A method, medium and system for three-dimensional quantitative analysis of infrastructure surface millimeter-scale cracks

Abstract

This invention provides a three-dimensional quantitative analysis method, medium, and system for millimeter-level cracks on infrastructure surfaces, belonging to the field of crack detection technology. The invention identifies crack contours using a minimum cut-maximum flow-based crack-background segmentation algorithm, extracts the crack skeleton centerline using an ant colony optimization algorithm, and performs network connectivity and branch identification of the crack network through a crack topology connectivity optimization model. Cracks are classified according to their direction angle and length characteristics. Inclined laser scanning lines are set in the crack cross-section area to form cross-scan coverage. Weak echo signals are amplified by adaptive gain and supplemented with neighborhood depth interpolation to complete missing depth data. This accurately calculates the cross-sectional crack width and depth, solving the technical problem of inaccurate crack depth measurement due to missing bottom point cloud data caused by laser occlusion in deep and narrow cracks.

Description

Technical Field

[0001] This invention belongs to the field of crack detection technology, and specifically relates to a three-dimensional quantitative analysis method, medium and system for millimeter-level cracks on infrastructure surfaces. Background Technology

[0002] Surface crack detection in infrastructure is a crucial basis for pavement maintenance decisions. Traditional 3D quantitative analysis of cracks uses vertical laser scanning to acquire pavement point cloud data, image segmentation algorithms to identify crack contours, and then calculates crack width and depth parameters based on point cloud elevation information. In existing applications, vertical laser beams can effectively capture the geometric features of shallow cracks when irradiated onto the pavement. However, for deep and narrow cracks, the laser beam is blocked by the crack's inner wall, resulting in weakened or even missing echo signals at the crack bottom. Existing techniques typically use neighborhood interpolation or empirical formulas to estimate the depth of the missing portion. However, this method relies on the statistical characteristics of the surrounding point cloud, and for cracks with complex inner wall morphology or large depth-to-width ratios, the interpolation results deviate significantly from the true depth. In other words, existing technologies suffer from the technical problem of inaccurate crack depth measurement due to missing bottom point cloud data caused by laser occlusion in deep and narrow cracks. Summary of the Invention

[0003] In view of this, the present invention provides a three-dimensional quantitative analysis method, medium and system for millimeter-level cracks on infrastructure surfaces, which can solve the technical problem in the prior art that the depth of deep and narrow cracks cannot be accurately measured due to the lack of bottom point cloud data caused by laser occlusion.

[0004] The present invention is implemented as follows: The first aspect of the present invention provides a three-dimensional quantitative analysis method for millimeter-level cracks on infrastructure surfaces. A high-frequency line structured light sensor is used to continuously scan the infrastructure surface to acquire a 1mm resolution laser point cloud elevation matrix, 3D depth image, and 2D grayscale image. An inertial navigation system is used to collect the attitude parameters of the scanning carrier, establishing a timestamp-to-spatial position mapping function for motion compensation and registration correction. A pixel-level undirected graph is constructed, and a crack-background segmentation algorithm based on minimum cut-maximum flow is used to identify the crack and background segmentation contours. An ant colony optimization algorithm is used to extract the crack skeleton centerline. A crack topology connectivity optimization model is used to output connectivity decision parameters and branch identification parameters for crack network connectivity and branch identification. Cracks are classified according to their direction angle and length characteristics. An auxiliary laser scanning line is set at a ±15 degree tilt angle in the scanning carrier's driving direction in the crack cross-section area to form cross-scan coverage. For weak echo signals, adaptive gain amplification is used, combined with neighborhood depth interpolation to complete missing depth data. The crack width and crack depth of the cross-section are calculated.

[0005] Specifically, the step of establishing the timestamp-to-spatial-position mapping function involves the inertial navigation system acquiring the position coordinates, velocity vector, attitude angle, and angular velocity data of the scanning carrier at a frequency of 1000Hz, and performing position and attitude interpolation calculations on the timestamp of the 1mm resolution laser point cloud elevation matrix sampling time using the cubic spline interpolation method.

[0006] Specifically, the motion compensation registration correction step involves obtaining the scanning carrier's position offset and attitude rotation based on the timestamp-to-spatial position mapping function, performing attitude rotation matrix transformation on the 1mm resolution laser point cloud elevation matrix coordinates to compensate for the scanning carrier's attitude change, and using the nearest neighbor matching method for spatial registration.

[0007] Specifically, the pixel-level undirected graph construction step involves treating each pixel of the 3D depth image and the 2D grayscale image as a graph node and establishing undirected edge connections between each graph node and its 8 neighboring pixels. The edge weights are the weighted sums of the absolute values ​​of the depth difference and the absolute values ​​of the grayscale difference, weighted by weight coefficients of 0.6 and 0.4, respectively.

[0008] The step of selecting crack seed points specifically involves identifying depth abrupt change points in the 3D depth image. Depth abrupt change points are pixels whose absolute depth difference with their 8 neighboring pixels is greater than 0.5 mm. Connectivity analysis is then performed on these depth abrupt change points to select those with a connected component area greater than [missing value]. The set of deep mutation points is used as the seed point set of the crack.

[0009] The step of selecting background seed points involves identifying uniform grayscale road surface regions in a 2D grayscale image. A uniform grayscale road surface region is a region where the pixel grayscale variance is less than 10. Pixels are randomly selected from the uniform grayscale road surface region as the set of background seed points.

[0010] The specific steps of the crack-background segmentation algorithm based on minimum cut-maximum flow are as follows: connect the crack seed point set to the source point and connect the background seed point set to the sink point. Set the weight of the connecting edge to infinity. Use the Boykov-Kolmogorov algorithm to solve for the maximum flow from the source point to the sink point. Obtain the pixel-level undirected graph minimum cut set according to the maximum flow minimum cut theorem. The boundary of the minimum cut set forms the crack and background segmentation contour.

[0011] The pheromone concentration is defined as the initial pheromone concentration of each pixel being inversely proportional to the shortest distance from the pixel to the crack and background segmentation contour. The shortest distance is obtained by calculating the Euclidean distance from the pixel to the boundary of the crack and background segmentation contour. The initial pheromone concentration value is dimensionless by dividing the reciprocal of the shortest distance by the maximum value of the reciprocal of the shortest distance.

[0012] Specifically, the steps of the probability transfer strategy involve the virtual ant calculating the transfer probability of 8 neighboring pixels when choosing the next movement direction at the current pixel. The transfer probability is proportional to the product of the pheromone concentration of the 8 neighboring pixels and the depth gradient heuristic function, which are weighted by weighting coefficients of 0.7 and 0.3, respectively.

[0013] The step of dynamically updating the path pheromone concentration involves enhancing the pheromone of the virtual ants' walking path pixels after each iteration. The pheromone increment is inversely proportional to the walking path length, and the pheromone concentration of all pixels is subject to evaporation attenuation with an attenuation coefficient of 0.9.

[0014] The input layer of the crack topology connectivity optimization model receives the coordinates of the endpoints of the crack skeleton centerline, the crack orientation angle, the mean depth, and the mean grayscale features. These features are mapped to a 256-dimensional embedding vector through a fully connected layer and then input to a multi-layer graph convolutional network module. The multi-layer graph convolutional network module contains four graph convolutional layers, followed by a network depth dynamic adjustment module based on random depth.

[0015] Among them, the network depth dynamic adjustment module based on random depth skips part of the graph convolutional layer during the training stage according to the random depth probability, and uses the complete 4-layer graph convolutional layer during the inference stage. The random depth probability is dynamically calculated based on the crack network density parameter and the average length parameter of the crack segment. The higher the crack network density parameter, the greater the random depth probability; the longer the average length parameter of the crack segment, the smaller the random depth probability.

[0016] The breakthrough distance threshold is obtained by multiplying the predicted breakthrough distance threshold in the breakthrough decision parameters by a safety factor of 1.2, and the path consistency threshold is obtained by multiplying the predicted path consistency threshold in the breakthrough decision parameters by a safety factor of 1.1.

[0017] The crack classification process involves the following steps: cracks with a direction angle within ±20 degrees of the scanning carrier's travel direction and a crack length greater than 10 cm are classified as longitudinal cracks; cracks with a direction angle within ±20 degrees of the direction perpendicular to the scanning carrier's travel direction and a crack length greater than 10 cm are classified as transverse cracks.

[0018] The adaptive gain amplification step specifically involves detecting and identifying weak echo signals with an intensity lower than a preset echo intensity threshold within the cross-sectional area. The preset echo intensity threshold is set to 30% of the normal echo intensity, and the gain coefficient is dynamically calculated based on the ratio of the weak echo signal intensity to the preset echo intensity threshold. The upper limit of the gain coefficient is set to 5 times.

