Evolution tracking method and system for collapsible loess subgrade settlement crack image

Abstract

The application discloses a collapsible loess roadbed subsidence crack image evolution tracking method and system, relates to the technical field of image data processing, and identifies a road surface crack and extracts a crack skeleton and a crack direction of a periodically collected road surface image; adjacent period images are registered with a settlement invariant reference ground object as a benchmark, a registration transformation is decomposed into a rigid body deformation component and a difference deformation component, and only the difference deformation component is reserved; a crack direction tensor field is constructed based on the crack direction in a difference registration coordinate system established by the difference deformation component, and a non-homogeneous space point process intensity field is constructed based on a newly appeared crack position; a settlement groove axis position and a transverse migration rate of a collapsible settlement groove are inversed from a direction conversion position and a strength field space centroid migration, an individual crack conservation assumption is abandoned, a crack group is regarded as a direction field marked by a direction, and a settlement groove geometry is reorganized and inversed at a field level, so that the collapsible settlement space distribution and migration are objectively tracked under the premise of only relying on two-dimensional images.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, specifically to a method and system for tracking the image evolution of subsidence cracks in collapsible loess roadbeds in road engineering projects in collapsible loess areas. Background Technology

[0002] Collapsible loess is widely distributed in Northwest and North my country. This type of soil has high strength at its natural moisture content, but once wetted by surface water or groundwater, its skeletal structure is damaged, resulting in significant additional settlement, known as collapsible settlement. When collapsible loess is used as roadbed fill or as a natural foundation, the uneven collapsible settlement after water immersion will cause it to sink as a whole in the form of settlement basins or troughs, inducing cracks on the pavement surface. These cracks initially appear as fine, hairline cracks, gradually expanding into visible longitudinal and transverse cracks as settlement continues, eventually potentially leading to overall pavement slab settlement failure. Therefore, timely and objective tracking of pavement crack evolution and understanding the spatial distribution of roadbed collapsible settlement are of significant engineering value for the scientific decision-making regarding road maintenance timing.

[0003] Traditional roadbed settlement monitoring mainly relies on road administration personnel patrolling along the route and visually recording the location and width of cracks. This method is limited by the frequency of patrols and the ability to interpret visual information, making it difficult to detect early, minute cracks in a timely manner, often delaying remediation efforts. To improve the objectivity and efficiency of detection, existing research has introduced digital image processing technology into pavement crack detection. For example, Chinese invention patent CN109767423B discloses a crack detection method for asphalt pavement images. This method constructs a detection model combining a deep residual convolutional neural network and a region proposal network, extracting features from the input single-frame pavement image and classifying cracks and background to mark the location of cracks in the image. However, this scheme only performs static crack identification and localization on a single-frame pavement image, processing images at isolated moments and not involving the evolutionary tracking of adjacent periodic images of the same road segment. Therefore, it cannot reflect the spatial distribution and development process of roadbed subsidence settlement from the periodic evolution of crack images.

[0004] Regarding the objective measurement of foundation settlement, Chinese invention patent CN118258356B discloses a computer vision-based method for detecting building foundation settlement. This method acquires three-dimensional point cloud data of the building to be detected and its surrounding ground area, iteratively filters the original point cloud data to construct a current three-dimensional point cloud model, and compares the current three-dimensional point cloud model with a reference model before settlement occurs to obtain the degree of foundation settlement. However, this scheme detects only the three-dimensional point cloud of the building and its surrounding ground, relying on the overall acquisition of the three-dimensional point cloud and point-by-point comparison with the reference model to measure vertical settlement. It does not utilize the evolution information of pavement crack images, and therefore cannot reveal the spatial distribution pattern of roadbed subsidence settlement on the pavement plane and its migration law over time.

[0005] Further analysis of existing approaches to time-series tracking of pavement cracks reveals that the common practice involves registering images acquired in adjacent periods for the same road segment, performing inter-frame identity matching for each crack, and then quantifying the length growth rate and width expansion rate of each crack. This approach implicitly assumes that crack identity is conserved over time and treats newly appearing cracks, the bifurcation of old cracks, and the reorganization of their orientations during adjacent periods as interference that disrupts matching consistency and is therefore suppressed. However, for collapsible loess subgrades, surface cracks are topological projections of underground collapsible settlement troughs onto the pavement. Crack identity is not conserved during settlement development; new cracks continuously emerge along the settlement trough, and existing cracks reorganize their orientations as the trough migrates. Adopting the line-by-line matching paradigm based on crack identity conservation leads to a sharp decline in matching success rate as settlement progresses. Furthermore, it discards the most diagnostically valuable information—the location of newly emerging cracks and the reorganization of their orientations—as errors, ultimately failing to objectively infer the spatial distribution and lateral migration of collapsible settlement troughs from the periodic evolution of pavement crack images.

[0006] In summary, the core technical problem to be solved by this invention is how to objectively invert the spatial distribution and lateral migration of subsidence troughs from the periodic evolution of pavement crack images under the condition that the identity of cracks is not conserved due to subsidence settlement. Summary of the Invention

[0007] To address the core bottleneck in existing technologies for monitoring the settlement of collapsible loess subgrades, which relies on a frame-by-frame matching paradigm based on crack identity conservation and fails to objectively invert the spatial distribution and lateral migration of collapsible settlement troughs from the periodic evolution of pavement crack images, this invention provides a method and system for tracking the image evolution of cracks in collapsible loess subgrades. By abandoning the assumption of individual crack conservation and treating crack groups in adjacent periods as a spatial distribution of directional markers, and inverting the geometry of settlement troughs through the reorganization of this distribution, this invention achieves objective image measurement of the axial position and lateral migration rate of collapsible settlement troughs from the evolution mechanism of crack direction tensor field and the spatial distribution of newly formed cracks, without relying solely on two-dimensional pavement images or adding additional sensors.

[0008] The technical solution of this invention is: a method for tracking the image evolution of settlement cracks in collapsible loess roadbeds, comprising the following steps: identifying road surface cracks in periodically acquired road surface images of collapsible loess sections, extracting the crack skeleton and crack direction of each road surface crack; using a settlement-invariant reference feature in the road surface that does not evolve with collapsibility settlement as a benchmark, performing registration on road surface images of adjacent periods, decomposing the registration transformation obtained by registration into rigid deformation components characterizing the overall translation and rotation of the road surface and differential deformation components characterizing the differential settlement of the roadbed, removing the rigid deformation components and retaining the differential deformation components, and using the registration result after retaining the differential deformation components as the geometric basis. A differential quasi-coordinate system is established. In the differential quasi-coordinate system, the road surface is divided into a road surface spatial grid. Based on the crack orientation, a crack orientation tensor field is constructed on the road surface spatial grid. Based on the position of newly appearing road surface cracks in adjacent periods, a non-homogeneous spatial point process intensity field is constructed. From the orientation conversion position between the longitudinal orientation dominant area and the transverse orientation dominant area in the crack orientation tensor field, and the migration of the spatial centroid of the non-homogeneous spatial point process intensity field between adjacent periods, the position of the settlement trough axis and the transverse migration rate of the settlement trough are inverted, and the image measurement results of crack geometric parameters and evolution rate are output.

