obtain image data for each period. Motion restoration of the three-dimensional structure was performed on the image data of each period to generate orthophotos and digital surface models of the slope in each period; The orthophoto data is sorted according to time order to generate a time-series image database; the time-series image database contains several periods of crack data, and the crack data includes orthophotos, digital surface models and geographic coordinate information.

[0021] This invention utilizes a UAV equipped with a visible light camera to periodically inspect and map images along a fixed route and with a preset sampling cycle, ensuring the consistency and repeatability of image acquisition and providing a standardized data source for multi-period time-series analysis. Motion-reconstructed 3D reconstruction of the structure is performed on each period's images, generating orthophotos and digital surface models, and uniformly assigning geographic coordinate information. This converts the pixel coordinates in the original images into geographic coordinates with a true spatial scale, eliminating geometric distortions caused by differences in shooting posture and altitude between different periods. This allows the vertex coordinates of subsequent triangular meshes to be accurately mapped to the same spatial reference plane. Organized into a time-series image database based on temporal order, this ensures that each period's crack data includes orthophotos, digital surface models, and geographic coordinates, providing a high-precision, highly consistent data foundation for subsequent inter-period geometric matching and deformation analysis, fundamentally reducing quantization errors introduced by irregularities in the data acquisition process.

[0022] Furthermore, the preset instance segmentation network identifies each crack data point in the time-series image database to obtain the binary mask and crack type of each crack in each crack data point, including: For each period of crack data in the time-series image database, the crack data is preprocessed to generate a standardized image; The standardized image is input into a preset instance segmentation network to obtain the binary mask of the crack and the crack type; the crack type includes transverse cracks, longitudinal cracks and mesh cracks.

[0023] This invention preprocesses crack data from each period of a time-series image database to generate standardized images, effectively mitigating the differences in image grayscale under varying weather and lighting conditions. This results in more consistent image quality input to the instance segmentation network, thereby improving the robustness of the network segmentation. The standardized images are then input into a pre-defined instance segmentation network, which directly outputs binary masks and type labels for the cracks, achieving an end-to-end conversion from raw images to structured crack information. The crack type labels provide crucial information for the subsequent adaptive calculation of the spherical neighborhood radius and weight allocation in risk assessment. This process seamlessly integrates deep learning segmentation results with geometric modeling, ensuring accurate and reliable input for all subsequent operations based on binary masks and crack types. This comprehensively enhances the accuracy and applicability of the entire quantitative assessment method in complex environments.

[0024] Secondly, the present invention provides a quantitative assessment system for highway slope cracks based on image detection, comprising: a crack identification module, a mesh construction module, a feature extraction module, a neighborhood construction module, and an assessment module; The crack identification module is used to acquire a time-series image database of highway slope cracks, and to identify each period of crack data in the time-series image database based on a preset instance segmentation network, so as to obtain the binary mask and crack type of each crack in each period of crack data. The mesh construction module is used to extract the skeleton lines of each crack based on the binary mask, discretize the skeleton lines into several control points along the crack extension direction, determine the left and right boundary points along the normal direction at each control point, and construct a triangular mesh based on the left and right boundary points between two adjacent control points. The feature extraction module is used to perform geometric matching on the remaining crack data in the time-series image database, using the triangular mesh of the first crack data in the time-series image database as a reference, to obtain the global deformation features of each crack in different images. The neighborhood construction module is used to identify deformation key points based on the global deformation features, calculate the neighborhood range based on the width and side length of the triangular mesh where the deformation key points are located and the crack type, and construct a spherical neighborhood with the deformation key points as the center and the neighborhood range as the radius. The evaluation module is used to calculate the local expansion strength index within each spherical neighborhood based on the global deformation characteristics, and to generate an evaluation report on highway slope cracks based on the local expansion strength index, crack type, and global deformation characteristics. Attached Figure Description

[0025] Figure 1 A flowchart illustrating a method for quantitative assessment of highway slope cracks based on image detection, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of a highway slope crack quantitative assessment system based on image detection, provided as an embodiment of the present invention. Detailed Implementation

[0026] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0027] The terms "first" and "second," etc., in the specification, claims, and drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0028] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0029] Example 1 See Figure 1 , Figure 1 This is a flowchart illustrating a method for quantitatively assessing highway slope cracks based on image detection, provided in an embodiment of the present invention. The embodiment of the present invention provides a method for quantitatively assessing highway slope cracks based on image detection, including steps 101 to 105, as detailed below: Step 101: Obtain a time-series image database of highway slope cracks, and identify each crack data period in the time-series image database based on a preset instance segmentation network to obtain the binary mask and crack type of each crack in each crack data period. Step 102: Extract the skeleton lines of each crack based on the binary mask, discretize the skeleton lines into several control points along the crack extension direction, determine the left and right boundary points at each control point along the normal direction, and construct a triangular mesh based on the left and right boundary points between two adjacent control points. Step 103: Using the triangular mesh of the first crack data in the time-series image database as a reference, perform geometric matching on the remaining crack data in the time-series image database to obtain the global deformation features of each crack in different images. Step 104: Identify deformation key points based on the global deformation features, calculate the neighborhood range based on the width side length and crack type of the triangular mesh where the deformation key points are located, and construct a spherical neighborhood with the deformation key points as the center and the neighborhood range as the radius. Step 105: Calculate the local expansion strength index within each spherical neighborhood based on the global deformation characteristics, and generate an assessment report of highway slope cracks based on the local expansion strength index, crack type, and global deformation characteristics.

[0030] In this embodiment, a UAV equipped with a visible light camera is used to periodically collect images of highway slopes along a fixed route. Motion recovery structure 3D reconstruction is performed on the images from each period to generate orthophotos and digital surface models. After georegistration, a time-series image database containing geographic coordinates is constructed in chronological order.

[0031] In this embodiment, for each image, a deep learning-based instance segmentation network, such as a mask region convolutional neural network, is used for processing to output a binary mask for each crack and a crack type label. For each crack's binary mask, a single-pixel-wide skeleton line is first extracted using a morphological thinning algorithm. Then, based on the curvature change of the skeleton line, it is discretized into a series of ordered control points. The line connecting adjacent control points forms the crack's midline broken line. At each control point, the tangent direction of the midline broken line is calculated and rotated 90 degrees to obtain the normal direction. This normal is extended to both sides to the edge of the binary mask, and the two intersection points are determined as the left and right boundary points, respectively. The line connecting the left and right boundary points at the same control point is the width edge, representing the local width of the crack at that location.

[0032] Subsequently, for each pair of adjacent control points, the left and right boundary points of the previous control point are connected to the left boundary point of the next control point to form the first triangle. Then, the right boundary point of the previous control point is connected to the left and right boundary points of the next control point to form the second triangle. The two triangles corresponding to all adjacent control points together form a continuous triangular mesh covering the entire crack. Using the triangular mesh of the crack data in the first phase as a reference, all its vertices are used as geometric primitives and the topological connection relationship between vertices is recorded. For the same crack in subsequent phase images, geometric matching based on topological constraints is performed to obtain the matching vertices of the first phase mesh in subsequent phases. The displacement vector of each vertex is calculated and a displacement field is constructed. Then, the area change rate and side length change rate of each triangle are calculated based on the displacement field. The vertex displacement vector, area change rate, and side length change rate are used together as global deformation features.

[0033] In this embodiment, a cumulative deformation field is constructed by accumulating displacement vectors from each period. The displacement modulus, displacement gradient modulus, and triangle change rate (derived from area change rate and side length change rate) of each vertex are calculated throughout the monitoring period. Vertices whose displacement modulus, displacement gradient modulus, or triangle change rate exceeds a preset threshold are identified as deformation key points. For each deformation key point, the nearest control point is found in the triangular cell of its corresponding first-period triangular mesh. The width side length corresponding to the control point is obtained, and the radius coefficient is determined according to the type of crack. The width side length is multiplied by the radius coefficient to obtain the radius of the spherical neighborhood. A spherical neighborhood is constructed with the deformation key point as the center. Then, for each spherical neighborhood, the first-period triangular cells that intersect with it are selected as target triangles. The intersection area of ​​each target triangle and the spherical neighborhood is calculated. The area change rate of each target triangle is weighted and averaged using the intersection area as the weight to obtain the average area change rate as the local expansion intensity index of the spherical neighborhood.

[0034] Finally, by combining the local expansion intensity index of all spherical neighborhoods in each crack, the crack type, and the overall length expansion rate and average width change rate extracted from the global deformation characteristics, a comprehensive score is calculated through weighted fusion. Based on the comprehensive score, the crack risk level is classified, and an assessment report is generated that includes the crack risk level, the spatial location and expansion intensity of each spherical neighborhood, the overall length expansion rate, the average width change rate, and a heat map of slope crack distribution marked with high-risk areas.