[0019] The neighborhood depth interpolation step involves selecting pixels with valid depth data within the 8 neighborhoods of the pixel with missing depth data in the cross-sectional area, and calculating the average depth of the pixels with valid depth data within the 8 neighborhoods as the interpolation depth of the pixel with missing depth data.

[0020] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, are used to perform the above-described method for three-dimensional quantitative analysis of millimeter-level cracks on the surface of infrastructure.

[0021] A third aspect of the present invention provides a three-dimensional quantitative analysis system for millimeter-level cracks on the surface of infrastructure, comprising the aforementioned computer-readable storage medium, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0022] This invention proposes setting up auxiliary laser scanning lines with a ±15 degree tilt angle in the cross-sectional area perpendicular to the crack's direction of travel to form cross-scan coverage. For weak echo signals, adaptive gain amplification combined with neighborhood depth interpolation is used to complete missing depth data. This scheme uses multi-angle laser beams to illuminate the crack, with vertical and tilted laser beams entering the deep, narrow crack from different incident directions. When the vertical laser beam cannot reach the bottom of the crack due to obstruction by the inner wall, the tilted laser beam can bypass the obstruction area from a lateral angle, increasing the sampling probability of the crack bottom point cloud. Simultaneously, the adaptive gain amplification technology dynamically adjusts the amplification factor according to the echo signal intensity, restoring the weak echo signal caused by depth attenuation to a recognizable range, thus effectively acquiring the originally missing bottom depth data. Neighborhood depth interpolation is only used to complete a very small number of still missing data points, avoiding the cumulative error caused by large-scale reliance on interpolation estimation. In summary, this invention solves the technical problem mentioned in the background art where the bottom point cloud data of deep, narrow cracks is missing due to laser obstruction, making accurate measurement of crack depth impossible. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention.

[0024] Figure 2 This is an example image of laser point cloud data.

[0025] Figure 3 Example image of 3D depth image.

[0026] Figure 4 Example image of a 2D grayscale image.

[0027] Figure 5 This is a high-precision map for marking and identifying traditional surface cracks.

[0028] Figure 6 This is a map showing the accuracy of automatic image recognition at the millimeter level.

[0029] Figure 7 This is a diagram showing the connection lines between longitudinal and transverse cracks.

[0030] Figure 8 This is a diagram showing the treatment of longitudinal and transverse cracks.

[0031] Figure 9 This is a diagram for calculating the width of the longitudinal and transverse seams.

[0032] Figure 10 This is a diagram showing the calculation of the depth of the longitudinal and transverse seams.

[0033] Figure 11 This is a diagram showing the connection lines of the cracks.

[0034] Figure 12 This is a diagram showing the treatment of cracks and fissures.

[0035] Figure 13 This is a diagram showing the calculation of the width of the cracks.

[0036] Figure 14 This is a diagram showing the calculated depth of the cracks. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0038] like Figure 1 The diagram shown is a flowchart of a three-dimensional quantitative analysis method for millimeter-level cracks on the surface of infrastructure provided by the first aspect of the present invention. This method includes the following steps: S01. A high-frequency line structured light sensor is used to continuously scan the surface of the infrastructure, and simultaneously acquires a 1mm resolution laser point cloud elevation matrix, a 3D depth image and a 2D grayscale image. The attitude parameters of the scanning carrier are collected in real time through an inertial navigation system, a mapping function from timestamp to spatial position is established, and motion compensation registration correction is performed on the 1mm resolution laser point cloud elevation matrix, the 3D depth image and the 2D grayscale image. S02. Construct a pixel-level undirected graph based on the motion-compensated registration and correction of the 3D depth image and the 2D grayscale image. Treat each pixel of the 3D depth image and the 2D grayscale image as a graph node of the pixel-level undirected graph. Establish edge connections between adjacent pixels and calculate the edge weights of the edge connections. Select crack seed points in the 3D depth image and background seed points in the 2D grayscale image. Use a crack-background segmentation algorithm based on minimum cut-maximum flow to solve for the minimum cut set of the pixel-level undirected graph. The boundary of the minimum cut set is used as the segmentation contour between the crack and the background. S03. The pixels within the segmentation contour of the crack and the background are regarded as the ant colony optimization space. The pheromone concentration of each pixel is set. Multiple virtual ants start from the endpoints of the segmentation contour of the crack and the background and perform a walking path search according to the probability transfer strategy calculated by the pheromone concentration and the depth gradient heuristic function. The pheromone concentration of the path is dynamically updated. After multiple iterations, the set of paths with the highest pheromone concentration is used as the center line of the crack skeleton. S04. Input the crack skeleton centerline into the crack topology connectivity optimization model. The crack topology connectivity optimization model outputs connectivity decision parameters and branch identification parameters. Determine the connectivity distance threshold and orientation consistency threshold based on the connectivity decision parameters. Connect crack segments in the crack skeleton centerline whose distance is less than the connectivity distance threshold and whose orientation angle is less than the orientation consistency threshold. Mark the main crack and branch cracks based on the branch identification parameters. Filter crack segments whose length is less than the short crack threshold and branch cracks whose branch length is less than the branch threshold, and retain the effective crack topology network. S05. Cracks in the effective crack topology network are grouped according to crack orientation angle, crack length, and crack density characteristics. Cracks with an orientation angle within ±20° of the scanning carrier's travel direction and a crack length greater than 10cm are classified as longitudinal cracks. Cracks with an orientation angle within ±20° of the direction perpendicular to the scanning carrier's travel direction and a crack length greater than 10cm are classified as transverse cracks. Cracks with a multi-directional, interwoven distribution and an enclosed area less than 100 cm² are classified as transverse cracks. 2 )且裂缝密度大于4(m / m 2) Cracks are classified as crazing. Cracks with a total length greater than 3 times the diagonal length of the board and divided into 3 or more independent areas are classified as broken boards. The remaining cracks are classified as other cracks. S06. The classified cracks are divided into several crack segments with a length of 10cm. Sampling points are selected at 1cm intervals on each crack segment. The cross-sectional area perpendicular to the crack direction angle is extracted with the sampling point as the center. An auxiliary laser scanning line with a ±15° tilt angle of the scanning carrier in the direction of travel is set in the cross-sectional area to form cross-scan coverage. For weak echo signals, adaptive gain amplification is used and neighborhood depth interpolation is combined to complete the missing depth data. Based on the complete 1mm resolution laser point cloud elevation matrix in the cross-sectional area, the crack edge points are identified and the horizontal distance between the crack edge points is calculated as the cross-sectional crack width. The normal road surface point cloud on both sides of the crack is fitted to form a reference plane. The vertical distance between the bottom point of the crack and the reference plane is calculated as the cross-sectional crack depth. The average cross-sectional crack width and cross-sectional crack depth of all sampling points in the crack segment are averaged to obtain the segmented average crack width and segmented average crack depth.

[0039] The steps for establishing the mapping function from timestamp to spatial location specifically include: the inertial navigation system acquiring the position coordinates, velocity vector, attitude angle, and angular velocity data of the scanning carrier at a frequency of 1000Hz; performing position and attitude interpolation calculations on the timestamp of each 1mm resolution laser point cloud elevation matrix sampling time; using cubic spline interpolation for position and attitude interpolation calculations; and outputting the interpolated position coordinates and interpolated attitude angle corresponding to the timestamp as the mapping function from timestamp to spatial location, thereby establishing the mapping relationship from timestamp to spatial location.

[0040] The motion compensation registration correction steps specifically include: obtaining the scanning carrier position offset and scanning carrier attitude rotation at each sampling time of the 1mm resolution laser point cloud elevation matrix based on the mapping function from timestamp to spatial position; translating the origin of the 1mm resolution laser point cloud elevation matrix coordinate system to the initial position of the scanning carrier; performing attitude rotation matrix transformation on the 1mm resolution laser point cloud elevation matrix coordinates to compensate for the attitude change of the scanning carrier; and spatially registering the compensated 1mm resolution laser point cloud elevation matrix coordinates with the corresponding 2D grayscale image pixel coordinates. The spatial registration process uses the nearest neighbor matching method, and the spatial registration error is controlled within 0.5mm.

[0041] The specific steps for constructing the pixel-level undirected graph include: treating each pixel of the 3D depth image and the 2D grayscale image as a graph node of the pixel-level undirected graph; establishing undirected edge connections between each graph node and its 8 neighboring pixels; calculating the absolute values ​​of the depth difference and grayscale difference between adjacent pixels; performing dimensionless processing by dividing the absolute value of the depth difference by the standard value of the depth difference and the absolute value of the grayscale difference by the standard value of the grayscale difference; taking the standard value of the depth difference as 1 mm and the standard value of the grayscale difference as 20; and performing a weighted summation of the dimensionless absolute values ​​of the depth difference and the grayscale difference using weighting coefficients of 0.6 and 0.4, respectively, as the edge weights. The weighting coefficients are adjusted according to the complexity of the actual road surface texture to construct a complete pixel-level undirected graph.