[0009] In one implementation, when constructing the crack orientation tensor field and the non-homogeneous spatial point process intensity field, newly appearing pavement cracks and pavement cracks undergoing orientation reorganization during adjacent cycles are retained as signals reflecting the geometry of the collapsible settlement trough and included in the construction. Specifically, the crack orientation concentration and crack connectivity density are calculated for each pavement spatial grid, where the crack connectivity density is the ratio of the number of intersection nodes of the crack skeleton within the pavement spatial grid to the area of ​​the pavement spatial grid. Pavement spatial grids with crack orientation concentration below a concentration threshold and crack connectivity density above a connectivity density threshold are classified as network-like fatigue. The fatigue crack-dominated grid is then identified. A network of fatigue cracks is defined as a settlement-induced network crack convergence zone if the migration vector of the crack skeleton's spatial centroid between adjacent periods is in the same direction as the migration vector of the spatial centroid of the non-homogeneous spatial point process intensity field, and the absolute value of the difference between their migration amounts is less than the migration consistency threshold. A network of fatigue cracks is defined as a traffic load network fatigue crack zone and eliminated if the migration amount of the crack skeleton's spatial centroid between adjacent periods is less than the position stability threshold and the grid center falls within the road surface wheel track zone. The settlement-induced network crack convergence zone is then used as a candidate location area for the settlement trough axis. In one embodiment, when a collapsible loess road section is in a continuous longitudinal settlement zone, causing the settlement-invariant reference feature to undergo overall translation and rotation, multi-baseline constraints along the route direction and zero-mean constraints of the differential deformation components at the road section scale are applied during the registration transformation decomposition to decouple the rigid deformation components from the differential deformation components. In one implementation, when constructing the crack orientation tensor field, the orientation of cracks within each pavement spatial grid is statistically estimated using direction to determine the principal orientation direction and crack orientation concentration, and the dominant grid of network fatigue cracks is not included in the determination of orientation transition positions. In another implementation, when constructing the non-homogeneous spatial point process intensity field, an anisotropic kernel function with its kernel major axis along the principal orientation direction is used to estimate the intensity of newly appearing pavement cracks in adjacent periods, resulting in an anisotropic distribution of the non-homogeneous spatial point process intensity field along the principal orientation direction. In yet another implementation, the neighborhood of the settlement trough shoulder on both sides of the inverted settlement trough axis position is used as a priori and fed back to the step of identifying pavement cracks in the next period's pavement image. Within the settlement trough shoulder neighborhood, the grayscale threshold used for identifying pavement cracks is lowered or the identification sensitivity is increased to identify early longitudinal hairline cracks that were missed when a uniform grayscale threshold was used in the rest of the pavement area.

[0010] This invention also provides an image evolution tracking system for subsidence cracks in collapsible loess roadbeds, comprising: a crack identification module for identifying pavement cracks in periodically acquired pavement images of collapsible loess road sections, and extracting the crack skeleton and crack direction of each pavement crack; and a registration decomposition module for performing registration on pavement images of adjacent periods based on a settlement-invariant reference feature in the pavement that does not evolve with subsidence, decomposing the registration transformation obtained by registration into rigid deformation components characterizing the overall translation and rotation of the pavement and differential deformation components characterizing the differential settlement of the roadbed, removing the rigid deformation components and retaining the differential deformation components, and establishing a differential registration based on the registration result after retaining the differential deformation components as a geometric reference. The system includes: a quasi-coordinate system; a field construction module, used to divide the road surface into a road surface spatial grid in the differential quasi-coordinate system, aggregate and construct a crack orientation tensor field on the road surface spatial grid based on the crack orientation, and construct a non-homogeneous spatial point process intensity field based on the position of newly appearing road surface cracks in adjacent periods; and a settlement trough inversion module, used to invert the settlement trough axis position and lateral migration rate of the settlement trough from the orientation conversion position between the longitudinal orientation dominant area and the lateral orientation dominant area in the crack orientation tensor field, and the migration of the spatial centroid of the non-homogeneous spatial point process intensity field between adjacent periods, and output as image measurement results of crack geometric parameters and evolution rate.

[0011] The beneficial effects of this invention are as follows.

[0012] First, this invention uses a settlement-invariant reference feature as a benchmark to register adjacent periodic images and decomposes the registration transformation into rigid deformation components and differential deformation components, retaining only the differential deformation components. The mechanism is that the translation and rotation of the road surface as a whole under changes in the vehicle-mounted imaging perspective and the overall displacement of the roadbed are interferences unrelated to subsidence settlement. If not removed, these interferences would overlap with the deformations that truly reflect the differential settlement of the roadbed, causing subsequent field-level statistics to be overwhelmed by the overall displacement. By explicitly decomposing and stripping the rigid deformation components through the registration transformation, the differential deformation components are purified into the true deformation signal of subsidence settlement on the road surface. Existing methods rely on the overall comparison between 3D point clouds and reference models to measure vertical settlement, failing to distinguish between overall displacement and differential settlement on 2D road surface images. This invention achieves this separation on 2D images, laying a geometric benchmark for inverting the spatial distribution of settlement from crack images.

[0013] Second, this invention constructs a crack orientation tensor field based on crack orientation and a non-homogeneous spatial point process intensity field based on the location of newly emerging cracks, and combines the two for settlement trough inversion. The mechanism is that subsidence settlement is a spatially correlated continuous field process. Tension bands at the shoulder of the settlement trough induce longitudinal cracks along the road, abrupt settlement gradient boundaries induce transverse cracks, and a network of fatigue cracks converges in the central zone of the settlement trough. Therefore, the spatial distribution of the crack group's orientation is the only state variable determined by the geometry of the settlement trough. The orientation tensor field characterizes the spatial organization of the orientation on the road surface, while the spatial point process intensity field of newly emerging cracks characterizes the emergence intensity at the settlement leading edge. The two, from the complementary dimensions of direction and intensity, jointly constrain the location of the settlement trough. Existing single-frame crack identification schemes only output the location of cracks at isolated moments without any spatial distribution inversion capability. This invention, through the combination of the orientation tensor field and the intensity field, provides settlement trough geometric constraints that a single statistical quantity cannot provide. The synergy between the two produces an inversion effect greater than the sum of their individual effects.

[0014] Third, this invention abandons the assumption of individual crack conservation and retains newly emerging cracks and cracks undergoing orientation reorganization during adjacent cycles as signals reflecting the geometry of the subsidence trough, and participates in its construction. The mechanism is that in subsidence scenarios where crack identity is not conserved, the frame-by-frame matching paradigm discards newly emerging and reorganized cracks as matching errors, precisely discarding the most valuable signals carrying the migration information of the subsidence trough. This invention reverses these signals and uses them as the basis for localization and inversion, thus transforming the source of failure in the traditional paradigm into the source of information for this invention. This effect is theoretically impossible to achieve with the frame-by-frame matching paradigm. Furthermore, this invention quantitatively determines the dominant grid of network fatigue cracks using dual thresholds of crack orientation concentration and crack connectivity density, and distinguishes between the settlement-induced network crack convergence area and the traffic load network fatigue crack area based on the migration directionality and consistency of the spatial centroid of the crack skeleton and the spatial centroid of the process intensity field of non-homogeneous spatial points. The mechanism lies in the fact that the settlement-induced network cracks are formed by the bidirectional tensile-compressive redistribution at the bottom of the settlement trough, and migrate as a whole with the lateral migration of the settlement trough axis. Therefore, the migration of its spatial centroid is in the same direction and of similar magnitude as the migration of the intensity field centroid. In contrast, the conventional network fatigue cracks caused by traffic loads are formed by the accumulation of repeated wheel loads, and their spatial location is locked in the wheel track zone and does not migrate significantly between adjacent cycles. The temporal evolution behaviors of the two are fundamentally different. Based on this, the network fatigue crack zone caused by traffic loads can be eliminated from the candidate location area of ​​the settlement trough axis, so that the characteristic boundary of the candidate location area can be uniquely determined by a quantifiable threshold criterion.