[0035] In this embodiment, by constructing a triangular mesh, the binary mask of the cracks obtained from image segmentation is transformed into a continuous set of triangular units composed of control points, left and right boundary points, and topological connections. This makes the geometric properties of the cracks no longer dependent on independent and noise-sensitive pixel accumulation or skeleton sampling, but unified within the mesh. The length is determined by the connection of control points, the area is accumulated from the triangular units, and the average width is naturally derived from the ratio of area to length. These three constitute a geometrically self-consistent parameter system, thereby significantly reducing the direct disturbance of segmentation noise, jagged edges, and uneven illumination on the quantization parameters. Based on this, geometric matching is performed using the first-phase mesh as a reference. The matching process relies solely on vertex coordinates and topological relationships, completely independent of image grayscale values. Therefore, it can completely avoid the problem of inconsistent grayscale distribution caused by changes in illumination angle, shadow occlusion, vegetation replacement, and differences in road surface dryness and wetness, ensuring the stability and accuracy of multi-phase deformation tracking. Furthermore, based on global deformation features, key deformation points are identified, and a spherical neighborhood is constructed around these points to calculate local expansion intensity indices. This allows risk assessment to move beyond the overall length or average width of the crack and dynamically focus on the most severely deformed local areas. Simultaneously, the radius of the spherical neighborhood is adaptively determined by the width of the corresponding side and the crack type, ensuring that the scale of the analysis window matches the local crack morphology. This avoids the introduction of background noise in narrow areas or the loss of deformation details in wide areas, which is a problem with fixed windows. The triangular mesh provides a geometric carrier for deformation features and topological anchors for intertemporal matching. The key deformation points and spherical neighborhood mechanism, based on this carrier, enable precise local risk assessment. These two mechanisms are progressively linked and synergistically enhance each other, ultimately enabling the entire quantitative assessment method to effectively resist interference from lighting and morphological changes, significantly improving the accuracy of crack geometric parameter measurement and hazard assessment.

[0036] In this embodiment, the acquisition of the time-series image database of highway slope cracks includes: Based on the visible light camera mounted on the drone, and according to the preset sampling period and fixed route, the highway slope is regularly inspected and mapped to obtain image data for each period. Motion restoration of the three-dimensional structure was performed on the image data of each period to generate orthophotos and digital surface models of the slope in each period; The orthophoto data is sorted according to time order to generate a time-series image database; the time-series image database contains several periods of crack data, and the crack data includes orthophotos, digital surface models and geographic coordinate information.

[0037] In this embodiment, the acquisition of a time-series image database of highway slope cracks is first achieved using a UAV equipped with a high-resolution visible light camera. The UAV conducts regular inspections and mapping of the highway slope along a pre-set sampling period, such as monthly or quarterly, following a predetermined fixed flight path. During each inspection, the UAV automatically executes the flight path, acquiring continuous images covering the entire slope area and recording metadata such as the shooting position, attitude angle, and GPS coordinates for each image. During the acquisition process, both the forward overlap and lateral overlap are set to be no less than 70% to ensure the accuracy of subsequent 3D reconstruction.

[0038] In this embodiment, motion-reconstructed structural 3D reconstruction is performed on each acquired image data period. The process begins by extracting feature points from each image, establishing correspondences between corresponding points in adjacent images through feature matching, and then calculating the precise exterior orientation elements and sparse point cloud for each image based on these correspondences and camera parameters. Further, based on the sparse point cloud and image data, a multi-view stereo matching algorithm is used to densely match all pixels, generating a dense 3D point cloud. Finally, the dense point cloud is projected onto a horizontal plane and interpolated to generate an orthophoto of the slope for that period. Simultaneously, a corresponding digital surface model is generated based on the point cloud height information. Each image data period, after reconstruction, yields an orthophoto with geographic coordinate information and a digital surface model reflecting the surface undulations.

[0039] In this embodiment, the orthophotos and digital surface models generated in each period are sorted chronologically, with each period serving as a data node, forming a time-series image database. Each period's crack data in this database includes not only orthophotos and digital surface models, but also the image's geographic coordinates, acquisition timestamps, and corresponding UAV pose parameters. Subsequent crack identification and geometric analysis are based on the data from each period in this database, thus ensuring a unified spatial reference and continuous temporal series across multiple analysis periods.

[0040] In this embodiment, by utilizing a UAV equipped with a visible light camera to periodically inspect and map along a fixed route and with a preset sampling cycle, the consistency and repeatability of image acquisition are ensured, providing a standardized data source for multi-period time-series analysis. Motion-reconstructed 3D reconstruction of the structure is performed on each period's imagery, generating orthophotos and digital surface models, and uniformly assigning geographic coordinate information. This converts the pixel coordinates in the original images into geographic coordinates with a true spatial scale, eliminating geometric distortions caused by differences in shooting posture and altitude between different periods. This allows the vertex coordinates of subsequent triangular meshes to be accurately mapped to the same spatial reference plane. A time-series image database is organized based on temporal order, ensuring that each period's crack data includes orthophotos, digital surface models, and geographic coordinates. This provides a high-precision, highly consistent data foundation for subsequent inter-period geometric matching and deformation analysis, fundamentally reducing quantization errors introduced by irregularities in the data acquisition process.

[0041] In this embodiment, the preset instance segmentation network identifies each period of crack data in the time-series image database to obtain the binary mask and crack type of each crack in each period of crack data, including: For each period of crack data in the time-series image database, the crack data is preprocessed to generate a standardized image; The standardized image is input into a preset instance segmentation network to obtain the binary mask of the crack and the crack type; the crack type includes transverse cracks, longitudinal cracks and mesh cracks.

[0042] In this embodiment, when performing instance segmentation and identification on each period of crack data in the time-series image database, the orthophoto obtained in the current period is first preprocessed to generate a standardized image. The preprocessing steps include: adjusting the grayscale distribution of the image using histogram equalization to reduce overall brightness differences caused by shooting under different weather or lighting conditions; applying denoising filters such as Gaussian filtering or median filtering to eliminate Gaussian noise or salt-and-pepper noise that may be introduced during image acquisition; and using deconvolution algorithms for deblurring if the image has slight motion blur. After the above preprocessing, the image is converted into a standardized image with uniform size, consistent grayscale range, and suppressed noise, facilitating stable processing by the deep learning network.

[0043] In this embodiment, the generated standardized image is input into a pre-defined instance segmentation network. This instance segmentation network adopts a mask-based convolutional neural network architecture, with its backbone network being a feature pyramid network, capable of simultaneously extracting multi-scale features. The network is pre-trained, and the training data includes a large number of slope image samples labeled with pixel-level contours and crack types. After the standardized image is computed through forward propagation, the network outputs two types of results: first, a binary mask of the crack, i.e., determining whether each pixel in the image belongs to a crack, outputting a binary image of the same size as the original image, where foreground pixels have a value of 1, representing crack areas, and background pixels have a value of 0, representing non-crack areas; second, a crack type label, where the network outputs the probability distribution of whether the crack belongs to a transverse crack, longitudinal crack, or mesh crack through a classification branch, and takes the category corresponding to the highest probability as the crack type. For multiple cracks that may exist in the same image, the instance segmentation network can distinguish different crack instances, outputting an independent binary mask and type label for each crack. The obtained binary mask and crack type serve as the basic input data for subsequent triangular mesh construction and deformation analysis.

[0044] In this embodiment, by preprocessing each period of crack data in the time-series image database to generate standardized images, the differences in image grayscale under different weather and lighting conditions are effectively reduced, making the image quality input to the instance segmentation network more consistent, thereby improving the robustness of the network segmentation. The standardized images are input into the preset instance segmentation network, which directly outputs the binary mask and type label of the cracks, realizing an end-to-end conversion from raw images to structured crack information. The crack type label provides a crucial basis for the subsequent adaptive calculation of the spherical neighborhood radius and the weight allocation in risk assessment. This process seamlessly integrates deep learning segmentation results with geometric modeling, ensuring that all subsequent operations based on binary masks and crack types receive accurate and reliable input, thus comprehensively improving the accuracy and applicability of the entire quantitative assessment method in complex environments.

[0045] In this embodiment, the step of extracting the skeleton lines of each crack based on the binary mask, discretizing the skeleton lines into several control points along the crack extension direction, and determining the left and right boundary points along the normal direction at each control point includes: Extract a single-pixel-wide skeleton line from the binary mask. The skeleton line is composed of a set of continuously arranged pixels and represents the central axis of the crack. Along the extension direction of the skeleton line, the skeleton line is discretized into several control points arranged in sequence based on the curvature change of the skeleton line, and the line connecting two adjacent control points constitutes the midline broken line of the crack. For each control point, determine the normal direction of the midline broken line at the control point, extend the normal to both sides to the two edges of the binary mask, and determine the intersection of the normal with the two edges as the left boundary point and the right boundary point; wherein, the line connecting the left boundary point and the right boundary point at the same control point constitutes the width edge at that control point.

[0046] In this embodiment, the binary mask is first processed using a morphological thinning algorithm, such as iteratively stripping boundary pixels until a skeleton line with a width of one pixel is obtained. This skeleton line consists of a series of continuously arranged pixels, representing the central axis of the crack. Then, along the extension direction of the skeleton line, the curvature value at each skeleton point is calculated. Control points are densely selected at locations with large curvature changes and sparsely selected at locations with gentle curvature, thereby discretizing the skeleton line into several sequentially arranged control points. The lines connecting adjacent control points form the midline broken line of the crack.

[0047] In this embodiment, the morphological thinning algorithm is the Zhang Sen parallel thinning algorithm.

[0048] In this embodiment, when discretizing the skeleton line into several sequentially arranged control points based on the curvature change of the skeleton line, the curvature at each pixel on the skeleton line is first calculated. The curvature is calculated as the reciprocal of the radius of the circumcircle determined by three consecutive points. A curvature threshold T is set. When the curvature of a point exceeds T, the point is marked as a candidate control point. Then, the candidate control points are filtered according to a preset minimum spacing D, retaining the points with local curvature maxima as the final control points, and ensuring that the arc length of the curve between adjacent control points does not exceed the maximum allowable arc length L (L is set to 10 cm). For straight lines with gentle curvature, control points are evenly inserted according to the maximum arc length L to ensure the geometric accuracy of discretization.

[0049] In this embodiment, D is set to the number of pixels corresponding to a real distance of 5 centimeters.