[0042] The specific steps for selecting crack seed points include: identifying depth abrupt change points in the 3D depth image, whereby a depth abrupt change point is defined as a pixel whose absolute value of the depth difference with its 8 neighboring pixels is greater than 0.5 mm; performing connected component analysis on the depth abrupt change points; and selecting a set of depth abrupt change points with a connected component area greater than 5 (mm^2) as the crack seed point set.

[0043] The specific steps of selecting background seed points include: identifying road surface areas with uniform grayscale in a 2D grayscale image, where a road surface area with uniform grayscale is defined as an area where the pixel grayscale variance is less than 10, and randomly selecting pixels within the road surface area with uniform grayscale as a set of background seed points.

[0044] The edge weight is defined as a weighted combination of the absolute values ​​of the depth difference and the grayscale difference between pixels. The weight combination method is to multiply the dimensionless absolute value of the depth difference by 0.6 and the dimensionless absolute value of the grayscale difference by 0.4, and then sum the two to obtain the edge weight value.

[0045] The specific steps of the crack-background segmentation algorithm based on minimum cut-maximum flow include: connecting the crack seed point set to the source point with the weight of the connecting edge set to infinity; connecting the background seed point set to the sink point with the weight of the connecting edge set to infinity; using the Boykov-Kolmogorov algorithm to solve for the maximum flow from the source point to the sink point in the pixel-level undirected graph; obtaining the minimum cut set of the pixel-level undirected graph according to the maximum flow-minimum cut theorem; the two endpoints of the cut edge in the minimum cut set belong to the crack region and the background region, respectively; and the boundary of the minimum cut set forms the segmentation contour between the crack and the background.

[0046] The proposed minimum-cut-maximum-flow crack-background segmentation algorithm transforms the crack identification problem into a graph-theoretic minimum-cut problem. By constructing a pixel-level undirected graph and defining edge weights as a weighted combination of the absolute values ​​of depth and grayscale differences, it achieves globally optimized segmentation of the crack and background. The core principle of this algorithm lies in the maximum-flow minimum-cut theorem, which guarantees the optimality of the segmentation boundary. Specifically, the sum of edge weights corresponding to the segmentation boundary is minimized. This means the segmentation contour of the crack and background is located at the position where the absolute values ​​of depth and grayscale differences change most drastically, thus accurately capturing the crack contour. Compared to traditional threshold-based edge detection methods, the minimum-cut-maximum-flow crack-background segmentation algorithm optimizes the segmentation results globally, exhibiting stronger robustness to noise and local blurring, and avoiding the subjectivity of threshold selection and boundary breakage problems caused by local optimization. In textured road surface scenarios, although interference features such as aggregates and joints have certain absolute values ​​of depth difference and grayscale difference, their edge weight distribution does not form a continuous high-weight path. The crack-background segmentation algorithm based on minimum cut-maximum flow can effectively distinguish between real cracks and texture interference, significantly improving the accuracy and completeness of crack identification.

[0047] The pheromone concentration is defined as follows: the initial pheromone concentration of each pixel is inversely proportional to the shortest distance from the pixel to the segmentation contour of the crack and the background. The shortest distance is obtained by calculating the Euclidean distance from the pixel to the boundary of the segmentation contour of the crack and the background. The initial pheromone concentration value is dimensionless by dividing the reciprocal of the shortest distance by the maximum value of the reciprocal of the shortest distance. The initial pheromone concentration of pixels in the central region of the segmentation contour of the crack and the background is high, and the initial pheromone concentration of pixels in the edge region of the segmentation contour of the crack and the background is low.

[0048] The specific steps of the probability transfer strategy include: when the virtual ant selects the next movement direction at the current pixel, it calculates the transfer probability of the 8 neighboring pixels. The transfer probability is proportional to the product of the pheromone concentration of the 8 neighboring pixels and the depth gradient heuristic function. The depth gradient heuristic function is defined as the absolute value of the depth difference between the current pixel and the 8 neighboring pixels divided by the maximum absolute value of the depth difference, and then processed without dimension. The pheromone concentration and the depth gradient heuristic function are weighted by weight coefficients of 0.7 and 0.3, respectively. The weighted values ​​of the 8 neighboring pixels are normalized to obtain the transfer probability distribution. The virtual ant randomly selects the next movement direction according to the transfer probability distribution.

[0049] The step of dynamically updating the pheromone concentration of the path specifically includes: after all virtual ants complete the path search in each iteration, the pixels of the path traversed by the virtual ants are enhanced with pheromone. The pheromone increment is inversely proportional to the length of the virtual ant's path. The shorter the path, the greater the pheromone increment. At the same time, the pheromone concentration of all pixels is evaporated and attenuated, with an evaporation attenuation coefficient of 0.9. After multiple iterations, the pheromone concentration of pixels on frequently traversed paths gradually accumulates, while the pheromone concentration of pixels on low-frequency traversed paths gradually disappears. The set of paths with the highest pheromone concentration is the center line of the crack skeleton.

[0050] The ant colony optimization-based crack skeleton extraction and refinement algorithm treats the pixels within the segmentation contour of the crack and background as the foraging space of an ant colony. It utilizes the global search capability and positive feedback mechanism of the ant colony algorithm to extract the centerline of the crack skeleton. The principle of this algorithm lies in the fact that the centerline of the crack skeleton is the optimal path connecting the two ends of the segmentation contour of the crack and background, corresponding to the shortest path from the anthill to the food source in ant foraging. By setting the pheromone concentration to be inversely proportional to the shortest distance from the pixel to the boundary of the segmentation contour of the crack and background, virtual ants are guided to preferentially roam in the central region within the segmentation contour of the crack and background. Combined with a depth gradient heuristic function, the virtual ants move along a direction of continuously changing depth, avoiding crossing the boundary of the segmentation contour of the crack and background. After multiple iterations, the roaming path along the centerline of the crack skeleton accumulates a high concentration of pheromone due to the passage of a large number of virtual ants, forming a clear centerline contour of the crack skeleton. Compared with traditional skeleton extraction methods based on morphological refinement, the crack skeleton extraction and refinement algorithm based on ant colony optimization does not rely on fixed structural elements and has stronger adaptability to crack width changes and topological branches. The extracted crack skeleton centerline maintains a single pixel width and completely preserves the topological connectivity of the crack, avoiding the breakpoint and burr problems caused by morphological refinement. It provides an accurate centerline reference for subsequent analysis of cross-sectional crack width and depth.

[0051] The structure of the crack topology connectivity optimization model is as follows: The input layer receives the endpoint coordinates, crack orientation angle, mean depth, and mean grayscale features of the crack skeleton centerline. The feature dimension is 128-dimensional. These features are mapped to a 256-dimensional embedding vector through a fully connected layer. The embedding vector is then input to a multi-layer graph convolutional network module, which contains four graph convolutional layers. Each graph convolutional layer models the relationship between the endpoints of the crack skeleton centerline and aggregates the feature information of neighboring crack segments. After the four graph convolutional layers, a network depth dynamic adjustment module based on random depth is connected. During the training phase, some graph convolutional layers are skipped according to the random depth probability. During the inference phase, the complete four graph convolutional layers are used for forward propagation. The crack network density parameter and the average length parameter of the crack segment are dynamically calculated. The crack network density parameter is defined as the number of pixels on the center line of the crack skeleton per unit area. The higher the crack network density parameter, the greater the probability of random depth. The average length parameter of the crack segment is defined as the arithmetic mean of the lengths of all center line segments of the crack skeleton. The longer the average length parameter of the crack segment, the smaller the probability of random depth. The multilayer graph convolutional network module outputs a 256-dimensional feature vector. The 256-dimensional feature vector outputs the penetration decision parameter and the branch recognition parameter through two parallel fully connected branches. The penetration decision parameter includes the predicted value of the penetration distance threshold and the predicted value of the orientation consistency threshold. The branch recognition parameter includes the confidence of the main crack and the confidence of the branch crack.

[0052] The steps for establishing the training dataset for the crack topology connectivity optimization model specifically include: collecting 1mm resolution laser point cloud elevation matrices, 3D depth images, and 2D grayscale image data for different pavement types and crack morphologies; performing crack identification and skeleton extraction processing on the data; manually labeling the connectivity and branching relationships of the centerline segments of each crack skeleton; recording the pairs of crack skeleton centerline segments that need to be connected and the optimal connectivity distance when labeling connectivity; recording the attribution relationship between the main crack and the branch cracks when labeling branching relationships; converting the labeled data into graph structure data; using each crack skeleton centerline segment as a graph node; establishing edge connections between potentially connected crack skeleton centerline segment pairs; graph node features including endpoint coordinates, crack direction angle, mean depth, and mean grayscale; edge connection features including the distance between crack skeleton centerline segments and the angle between crack direction angles; and constructing a training dataset containing 5000 crack network samples.