[0015] Fourth, this invention uses the directional transformation position of the directional tensor field and the migration of the spatial centroid of the non-homogeneous spatial point process intensity field between adjacent periods to jointly invert the axial position and lateral migration rate of the settlement trough. The mechanism is that the directional transformation position marks the spatial boundary between the shoulder and bottom of the settlement trough, and the centroid of the intensity field marks the spatial concentration position of the settlement front. The relative displacement of these two points between adjacent periods corresponds to the lateral migration of the settlement trough. Through joint inversion, this invention can still output objective image measurement results of the axial position and lateral migration rate of the settlement trough even under the condition that crack identity is not conserved. This result can be used by downstream roadbed maintenance engineers for comprehensive judgment based on geological conditions, and is strictly limited to image geometric measurement results, not directly providing maintenance decisions.

[0016] Fifth, this invention uses the neighborhood of the settlement trough shoulder obtained from the inversion as a priori feedback to the crack identification step in the next cycle, reducing the grayscale threshold or increasing the identification sensitivity within the shoulder neighborhood. The mechanism is that the settlement trough shoulder is a high-incidence tension zone for early longitudinal hairline cracks, but the grayscale contrast of hairline cracks is weak, making them easily missed when a uniform grayscale threshold is used across the entire road surface. By feeding back the inversion results of the settlement trough geometry into the identification process, the identification sensitivity is specifically improved in the shoulder zone where early cracks are most likely to occur, thereby significantly increasing the detection rate of early hairline cracks. This improvement in early detection rate stems from the closed loop formed by the inversion results and the identification process, a nonlinear synergistic effect that cannot be achieved by simply connecting the identification and inversion processes in series. Attached Figure Description

[0017] Figure 1 This is a flowchart of the image evolution tracking method for subsidence cracks in collapsible loess roadbed according to the present invention;

[0018] Figure 2 This is an architecture diagram of the image evolution tracking system for subsidence cracks in collapsible loess roadbeds according to the present invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0020] like Figure 1As shown, the image evolution tracking method for settlement cracks in collapsible loess subgrade provided in this embodiment includes the following steps: S1, pavement crack identification; S2, settlement-invariant reference registration decomposition; S3, orientation tensor field and intensity field construction; and S4, settlement trough inversion. The inversion result is fed back to S1 after S4, forming a closed-loop structure from S1 to S4 and back to S1. Each step is deeply coupled, with the output of the previous step serving as the key input for the next: the crack skeleton and crack orientation output in S1 provide the processing objects for S2 and S3; the differential deformation components output in S2 provide the registration coordinate system for S3; the crack orientation tensor field and non-homogeneous spatial point process intensity field output in S3 provide the inversion basis for S4; and the settlement trough axis position output in S4 is fed back to S1 as a priori information. Each step will be explained in detail below. The road surface images described in this embodiment are obtained by performing periodic high-resolution imaging of the road surface of collapsible loess sections using a vehicle-mounted or fixed industrial camera. The time interval between two adjacent imaging sessions is recorded as the acquisition period. Images of the same road section obtained in different acquisition periods constitute road surface images of adjacent periods.

[0021] Step S1: Road Crack Identification. Step S1 identifies road cracks in periodically acquired images of collapsible loess road sections, extracting the crack skeleton and direction of each crack. First, illumination homogenization and contrast enhancement are performed on the road images to suppress grayscale unevenness caused by sunlight angle, road surface wet spots, and shadows in vehicle imaging, thus achieving a stable enhancement of the grayscale contrast between the road cracks and the road background. Illumination homogenization is achieved through low-frequency estimation and pixel-by-pixel normalization of the image background grayscale field. Contrast enhancement is achieved through adaptive histogram equalization with contrast limitation. The above image preprocessing is well-known in the art, and its key parameters include the morphological structural element size used for background estimation and the contrast limitation coefficient of the adaptive histogram equalization. In this embodiment, the morphological structural element size is set to 3 to 5 times the maximum crack width in the road image, and the contrast limitation coefficient is set to 2.0 to 3.0.

[0022] Subsequently, based on the low-grayscale, elongated connected region characteristics of road surface cracks, a crack skeleton is extracted using a combination of directional gradient field and connected region growth. Specifically, the grayscale directional gradient of each pixel is first calculated on the enhanced road surface image to obtain a directional gradient field reflecting the orientation of the local dark, elongated structure. Then, pixels with significant gradient magnitudes and consistent directions in the directional gradient field are used as seed points. Connected region growth is performed along the crack extension direction indicated by the directional gradient field, aggregating pixels with grayscale values ​​below the crack grayscale threshold and connected to the seed points into the same crack connected region. Finally, each crack connected region is thinned to obtain a crack skeleton with a single pixel width. This method of extracting the crack skeleton using a combination of directional gradient field and connected region growth is well-adapted to the elongated connected characteristics of road surface cracks and can robustly preserve the crack skeleton structure amidst road surface texture noise.

[0023] After obtaining the crack skeleton, the length, maximum width, and direction of each road surface crack are quantified. The crack length is taken as the pixel chain length of the crack skeleton and converted to a physical length in mm according to the camera imaging scale; the maximum crack width is taken as the maximum Euclidean distance between pairs of crack boundary points perpendicular to the skeleton tangent at each skeleton point, converted to mm according to the scale; the crack direction is taken as the angle between the principal axis of the crack skeleton and the road direction, recorded in degrees. In this embodiment, the camera imaging scale is calibrated so that 1 pixel corresponds to a road surface physical dimension of 0.5 mm to 1.0 mm, thus giving the quantification of length and maximum width a clear physical dimension. The above-mentioned length, maximum width, and direction constitute the crack geometry parameters of each road surface crack, and are output together with the crack skeleton to steps S2 and S3.

[0024] It is worth noting that for the settlement trough shoulder neighborhood reported in step S4, step S1 lowers the grayscale threshold used to identify pavement cracks or increases the identification sensitivity within this neighborhood. The specific feedback mechanism will be explained in detail in the closed-loop feedback section following step S4. When there is no feedback information in the first acquisition cycle, step S1 performs identification using a uniform grayscale threshold across the entire pavement.

[0025] Step S2: Settlement-invariant reference registration decomposition. Step S2 uses a settlement-invariant reference feature in the road surface that does not evolve with subsidence settlement as a reference. It performs registration on the road surface images of adjacent periods, decomposes the registration transformation obtained by registration into rigid deformation components that characterize the overall translation and rotation of the road surface and differential deformation components that characterize the differential settlement of the subgrade. The rigid deformation components are removed and the differential deformation components are retained. The differential registration coordinate system is established based on the registration result after retaining the differential deformation components as the geometric reference.

[0026] Settlement-invariant reference features refer to road features whose positions do not significantly change with subgrade subsidence settlement. In this embodiment, settlement-invariant reference features include lane markings, curbs, and road expansion joints. Several control points on the settlement-invariant reference features are extracted from the road surface images of adjacent periods, and a correspondence is established between control points in adjacent periods to obtain the displacement of each control point between adjacent periods. It should be noted that during the development of subsidence settlement, although settlement-invariant reference features do not settle due to their own materials, when the entire road section where they are located experiences subsidence or is affected by changes in the vehicle's viewing angle, the control points will still observe displacements caused by the overall translation and rotation of the road surface. This part of the displacement is not a reflection of the subgrade differential settlement and must be separated from the registration transformation.