[0050] In this embodiment, for each control point, the tangent direction of the midline broken line at that point is first determined based on the line connecting adjacent control points. Then, the tangent direction is rotated 90 degrees to obtain the normal direction. The search proceeds pixel-by-pixel along the normal direction until the edge of the binary mask is reached. The pixels where the normal first intersects the two edges are designated as the left and right boundary points, respectively. The line connecting the left and right boundary points at the same control point is the width edge at that control point, and its length directly represents the local width of the crack at that section. Through this method, the left and right boundary points and the width edge corresponding to each control point can be obtained, providing accurate geometric primitives for the subsequent construction of the triangular mesh.

[0051] In this embodiment, by extracting single-pixel-wide skeleton lines from a binary mask and discretizing them into ordered control points based on curvature changes, the distribution of control points is ensured to adapt to the actual bending shape of the crack, avoiding the loss of details caused by uniform sampling. At each control point, the left and right boundary points are determined by extending along the normal direction to the mask edge, accurately locking the intersection of the normal and the edge, thus forming the width edge. This allows the local width of the crack to be explicitly defined as the length of the line connecting the left and right boundary points. By segmenting the geometric edges of the mask, the stability of width measurement is significantly improved without relying on subsequent complex distance transformations or normal resampling. Especially in complex scenarios such as irregular edges and mesh-like cracks, the width at each control point can be clearly determined, avoiding measurement jumps caused by skeleton jitter or overlapping in traditional normal sampling.

[0052] In this embodiment, the construction of a triangular mesh based on the left and right boundary points between two adjacent control points includes: Traverse any two adjacent control points and connect the left boundary point of the previous control point, the right boundary point of the previous control point, and the left boundary point of the next control point in sequence to form the first triangle; Connect the right boundary point of the previous control point, the left boundary point of the next control point, and the right boundary point of the next control point in sequence to form the second triangle; A triangular grid is formed based on the first and second triangles between all adjacent control points.

[0053] In this embodiment, the three vertices of the triangle originate from the left and right boundary points corresponding to any two adjacent control points. Specifically, for any two adjacent control points... and The corresponding left and right boundary points are respectively and To maintain continuous coverage of the crack and ensure the width of the edge... To become one side of a triangle, the following two triangles are typically constructed: First triangle: The sides of the triangle That is, control point The local width of the edge at that location.

[0054] Second triangle: The sides of the triangle That is, control point The local width of the edge at that location.

[0055] In this embodiment, the entire crack region is discretized into a series of adjacent triangles, forming triangular strips, i.e., a triangular mesh. The width edge at each control point... Each is explicitly stored as a side of a triangle, while the control points... It does not directly function as a mesh vertex; its position is determined by... and The midpoint is implicitly represented.

[0056] In practical applications, highway slope images often exhibit significant variations due to factors such as lighting angle, shadows, vegetation growth / withering, and road surface moisture levels. This results in completely different grayscale histogram distributions for the same crack in images from different periods, making traditional registration methods based on mutual information or normalized cross-correlation prone to failure. However, in this embodiment, the triangular mesh provides a crack length derived from the sum of control point connections, an area derived from the sum of all triangular faces, and an average width derived from the area divided by the length. These three elements are embedded within the same geometric model, avoiding the logical contradictions that easily arise when traditional pixel counting, skeleton accumulation, and distance transformation are calculated independently. This significantly reduces the interference of segmentation noise and jagged edges on the quantization parameters.

[0057] Meanwhile, the triangular mesh not only contains vertex coordinates but also records the connections between vertices. This topology provides strong constraints for geometric matching of multi-phase images. During matching, it is necessary not only to look at the position of a single vertex but also to verify whether the edges, included angles, and triangle areas between adjacent vertices are consistent. This effectively prevents erroneous matching caused by local occlusion, crack propagation, or changes in illumination, making deformation tracking more robust.

[0058] Since a crack typically has only a few to a dozen keypoints, it can only reflect the local structure of the crack, losing a significant amount of information about the crack's outline, width variations, and continuous curvature. In traditional keypoint-based matching of image data from different periods, the matching relationship for the entire crack can easily break down when keypoints are missed or shifted due to segmentation noise or illumination changes. Furthermore, keypoint matching is essentially a point-to-point correspondence, without considering the spatial relationships between points or the overall topology. For example, two endpoints from different periods may be far apart due to local expansion, making it difficult to determine whether they belong to the same crack based solely on point location, easily leading to mismatches. In addition, when the crack width increases, branches emerge, or local curvature changes, the position and number of keypoints may change. Direct matching methods struggle to adapt to these topological changes, often requiring manual intervention or complex heuristic rules. In contrast, triangular mesh-based matching structures cracks as a whole. It not only utilizes all mesh vertices but also enforces local geometric consistency through topological connectivity. Even if a vertex shifts due to noise, the constraints of its neighboring vertices help correct the matching result; when new branches appear in the crack, the topological changes in the mesh are systematically captured. Therefore, constructing a triangular mesh is not simply "drawing cracks as many triangles," but rather establishing a unified, measurable, and interference-resistant geometric surrogate model for the quantitative assessment and intertemporal tracking of cracks.

[0059] In this embodiment, the triangular mesh constructed in the above manner completely digitizes the geometry and topological relationships of the cracks. The area, side length, and normal direction of each triangular unit can be directly calculated, while adjacent triangles share vertices and edges, forming a continuous, non-overlapping discrete representation of the crack region. Compared with traditional pixel-based or skeleton-line-based models, the triangular mesh allows the length, area, and average width of the crack to be naturally derived from the geometric properties of the mesh. The three satisfy the self-consistent relationship that the area equals the length multiplied by the average width, fundamentally eliminating the logical contradiction between pixel accumulation and skeleton sampling. Furthermore, the topological connectivity of the mesh provides rigid constraints for inter-period matching, enabling the matching algorithm to not only consider the position of individual vertices but also force the displacement vectors of adjacent vertices to remain smooth, greatly reducing the false matching rate. At the same time, the width edge at each control point is explicitly stored and can be directly used for the calculation of local expansion intensity without the need for repeated distance transformation or edge detection.

[0060] In this embodiment, by traversing adjacent control points, the left and right boundary points of the previous control point are connected to the left boundary point of the next control point to form a first triangle. Then, the right boundary point of the previous control point and the left and right boundary points of the next control point are connected to form a second triangle, thus forming a continuous triangular mesh covering the entire crack region. This ensures that the width edge at each control point is explicitly preserved as an edge of the triangle, and adjacent triangles share boundary vertices, guaranteeing the topological continuity of the mesh. This triangular mesh structure not only represents the crack as a set of triangular elements but also provides natural adjacency constraints for vertex matching between subsequent meshes. This allows the adjacency relationship of triangles to be used to verify the rationality of vertex correspondence during cross-period matching, effectively preventing erroneous matching caused by local occlusion or crack expansion, thereby improving the reliability of multi-period deformation tracking.

[0061] In this embodiment, the time-series image database includes crack data from various periods ordered chronologically; the step of using the triangular mesh of the first crack data in the time-series image database as a reference to perform geometric matching on the remaining crack data in the time-series image database to obtain the global deformation features of each crack across different periods of images includes: Obtain the triangular mesh of each crack in the first phase crack data of the time-series image database, take all vertices of the triangular mesh as geometric primitives to be matched, the vertices include the left boundary point and the right boundary point corresponding to each control point, and obtain the topological connection relationship between each vertex; Based on the topological connection relationship, the crack data of each period are matched to obtain the matching vertices in the crack data of each period that correspond to the geometric primitives; For each phase of crack data, calculate the displacement vector between the geometric primitive and the matching vertex, and construct a displacement field based on the displacement vectors of all matching vertices; The area change rate and side length change rate of each triangle in the triangular mesh are calculated based on the displacement field, and a global deformation feature is formed based on the displacement vector of each matching vertex, the area change rate and the side length change rate.

[0062] In this embodiment, the time-series image database contains multiple crack data sets arranged chronologically, with each set containing a triangular mesh constructed using the aforementioned method. To achieve cross-period deformation tracking, the triangular mesh of each crack in the first set of crack data is first read from the time-series image database. All vertices of this triangular mesh are used as geometric primitives to be matched, including the left and right boundary points corresponding to each control point. Simultaneously, the topological connections between the vertices are recorded, i.e., which three vertices constitute each triangle, which vertices are adjacent via edges, and the adjacency relationships between triangles.

[0063] In this embodiment, a geometric matching algorithm based on topological constraints is used for each subsequent crack data phase. This matching process uses the topological connectivity of the first phase mesh as a constraint condition, and searches for the corresponding point in the subsequent phase mesh that best matches each vertex of the first phase.

[0064] In this embodiment, during matching, the georegistration parameters of each image period are first used to transform the vertex coordinates of subsequent grids to the unified spatial reference plane of the first grid. Then, using the vertex coordinates of the first grid as a reference, the point with the closest Euclidean distance in the vertex set of subsequent grids is found as the initial match. Next, through an iterative nearest-point algorithm or graph-based optimization, the matching result is optimized while maintaining consistent topological connectivity, ensuring smooth displacement vectors between adjacent vertices without disrupting the grid's topological structure. After matching is completed, the matching vertices corresponding to the geometric primitives of the first period in each subsequent crack data period are obtained.

[0065] In this embodiment, it is assumed that the orthophoto plane coordinate system of the first phase image and the affine transformation parameters or projection transformation matrices of subsequent phase images relative to this coordinate system have been obtained through motion-reconstructed 3D reconstruction and georegistration. For a subsequent phase of crack data, the coordinates of all vertices of the triangular mesh for that phase are first read. These coordinates were originally located in the pixel coordinate system of the image itself for that phase. Then, according to the georegistration parameters of that phase image, the pixel coordinates of each vertex are multiplied by the corresponding affine transformation matrix to convert them into coordinates on the unified spatial reference plane where the first phase image is located. This transformation eliminates geometric distortions caused by changes in UAV flight attitude, altitude, and terrain undulations, allowing the mesh vertices of the same physical crack in different phase images to fall in similar spatial locations.