[0053] The specific steps for training the crack topology connectivity optimization model include: dividing the training dataset into a training set and a validation set, with the training set accounting for 80% and the validation set accounting for 20%; using the cross-entropy loss function to optimize the prediction accuracy of connectivity decision parameters and branch recognition parameters; calculating crack network density parameters and average crack segment length parameters for each sample during training; dynamically calculating random depth probabilities based on crack network density parameters and average crack segment length parameters; randomly skipping some graph convolutional layers during the training phase according to the random depth probabilities, with the skipped graph convolutional layers directly passed to the next graph convolutional layer; using a complete 4-layer graph convolutional layer for forward propagation during the inference phase; using the Adam optimizer for gradient descent optimization; setting the learning rate to 0.001, the batch size to 32, and the number of training epochs to 100; evaluating the connectivity accuracy and branch recognition accuracy of the crack topology connectivity optimization model on the validation set; and selecting the crack topology connectivity optimization model parameters with the best performance on the validation set as the final crack topology connectivity optimization model.

[0054] The proposed stochastic depth-based dynamic network depth adjustment module effectively mitigates the vanishing gradient problem in deep network training by randomly skipping some graph convolutional layers during the training phase and using full graph convolutional layers during the inference phase. The core principle of this module lies in the stochastic depth mechanism, which randomly selects different network depths in each training iteration. This allows both shallow and deep networks to participate in training simultaneously. Shallow networks have shorter gradient propagation paths, making them less prone to vanishing gradients, while deep networks have stronger expressive power and can learn more complex features. This combination avoids the limitations of a single-depth network. In crack topology connectivity tasks, the complexity of crack network structures varies greatly. Simple longitudinal or transverse cracks can be accurately identified using only shallow networks, while cracks or broken plates require deep networks to capture multi-scale topological features. The stochastic depth mechanism dynamically adjusts the skipping probability based on the crack network density parameter and the average crack segment length parameter. For networks with high crack network density parameters and short average crack segment length parameters, the skipping probability is increased to introduce more shallow paths and accelerate training convergence. For networks with low crack network density parameters and long average crack segment length parameters, the skipping probability is decreased to retain deep paths and improve feature representation capabilities. The network depth dynamic adjustment module based on random depth significantly improves the training efficiency and generalization performance of the crack topology connection optimization model, enabling the crack topology connection optimization model to adapt to crack networks of different complexities and accurately output connection decision parameters and branch identification parameters.

[0055] The breakthrough distance threshold and the orientation consistency threshold are determined based on the breakthrough decision parameters. The breakthrough distance threshold is obtained by multiplying the predicted breakthrough distance threshold value in the breakthrough decision parameters by a safety factor of 1.2, and the orientation consistency threshold is obtained by multiplying the predicted orientation consistency threshold value in the breakthrough decision parameters by a safety factor of 1.1. The safety factor is used to increase the robustness of the breakthrough connection decision and avoid misjudgment in boundary cases.

[0056] The crack orientation angle is defined as the directional angle of the center line segment of the crack skeleton. The crack orientation angle is obtained by calculating the angle between the line connecting the coordinates of the two ends of the center line segment of the crack skeleton and the travel direction of the scanning carrier.

[0057] The traveling direction of the scanning carrier is defined as the forward direction of the scanning carrier when the high-frequency line structured light sensing device scans. The traveling direction of the scanning carrier is determined by the direction of the scanning carrier's velocity vector collected by the inertial navigation system.

[0058] The specific steps of the through-connection include: traversing all endpoints of the center line segments of the crack skeleton in the effective crack topology network; for each endpoint of the center line segment of the crack skeleton, finding the nearest other endpoint of the center line segment of the crack skeleton in the fan-shaped search area in the crack direction angle direction, where the angle range of the fan-shaped search area is ±60° in the crack direction angle direction; calculating the straight-line distance and the crack direction angle between the endpoints of the two endpoints of the center line segments of the crack skeleton; determining whether the straight-line distance is less than the through-connection distance threshold and whether the crack direction angle is less than the direction consistency threshold; if both conditions are met, then the two crack skeleton center lines are through-connected, establishing a connecting line segment between the endpoints of the two crack skeleton center lines; the connecting line segment inherits the depth mean and grayscale mean attributes of the two crack skeleton center lines, forming a complete effective crack topology network.

[0059] The short crack threshold is set to 10cm, and the branch threshold is set to 1cm.

[0060] The steps for filtering short cracks and branch cracks specifically include: traversing all crack skeleton centerline segments in the effective crack topology network, calculating the length of each crack skeleton centerline segment, determining whether the length of the crack skeleton centerline segment is less than the short crack threshold, and if it is less than the short crack threshold, marking it as a noise crack for filtering; for crack skeleton centerline segments with a length greater than the short crack threshold, identifying their branch cracks, defined as crack skeleton centerline segments branching from the main crack node, calculating the length of the branch crack, determining whether the length of the branch crack is less than the branch threshold, and if it is less than the branch threshold, marking it as a noise branch for filtering; and retaining the effective crack topology network where the crack skeleton centerline segment length is greater than the short crack threshold and the branch crack length is greater than the branch threshold.

[0061] The crack length is defined as the geometric length of the center line segment of the crack skeleton, and the crack length is obtained by accumulating the Euclidean distance between adjacent pixels of the center line segment of the crack skeleton.

[0062] The crack density is defined as the total length of the centerline of the crack skeleton within a unit area. The crack density is obtained by summing the lengths of all the centerline segments of the crack skeleton within a unit area and dividing by the unit area value.

[0063] The steps for setting the auxiliary laser scanning lines specifically include: in the cross-sectional area perpendicular to the crack direction angle, setting two inclined laser scanning lines along the traveling direction of the scanning carrier, with inclination angles of +15° and -15° respectively. The inclined laser scanning lines and the vertical laser scanning lines form an intersecting scanning coverage. For deep and narrow cracks, the bottom echo signal is lost after the vertical laser beam enters the crack due to the blockage of the crack inner wall. The inclined laser beam enters the crack from different angles, increasing the sampling probability of the crack bottom point cloud and improving the integrity of the cross-sectional crack depth data.

[0064] The adaptive gain amplification step specifically includes: intensity detection of the laser echo signal in the cross-sectional area, identification of weak echo signals with intensity lower than a preset echo intensity threshold (set to 30% of normal echo intensity), amplification of the weak echo signal using an adaptive gain algorithm, with the gain coefficient dynamically calculated based on the ratio of the weak echo signal intensity to the preset echo intensity threshold (the lower the weak echo signal intensity, the larger the gain coefficient, with an upper limit of 5 times), and the amplified weak echo signal used to extract point cloud data at the bottom of the crack, improving the accuracy of crack depth measurement in deep and narrow crack cross-sections.

[0065] The neighborhood depth interpolation step specifically includes: for pixels in the cross-sectional area that still lack depth data, a neighborhood depth interpolation algorithm is used to complete the data. Pixels with valid depth data within an 8-neighborhood of the pixel with missing depth data are selected, and the average depth of the pixels with valid depth data within the 8-neighborhood is calculated as the interpolation depth of the pixel with missing depth data. If the number of pixels with valid depth data within the 8-neighborhood is less than 3, the neighborhood range is expanded to 16 or 24 neighborhoods until the number of pixels with valid depth data meets the interpolation requirements. The interpolation depth is used for subsequent calculations of the cross-sectional crack width and crack depth to ensure the integrity of the depth data within the cross-sectional area.

[0066] The specific steps of crack edge point identification include: scanning pixel by pixel in the cross-sectional area along a direction perpendicular to the crack direction angle, calculating the depth gradient of each pixel and its neighboring pixels, the depth gradient being defined as the absolute value of the depth difference between neighboring pixels, identifying pixels with a depth gradient greater than 0.3 mm as candidate crack edge points, performing connectivity analysis on the candidate crack edge points, and selecting two candidate crack edge points located on both sides of the crack with opposite depth gradient directions as crack edge points, the crack edge point being defined as the boundary position between the crack and the normal road surface.