[0027] Therefore, the observed displacement of the control point is decomposed into rigid body deformation components and differential deformation components. The rigid body deformation components are described by the overall translation of the road surface and the rotation around the centroid of the road segment. For each control point, the estimated rigid body deformation component is obtained from the global translation vector and the rotational motion acting on the position vector of that control point relative to the centroid of the road segment. The rigid body deformation component is obtained by solving for the global translation vector and rotation angle that minimize the sum of squared weighted residuals between the observed displacement and the estimated rigid body deformation component for all control points. The solution model is as follows:

[0028] ,

[0029] in: The vector is a global translation vector, a two-dimensional column vector, whose two components are the horizontal and vertical translations along the road surface image, respectively, with values ​​ranging from [value range missing]. to The unit is The value obtained by solving this minimization problem represents the overall translation of the road surface. Let be the rotation angle, and be a scalar with a range of values. to The unit is This is obtained by solving the minimization problem, and it represents the rotation of the entire road surface around the centroid of the road segment;

[0030] For the rotation angle The constructed two-dimensional rotation matrix is A matrix whose elements are the sine and cosine of the rotation angles, with dimensions of one, is formed by... Determine, characterize rotational action; Let be a second-order identity matrix. A matrix, with one dimension, is a constant and is used to construct relative displacements during rotation;

[0031] For the first The observed displacements of each control point between adjacent periods are represented by a two-dimensional column vector, with values ​​ranging from [value range missing]. to The unit is The total displacement of the control point is obtained by converting the coordinate difference of the same control point on adjacent periodic images according to the scale.

[0032] For the first The position vector of each control point is a two-dimensional column vector, with units of _____. The coordinates of the control points in the road surface image are converted according to the scale, which represents the position of the control points;

[0033] The position vector of the centroid of the road segment is a two-dimensional column vector with units of 1. It is obtained by the arithmetic mean of the position vectors of all control points and represents the center of rotation;

[0034] For the first The weight of each control point is a scalar, and its value ranges from... to The dimensionless value is assigned according to the reliability of the settlement-invariant reference ground feature where the control point is located, and it represents the contribution of the control point to the estimation of rigid body deformation components.

[0035] The total number of control points is a positive integer, and its value range is... to It has one dimension and is determined by the number of extracted control points, representing the scale of control points involved in the solution; Here is the control point number, which is a positive integer and its value range is... to , with one dimension, is a summation index;

[0036] The L2 norm of a two-dimensional vector has dimensions consistent with its independent variable.

[0037] Summation sign in constraint terms Indicates all The differential deformation components at each control point are summed and constrained to be zero, i.e., a zero-mean constraint is applied to the differential deformation components at the road segment scale. This minimization problem is a weighted least squares problem with equality constraints, and the optimal solution for the global translation vector and rotation angle can be obtained using the Lagrange multiplier method. and .

[0038] Specifically, when a collapsible loess road section is located in a continuous longitudinal settlement zone, and the reference ground feature with constant settlement undergoes overall translation and rotation, the overall translation of the pavement and the differential settlement of the subgrade are in the same direction along the route. Simply relying on control point displacements to solve for the rigid deformation components will result in an ill-conditioned situation where the overall translation and differential deformation are difficult to distinguish. Therefore, a multi-baseline constraint is applied along the route in the above solution. This involves setting up multiple sets of control point baselines distributed along the route at different mileage locations along the road section. This ensures that the overall translation and rotation are consistently interpreted across multiple baselines, while the differential settlement exhibits differences across each baseline, thereby improving the distinguishability of the rigid deformation component solution. Simultaneously, a zero-mean constraint is applied to the differential deformation components at the road section scale, as shown in the above constraint terms. This ensures that the overall translation is completely attributed to the rigid deformation components, and the zero-order component of the differential settlement is not misjudged as overall translation. After obtaining the optimal solution for the overall translation vector and rotation angle, the differential deformation components of each control point are obtained as follows: ,in: For the first The differential deformation components of each control point are represented by a two-dimensional column vector, with values ​​ranging from [value range missing]. to The unit is The value obtained by this formula represents the residual deformation at the control point caused by the differential settlement of the subgrade after removing the overall translation and rotation of the road surface. This is the optimal solution for the overall translation vector. The optimal solution for the rotation angle is defined as the same as the aforementioned minimization problem, and is obtained from the aforementioned constrained least squares problem. , , , , The definition is the same as the aforementioned formula. The differential deformation components of each control point are interpolated onto the pavement mesh to obtain a registration coordinate system that retains the differential deformation components, denoted as the differential registration coordinate system. This system serves as the geometric reference for constructing the crack orientation tensor field and the intensity field of the non-homogeneous spatial point process in subsequent steps. The discarded rigid body deformation components are no longer involved in subsequent processing.

[0039] Step S3: Construction of the Crack Orientation Tensor Field and Intensity Field. In Step S3, within the differential calibration coordinate system established in Step S2, a crack orientation tensor field is constructed on the pavement spatial grid based on crack orientation. A non-homogeneous spatial point process intensity field is constructed based on the locations of newly appearing pavement cracks in adjacent cycles. When constructing these two fields, newly appearing pavement cracks during adjacent cycles and pavement cracks undergoing orientation reorganization are retained as signals reflecting the geometry of the collapsible settlement trough and are included in the construction, rather than being discarded as matching errors.

[0040] The road surface is divided into regular road space grids in the differential reference coordinate system. For each road space grid, the crack orientation of all crack skeleton segments falling within that grid is collected and aggregated into a orientation tensor by weighting the crack skeleton segment length. Since the crack orientation is axial data that does not distinguish between positive and negative directions, the following orientation tensor is used for aggregation:

[0041] ,in: For the first The orientation tensor of each road surface spatial grid is... A symmetric matrix whose elements have a dimension of one and whose values ​​range from 1 to 1. to The formula is used to calculate and characterize the spatial distribution and concentration of crack orientation within the grid. The index of the road surface spatial grid is a positive integer with a dimension of one, used to identify the grid. For the first in this grid The length of each crack skeleton segment is a scalar, and its value ranges from greater than 1. The unit is The length of the pixel chain of the crack skeleton segment is converted according to the scale bar and used as a weighting coefficient to characterize the greater contribution of the long skeleton segment to the orientation. For the first in this grid The crack orientation of a crack skeleton segment is a scalar, with a value range of [value missing]. to The unit is The angle between the main axis of the crack skeleton segment and the direction of the route is determined, which represents the orientation of the skeleton segment; The total number of crack skeleton segments within the grid is a non-negative integer with a dimension of one, determined by the number of skeleton segments falling into the grid. Here is the skeleton segment number, which is a positive integer, and its value range is... to , is the summation index; This is the sum of the lengths of all crack skeleton segments within the grid, a scalar whose value ranges from greater than 1. The unit is This is obtained by summing the right side of the expression and is used for normalization. and These are cosine and sine functions, respectively, acting on the crack orientation, with dimensions of one. Eigenvalue decomposition is performed on the orientation tensor to obtain its largest and smallest eigenvalues, and the orientation concentration is calculated based on these values. Simultaneously, the principal orientation direction is calculated from the elements of the orientation tensor. ,in: For the first The orientation concentration of each road surface spatial grid is a scalar quantity with a value range of [value missing]. to The dimensionless value, calculated by this formula, characterizes the consistency of crack orientation within the grid. The closer the value is to the value of the crack orientation, the better. The more consistent and similar the trends, the better. This indicates a more dispersed trend; For the first The main orientation of each road surface spatial grid is a scalar, with a value range of [value range missing]. to The unit is The direction of crack orientation within the grid is obtained by calculation using this formula. and respectively the direction tensor The larger and smaller eigenvalues ​​are scalars, satisfying the following... and Dimensionless, derived from the pair Eigenvalue decomposition yields a characterization of energy distribution along primary and secondary directions; , , respectively the direction tensor The elements in the first row and first column, the first row and second column, and the second row and second column have a dimension of one, as determined by the previous formula; The arctangent function has a dimensionless value of one. In terms of dimensions, the directional concentration is obtained by dividing the difference of the dimensionless eigenvalues ​​by their sum. Road surface spatial grids with a directional concentration below the concentration threshold are identified as grids dominated by network fatigue cracks. In this embodiment, the concentration threshold is set to 0.35 to 0.45. These grids dominated by network fatigue cracks are not included in the subsequent determination of directional conversion positions due to their dispersed directional distribution, but their positions and connectivity densities are retained as candidate locations for the settlement trough axis.