[0066] In this embodiment, after coordinate unification, the triangular mesh of the first-stage crack is used as a reference template, and all its vertices are taken as the target point set. For each vertex in the subsequent-stage mesh, the Euclidean distance between it and each vertex of the first stage is calculated on the unified spatial reference plane, and the first-stage vertex with the closest distance is taken as the initial matching point of the subsequent-stage vertex. For example, if a left boundary point in the subsequent-stage mesh is located at a point in space after coordinate transformation, and the vertex with the closest distance to it in the first-stage mesh happens to be the left boundary point of the corresponding position of the same crack, then the two are initially considered to be matched.

[0067] In this embodiment, relying solely on the initial matching established by the nearest point may contain a few errors, such as mismatching a left boundary point to a neighboring right boundary point or adjacent control point. To correct these errors and maintain the consistency of the mesh topology, an iterative nearest-point algorithm is introduced, and topological constraints are imposed. Specifically, the topological connectivity of the first-phase mesh is represented as an undirected graph, where nodes are vertices and edges are triangle edges. Similarly, subsequent-phase meshes have the same topological structure. During the iteration process, for each subsequent-phase vertex, not only its distance to the first-phase vertex is considered, but also the matching relationships already established with its neighboring vertices in its neighborhood. By minimizing the matching error function of the entire mesh, which includes vertex distance error and displacement vector smoothness error between adjacent vertices, gradient descent or Gauss-Newton's method is used to iteratively optimize the matching results. After several iterations, all subsequent-phase vertices find a unique corresponding first-phase vertex, and the matching relationships of the three vertices within the same triangle remain continuous, with no abrupt changes or unreasonable intersections in the displacement vectors between adjacent vertices.

[0068] In this embodiment, after matching is completed, matching vertices corresponding one-to-one with the geometric primitives of the first phase are obtained in each subsequent crack data phase. For example, the right boundary point numbered 10 in the first phase mesh can be accurately matched with a corresponding point in the subsequent second phase mesh, which is also the right boundary point of the crack at the same physical location. This correspondence lays the foundation for subsequent calculation of displacement vectors, construction of displacement fields, and calculation of the rate of change of triangle area and side length.

[0069] In this embodiment, if new vertices not included in the first-phase mesh topology appear in subsequent crack data, such as due to crack branch extension or the generation of new cracks, then during the matching process, candidate vertices that cannot establish a correspondence with vertices in the first-phase mesh data are marked as "new vertices." These new vertices do not participate in displacement vector calculations but are retained in the mesh for subsequent risk assessment. Conversely, if a vertex in the first phase cannot find a reasonable corresponding point in the subsequent mesh, such as due to the partial disappearance or occlusion of cracks, then the displacement vector of that vertex is marked as missing, and this outlier is excluded when calculating the overall deformation characteristics. Through this fault-tolerant mechanism, the natural evolution of cracks can be adapted, and matching will not fail due to topology changes.

[0070] In this embodiment, for each phase of crack data, the displacement vector between each vertex in the first phase and its matching vertex is calculated. The displacement vectors of all vertices together constitute the displacement field of the crack in this phase relative to the crack in the first phase. Each component of this displacement field records the spatial distance and direction of the corresponding mesh vertex from the first phase to this phase.

[0071] In this embodiment, it is assumed that a crack exists in the triangular mesh constructed in the first phase of the image, which contains multiple vertices. To illustrate the specific calculation process, one triangle is used as an example. The three vertices of this triangle are the left boundary point. Right boundary point and the left boundary point corresponding to the second control point After the aforementioned geometric matching, the corresponding matching points of these vertices were determined in the second phase of imagery, and are denoted as follows: , and On a unified spatial reference plane, The coordinates are A. If the coordinates of the vertex are B, then the vertex... The displacement vector is obtained by subtracting A from B, and the length of this displacement vector represents... The Euclidean distance from the first phase to the second phase is in the direction of... point to The unit vector. Similarly, calculate respectively. and The displacement vectors of these three vertices together constitute the displacement field of the triangular region.

[0072] In this embodiment, the displacement vectors of all vertices are organized into a vector list according to vertex numbers. Each vector contains a horizontal component and a vertical component, thus constituting the displacement field of the entire crack. For example, if the horizontal component of the displacement vector of a left boundary point is positive and the vertical component is close to zero, it indicates that the point mainly moves horizontally outward from the crack, corresponding to an increase in the width of the crack at that section. If the displacement vector of a control point is along the crack extension direction and has a large length, it indicates that the crack has significantly extended in that segment. This displacement field records the cumulative motion information of each local location of the crack from the reference period to the current period, providing a direct geometric basis for subsequent calculations of the area change rate and side length change rate of the triangle.

[0073] In practical applications, a crack may contain dozens or even hundreds of vertices, and the displacement vectors of all the vertices together constitute the complete displacement field of the crack within the corresponding time interval.

[0074] In this embodiment, the area change rate and side length change rate of each triangle in the triangular mesh are calculated based on the displacement field. Specifically, for each triangle, the lengths of the three sides and the area of ​​the triangle after deformation are calculated based on the coordinates of its three vertices in the first phase and the matching coordinates in subsequent phases. These are then compared with the original side lengths and areas in the first phase to obtain the side length change rate and area change rate. The displacement vectors of each matching vertex, the area change rate of each triangle, and the side length change rate are collectively organized into the global deformation feature of the crack. This global deformation feature includes both discrete displacement information at the vertex level and expansion and contraction deformation information at the triangle surface level, providing a complete data foundation for subsequent identification of key deformation points and calculation of local expansion intensity.

[0075] In this embodiment, using the triangular mesh of the first-phase crack data as a reference, all vertices are treated as geometric primitives, and the topological connections between vertices are recorded. Subsequent crack data are matched while maintaining the same topological structure to obtain matching vertices corresponding to the geometric primitives. Utilizing the inherent topological constraints of the mesh, and independent of the image's grayscale distribution, the problem of grayscale inconsistencies caused by lighting, shadows, and seasonal changes can be completely avoided. The displacement vector is calculated based on the matching results, and a displacement field is constructed to obtain the area change rate and side length change rate of each triangle, forming global deformation features. Since the displacement field is directly established on the mesh vertices, the deformation features correspond one-to-one with the physical expansion of the cracks, avoiding the drift caused by insufficient texture or large deformation at the pixel level in traditional optical flow methods, and significantly improving the quantitative accuracy of crack deformation parameters in complex environments.

[0076] In this embodiment, the step of identifying deformation key points based on the global deformation features, and constructing a spherical neighborhood with the deformation key points as the center and the width side length of the triangular mesh containing the deformation key points as the radius, includes: A cumulative deformation field is constructed based on the global deformation characteristics corresponding to crack data from each period; the cumulative deformation field is calculated by accumulating the displacement vectors between adjacent crack data periods period by period. The displacement modulus, displacement gradient modulus, and triangle change rate of the geometric primitive are calculated based on the cumulative deformation field. Vertices whose displacement modulus is higher than a preset first threshold, or whose displacement gradient modulus is higher than a second threshold, or whose triangle change rate is higher than a preset third threshold are identified as deformation key points. The triangle change rate is calculated based on the area change rate and the side length change rate. For each deformation key point, determine the triangular cell to which the deformation key point belongs in the triangular mesh of the first phase crack data, and take the control point in the triangular cell that is closest to the deformation key point as the target control point; Calculate the width length between the left and right boundary points corresponding to the target control point; The radius coefficient is determined based on the crack type, and the neighborhood range is calculated based on the radius coefficient and the width side length.

[0077] In this embodiment, a cumulative deformation field is constructed based on the displacement vector of each vertex in the crack data obtained from the aforementioned geometric matching. Using the first crack data period as the baseline, the cumulative displacement vector of each subsequent crack data period relative to the first period is calculated. This cumulative displacement vector is not directly composed of the displacement vector of that period itself, but is obtained by progressively accumulating the displacement vectors of adjacent periods from the first to the second period, from the second to the third period, and up to the current period. For example, the cumulative displacement vector of the third period relative to the first period is equal to the displacement vector from the first to the second period plus the displacement vector from the second to the third period, and so on. This progressive accumulation method effectively reduces the cumulative effect of single-period matching errors and comprehensively reflects the overall deformation trend of each vertex of the crack throughout the entire monitoring period. The cumulative displacement vectors of all vertices in each period are organized according to the spatial position and topological relationship of the vertices, thus forming a cumulative deformation field covering the entire crack region. Each record in this cumulative deformation field contains the cumulative displacement magnitude and direction of a specific vertex, as well as the adjacency information of that vertex in the triangular mesh.

[0078] In this embodiment, based on the cumulative deformation field, three quantitative indicators are calculated for each vertex in the triangular mesh of the first-phase crack data: displacement modulus, displacement gradient modulus, and triangle change rate. The displacement modulus is directly taken as the length of the cumulative displacement vector of that vertex; the displacement gradient modulus is calculated by comparing the cumulative displacement vector differences between that vertex and its neighboring vertices, reflecting the spatial drasticness of displacement change in the local area; the triangle change rate is determined by combining the area change rate and side length change rate of the triangle to which the vertex belongs, for example, by taking the weighted average of the absolute value of the triangle area change rate and the absolute values ​​of the three side length change rates. Subsequently, the above three indicators are compared with preset first, second, and third thresholds, respectively. If the displacement modulus of a vertex is greater than the first threshold, or its displacement gradient modulus is greater than the second threshold, or the triangle change rate to which it belongs is greater than the third threshold, then that vertex is marked as a deformation key point. This identification strategy does not rely on the inherent geometric features of the crack, such as endpoints or bifurcation points, but rather dynamically selects the spatial locations with the most concentrated stress or the most severe deformation based entirely on the actual degree of deformation, making the key point location objective and adaptable to the deformation patterns of different types of cracks.