[0067] The specific steps for calculating the width of the cross-section crack include: obtaining the coordinates of the crack edge points identified within the cross-section area; calculating the Euclidean distance between two crack edge points on the horizontal plane; the Euclidean distance value is the width of the cross-section crack; the accuracy of the cross-section crack width is 0.1 mm; for micro-cracks with a width less than 1 mm, the accuracy of the crack edge point coordinates is improved by using a sub-pixel edge positioning algorithm; the sub-pixel edge positioning uses a Gaussian fitting method to fit the depth gradient of the neighborhood of the crack edge point; and the peak position of the fitted curve is used as the accurate coordinates of the crack edge point.

[0068] The specific steps of the reference plane fitting include: selecting normal road surface point cloud data within a 5mm range on both sides of the crack in the cross-sectional area. Normal road surface point cloud data is defined as a point cloud area with a smooth depth change and no obvious abrupt change. The least squares method is used to perform plane fitting on the selected normal road surface point cloud data. The fitted plane equation is a plane equation in three-dimensional space. The normal vector of the plane equation is consistent with the vertical direction of the road surface. The fitted plane equation serves as the reference plane of the cross-sectional area and is used as a reference benchmark for calculating the depth of the cross-sectional crack.

[0069] The specific steps for calculating the crack depth in the cross-section include: identifying the bottom point of the crack within the cross-section area, where the bottom point is defined as the point cloud point with the largest depth value; calculating the three-dimensional coordinates of the bottom point; substituting the three-dimensional coordinates of the bottom point into the reference plane equation; calculating the vertical distance from the bottom point to the reference plane using the point-to-plane distance formula, where the three-dimensional coordinates of the bottom point and the coefficients of the reference plane equation are both in standard SI units; and the vertical distance is the crack depth in the cross-section, with an accuracy of 0.1 mm.

[0070] The steps for calculating the segmented average crack width and segmented average crack depth specifically include: performing an arithmetic mean calculation on the cross-sectional crack width data of all sampling points within each crack segment, with the arithmetic mean calculated to an average crack width value accurate to 0.1 mm; performing an arithmetic mean calculation on the cross-sectional crack depth data of all sampling points within each crack segment, with the arithmetic mean calculated to an average crack depth value accurate to 0.1 mm; and using the segmented average crack width and segmented average crack depth as quantitative parameters of the crack segment for crack damage assessment and maintenance decisions.

[0071] The specific implementation methods of the above steps are described in detail below.

[0072] The specific implementation of step S01 involves continuously scanning the infrastructure surface using a high-frequency line structured light sensor. First, the high-frequency line structured light sensor is mounted on a scanning carrier, which travels along the infrastructure surface at a preset speed. The high-frequency line structured light sensor scans the road surface line by line at a kilohertz sampling frequency, acquiring point cloud data for one transverse laser line each time. Continuous scanning forms a 1mm resolution laser point cloud elevation matrix covering the entire road surface area. The inertial navigation system collects the scanning carrier's position coordinates, velocity vector, attitude angle, and angular velocity data in real time at a high frequency of 1000Hz. This data is used to compensate for point cloud position deviations caused by bumps and tilting of the scanning carrier during travel. A cubic spline interpolation method is used to establish the mapping function from timestamp to spatial position. This method smoothly interpolates the discrete inertial navigation data by constructing a cubic polynomial function, ensuring the continuity and smoothness of the interpolation curve and enabling accurate calculation of the spatial position and attitude angle of the scanning carrier at any given time when the point cloud is sampled. The motion compensation registration and correction process adopts a coordinate transformation method. First, the origin of the point cloud coordinate system is translated according to the position offset of the scanning carrier. Then, a rotation matrix is ​​constructed according to the rotation of the scanning carrier attitude to rotate the point cloud coordinates. Finally, the nearest neighbor matching method is used to spatially register the transformed point cloud coordinates with the pixel coordinates of the 2D grayscale image to ensure that the spatial registration error is controlled within 0.5mm. This accuracy meets the requirements of millimeter-level crack analysis.

[0073] The specific implementation of step S02 is a process of segmenting cracks and background based on the registered 3D depth image and 2D grayscale image. First, a pixel-level undirected graph is constructed, with each pixel as a graph node. Edge connections are established between adjacent pixels to form a mesh-like graph structure. When calculating edge weights, the absolute values ​​of the depth difference and grayscale difference between adjacent pixels are calculated separately. These two differences are then dimensionless to eliminate the influence of dimensions. The standard value for the depth difference is 1 mm because millimeters are the accuracy requirement for road crack detection, and the standard value for the grayscale difference is 20 based on the grayscale variation range of common road textures. The dimensionless depth difference and grayscale difference are weighted and summed with weight coefficients of 0.6 and 0.4, respectively, to obtain the edge weights. The weight coefficient of 0.6 indicates the dominant role of depth features in crack identification. When selecting crack seed points, depth abrupt change points are identified. The depth abrupt change threshold of 0.5 mm is set based on the typical depth difference between cracks and normal road surfaces, and the connected component area threshold of 5 (mm²) is used to filter isolated noise points. The Boykov-Kolmogorov algorithm is used to solve the minimum cut problem. This algorithm is an efficient maximum flow algorithm. It constructs a flow network from the source to the sink and uses the augmenting path method to continuously increase the flow until the maximum flow is reached. According to the maximum flow minimum cut theorem, the cut set corresponding to the maximum flow is the optimal boundary for separating the crack from the background. The time complexity of this algorithm is linearly related to the number of edges and is suitable for processing large-scale pixel-level graph structures.

[0074] Step S03 is specifically implemented as a process of extracting the center line of the crack skeleton based on the ant colony optimization algorithm. First, the pheromone concentration is calculated based on the Euclidean distance from each pixel to the boundary of the crack and background segmentation contour. Pixels closer to the crack have lower initial pheromone concentrations, while pixels further away in the central region have higher initial pheromone concentrations. This setting guides the virtual ants to preferentially roam in the central region of the crack. The depth gradient heuristic function utilizes the continuity of depth within the crack to guide the virtual ants to move along a direction with a gentle depth change, avoiding crossing the crack boundary. When calculating the probability transition strategy, the pheromone concentration and the depth gradient heuristic function are weighted with weights of 0.7 and 0.3, respectively, representing the dominant role of pheromone concentration and the auxiliary role of depth gradient. The normalized transition probability distribution ensures that the sum of probabilities is 1. The virtual ant's path search uses a roulette wheel selection method, randomly selecting the next movement direction based on the transition probability distribution. The pheromone dynamic update mechanism includes two processes: pheromone enhancement and pheromone evaporation. The inverse relationship between pheromone increment and path length reflects the positive feedback mechanism of the ant colony algorithm, meaning that shorter paths acquire more pheromones, attracting more virtual ants. The evaporation decay coefficient of 0.9 indicates that 90% of the pheromone is retained after each iteration. This setting ensures the utilization of historical information while avoiding premature convergence. After multiple iterations, the set of paths with the highest pheromone concentration forms the centerline of the crack skeleton, which is one pixel wide and fully preserves the topological connectivity of the crack.

[0075] Step S04 is implemented by using a crack topology connectivity optimization model to identify crack connectivity and branches. The crack topology connectivity optimization model employs a graph convolutional network structure. The input layer receives the endpoint coordinates of the crack skeleton centerline, the crack direction angle, the mean depth, and the mean grayscale value. These features comprehensively characterize the spatial location, direction, and intensity information of the crack segment. The fully connected layer maps the 128-dimensional input features to a 256-dimensional embedding vector, achieving nonlinear transformation and dimensional expansion of the features. The multi-layer graph convolutional network module contains four graph convolutional layers. Each layer models the crack topology by aggregating the feature information of neighboring crack segments. The core of the graph convolution operation is the weighted summation of the features of the node's neighborhood, with the weights determined by the features connected by the edges. This mechanism enables the model to learn the spatial relationships and topological connection patterns between crack segments. A network depth dynamic adjustment module based on random depth dynamically calculates the random depth probability according to the crack network density parameter and the average length parameter of the crack segment. For complex networks with high-density short cracks, the skip probability is increased to accelerate training convergence; for simple networks with low-density long cracks, the skip probability is decreased to preserve the expressive power of deep features. The mechanism of randomly skipping some graph convolutional layers during training alleviates the gradient vanishing problem in deep networks, while using a full four-layer graph convolutional layer during inference ensures the model's complete expressive power. When establishing a connection, the system searches for the nearest endpoint of the centerline segment of another crack skeleton within a fan-shaped search region. The fan-shaped angle range of ±60° is set based on a reasonable deviation range of the crack orientation. The connection distance threshold is obtained by multiplying the predicted value by a safety factor of 1.2, which increases the robustness of the decision. A short crack threshold of 10cm and a branch threshold of 1cm are used to filter background noise and invalid branches, retaining cracks with practical maintenance significance.