[0042] When constructing the intensity field of a non-homogeneous spatial point process, the location of newly appearing pavement cracks in adjacent cycles is taken as the event point of the spatial point process. An anisotropic kernel function with the kernel major axis along the main strike direction is used to estimate the intensity of the event point, so that the intensity field of the non-homogeneous spatial point process exhibits anisotropic distribution along the main strike direction.

[0043] .

[0044] .

[0045] in: For the location on the road surface The intensity of a nonhomogeneous point process at a given location is a scalar, and its value ranges from 0 to 1. The unit is The value obtained from this formula represents the spatial emergence intensity of newly appearing pavement cracks at that location. The road surface position vector is a two-dimensional column vector with units of 1. The coordinates in the differential quasi-coordinate system determine the location of the intensity to be estimated. For the first The location vector of each newly appearing pavement crack is a two-dimensional column vector, with units of [missing information]. The coordinates of the centroid of the crack skeleton in the differential quasi-coordinate system are used as the event point of the spatial point process. The total number of new road surface cracks that appear between adjacent weeks is a non-negative integer with a dimension of one, and is determined by the number of new cracks obtained from inter-frame comparison.

[0046] The index of the newly appearing pavement crack is a positive integer, with a value range of 1. to , is the summation index; For the first The anisotropic bandwidth matrix at each event point is: A symmetric positive definite matrix whose element unit is . The expression is constructed from the right side of this equation and represents the anisotropic distribution of the kernel function at the event point.

[0047] The main direction of movement from the grid where the event point is located. The constructed two-dimensional rotation matrix is The matrix has one dimension and is determined by the direction of the main axis, so that the major axis of the bandwidth matrix is ​​aligned with the direction of the main axis. For the first The main direction of the grid where each event point is located is defined in the previous formula. It is determined by the previous formula; Let be the bandwidth of the anisotropic kernel function along the principal direction, and let be a scalar with a range of values. to The unit is Selected according to the road surface grid scale, representing the smoothness scale along the direction of travel;

[0048] Let be the bandwidth of the anisotropic kernel function perpendicular to the principal direction, and let be a scalar with a range of values. to The unit is Selected according to the road surface grid scale and meeting the requirements , representing the smooth scale in the vertical direction; The determinant of the bandwidth matrix, in units of ; This is the inverse of the bandwidth matrix, and its element units are... ; superscript Represents the transpose of a vector or matrix; It is an exponential function with the natural constant as its base and has one dimension; Pi is a constant with dimensions of one. The anisotropic core smoothly connects newly emerging cracks into a continuous high-strength band along the direction of the settlement trough, thus highlighting the spatial distribution pattern of the settlement front in the intensity field.

[0049] Step S4: Settlement trough inversion. Step S4 uses the orientation transition position between the longitudinal and transverse orientation dominant regions in the crack orientation tensor field, and the migration of the spatial centroid of the non-homogeneous spatial point process intensity field between adjacent periods, to invert the settlement trough axis position and transverse migration rate of the settlement trough, and outputs the image measurement results of crack geometric parameters and evolution rate.

[0050] The road surface is divided into several mileage sections along the route. For each mileage section, the dominant directional direction is examined along the transverse coordinate perpendicular to the route. Grids whose dominant directional direction is close to the route direction are identified as longitudinally dominant areas, and grids whose dominant directional direction is close to the route direction are identified as transversely dominant areas. The transverse position between these two is the directional transition position. The tension zone at the shoulder of the settlement trough induces longitudinal cracks along the route, and transverse cracks and network fatigue cracks converge in the center zone of the settlement trough. Therefore, the directional transition position marks the boundary between the settlement trough shoulder and the bottom of the trough. To improve the robustness of the directional transition position location, this embodiment sets a sliding window along the transverse coordinate. Within the window, the proportion of longitudinally dominant grids and transversely dominant grids is counted. The transverse coordinate where the proportion flips from longitudinally dominant to transversely dominant is taken as the directional transition position. At the same time, a continuity constraint along the route direction is applied to the directional transition positions of several adjacent mileage sections to eliminate isolated jumps caused by local texture noise, so that the boundary of the settlement trough shoulder is a continuous curve along the route direction.

[0051] Considering that settlement troughs typically have a shoulder boundary on each side of the route, this embodiment calculates the directional transition positions on both sides of the route center to obtain the two shoulder boundaries of the settlement trough and the trough bottom area enclosed by them. When determining the directional transition positions, only grids with a directional concentration higher than the concentration threshold are used to ensure that the grid dominated by network fatigue cracks does not contaminate the shoulder boundary. This grid dominated by network fatigue cracks, after being screened using the distinction criteria between settlement-induced network cracks and traffic load network fatigue cracks (described later), is used as a candidate location area for determining the trough bottom position.

[0052] Furthermore, to ensure a clear quantitative boundary for determining the dominant grid of network fatigue cracks, and to distinguish network cracks induced by subsidence settlement from conventional network fatigue cracks caused by traffic loads, this embodiment quantifies the crack connectivity density of each pavement spatial grid and applies a dual-threshold criterion. Crack connectivity density is defined as the ratio of the number of intersection nodes of the crack skeleton within the pavement spatial grid to the area of ​​the pavement spatial grid, and its calculation formula is: ρ g =Ng / A g Wherein: ρ g Let N be the crack connectivity density of the g-th pavement spatial grid, a scalar value ranging from 0 to 200, with units of 1 / m², calculated by this formula, characterizing the degree of bifurcation and interweaving of the crack skeleton within this grid; g denoted as the number of intersection nodes of the crack skeleton within the g-th pavement spatial grid, a non-negative integer ranging from 0 to 200, with a dimension of one. It is obtained by counting the number of skeleton points with at least 3 skeleton pixels in their eight neighborhoods for each refined single-pixel width crack skeleton, representing the number of skeleton intersections; A g Let be the area of ​​the g-th road surface spatial grid, where is a scalar with a value ranging from 0.01 to 4, and the unit is m². 2 The grid size is calculated based on the camera imaging scale, representing the normalized spatial area benchmark; g is the index of the pavement spatial grid, a positive integer with a dimension of one, defined as in the previous formula. Based on this, the crack orientation concentration is set below the concentration threshold c. th And the crack connectivity density ρ g Above the connectivity density threshold ρ th The pavement spatial grid was determined to be a grid dominated by network fatigue cracks. Where: c th ρ is the concentration threshold, a scalar value ranging from 0.20 to 0.60, with one dimension. It is calibrated according to the valley value of the histogram of the distribution of orientation concentration across the entire road surface. In this embodiment, it is set to 0.40, representing the lower bound for determining orientation consistency; th The connectivity density threshold is a scalar value ranging from 4 to 30, with units of 1 / m. 2 The boundary value for the connectivity density distribution between the network crack area and the non-network crack area, as manually interpreted on the calibrated road section, is used. In this embodiment, 12 is taken as the lower bound for determining the degree of skeleton interweaving. The orientation concentration is lower than c. th This indicates that the crack orientation within the grid is diffuse, and the crack connectivity density is higher than ρ. th The mesh is characterized by densely intersecting crack skeletons. Only when both conditions are met can it be determined as a mesh dominated by network fatigue cracks, thus excluding isolated short crack meshes with diffuse orientation but sparse skeletons, as well as meshes with dense skeletons but consistent orientation of longitudinal cracks.