[0079] In this embodiment, the first threshold, the second threshold, and the third threshold can be determined using an adaptive method, without the need for manually setting fixed values.

[0080] In one optional embodiment, the displacement magnitudes of all vertices are statistically analyzed, and the sum of the mean and standard deviation is taken as a first threshold. The magnitude of the displacement vector difference between each vertex and all its adjacent vertices is calculated, and the average of these differences for all vertices is multiplied by 1.5 to obtain a second threshold. A third threshold is set within a reference range based on the crack type: for transverse and longitudinal cracks, it can be set to 5% to 10%, and for network cracks, it can be set to 3% to 5%. In practice, specific values ​​can also be determined through experimental calibration or statistical analysis of historical deformation data.

[0081] In this embodiment, the triangle change rate is obtained by weighted summation of the triangle area change rate and the side length change rate of the three sides, wherein the weight of the area change rate is 0.5, the sum of the weights of each side length change rate is 0.5, and each side length change rate is taken as its absolute value.

[0082] In this embodiment, the deformation of the crack exhibits significant spatial non-uniformity. Actual monitoring shows that different segments of the crack often show varying degrees of opening, displacement, or extension. For example, local stress concentration points, branch intersections, or abrupt width changes may be the locations of the most severe deformation. If risk assessment relies solely on global parameters such as the overall length or average width of the crack, these locally high-risk areas are easily masked by the overall average value, leading to missed alarms or false alarms. Secondly, traditional pixel-based or fixed-window methods lack targeted focus on deformation hotspots. If uniform calculations are performed across the entire crack or the entire map, not only is the computational load high, but the large number of areas with gentle deformation introduces redundant information and noise, interfering with the identification of truly dangerous areas. Therefore, a mechanism is needed that can automatically identify the most significant deformation locations and construct local analysis windows centered on them to achieve accurate location and efficient assessment of crack risks.

[0083] In this embodiment, the deformation key points identified by the accumulated deformation field directly correspond to the locations where the displacement modulus, displacement gradient modulus, or triangular rate of change exceeds the threshold. These points are precisely the locations where crack deformation is most severe and instability is most likely to occur. Constructing a spherical neighborhood around these key points ensures that subsequent calculations of local expansion intensity are entirely focused on high-risk areas, avoiding wasting computational resources in stable regions and significantly improving the targeting and efficiency of the assessment.

[0084] In this embodiment, traditional methods rely solely on the overall length or average width of the crack, easily obscuring intense local deformations within the global average. This embodiment uses the accumulated deformation field to filter out vertices where the displacement modulus, displacement gradient modulus, or triangle rate of change exceeds a threshold as key deformation points. These points are precisely where the crack experiences the most concentrated stress and most active deformation. Centered on these key points, the radius of a spherical neighborhood is adaptively determined based on the crack type and local width, ensuring that the scale of the analysis window matches the physical morphology: a small radius is used for mesh-like cracks to avoid branch interference, while a large radius is used for transverse cracks to capture macroscopic trends. This "key point localization and adaptive window" strategy focuses computational resources on high-risk areas, significantly improving the relevance and computational efficiency of the assessment.

[0085] In this embodiment, for each vertex marked as a deformation keypoint, it is first necessary to determine the triangular cell it belongs to in the triangular mesh of the first-phase crack data. Since the triangular mesh records which three vertices each triangle is composed of, it can be quickly located by traversing all triangles and checking whether the vertex is one of the vertices of a triangle. Since the deformation keypoint itself is a vertex in the mesh, it must belong to one or more triangular cells sharing that vertex. For subsequent calculation of the neighborhood radius, it is necessary to associate the deformation keypoint with the crack control points. Since the control points are located on the skeleton line, and the deformation keypoint may be the left boundary point, the right boundary point, or the control point itself, it is necessary to find the control point closest to the deformation keypoint. Specifically, it is done by traversing all control points in the triangular cell to which the deformation keypoint belongs, calculating the Euclidean distance between the deformation keypoint and each control point, and taking the control point with the smallest distance as the target control point. If the deformation keypoint itself is a control point, then the target control point is itself.

[0086] In this embodiment, after determining the target control point, the coordinates of the left and right boundary points corresponding to the control point are directly read from the previously constructed triangular mesh data, and the Euclidean distance between the two points is calculated. This distance is the width edge length at the control point. This width edge length represents the local width of the crack at that section. Since the width edge is precisely determined based on the binary mask edge when constructing the triangular mesh and is verified by the segmentation mask of the first phase image, it has high geometric accuracy. The value of this width edge length will be directly used for the subsequent calculation of the spherical neighborhood radius, so that the size of the spherical neighborhood is related to the physical scale of the local crack width.

[0087] In this embodiment, a radius coefficient is finally determined based on the crack type (transverse crack, longitudinal crack, or network crack). Different types of cracks have different preset radius coefficient values: for transverse cracks, due to their gentle direction and inconsistent width, a larger radius coefficient is set, allowing the spherical neighborhood to cover a wider context area for better matching and deformation analysis; for longitudinal cracks, a moderate radius coefficient is used; for network cracks, due to their complex structure and dense branches, a smaller radius coefficient is set to avoid interference from the spherical neighborhood crossing into adjacent branches. The radius coefficient is multiplied by the aforementioned width side length to obtain the radius of the spherical neighborhood. Then, a spherical neighborhood is constructed with this radius, centered on the deformation key point. This spherical neighborhood is a spherical region in three-dimensional space, which is typically projected onto the crack plane to form a circular region in actual calculations. Each deformation key point corresponds to an independent spherical neighborhood, and all spherical neighborhoods together constitute the analysis unit for subsequent local expansion strength calculations. In this way, the spatial range of the spherical neighborhood adaptively matches the local width and overall morphological characteristics of the crack, providing a spatial basis for refined risk assessment.

[0088] In this embodiment, the radius of the spherical neighborhood is not a fixed value, but is determined by the width of the key point and the crack type. The width reflects the local physical scale of the crack, ensuring that the neighborhood size is adapted to the width of the crack itself. A wider neighborhood accommodates sufficient deformation information, while a narrower neighborhood avoids introducing background noise. Furthermore, different radius coefficients are set for transverse, longitudinal, and network cracks, further utilizing the crack morphology prior to ensure that the geometric characteristics of the analysis window match the structural features of the crack.

[0089] In this embodiment, deformation key points are stable features identified based on the cumulative deformation field. These are typically regions with large displacement gradients and are more easily identifiable in images. A spherical neighborhood centered on these key points serves as a matching template, containing richer local geometric information than simple key point coordinates or fixed pixel windows. This provides stronger constraints during subsequent inter-period matching, effectively addressing matching difficulties caused by illumination variations, local occlusion, and crack morphological evolution. Each spherical neighborhood independently calculates local expansion intensity indicators (such as average area change rate), which are ultimately integrated into the crack's comprehensive risk score. This hierarchical analysis framework of "deformation key point localization and fine-grained spherical neighborhood quantification" allows the assessment report to clearly indicate the risk level differences in different sections of the crack, providing highway maintenance personnel with an intuitive decision-making basis for "where is most dangerous and where to prioritize treatment," thus achieving a leap from "qualitative overall judgment" to "quantitative local diagnosis."

[0090] In this embodiment, by accumulating the displacement vectors between adjacent periods to construct a cumulative deformation field, the overall deformation trend of the crack during the entire monitoring period can be comprehensively reflected, avoiding interference from local noise in a single period on the identification of key deformation points. Based on the cumulative deformation field, the displacement modulus, displacement gradient modulus, and triangle change rate obtained by combining the area change rate and the side length change rate of each grid vertex are calculated, and vertices exceeding a preset threshold are identified as key deformation points. This identification strategy does not rely on the inherent geometric features of the crack, such as endpoints or bifurcation points, but dynamically selects the locations with the most concentrated stress or the most severe deformation based on the actual degree of deformation, making the key point location more accurate. Subsequently, with the key deformation point as the center, the neighborhood radius is determined according to the width side length of its location and the crack type. The width side length reflects the local structural scale, and the crack type provides morphological priors. The combination of the two allows the spatial range of the spherical neighborhood to adaptively match the local features of the crack, laying a spatial foundation for the subsequent refined calculation of local expansion intensity.

[0091] In this embodiment, calculating the local expansion intensity index within each spherical neighborhood based on the global deformation features includes: For each spherical neighborhood, the triangular units that intersect with the spherical neighborhood are determined as target triangular units, and the intersection area of ​​each target triangle with the spherical neighborhood is calculated. The weighting coefficients of each target triangle are determined based on the intersection area, and the weighted change rate is calculated based on the weighting coefficients and the area change rate of each target triangle. The total intersection area of ​​the spherical neighborhood is calculated based on the intersection areas of all target triangles, and the average area change rate is calculated based on the weighted change rate and the total intersection area. The average area change rate is used as an index of local expansion intensity within the spherical neighborhood.

[0092] In this embodiment, for each constructed spherical neighborhood, it is first necessary to determine the triangular elements that intersect with it. Since the center of the spherical neighborhood is the deformation key point, its radius is determined by the width side length and the crack type coefficient. The triangular elements originate from the triangular mesh of the first phase of crack data and have been mapped onto a unified spatial reference plane. Using a 3D algorithm to determine the intersection of a sphere and a planar triangle, each triangle is checked one by one to see if it intersects with the spherical neighborhood. If at least one of the three vertices of a triangle is located inside the sphere, or if the triangle intersects the surface of the sphere, then the triangle is identified as a target triangular element. For each target triangle, its intersection area with the spherical neighborhood is further calculated. Since the spherical neighborhood projects as a circular region on the crack plane, and the triangle is a planar polygon, the intersection area can be obtained by clipping the triangle to the circle and calculating the polygon area. The intersection area of ​​each target triangle with the spherical neighborhood is recorded.