[0076] Step S05 is specifically implemented as a process of classifying the effective crack topology network. The crack orientation angle is obtained by calculating the angle between the line connecting the two endpoints of the crack skeleton centerline segment and the travel direction of the scanning carrier. The travel direction of the scanning carrier is determined by the velocity vector direction acquired by the inertial navigation system. The criteria for determining a longitudinal crack are that the crack orientation angle is within ±20° of the travel direction and the crack length is greater than 10cm. This condition ensures that the identified longitudinal cracks are basically parallel to the travel direction and have a certain length. The criteria for determining a transverse crack are that the crack orientation angle is within ±20° of the direction perpendicular to the travel direction and the crack length is greater than 10cm. This condition ensures that the identified transverse cracks are basically perpendicular to the travel direction. The criteria for determining a crazing crack include multi-directional interlaced distribution and an enclosed area of ​​less than 100. and crack density greater than 4 These three characteristics comprehensively reflect the network distribution and high density of the cracks. The criterion for classifying a slab as broken is that the total crack length is greater than three times the diagonal length of the slab and it is divided into more than three independent regions; this condition reflects the severity of the damage. The classification process employs a rule-based method, determining each crack type sequentially according to priority. Cracks not classified into the above categories are categorized as other cracks.

[0077] Step S06 involves the quantitative analysis of the cross-sectional width and depth of the classified cracks. First, the cracks are divided into several segments of 10cm length. This 10cm segment length reflects local changes in the cracks while ensuring statistical stability. Sampling points are selected at 1cm intervals on each crack segment. This 1cm sampling interval captures subtle changes in crack width and depth. When extracting the cross-sectional area, an area approximately 10mm wide is extracted perpendicular to the crack's direction, centered on the sampling point. This area covers the crack and the surrounding normal road surface. An auxiliary laser scanning line is set with a ±15° tilt angle to create cross-scan coverage. This setting addresses the issue of missing echo signals at the bottom of deep and narrow cracks; the tilted laser beam entering the crack from different angles increases the sampling probability of the bottom point cloud. An adaptive gain amplification algorithm dynamically adjusts the gain coefficient based on the echo signal intensity. A preset echo intensity threshold of 30% of the normal echo intensity is used to identify weak echo signals, and the upper limit of the gain coefficient is 5 times to prevent excessive noise amplification. Neighborhood depth interpolation uses the mean interpolation method to complete the missing depth data. This method is simple, effective, and maintains the local continuity of the depth data. A depth gradient threshold of 0.3 mm is used to distinguish crack boundaries from normal road surfaces during crack edge point identification, a threshold set based on millimeter-level accuracy requirements. The cross-sectional crack width is calculated using the Euclidean distance formula to determine the horizontal distance between two crack edge points. For micro-cracks with a width less than 1 mm, a sub-pixel edge localization algorithm using Gaussian fitting is employed to improve accuracy. The reference plane fitting uses the least squares method to fit the point cloud of normal road surfaces within a 5 mm radius on both sides of the crack. The least squares method solves for the optimal plane parameters by minimizing the sum of squares of the point-to-plane distances, demonstrating good robustness to noise. The cross-sectional crack depth is calculated using the point-to-plane distance formula to determine the vertical distance between the bottom point of the crack and the reference plane; this distance reflects the true crack depth. The segmented average crack width and segmented average crack depth are calculated using arithmetic mean, both with an accuracy of 0.1 mm, meeting the data requirements for refined maintenance decision-making.

[0078] It should be noted that the key technical ideas of this invention include a crack background segmentation algorithm based on minimum cut-maximum flow, a crack skeleton extraction algorithm based on ant colony optimization, and crack topology connectivity optimization based on random depth graph convolutional networks. The crack background segmentation algorithm based on minimum cut-maximum flow transforms crack identification into a graph-theoretic minimum cut problem. Global optimization ensures the optimality of the segmentation boundary, exhibiting stronger noise resistance and more accurate boundary localization compared to traditional threshold segmentation methods. It can effectively distinguish between real cracks and background interference in complex textured road scenes. The crack skeleton extraction algorithm based on ant colony optimization utilizes the global search and positive feedback mechanism of the ant colony algorithm to extract the centerline of cracks with a single pixel width. Compared to traditional morphological thinning methods, it does not rely on fixed structural elements and has stronger adaptability to crack width variations and topological branches. The extracted skeleton completely preserves the crack topological connectivity without breaks or burrs. The crack topology connectivity optimization based on random depth graph convolutional networks alleviates the gradient vanishing problem by dynamically adjusting the network depth. It adaptively adjusts the model's expressive power according to the complexity of the crack network, exhibiting better training efficiency and generalization performance compared to fixed-depth networks. It can accurately identify the connectivity relationships of crack networks with different complexities. The synergistic effect of these three key technological approaches forms a complete technological chain from crack identification to skeleton extraction and topology optimization. In the crack identification stage, global optimization ensures boundary accuracy; in the skeleton extraction stage, adaptive search ensures topological integrity; and in the topology optimization stage, deep learning ensures the accuracy of the connection decision. The three work together to achieve precise quantitative analysis of millimeter-level cracks, which significantly improves the accuracy, integrity, and robustness of crack analysis compared to existing technologies.

[0079] It should be noted that this invention also solves the following technical problem: In existing crack recognition technologies, the aggregate and joint features of textured pavements are similar to real cracks in terms of depth and grayscale differences. Traditional threshold-based edge detection methods struggle to distinguish between cracks and texture interference, leading to a decrease in crack recognition accuracy. This invention constructs a pixel-level undirected graph and employs a crack-background segmentation algorithm based on minimum cut and maximum flow. This transforms the crack recognition problem into a graph theory minimum cut problem. Edge weights are defined as a weighted combination of the absolute values ​​of depth and grayscale differences. The maximum flow minimum cut theorem is used to optimize the segmentation boundary globally, ensuring that the segmentation contour is located at the position where the depth and grayscale differences change most drastically. Although aggregate and joint texture interferences have certain edge weights, their edge weight distribution does not form continuous high-weight paths. The algorithm can effectively distinguish between real cracks and texture interferences, avoiding the subjectivity of local threshold selection and boundary breakage problems, and significantly improving the crack recognition accuracy in complex textured pavement scenarios.

[0080] A second aspect of the present invention provides a computer-readable storage medium storing program instructions that, when executed in a computer, are used to perform the above-described method for three-dimensional quantitative analysis of millimeter-level cracks on the surface of infrastructure.

[0081] A third aspect of the present invention provides a three-dimensional quantitative analysis system for millimeter-level cracks on the surface of infrastructure, comprising the aforementioned computer-readable storage medium. The system is any one of a computer, a server, or a microcontroller. The computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

[0082] Specifically, the principle of this invention is as follows: This invention employs a technical solution combining auxiliary laser scanning line cross-coverage and adaptive gain amplification to solve the problem of missing point cloud data at the bottom of deep and narrow cracks. The principle lies in the close relationship between the intensity of the laser echo signal and the incident angle and propagation distance. For deep and narrow cracks, when a vertical laser beam is incident, it forms a nearly parallel geometric relationship with the inner wall of the crack. The inner wall surface has a significant blocking effect on the laser, causing a sharp attenuation of the laser energy reaching the bottom of the crack. However, when an inclined laser beam is incident at an angle of ±15 degrees, the incident light path forms an intersection angle with the inner wall of the crack, allowing the beam to propagate along the gap in the inner wall to a deeper location. The blocking effect is geometrically dispersed, increasing the laser energy received at the bottom of the crack. The adaptive gain amplification dynamically calculates the amplification factor based on the ratio of the echo signal intensity to a preset threshold. For weak echo signals caused by depth attenuation, the gain coefficient is increased to raise the signal amplitude to the effective detection range of the sensor, thereby achieving accurate acquisition of the point cloud at the bottom of the crack. This technical solution, based on the geometric optics principle of laser propagation, solves the physical limitations of a single vertical scanning method in deep and narrow crack scenarios by combining multi-angle scanning and signal processing.