[0053] The dominant meshes of network fatigue cracks fall into two categories based on their origin, requiring further differentiation: network cracks induced by subsidence settlement form in the bidirectional tensile-compression redistribution zone at the bottom of the settlement trough, migrating as a whole with the lateral migration of the settlement trough axis; conventional network fatigue cracks caused by traffic loads form in the wheel track zone subjected to repeated wheel loads, their spatial position locked by lane geometry, and they essentially do not migrate between adjacent cycles. Based on this, the spatial centroid of the internal crack skeleton pixels in the differential criterion coordinate system is calculated for each dominant mesh of network fatigue cracks, and its migration vector Δc between adjacent cycles is taken. g Simultaneously, the spatial centroid of the intensity field is obtained by weighting the intensity of non-homogeneous spatial point processes, and its migration vector Δc between adjacent periods is taken. λ The following three criteria are applied: Criterion one is the same-direction criterion, i.e., Δc g With Δc λ The inner product is greater than 0; criterion two is the migration consistency criterion, i.e., Δc g The modulus and Δc λ The absolute value of the difference between the moduli is less than the migration consistency threshold ε. m Criterion three is the positional stability and wheel track attribution criterion, i.e., Δc g The magnitude is less than the position stability threshold ε s Furthermore, the lateral coordinates of the center of the grid fall within the lateral coordinate range of the road surface wheel track. A grid dominated by network fatigue cracks that simultaneously satisfies criteria one and two is determined to be a settlement-induced network crack convergence area, serving as a candidate location area for the settlement trough axis. A grid dominated by network fatigue cracks that satisfies criterion three is determined to be a traffic load network fatigue crack area, and is therefore removed from the candidate location area and does not participate in the strength-weighted centroid determination of the lateral position of the settlement trough axis.

[0054] Where: Δc g Δc is the migration vector of the spatial centroid of the crack skeleton pixel within the dominant mesh of the g-th mesh fatigue crack between adjacent periods. It is a two-dimensional column vector with a value range of -2 to 2, in meters. It is obtained by converting the difference of the arithmetic mean of the crack skeleton pixel coordinates within this mesh in the differential standard coordinate system according to the scale, and represents the overall spatial displacement of the crack group within this mesh; λ ε is the migration vector of the spatial centroid of the non-homogeneous spatial point process intensity field between adjacent periods. It is a two-dimensional column vector with a value range of -2 to 2, and the unit is m. It is obtained by subtracting the spatial centroids obtained by intensity-weighted summation of the intensity fields of adjacent periods, and represents the spatial displacement of the settlement front. m ε is the migration consistency threshold, a scalar value ranging from 0.05 to 0.50, in meters. It is selected based on the order of magnitude of the product of the lateral migration rate in the settling tank and the acquisition period. In this embodiment, it is set to 0.20, representing the upper limit of the allowable migration deviation; sThe positional stability threshold is a scalar value ranging from 0.01 to 0.10, in meters, selected based on the standard deviation of the registration residual. In this embodiment, it is set to 0.05, representing the upper limit of the allowable migration amount for determining positional stability. The lateral coordinate interval of the road wheel track is a scalar interval in meters, determined by the lane width and the standard wheel track position. In this embodiment, two symmetrical intervals are selected: 0.6 to 1.6 meters and 2.4 to 3.4 meters from the route center. Based on this, the dominant lateral grid on this mileage section and the grid determined to be the convergence zone of settlement-induced network cracks are weighted by the intensity of non-homogeneous spatial point processes, and the intensity-weighted centroid of its lateral coordinate is taken as the lateral position of the settlement trough axis on this mileage section. ,in: Mileage section The transverse position of the settling trough axis at the location is a scalar quantity, with units of . The value obtained by this formula represents the lateral coordinate of the settlement trough axis deviating from the center of the route on the cross section at this mileage. The location of the mileage section is given by [value], and [value] is a scalar quantity in units of [unit]. The value is taken along the route direction to represent the mileage; It is a set of the dominant transversely oriented grid and the grid of the settlement-induced network crack convergence zone at the mileage section. It consists of grids whose main orientation is approximately perpendicular to the route direction and grids that have been identified as settlement-induced network crack convergence zones based on the differentiation criteria. , representing the candidate mesh set for the central zone of the settling trough; , where is the grid index, is a positive integer, is a unit of 1, and is the summation index; iterate through the set. ; For the first The intensity of a non-homogeneous spatial point process at the center of a grid is a scalar, with units of . From the previous formula The value is obtained at the center of the grid and used as a weighting coefficient; For the first The horizontal coordinates of the center of each grid are scalars, in units of . The lateral position of the grid is determined by the differential calibrated coordinate system. The lateral position of the settlement trough axis is obtained by calculating the lateral position of the settlement trough axis for all mileage sections.

[0055] The lateral migration rate of the settling tank is obtained by subtracting the lateral positions of the settling tank axis obtained from adjacent periods and dividing by the acquisition period. ,in: Mileage section The lateral migration rate of the settling trough at the location is a scalar, with units of . The value obtained from this formula represents the rate at which the settling trough axis migrates laterally between adjacent cycles. and The first The first collection cycle and the first Each data collection cycle is performed at the mileage section. The transverse position of the settling trough axis at the location is defined as in the previous formula. The unit is These are calculated from images acquired in two adjacent acquisition cycles, respectively. The time interval between adjacent acquisition cycles is denoted as , where is a scalar and its value range is . to The unit is The periodic imaging interval of the industrial camera determines the acquisition period; The sequence number of the acquisition cycle is a positive integer with a dimension of one, representing which acquisition cycle it is. The lateral migration rate of the settling trough is consistent with the migration of the spatial centroid of the non-homogeneous spatial point process intensity field between adjacent cycles, and the two corroborate each other.

[0056] Finally, the location of the settlement trough axis, the lateral migration rate of the settlement trough, along with the length, maximum width, direction, and the length growth rate and width expansion rate of each pavement crack during adjacent cycles, are output as image measurement results of crack geometric parameters and evolution rate. The above output is strictly limited to image measurement results and is intended for use by downstream roadbed maintenance engineers in conjunction with geological conditions for comprehensive judgment.