[0093] In this embodiment, since each crack data point has been georeferenced and mapped to a unified spatial reference plane, crack and deformation analysis is primarily performed on this two-dimensional plane. A "spherical neighborhood" refers to a three-dimensional sphere centered on a deformation key point and with a calculated radius as its length. The projection of this sphere onto the crack plane is a circular region. In actual calculations, since the crack plane is approximately horizontal and deformation mainly occurs within the plane, the spherical neighborhood can be directly simplified to a circular neighborhood on the plane, with the sphere's center coordinates taken as the planar coordinates of the deformation key point. Therefore, the "spherical neighborhood" refers to this circular neighborhood or its three-dimensional equivalent representation.

[0094] In this embodiment, the weighting coefficients for each target triangle are determined based on the aforementioned intersection area. Specifically, the intersection area of ​​each target triangle is used as the weight of its area change rate; that is, a triangle with a larger intersection area contributes more to the calculation of local expansion intensity because it has a higher degree of spatial overlap with the spherical neighborhood and is more representative of the deformation characteristics within that neighborhood. Next, the weighted change rate is calculated by multiplying the area change rate of each target triangle by its corresponding intersection area and summing all products to obtain the sum of the weighted change rates.

[0095] In this embodiment, the intersection areas of all target triangles are calculated and summed to obtain the total intersection area of ​​the spherical neighborhood. This total area reflects the overall size of the region covered by all considered triangular units within the spherical neighborhood and is used for subsequent averaging.

[0096] In this embodiment, the intersection determination of a spherical neighborhood and a planar triangle in three-dimensional space is as follows: First, the coordinates of the sphere's center are projected onto the plane containing the triangle to obtain the projection point; the distance from the projection point to the nearest point of the triangle is calculated; if this distance is less than the radius of the sphere, and the distance from the center of the sphere to the plane is less than the radius of the sphere, then an intersection is determined. The intersection area can be calculated by projecting the vertices of the triangle onto the coordinate system of the sphere's center, dividing it into several sub-regions, and then numerically integrating, or by using existing geometric library functions.

[0097] In this embodiment, the sum of weighted rates of change is divided by the total intersecting area to obtain the average area change rate. This average area change rate is the local expansion intensity index within the spherical neighborhood. If the value is positive, it indicates that the local area is generally in an expanding state, and the crack in the local area is either opening or expanding; if it is negative, it indicates contraction or closure; the larger the absolute value, the more severe the deformation. This calculation method directly utilizes the area change rate already obtained from the global deformation characteristics, without the need to reconstruct the displacement field or perform complex interpolation, making the calculation efficient and physically meaningful. At the same time, weighted averaging with the actual intersecting area as the weight makes the triangles with a higher degree of overlap with the spherical neighborhood contribute more, avoiding the bias that may be introduced by simple averaging, thereby improving the stability and accuracy of the local expansion intensity index. The local expansion intensity indices calculated for all spherical neighborhoods will be output together for subsequent risk assessment and thermal chart annotation.

[0098] As a specific example of an embodiment of the present invention, suppose that in a certain monitoring period, a spherical neighborhood is constructed for a deformation key point. The projection of this spherical neighborhood onto the crack plane is a circular region with a radius determined based on the width side length and the crack type coefficient. Within this circular region, three triangular elements intersect with the circular region, denoted as triangle A, triangle B, and triangle C, respectively.

[0099] Calculations show that the intersection area of ​​triangle A and the circular region is 2.5 square centimeters, and the area change rate of triangle A is 12%, meaning that the area of ​​the triangle has increased by 12% from the first period to the current period. The intersection area of ​​triangle B and the circular region is 1.8 square centimeters, and its area change rate is 8%. The intersection area of ​​triangle C and the circular region is 3.2 square centimeters, and its area change rate is -3%, indicating that the area of ​​the triangle has shrunk by three percent.

[0100] Following the calculation method for the local expansion intensity index, the area change rate of each triangle is first multiplied by its intersection area to obtain a weighted contribution value: the contribution value of triangle A is 12% multiplied by 2.5, which equals 0.3; the contribution value of triangle B is 8% multiplied by 1.8, which equals 0.144; and the contribution value of triangle C is -3% multiplied by 3.2, which equals -0.096. These weighted contribution values ​​are then summed to obtain a weighted change rate sum of 0.384.

[0101] Then calculate the total area of ​​the intersection of the spherical neighborhood, which is the sum of the areas of the intersection of the three triangles, which is 7.5 square centimeters.

[0102] Finally, dividing the sum of weighted change rates, 0.384, by the total intersecting area of ​​7.5 square centimeters yields an average area change rate of approximately 0.0464, or 4.64%. This positive value indicates that the spherical neighborhood is generally in a state of expansion, with the crack locally expanding by approximately 4.64%. This average area change rate serves as an indicator of the local expansion intensity of the spherical neighborhood and is used for subsequent risk assessment.

[0103] If the indicator exceeds a preset warning threshold, such as 5%, the local area is considered to have a high risk and should be marked in the assessment report. If the indicator is far below the threshold or even negative, it indicates that the deformation in the area is not significant or is in a contraction state, and the risk is low. In this way, each spherical neighborhood can obtain a quantified local expansion intensity index, providing data support for refined crack risk assessment.

[0104] In this embodiment, for each spherical neighborhood, intersecting triangular units are selected as target triangles. The intersection area of ​​each target triangle with the spherical neighborhood is calculated and used as a weighting factor to calculate the weighted sum of the area change rates. This sum is then divided by the total intersection area to obtain the average area change rate, which is used as an indicator of local expansion intensity. This calculation method directly utilizes the area change rate of each triangle already obtained from the global deformation features, without the need for re-interpolation or the construction of a continuous displacement field. It is computationally efficient and has clear physical meaning. Furthermore, using the actual intersection area as the weight ensures that triangles with a higher degree of overlap with the spherical neighborhood contribute more to the local expansion intensity, avoiding the bias caused by simple averaging. Since the average area change rate integrates the overall expansion or contraction information of multiple triangles, it has a smoothing and suppressing effect on noise at the edges of individual triangles. This improves the stability and reliability of the local expansion intensity indicator while maintaining high sensitivity, accurately reflecting the net expansion degree of cracks in the region.

[0105] In this embodiment, generating an assessment report for highway slope cracks based on the local expansion strength index, crack type, and global deformation characteristics includes: The average expansion strength index is calculated based on all local expansion strength indices in the crack. Calculate the overall length propagation rate and average width change rate of each crack based on the global deformation characteristics; Based on the crack type, a risk weight coefficient is determined, and a comprehensive score is calculated based on the risk weight coefficient, the average expansion strength index, the overall length expansion rate, and the average width change rate. The crack level is then determined based on the comprehensive score. A heat map of slope crack distribution is generated based on the triangular mesh of the cracks. The heat map is then labeled based on the crack level, crack type, overall length expansion rate, average width change rate, spatial location of each spherical neighborhood and its local expansion intensity index, and an assessment report of highway slope cracks is generated.

[0106] In this embodiment, the local expansion intensity index refers to the average area change rate calculated by area weighting within a single spherical neighborhood, reflecting the net expansion or contraction degree of that local region (e.g., around a deformation critical point). The average expansion intensity index refers to the arithmetic mean of the local expansion intensity indices of all spherical neighborhoods on the same crack, used to comprehensively characterize the overall expansion level of each high-risk area on the entire crack. In subsequent risk assessment, the average expansion intensity index, along with the overall length expansion rate and the average width change rate, participates in the calculation of the comprehensive score, while each local expansion intensity index is used to mark the risk differences of different segments on the heat map, thereby achieving a complete evaluation from local to multi-scale.

[0107] In this embodiment, for each crack, after calculating the local expansion intensity index of all spherical neighborhoods, these local expansion intensity indices are first summarized. Specifically, all spherical neighborhoods contained in the crack are traversed, and the local expansion intensity index (i.e., the average area change rate) of each spherical neighborhood is read. These values ​​are summed and then divided by the total number of spherical neighborhoods to obtain the average expansion intensity index of the crack. This index reflects the average degree of expansion or contraction of local areas within the overall crack range, providing a quantitative description of the overall level of crack deformation. Since each spherical neighborhood focuses on the high-risk area where the key deformation points are located, the average expansion intensity index can comprehensively reflect the overall deformation trend of these key areas, avoiding the one-sidedness that may result from a single local index.

[0108] In this embodiment, the overall length expansion rate and average width change rate of each crack are extracted from the obtained global deformation features. The overall length expansion rate is obtained by comparing the total length of the crack skeleton line in the first phase with the total length of the crack skeleton line in the current phase, and dividing the length difference by the monitoring time interval. Since the skeleton line is composed of control points, the length calculation is based on the Euclidean distance between the control points and is not affected by segmentation noise. The average width change rate is obtained by comparing the length of the width edge at each control point in the first phase and the current phase, and calculating the average width change rate at all control points. This average value can reflect the overall trend of crack width change; a positive value indicates that the crack is opening up, and a negative value indicates contraction.