[0083] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0084] The specific implementation of step S01 involves continuously scanning the surface of infrastructure using a high-frequency line structured light sensor, simultaneously acquiring a 1mm resolution laser point cloud elevation matrix, a 3D depth image, and a 2D grayscale image. An inertial navigation system is used to collect the attitude parameters of the scanning platform in real time, establishing a mapping function from timestamp to spatial position. Motion compensation registration correction is then performed on the 1mm resolution laser point cloud elevation matrix, the 3D depth image, and the 2D grayscale image. During the establishment of the mapping function from timestamp to spatial position, the inertial navigation system collects the position coordinates, velocity vector, attitude angle, and angular velocity data of the scanning platform at a frequency of 1000Hz. Position and attitude interpolation calculations are performed on the timestamp of each 1mm resolution laser point cloud elevation matrix sampling moment. The position interpolation calculation formula is expressed as follows: ; In the formula, For timestamps The corresponding interpolation position coordinate vector, in meters; This is the current timestamp, in seconds. This is the start timestamp of the interpolation interval, in seconds. This is the vector of coefficients for the zeroth-order term of cubic spline interpolation, in units of m; This is the vector of coefficients for the first-order term in cubic spline interpolation, in units of... ; This is the vector of quadratic coefficients from cubic spline interpolation, in units of... ; This is the vector of cubic term coefficients for cubic spline interpolation, in units of... interpolation coefficients The position coordinates and their derivatives at adjacent time points are used to determine the position. Attitude interpolation is calculated using the same cubic spline interpolation method, outputting the interpolated position coordinates and interpolated attitude angle corresponding to the timestamp as a mapping function from timestamp to spatial position, thus establishing the mapping relationship between timestamp and spatial position. During motion compensation registration and correction, the scanning carrier position offset and scanning carrier attitude rotation at each 1mm resolution laser point cloud elevation matrix sampling time are obtained based on the timestamp-to-spatial position mapping function. The coordinate transformation formula is expressed as follows: ; In the formula, The coordinate vector of the point cloud after motion compensation, in meters; Let be the attitude rotation matrix, which is dimensionless; The pitch angle is expressed in rad. This is the roll angle, measured in rad. This is the heading angle, measured in rad. This is the original point cloud coordinate vector, in meters; This is the vector of the scanning carrier's position offset, in meters (m). The spatial registration process uses the nearest neighbor matching method, and the spatial registration error is controlled within 0.5 mm.

[0085] The specific implementation of step S02 is to construct a pixel-level undirected graph based on the motion-compensated registration and correction of the 3D depth image and the 2D grayscale image. Each pixel of the 3D depth image and the 2D grayscale image is taken as a graph node of the pixel-level undirected graph, and an undirected edge is established between each graph node and its 8 neighboring pixels. The edge weight calculation formula is expressed as follows: ; In the formula, For pixels and pixels The edge weights between them are dimensionless. For pixels The depth value, in mm; For pixels The depth value, in mm; For pixels The grayscale value is dimensionless; For pixels The grayscale value is dimensionless; The standard value for depth difference is 1 mm. The grayscale difference standard value is set to 20. Crack seed points are selected in the 3D depth image. A crack seed point is defined as a pixel whose absolute depth difference with its 8 neighboring pixels is greater than 0.5 mm. Connectivity analysis is performed on depth abrupt change points, and connected component areas greater than [value missing] are selected. The set of depth abrupt change points is used as the seed point set for cracks. Uniform grayscale road surface regions are identified in the 2D grayscale image, defined as regions where the pixel grayscale variance is less than 10. Pixels are randomly selected within these uniform grayscale road surface regions as the background seed point set. The Boykov-Kolmogorov algorithm is used to solve for the maximum flow from the source to the sink in the pixel-level undirected graph. The minimum cut set of the pixel-level undirected graph is obtained according to the maximum flow minimum cut theorem, and the boundary of the minimum cut set forms the segmentation contour between the crack and the background.

[0086] The specific implementation of step S03 involves treating the pixels within the segmentation contour of the crack and background as an ant colony optimization space, and setting the pheromone concentration for each pixel. The initial pheromone concentration calculation formula is expressed as follows: ; In the formula, For pixels The initial pheromone concentration is dimensionless; For pixels The shortest Euclidean distance to the boundary of the dividing contour between the crack and the background, in mm; This represents the maximum value of the shortest distance among all pixels, in mm. The formula for calculating the transition probability when the virtual ant chooses its next movement direction at the current pixel is as follows: ; In the formula, For virtual ants from pixels Move to neighboring pixels The transition probability is dimensionless. For neighboring pixels The pheromone concentration is dimensionless. The depth gradient heuristic function is dimensionless. For pixels The set of 8 neighboring pixels; For 8 neighboring pixels medium pixel The pheromone concentration is dimensionless. For 8 neighboring pixels medium pixel The depth gradient heuristic function is dimensionless. The formula for calculating the depth gradient heuristic function is as follows: ; In the formula, For pixels The depth value, in mm; For neighboring pixels The depth value, in mm; For pixels 8-neighbor pixel set any pixel The depth value is expressed in mm. The formula for dynamically updating the pheromone concentration after each iteration is as follows: ; In the formula, For pixels In the The pheromone concentration after rounds of iteration is dimensionless. For pixels In the The pheromone concentration after rounds of iteration is dimensionless. This represents the pheromone increment, which is dimensionless. The formula for calculating the pheromone increment is as follows: ; In the formula, For passing pixels The number of virtual ants; This is the pheromone intensity coefficient, in mm, with an empirical value of 100 mm. For the first The path length of a virtual ant is measured in mm. After multiple iterations, the set of paths with the highest pheromone concentration is used as the centerline of the crack skeleton.

[0087] The specific implementation of step S04 involves inputting the crack skeleton centerline into the crack topology connectivity optimization model, which then outputs connectivity decision parameters and branch identification parameters. The connectivity distance threshold and orientation consistency threshold are determined based on the connectivity decision parameters, and the calculation formulas are as follows: , ; In the formula, This is the threshold for the distance to penetrate, in mm. The threshold value for the penetration distance output by the crack topology penetration optimization model is in mm. To reach the consistency threshold, the unit is degrees; This is the predicted strike orientation consistency threshold value output by the crack topology connectivity optimization model, in degrees. The formula for calculating the crack strike angle is as follows: ; In the formula, The angle representing the direction of the crack is in degrees. The x-coordinate of the starting point of the centerline segment of the crack skeleton is in mm; The ordinate of the starting point of the centerline segment of the crack skeleton is in mm; The x-coordinate of the endpoint of the centerline segment of the crack skeleton, in mm; The vertical coordinate of the endpoint of the centerline segment of the crack skeleton is in mm. The direction of travel of the scanning carrier is determined by the velocity vector direction of the scanning carrier, measured in degrees, collected by the inertial navigation system. The value is pi, taken as 3.14159. Crack segments with a distance less than the penetration distance threshold and a strike angle less than the strike consistency threshold along the crack skeleton centerline are connected. Main cracks and branch cracks are marked according to branch identification parameters. Crack segments with a length less than 10cm and branch cracks with a branch length less than 1cm are filtered out, retaining the effective crack topology network. The crack length calculation formula is expressed as follows: ; In the formula, The crack length is in mm. This represents the total number of pixels along the center line segment of the crack skeleton. For the first The x-coordinate of each pixel is in mm. For the first The ordinate of each pixel is in mm. For the first The x-coordinate of each pixel is in mm. For the first The ordinate of each pixel is given in mm.

[0088] The specific implementation of step S05 involves grouping the cracks in the effective crack topology network according to their orientation angle, length, and density characteristics. The crack density calculation formula is as follows: ; In the formula, Crack density, in units of ; This represents the total number of centerline segments of the crack skeleton per unit area. For the first The length of the centerline segment of the crack skeleton, in meters; The value per unit area is [value], and it takes the value [value]. Cracks with an azimuth angle within ±20 degrees of the scanning carrier's travel direction and a crack length greater than 10 cm are classified as longitudinal cracks. Cracks with an azimuth angle within ±20 degrees of the direction perpendicular to the scanning carrier's travel direction and a crack length greater than 10 cm are classified as transverse cracks. Cracks that are multi-directionally interwoven and whose enclosed area is less than [a certain size] are classified as transverse cracks. And the crack density is greater than Cracks are classified as crazing. Cracks with a total length greater than three times the diagonal length of the board and divided into three or more independent areas are classified as broken boards. The remaining cracks are classified as other cracks.