[0057] After completing the settlement trough inversion in step S4, the neighborhood of the settlement trough shoulder on both sides of the obtained settlement trough axis position is used as prior feedback to step S1. For the next cycle's pavement image, the grayscale threshold used to identify pavement cracks within the neighborhood of the settlement trough shoulder is lowered or the identification sensitivity is increased. The settlement trough shoulder is a high-incidence tension zone for early longitudinal hairline cracks, but the grayscale contrast of hairline cracks is weak, making them easily missed when a uniform grayscale threshold is used across the entire pavement. Therefore, the grayscale threshold is spatially adjusted within the shoulder neighborhood as follows:

[0058] ,in: For the next cycle at the road surface position The grayscale threshold used to identify road surface cracks is a scalar, with a value range of [value range missing]. to The dimensionless value is calculated by this formula and represents the grayscale threshold for crack identification at this location. The uniform grayscale threshold used for the remaining areas of the road surface is denoted as , which is a scalar and has a value range of . to It has one dimension and is calibrated according to the grayscale distribution of the entire road surface, representing the benchmark threshold. The maximum reduction in grayscale threshold within the neighborhood of the settling tank shoulder is denoted by , which is a scalar and has a value range of . to The dimensionless value is selected based on the grayscale contrast of early hairline cracks to characterize the degree of sensitivity improvement. Road surface location The horizontal coordinate is a scalar, and the unit is . Determined by the differential quasi-coordinate system; Mileage section The lateral position at the shoulder of the settling tank is a scalar, and the unit is . This is derived from the right side of the equation and represents the center of the shoulder band. Mileage section The transverse position of the settling trough axis at the location is defined in the same way as before; Let be the lateral offset of the shoulder of the settling tank relative to the axis of the settling tank, a scalar quantity, with a value range of . to The unit is Selected according to the width of the settling trough, it represents the distance between the shoulder and the axis, and the positive and negative signs indicate the two shoulder bands on both sides of the axis; Let be the lateral bandwidth of the neighborhood of the settling tank shoulder, and let be a scalar with a value range of . to The unit is Selected according to the width of the shoulder tension band, representing the range of effect of threshold reduction in the lateral direction; It is an exponential function with the natural constant as its base and a dimension of one. As can be seen from the above formula, the grayscale threshold reduction is the greatest and the recognition sensitivity is the highest at the center of the shoulder zone of the settling trough. It gradually recovers to the reference threshold towards both sides, thereby specifically improving the detection rate of early longitudinal hairline cracks in the shoulder zone where early cracks are most likely to occur, forming a closed loop between the inversion result of step S4 and the recognition link of step S1.

[0059] In one numerical embodiment, a collapsible loess road section was continuously imaged for 12 cycles using a vehicle-mounted industrial camera with a 7-day acquisition cycle. The camera imaging scale was calibrated to 1 pixel corresponding to 0.8 mm. The road surface spatial grid size was 0.5 m × 0.5 m, the concentration threshold was 0.40, the anisotropic core bandwidth along the strike was 1.2 m, the bandwidth perpendicular to the strike was 0.4 m, the uniform grayscale threshold was 120, the maximum threshold reduction at the shoulder was 25, the lateral offset at the shoulder was 1.5 m, and the shoulder bandwidth was 0.8 m. In this embodiment, the lateral deviation between the settlement trough axis position obtained by the inversion of this invention and the settlement trough axis position obtained by on-site leveling measurement was within 0.6 m, and the relative deviation between the lateral migration rate of the settlement trough obtained by the inversion and the on-site observation value was within 12%.

[0060] As a comparison, the traditional paradigm of matching crack identities frame by frame and quantifying the length growth rate and width expansion rate of each crack was used to process the same set of images. As settlement progressed to the 8th cycle, a large number of new cracks and directional reconstructed cracks were generated. The success rate of crack inter-frame matching in the traditional paradigm dropped from approximately 90% initially to approximately 55%. Furthermore, because it discarded new cracks and directional reconstructed cracks as matching errors, it could not provide the location of the settlement trough axis and the lateral migration rate. Regarding the detection of early hairline cracks, the detection rate of early longitudinal hairline cracks was approximately 68% without closed-loop feedback, but increased to approximately 89% after enabling prior feedback on the settlement trough shoulder. The above comparison shows that the present invention can still stably output the spatial distribution and lateral migration of the settlement trough under conditions of non-conservation of crack identities, and significantly improve the detection rate of early cracks.

[0061] like Figure 2 As shown, this embodiment also provides a system for tracking the image evolution of settlement cracks in collapsible loess subgrade. This system includes a crack identification module, a registration and decomposition module, a field construction module, and a settlement trough inversion module connected in sequence. The output of the settlement trough inversion module is fed back to the crack identification module, forming a closed loop. Each module corresponds one-to-one with each step in the aforementioned method embodiment. The following describes each module; the specific processing performed by the module is the same as the processing of the corresponding steps in the aforementioned method embodiment, and the formula details will not be repeated.

[0062] The crack identification module is used to identify pavement cracks in periodically acquired images of collapsible loess road sections, extracting the crack skeleton and direction of each crack. This module receives pavement images obtained through periodic high-resolution imaging by vehicle-mounted or fixed industrial cameras, performs illumination homogenization and contrast enhancement on the images, and extracts the crack skeleton using a combination of directional gradient field and connected domain growth based on the low-grayscale, elongated connected domain characteristics of the pavement cracks. It also quantifies the length, maximum width, and direction of each pavement crack, processing the same as step S1 in the aforementioned method embodiment. The crack identification module also receives the settlement trough shoulder neighborhood from the settlement trough inversion module, lowering the grayscale threshold used for crack identification or increasing the identification sensitivity within this neighborhood. The crack identification module outputs the crack skeleton and direction to the registration decomposition module and the field construction module.

[0063] The registration decomposition module is used to register pavement images of adjacent periods based on settlement-invariant reference features in the pavement that do not evolve with subsidence. The registered transformation is decomposed into rigid deformation components representing the overall translation and rotation of the pavement and differential deformation components representing the differential settlement of the subgrade. The rigid deformation components are removed, and the differential deformation components are retained. A differential registration coordinate system is established using the registration result with retained differential deformation components as the geometric datum. This module extracts control points on settlement-invariant reference features in adjacent period images and establishes correspondences. The rigid deformation components are solved using constrained weighted least squares to obtain the differential deformation components. Multiple baseline constraints and zero-mean constraints on the differential deformation components are applied along the route direction in continuous longitudinal settlement sections, the processing being the same as step S2 in the aforementioned method embodiment. The registration decomposition module outputs the differential registration coordinate system with retained differential deformation components to the field construction module.

[0064] The field construction module is used to divide the road surface into a road surface spatial grid in the differential calibration coordinate system, aggregate and construct a crack orientation tensor field on the road surface spatial grid based on crack orientation, and construct a non-homogeneous spatial point process intensity field based on the location of newly appearing road surface cracks in adjacent periods. This module divides the road surface into a road surface spatial grid, aggregates the crack orientations within each grid into an orientation tensor weighted by the skeleton segment length, and calculates the dominant orientation direction and orientation concentration. Grids with orientation concentration below the concentration threshold and crack connectivity density above the connectivity density threshold are identified as grids dominated by network fatigue cracks. Based on the isotropic and consistent migration of the spatial centroid, these grids are further divided into settlement-induced network crack convergence areas and traffic load network fatigue crack areas. An anisotropic kernel function with the kernel major axis along the dominant orientation direction is used to estimate the intensity of newly appearing road surface cracks to obtain the non-homogeneous spatial point process intensity field, which is processed in the same way as step S3 of the aforementioned method embodiment. When constructing the above two fields, the field construction module retains newly appearing road surface cracks and road surface cracks undergoing orientation reorganization during adjacent periods as signals reflecting the geometry of the subsidence settlement trough and participates in the construction. The field construction module outputs the crack orientation tensor field and the non-homogeneous spatial point process intensity field to the settlement trough inversion module.

[0065] The settlement trough inversion module is used to invert the settlement trough axis position and lateral migration rate of the settlement trough by analyzing the orientation transition position between the longitudinally dominant and laterally dominant regions in the crack orientation tensor field, as well as the migration of the spatial centroid of the non-homogeneous spatial point process intensity field between adjacent periods. The module outputs image measurement results of crack geometric parameters and evolution rate. This module calculates the lateral position of the settlement trough axis at each mileage section and obtains the lateral migration rate of the settlement trough by dividing the difference in lateral position of the settlement trough axis between adjacent periods by the acquisition period. This processing is the same as step S4 in the aforementioned method embodiment. The settlement trough inversion module uses the settlement trough shoulder neighborhood on both sides of the inverted settlement trough axis position as prior feedback to the crack identification module, forming a closed loop.