[0109] In this embodiment, the corresponding risk weight coefficient is read from a preset type weight table according to the crack type. Different types of cracks have different risk weight coefficients. For example, network cracks, due to their complex structure and tendency to cause slope instability, have a higher risk weight coefficient; transverse cracks have a lower coefficient; and longitudinal cracks have a relatively lower coefficient. The risk weight coefficients are... Average expansion intensity index Overall length expansion rate and average width change rate A comprehensive score is obtained through fusion calculation. The fusion method uses weighted summation, where the average expansion intensity index reflects the combined effect of local deformation, the overall length expansion rate describes the development trend of the crack in the extension direction, and the average width change rate characterizes the overall change in crack width. Each of the three indices is multiplied by its respective preset weight and then multiplied by a risk weight coefficient to obtain the comprehensive score. The higher the comprehensive score, the higher the risk level of the crack. Based on the numerical range of the comprehensive score, cracks are divided into four levels: safe, attention, warning, and danger, each corresponding to different maintenance response measures.

[0110] As an optional implementation method, the overall score is... Specifically: (1) Where α, β, and γ are preset weighting coefficients, satisfying α + β + γ = 1. For example, α = 0.5, β = 0.3, and γ = 0.2, meaning that local expansion intensity contributes the most to the score, followed by the overall length expansion rate, with the average width change rate serving as an auxiliary factor. Each indicator should be normalized based on field experience before being substituted into the formula to ensure comparability. The numerical range of the comprehensive score can be divided into four levels—safe, cautious, warning, and dangerous—based on historical calibration data. For example, a score less than 0.2 indicates safety, 0.2 to 0.4 indicates caution, 0.4 to 0.7 indicates warning, and greater than 0.7 indicates danger.

[0111] In this embodiment, a slope crack distribution heatmap is generated based on a triangular mesh of cracks. This heatmap uses the triangular mesh of the first phase crack data as its base map, assigning different colors based on the area change rate of each triangle or the local expansion intensity index of each spherical neighborhood. For example, low expansion areas are displayed in green, medium expansion areas in yellow, and high expansion areas in red. Based on this, information such as crack grade, crack type, overall length expansion rate, average width change rate, and the spatial location and local expansion intensity index of each spherical neighborhood are superimposed on the heatmap in labeled form. The labeled content includes the center location of each spherical neighborhood, the expansion intensity value, and the risk level to which the neighborhood belongs. The final evaluation report is presented in a combined text and graphic format. The report text lists the comprehensive evaluation conclusions and various quantitative parameters for each crack, while the attached figure shows the fully labeled slope crack distribution heatmap, providing highway maintenance personnel with intuitive decision-making basis, enabling them to quickly locate high-risk crack sections and formulate targeted maintenance plans.

[0112] In this embodiment, the average expansion intensity index is calculated based on the local expansion intensity index of all spherical neighborhoods within the crack. Simultaneously, the overall length expansion rate and average width change rate of each crack are extracted from the global deformation characteristics, achieving multi-scale deformation quantification from local to global levels. Risk weight coefficients are determined according to the crack type, and the average expansion intensity, overall length expansion rate, and average width change rate are fused to obtain a comprehensive score, which is used to classify crack levels. This assessment method not only considers the overall development trend of the crack but also highlights the comprehensive impact of local deformation through the average expansion intensity index, avoiding the potential for a single global parameter to mask high-risk local areas. Finally, a heat map of slope crack distribution is generated based on a triangular mesh, and the heat map is annotated with the comprehensive level, type, expansion rate, location of each spherical neighborhood, and expansion intensity, making the assessment report comprehensive and clearly structured, providing intuitive and accurate data support for maintenance decisions.

[0113] Please refer to Figure 2 , Figure 2 A schematic diagram of a highway slope crack quantitative assessment system based on image detection provided in an embodiment of the present invention includes: a crack identification module 201, a mesh construction module 202, a feature extraction module 203, a neighborhood construction module 204, and an assessment module 205; The crack identification module 201 is used to acquire a time-series image database of highway slope cracks, and to identify each period of crack data in the time-series image database based on a preset instance segmentation network, so as to obtain the binary mask and crack type of each crack in each period of crack data. The mesh construction module 202 is used to extract the skeleton lines of each crack based on the binary mask, discretize the skeleton lines into several control points along the crack extension direction, determine the left and right boundary points along the normal direction at each control point, and construct a triangular mesh based on the left and right boundary points between two adjacent control points. The feature extraction module 203 is used to perform geometric matching on the remaining crack data in the time series image database, using the triangular grid of the first crack data in the time series image database as a reference, to obtain the global deformation features of each crack in different period images. The neighborhood construction module 204 is used to identify deformation key points based on the global deformation features, calculate the neighborhood range based on the width side length and crack type in the triangular mesh where the deformation key points are located, and construct a spherical neighborhood with the deformation key points as the center and the neighborhood range as the radius. The evaluation module 205 is used to calculate the local expansion strength index in each spherical neighborhood based on the global deformation characteristics, and to generate an evaluation report on highway slope cracks based on the local expansion strength index, crack type and global deformation characteristics.

[0114] In this embodiment, the mesh construction module is used to extract the skeleton lines of each crack based on the binary mask, discretize the skeleton lines into several control points along the crack extension direction, and determine the left and right boundary points along the normal direction at each control point, including: Extract a single-pixel-wide skeleton line from the binary mask. The skeleton line is composed of a set of continuously arranged pixels and represents the central axis of the crack. Along the extension direction of the skeleton line, the skeleton line is discretized into several control points arranged in sequence based on the curvature change of the skeleton line, and the line connecting two adjacent control points constitutes the midline broken line of the crack. For each control point, determine the normal direction of the midline broken line at the control point, extend the normal to both sides to the two edges of the binary mask, and determine the intersection of the normal with the two edges as the left boundary point and the right boundary point; wherein, the line connecting the left boundary point and the right boundary point at the same control point constitutes the width edge at that control point.

[0115] In this embodiment, the mesh construction module is used to construct a triangular mesh based on the left and right boundary points between two adjacent control points, including: Traverse any two adjacent control points and connect the left boundary point of the previous control point, the right boundary point of the previous control point, and the left boundary point of the next control point in sequence to form the first triangle; Connect the right boundary point of the previous control point, the left boundary point of the next control point, and the right boundary point of the next control point in sequence to form the second triangle; A triangular grid is formed based on the first and second triangles between all adjacent control points.

[0116] In this embodiment, the time-series image database includes crack data from various periods ordered chronologically; the feature extraction module is used to perform geometric matching on the remaining crack data in the time-series image database, using the triangular mesh of the first crack data in the time-series image database as a reference, to obtain the global deformation features of each crack across different periods of images, including: Obtain the triangular mesh of each crack in the first phase crack data of the time-series image database, take all vertices of the triangular mesh as geometric primitives to be matched, the vertices include the left boundary point and the right boundary point corresponding to each control point, and obtain the topological connection relationship between each vertex; Based on the topological connection relationship, the crack data of each period are matched to obtain the matching vertices in the crack data of each period that correspond to the geometric primitives; For each phase of crack data, calculate the displacement vector between the geometric primitive and the matching vertex, and construct a displacement field based on the displacement vectors of all matching vertices; The area change rate and side length change rate of each triangle in the triangular mesh are calculated based on the displacement field, and a global deformation feature is formed based on the displacement vector of each matching vertex, the area change rate and the side length change rate.

[0117] In this embodiment, the neighborhood construction module is used to identify deformation key points based on the global deformation features, and construct a spherical neighborhood with the deformation key points as the center and the width side length of the triangular mesh containing the deformation key points as the radius, including: A cumulative deformation field is constructed based on the global deformation characteristics corresponding to crack data from each period; the cumulative deformation field is calculated by accumulating the displacement vectors between adjacent crack data periods period by period. The displacement modulus, displacement gradient modulus, and triangle change rate of the geometric primitive are calculated based on the cumulative deformation field. Vertices whose displacement modulus is higher than a preset first threshold, or whose displacement gradient modulus is higher than a second threshold, or whose triangle change rate is higher than a preset third threshold are identified as deformation key points. The triangle change rate is calculated based on the area change rate and the side length change rate. For each deformation key point, determine the triangular cell to which the deformation key point belongs in the triangular mesh of the first phase crack data, and take the control point in the triangular cell that is closest to the deformation key point as the target control point; Calculate the width length between the left and right boundary points corresponding to the target control point; The radius coefficient is determined based on the crack type, and the neighborhood range is calculated based on the radius coefficient and the width side length.

[0118] In this embodiment, the evaluation module is used to calculate the local expansion intensity index within each spherical neighborhood based on the global deformation features, including: For each spherical neighborhood, the triangular units that intersect with the spherical neighborhood are determined as target triangular units, and the intersection area of ​​each target triangle with the spherical neighborhood is calculated. The weighting coefficients of each target triangle are determined based on the intersection area, and the weighted change rate is calculated based on the weighting coefficients and the area change rate of each target triangle. The total intersection area of ​​the spherical neighborhood is calculated based on the intersection areas of all target triangles, and the average area change rate is calculated based on the weighted change rate and the total intersection area. The average area change rate is used as an index of local expansion intensity within the spherical neighborhood.

[0119] In this embodiment, the evaluation module is used to generate an evaluation report on highway slope cracks based on the local expansion strength index, crack type, and global deformation characteristics, including: The average expansion strength index is calculated based on all local expansion strength indices in the crack. Calculate the overall length propagation rate and average width change rate of each crack based on the global deformation characteristics; Based on the crack type, a risk weight coefficient is determined, and a comprehensive score is calculated based on the risk weight coefficient, the average expansion strength index, the overall length expansion rate, and the average width change rate. The crack level is then determined based on the comprehensive score. A heat map of slope crack distribution is generated based on the triangular mesh of the cracks. The heat map is then labeled based on the crack level, crack type, overall length expansion rate, average width change rate, spatial location of each spherical neighborhood and its local expansion intensity index, and an assessment report of highway slope cracks is generated.