[0066] In the above system embodiments, each module can be implemented by a program deployed on an industrial computer or server. The modules exchange data such as crack skeleton, crack orientation, differential calibration coordinate system, crack orientation tensor field, non-homogeneous spatial point process intensity field, and settlement trough axis position via a data interface. The system embodiments and method embodiments of the present invention are based on the same inventive concept, and their beneficial effects are the same as those of the method embodiments, and will not be repeated here.

[0067] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for tracking the image evolution of settlement cracks in collapsible loess subgrade, characterized in that, include: Identify pavement cracks in periodically collected images of collapsible loess road sections, and extract the crack skeleton and crack direction for each pavement crack. Using a settlement-invariant reference feature in the road surface that does not evolve with subsidence as a benchmark, the road surface images of adjacent periods are registered. The registration transformation obtained by registration is decomposed into rigid body deformation components that characterize the overall translation and rotation of the road surface and differential deformation components that characterize the differential settlement of the subgrade. The rigid body deformation components are removed and the differential deformation components are retained. A differential registration coordinate system is established based on the registration result after retaining the differential deformation components as the geometric benchmark. In the differential calibrated coordinate system, the road surface is divided into a road surface spatial grid. Based on the crack orientation, a crack orientation tensor field is constructed by aggregating the crack orientation on the road surface spatial grid. A non-homogeneous spatial point process intensity field is constructed based on the position of newly appearing road surface cracks in adjacent periods. The location of the orientation transition between the longitudinal and transverse orientation dominant regions in the crack orientation tensor field, and the migration of the spatial centroid of the non-homogeneous spatial point process intensity field between adjacent periods, are used to invert the position of the sinking trough axis and the transverse migration rate of the sinking trough, and output as image measurement results of crack geometric parameters and evolution rate.

2. The method for tracking the image evolution of settlement cracks in collapsible loess subgrade according to claim 1, characterized in that, When constructing the crack orientation tensor field and the non-homogeneous spatial point process intensity field, newly appearing pavement cracks and pavement cracks undergoing orientation reorganization during adjacent cycles are retained as signals reflecting the geometry of the collapsible settlement trough and included in the construction. Specifically, the crack orientation concentration and crack connectivity density are calculated for each pavement spatial grid, where the crack connectivity density is the ratio of the number of intersection nodes of the crack skeleton within the pavement spatial grid to the area of ​​the pavement spatial grid. Pavement spatial grids with crack orientation concentration below a concentration threshold and crack connectivity density above a connectivity density threshold are identified as grids dominated by network fatigue cracks. In the fatigue crack-dominated grid, a network fatigue crack-dominated grid is identified as a settlement-induced network crack convergence area if the migration vector of the spatial centroid of the crack skeleton between adjacent periods is in the same direction as the migration vector of the spatial centroid of the non-homogeneous spatial point process intensity field between adjacent periods, and the absolute value of the difference between their migration amounts is less than the migration consistency threshold. A network fatigue crack-dominated grid is identified as a traffic load network fatigue crack area and is removed if the migration amount of the spatial centroid of the crack skeleton between adjacent periods is less than the position stability threshold and the grid center falls within the road wheel track. The settlement-induced network crack convergence area is used as a candidate location area for the settlement trough axis position.

3. The method for tracking the image evolution of settlement cracks in collapsible loess subgrade according to claim 2, characterized in that, When the collapsible loess road section is in a continuous longitudinal settlement zone, causing the settlement-unchanged reference feature to undergo overall translation and rotation, multiple baseline constraints along the route direction and zero mean constraints of the differential deformation component at the road section scale are applied in the decomposition of the registration transformation to decouple the rigid deformation component from the differential deformation component.

4. The method for tracking the image evolution of settlement cracks in collapsible loess subgrade according to claim 3, characterized in that, When constructing the crack orientation tensor field, the orientation of cracks in each of the road surface spatial grids is statistically estimated using direction to determine the main orientation direction and the concentration of the crack orientation, and the dominant grid of the network fatigue crack is not included in the determination of the orientation transition position.

5. The method for tracking the image evolution of settlement cracks in collapsible loess subgrade according to claim 4, characterized in that, When constructing the non-homogeneous spatial point process intensity field, an anisotropic kernel function with the kernel major axis along the main direction of the strike is used to estimate the intensity of newly appearing pavement cracks in adjacent periods, so that the non-homogeneous spatial point process intensity field is anisotropically distributed along the main direction of the strike.

6. The method for tracking the image evolution of settlement cracks in collapsible loess subgrade according to claim 5, characterized in that, The settlement trough shoulder neighborhood on both sides of the settlement trough axis obtained by inversion is used as a priori and fed back to the step of identifying road surface cracks in the next cycle of road surface image. In the settlement trough shoulder neighborhood, the grayscale threshold used to identify road surface cracks is reduced or the identification sensitivity is increased in order to identify early longitudinal hairline cracks that were missed when a uniform grayscale threshold was used in the rest of the road surface area.

7. The method for tracking the image evolution of settlement cracks in collapsible loess subgrade according to claim 1, characterized in that, Identifying road surface cracks includes: performing illumination homogenization and contrast enhancement on the road surface image; extracting the crack skeleton by combining directional gradient field and connected domain growth based on the low grayscale elongated connected domain features of the road surface cracks; and quantifying the length, maximum width, and direction of each road surface crack.

8. The method for tracking the image evolution of settlement cracks in collapsible loess subgrade according to claim 1, characterized in that, It also includes classifying road surface cracks into three categories based on the crack direction: longitudinal settlement cracks, transverse settlement cracks, and network fatigue cracks, and performing statistics on the location of longitudinal settlement cracks along the route and the distribution spacing of transverse settlement cracks.

9. The method for tracking the image evolution of settlement cracks in collapsible loess subgrade according to claim 1, characterized in that, The road surface images are obtained by periodic high-resolution imaging of the road surface using a vehicle-mounted or fixed industrial camera, with the time interval between two adjacent images being the acquisition period; the image measurement results are strictly limited to crack geometric parameters and evolution rate.

10. A system for tracking the image evolution of settlement cracks in collapsible loess subgrade, used to implement the method for tracking the image evolution of settlement cracks in collapsible loess subgrade as described in any one of claims 1-9, characterized in that, include: The crack recognition module is used to identify road surface cracks in periodically acquired images of collapsible loess road sections and extract the crack skeleton and crack direction of each road surface crack. The registration decomposition module is used to perform registration on pavement images of adjacent periods based on a settlement-invariant reference feature in the pavement that does not evolve with the subsidence settlement. The registration transformation obtained by registration is decomposed into rigid deformation components that characterize the overall translation and rotation of the pavement and differential deformation components that characterize the differential settlement of the subgrade. The rigid deformation components are removed and the differential deformation components are retained. The differential registration coordinate system is established based on the registration result after retaining the differential deformation components as the geometric reference. The field construction module is used to divide the road surface into a road surface spatial grid in the differential calibrated coordinate system, aggregate and construct a crack direction tensor field on the road surface spatial grid based on the crack direction, and construct a non-homogeneous spatial point process intensity field based on the position of newly appearing road surface cracks in adjacent periods. The settlement trough inversion module is used to invert the settlement trough axis position and the settlement trough lateral migration rate by taking the orientation conversion position between the longitudinal orientation dominant region and the transverse orientation dominant region in the crack orientation tensor field, and the migration of the spatial centroid of the non-homogeneous spatial point process intensity field between adjacent periods, and outputs the image measurement results of crack geometric parameters and evolution rate.

Images