[0120] The crack identification module described in this embodiment is used to acquire a time-series image database of highway slope cracks, including: Based on the visible light camera mounted on the drone, and according to the preset sampling period and fixed route, the highway slope is regularly inspected and mapped to obtain image data for each period. Motion restoration of the three-dimensional structure was performed on the image data of each period to generate orthophotos and digital surface models of the slope in each period; The orthophoto data is sorted according to time order to generate a time-series image database; the time-series image database contains several periods of crack data, and the crack data includes orthophotos, digital surface models and geographic coordinate information.

[0121] In this embodiment, the crack identification module is used to identify crack data in each period of the time-series image database based on a preset instance segmentation network, and to obtain the binary mask and crack type of each crack in each period of crack data, including: For each period of crack data in the time-series image database, the crack data is preprocessed to generate a standardized image; The standardized image is input into a preset instance segmentation network to obtain the binary mask of the crack and the crack type; the crack type includes transverse cracks, longitudinal cracks and mesh cracks.

[0122] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for quantitative assessment of cracks in highway slopes based on image detection, characterized in that, include: A time-series image database of highway slope cracks is obtained, and each crack data period in the time-series image database is identified based on a preset instance segmentation network to obtain the binary mask and crack type of each crack in each crack data period. Based on the binary mask, the skeleton lines of each crack are extracted, and the skeleton lines are discretized into several control points along the crack extension direction. At each control point, the left and right boundary points are determined along the normal direction. A triangular mesh is constructed based on the left and right boundary points between two adjacent control points. Using the triangular mesh of the first crack data in the time-series image database as a reference, geometric matching is performed on the remaining crack data in the time-series image database to obtain the global deformation characteristics of each crack in different images. Based on the global deformation features, key deformation points are identified, and the neighborhood range is calculated based on the width and side length of the triangular mesh where the key deformation points are located and the crack type. A spherical neighborhood is constructed with the key deformation points as the center and the neighborhood range as the radius. Based on the global deformation characteristics, the local expansion strength index within each spherical neighborhood is calculated, and an assessment report on highway slope cracks is generated based on the local expansion strength index, crack type, and global deformation characteristics.

2. The method for quantitative assessment of highway slope cracks based on image detection as described in claim 1, characterized in that, The process of extracting the skeleton lines of each crack based on the binary mask, discretizing the skeleton lines into several control points along the crack extension direction, and determining the left and right boundary points along the normal direction at each control point includes: Extract a single-pixel-wide skeleton line from the binary mask. The skeleton line is composed of a set of continuously arranged pixels and represents the central axis of the crack. Along the extension direction of the skeleton line, the skeleton line is discretized into several control points arranged in sequence based on the curvature change of the skeleton line, and the line connecting two adjacent control points constitutes the midline broken line of the crack. For each control point, determine the normal direction of the midline broken line at the control point, extend the normal to both sides to the two edges of the binary mask, and determine the intersection of the normal with the two edges as the left boundary point and the right boundary point; wherein, the line connecting the left boundary point and the right boundary point at the same control point constitutes the width edge at that control point.

3. The method for quantitative assessment of highway slope cracks based on image detection as described in claim 2, characterized in that, The construction of a triangular mesh based on the left and right boundary points between two adjacent control points includes: Traverse any two adjacent control points and connect the left boundary point of the previous control point, the right boundary point of the previous control point, and the left boundary point of the next control point in sequence to form the first triangle; Connect the right boundary point of the previous control point, the left boundary point of the next control point, and the right boundary point of the next control point in sequence to form the second triangle; A triangular grid is formed based on the first and second triangles between all adjacent control points.

4. The method for quantitative assessment of highway slope cracks based on image detection as described in claim 3, characterized in that, The time-series image database includes crack data from various periods ordered chronologically. The step of using the triangular mesh of the first crack data in the time-series image database as a reference to perform geometric matching on the remaining crack data in the database, obtaining the global deformation features of each crack across different periods of images, includes: Obtain the triangular mesh of each crack in the first phase crack data of the time-series image database, take all vertices of the triangular mesh as geometric primitives to be matched, the vertices include the left boundary point and the right boundary point corresponding to each control point, and obtain the topological connection relationship between each vertex; Based on the topological connection relationship, the crack data of each period are matched to obtain the matching vertices in the crack data of each period that correspond to the geometric primitives; For each phase of crack data, calculate the displacement vector between the geometric primitive and the matching vertex, and construct a displacement field based on the displacement vectors of all matching vertices; The area change rate and side length change rate of each triangle in the triangular mesh are calculated based on the displacement field, and a global deformation feature is formed based on the displacement vector of each matching vertex, the area change rate and the side length change rate.

5. The method for quantitative assessment of highway slope cracks based on image detection as described in claim 4, characterized in that, The step of identifying deformation key points based on the global deformation features, and constructing a spherical neighborhood with the deformation key points as the center and the width side length of the triangular mesh containing the deformation key points as the radius, includes: A cumulative deformation field is constructed based on the global deformation characteristics corresponding to crack data from each period; the cumulative deformation field is calculated by accumulating the displacement vectors between adjacent crack data periods period by period. The displacement modulus, displacement gradient modulus, and triangle change rate of the geometric primitive are calculated based on the cumulative deformation field. Vertices whose displacement modulus is higher than a preset first threshold, or whose displacement gradient modulus is higher than a second threshold, or whose triangle change rate is higher than a preset third threshold are identified as deformation key points. The triangle change rate is calculated based on the area change rate and the side length change rate. For each deformation key point, determine the triangular cell to which the deformation key point belongs in the triangular mesh of the first phase crack data, and take the control point in the triangular cell that is closest to the deformation key point as the target control point; Calculate the width length between the left and right boundary points corresponding to the target control point; The radius coefficient is determined based on the crack type, and the neighborhood range is calculated based on the radius coefficient and the width side length.

6. The method for quantitative assessment of highway slope cracks based on image detection as described in claim 5, characterized in that, The calculation of the local expansion intensity index within each spherical neighborhood based on the global deformation features includes: For each spherical neighborhood, the triangular units that intersect with the spherical neighborhood are determined as target triangular units, and the intersection area of ​​each target triangle with the spherical neighborhood is calculated. The weighting coefficients of each target triangle are determined based on the intersection area, and the weighted change rate is calculated based on the weighting coefficients and the area change rate of each target triangle. The total intersection area of ​​the spherical neighborhood is calculated based on the intersection areas of all target triangles, and the average area change rate is calculated based on the weighted change rate and the total intersection area. The average area change rate is used as an index of local expansion intensity within the spherical neighborhood.

7. The method for quantitative assessment of highway slope cracks based on image detection as described in claim 6, characterized in that, The assessment report for highway slope cracks, generated based on the local expansion strength index, crack type, and global deformation characteristics, includes: The average expansion strength index is calculated based on all local expansion strength indices in the crack. Calculate the overall length propagation rate and average width change rate of each crack based on the global deformation characteristics; Based on the crack type, a risk weight coefficient is determined, and a comprehensive score is calculated based on the risk weight coefficient, the average expansion strength index, the overall length expansion rate, and the average width change rate. The crack level is then determined based on the comprehensive score. A heat map of slope crack distribution is generated based on the triangular mesh of the cracks. The heat map is then labeled based on the crack level, crack type, overall length expansion rate, average width change rate, spatial location of each spherical neighborhood and its local expansion intensity index, and an assessment report of highway slope cracks is generated.

8. The method for quantitative assessment of highway slope cracks based on image detection as described in claim 1, characterized in that, The time-series image database for acquiring highway slope cracks includes: Based on the visible light camera mounted on the drone, and according to the preset sampling period and fixed route, the highway slope is regularly inspected and mapped to obtain image data for each period. Motion restoration of the three-dimensional structure was performed on the image data of each period to generate orthophotos and digital surface models of the slope in each period; The orthophoto data is sorted according to time order to generate a time-series image database; the time-series image database contains several periods of crack data, and the crack data includes orthophotos, digital surface models and geographic coordinate information.

9. A method for quantitative assessment of highway slope cracks based on image detection as described in any one of claims 1 to 8, characterized in that, The preset instance segmentation network identifies each crack data point in the time-series image database, obtaining the binary mask and crack type of each crack in each crack data point, including: For each period of crack data in the time-series image database, the crack data is preprocessed to generate a standardized image; The standardized image is input into a preset instance segmentation network to obtain the binary mask of the crack and the crack type; the crack type includes transverse cracks, longitudinal cracks and mesh cracks.

10. A quantitative assessment system for highway slope cracks based on image detection, characterized in that, include: The system includes a crack identification module, a mesh construction module, a feature extraction module, a neighborhood construction module, and an evaluation module. The crack identification module is used to acquire a time-series image database of highway slope cracks, and to identify each period of crack data in the time-series image database based on a preset instance segmentation network, so as to obtain the binary mask and crack type of each crack in each period of crack data. The mesh construction module is used to extract the skeleton lines of each crack based on the binary mask, discretize the skeleton lines into several control points along the crack extension direction, determine the left and right boundary points along the normal direction at each control point, and construct a triangular mesh based on the left and right boundary points between two adjacent control points. The feature extraction module is used to perform geometric matching on the remaining crack data in the time-series image database, using the triangular mesh of the first crack data in the time-series image database as a reference, to obtain the global deformation features of each crack in different images. The neighborhood construction module is used to identify deformation key points based on the global deformation features, calculate the neighborhood range based on the width and side length of the triangular mesh where the deformation key points are located and the crack type, and construct a spherical neighborhood with the deformation key points as the center and the neighborhood range as the radius. The evaluation module is used to calculate the local expansion strength index within each spherical neighborhood based on the global deformation characteristics, and to generate an evaluation report on highway slope cracks based on the local expansion strength index, crack type, and global deformation characteristics.

Images