[0089] The specific implementation of step S06 involves dividing the classified cracks into several crack segments of 10cm length. Sampling points are selected at 1cm intervals on each crack segment, and a cross-sectional area perpendicular to the crack's direction is extracted centered on these sampling points. Within the cross-sectional area, auxiliary laser scanning lines with a ±15° tilt angle along the scanning carrier's travel direction are set to form a cross-scan coverage. Adaptive gain amplification is used for weak echo signals. The formula for calculating the adaptive gain coefficient is as follows: ; In the formula, The gain coefficient is dimensionless. The preset echo intensity threshold is a dimensionless relative intensity value, set at 30% of the normal echo intensity. The intensity of the weak echo signal is a dimensionless relative intensity value. For pixels within the cross-sectional area that still lack depth data, a neighborhood depth interpolation algorithm is used to complete the data. The formula for calculating the neighborhood depth interpolation is as follows: ; In the formula, Interpolated depth for pixels with missing depth data, in mm; The number of pixels with valid depth data in an 8-neighborhood; For the first The depth value of each effective depth data pixel is expressed in mm. Within the cross-sectional area, the crack edge points are identified by scanning pixel by pixel along a direction perpendicular to the crack's direction. The depth gradient calculation formula is as follows: ; In the formula, For pixels The depth gradient, in mm; For pixels The depth value, in mm; Adjacent pixels The depth value is expressed in mm. Pixels with a depth gradient greater than 0.3 mm are identified as candidate crack edge points. Connectivity analysis is performed on these candidate crack edge points, and two candidate crack edge points located on opposite sides of the crack with opposite depth gradient directions are selected as the crack edge points. The formula for calculating the cross-sectional crack width is as follows: ; In the formula, The width of the cross-sectional crack is in mm. The horizontal coordinate of the first crack edge point is in mm. The horizontal coordinate of the first crack edge point is shown in mm. The horizontal coordinate of the second crack edge point is in mm. Here is the horizontal ordinate of the second crack edge point, in mm. The reference plane is fitted using the least squares method, and the plane equation is expressed as follows: ; In the formula, The normalized transverse direction coefficients of the plane equation are dimensionless and satisfy the following conditions: ; The normalized longitudinal direction coefficient for the plane equation is dimensionless. The normalized vertical direction coefficient for the plane equation is dimensionless. represents the intercept coefficient of the plane equation, in mm; The x-coordinate of the point cloud is in mm. The vertical coordinate of the point cloud is in mm. Here are the vertical coordinates of the point cloud, in mm. The formula for calculating the depth of a cross-section crack is as follows: ; In the formula, The depth of the cross-sectional crack is expressed in mm. The x-coordinate of the bottom point of the crack is in mm; The vertical coordinate of the bottom point of the crack is in mm; Here are the three-dimensional vertical coordinates of the point at the bottom of the crack, in mm. Because... Therefore, the denominator is 1, and the formula simplifies to The formulas for calculating the average crack width and average crack depth per segment are as follows: , ; In the formula, The average crack width in segments is expressed in mm. The average crack depth is expressed in mm. This represents the total number of sampling points within the crack segment. The average crack width and average crack depth per segment are used as quantitative parameters for crack segment assessment and maintenance decisions.

[0090] To better understand and implement this invention, the following is a specific application scenario of embodiment 2: To verify the practical application effect of this invention, technicians conducted crack detection tests on a main road in a city. The test section was an asphalt concrete pavement, 300m long and 3.75m wide, with a service life of 8 years, and exhibited various types of cracks. Technicians used a detection vehicle equipped with a high-frequency line structured light sensor at a speed of 20... The high-frequency line structured light sensor continuously scans the road section at a speed of 50 Hz. The horizontal scanning field of view is 120°, simultaneously acquiring a 1mm resolution laser point cloud elevation matrix, 3D depth image, and 2D grayscale image. For example... Figure 2 As shown, the laser point cloud data achieves a resolution of 1mm in both the horizontal and vertical directions, and an elevation accuracy of 0.25mm, clearly capturing minute undulations and crack details on the road surface. The inertial navigation system operates at 1000... The scanning carrier's position coordinates, velocity vector, attitude angle, and angular velocity data are acquired via frequency acquisition. Cubic spline interpolation is performed on the timestamp of each 1mm resolution laser point cloud elevation matrix sampling moment to establish a mapping function from timestamp to spatial position. Based on this mapping function, the scanning carrier's position offset and attitude rotation are obtained. Motion compensation registration correction is then performed on the 1mm resolution laser point cloud elevation matrix, using the nearest neighbor matching method for spatial registration, with the spatial registration error controlled within 0.3mm.

[0091] A pixel-level undirected graph is constructed based on the motion-compensated registration and correction of the 3D depth image and the 2D grayscale image, such as... Figure 3As shown, the 3D depth image presents the road surface depth information in a pseudo-color manner, with crack areas displaying obvious depression features due to abrupt changes in depth. Figure 4 As shown, the 2D grayscale image reflects the texture features of the road surface, with a significant grayscale difference between cracked areas and normal road surfaces. Each pixel is treated as a graph node, and undirected edges are established between each node and its 8 neighboring pixels. The absolute values ​​of the depth difference and grayscale difference between adjacent pixels are calculated. The absolute values ​​of the depth difference and grayscale difference are both dimensionless, divided by 1 mm and 20 respectively. The dimensionless absolute values ​​of the depth difference and grayscale difference are then weighted and summed using weights of 0.6 and 0.4, respectively, to determine the edge weights. Depth abrupt change points are identified in the 3D depth image, selecting connected regions with an area greater than 5. The set of depth abrupt change points was used as the crack seed point set, and a total of 216 crack seed points were identified. A uniform grayscale road surface region was identified in the 2D grayscale image, and 358 pixels were randomly selected within this region as the background seed point set. The Boykov-Kolmogorov algorithm was used to solve for the minimum cut set of the pixel-level undirected graph, and the boundary of the minimum cut set formed the segmentation contour between the crack and the background.

[0092] Comparing the differences in recognition accuracy between traditional methods and the method of this invention, such as... Figure 5 As shown, traditional surface crack marking and identification methods use a 10cm grid precision for crack location, which can only roughly mark the approximate location and direction of the crack, and cannot accurately capture the detailed features and true shape of the crack. Figure 6 As shown, this invention employs millimeter-level image automatic recognition accuracy, with 1 pixel corresponding to an actual size of 1mm. It can accurately depict the edge contour, width variation, and topological structure of cracks, achieving a 100-fold improvement in recognition accuracy compared to traditional methods. Pixels within the segmentation contour of the crack and background are considered as the ant colony optimization space. The initial pheromone concentration of each pixel is set to be inversely proportional to the shortest distance from the pixel to the boundary of the segmentation contour. 128 virtual ants start from the endpoints of the segmentation contour of the crack and background. When a virtual ant selects its next movement direction at the current pixel, the transition probability of its 8 neighboring pixels is calculated. The transition probability is directly proportional to the product of the pheromone concentration of the 8 neighboring pixels and the depth gradient heuristic function, which are weighted by coefficients of 0.7 and 0.3, respectively. In each iteration, after all virtual ants complete their path search, pheromone enhancement is applied to the pixels along the virtual ant's path, while pheromone concentration attenuation is applied to all pixels using a volatile attenuation coefficient of 0.9. After 50 iterations, the set of paths with the highest pheromone concentration was used as the center lines of the crack skeleton, resulting in 127 initial crack skeleton center lines, such as... Figure 7 As shown, the crack lines clearly demonstrate the skeletal structure and topological relationship of the road surface cracks.

[0093] The crack skeleton centerline is input into the crack topology connectivity optimization model. The input layer of this model receives the endpoint coordinates, crack orientation angle, mean depth, and mean grayscale features of the crack skeleton centerline. These features have a 128-dimensional feature dimension and are mapped to a 256-dimensional embedding vector through a fully connected layer. The embedding vector is then input into a multi-layer graph convolutional network module containing four graph convolutional layers, and the crack network density parameter is calculated to be 3.8. The average length of the crack segment is 42cm, and the random depth probability is dynamically calculated to be 0.15 based on these two parameters. The model outputs a predicted breakthrough distance threshold of 8.3cm and a predicted orientation consistency threshold of 24°. By multiplying by safety factors of 1.2 and 1.1, the breakthrough distance threshold of 10cm and the orientation consistency threshold of 26.4° are obtained, respectively. All crack skeleton centerline segment endpoints in the effective crack topology network are traversed. Within a fan-shaped search area of ​​±60° in the crack orientation angle direction, the nearest other endpoints are found. Crack segments with a distance less than 10cm and an orientation angle less than 26.4° are connected, completing 15 breakthrough connections. Main cracks and branch cracks are marked according to the branch identification parameters. 16 short cracks with a length less than 10cm and 8 branch cracks with a branch length less than 1cm are filtered out, retaining an effective crack topology network containing 88 cracks. Figure 8 As shown, a complete and continuous crack topology network is formed during the crack treatment process through through-connection and branch filtering.

[0094] Eighty-eight cracks in the effective crack topology network were grouped according to their orientation angle, length, and density. Cracks with an orientation angle within ±20° of the scanning carrier's travel direction and a length greater than 10cm were classified as longitudinal cracks; 23 longitudinal cracks were identified, with an average length of 1.8m. Cracks with an orientation angle within ±20° of the direction perpendicular to the scanning carrier's travel direction and a length greater than 10cm were classified as transverse cracks; 31 transverse cracks were identified, with an average length of 2.1m. The cracks were multi-directionally interwoven and enclosed areas less than 100... And the crack density is greater than 4 The cracks were classified as crazing, and 12 crazing areas were identified, with an average area of ​​68 mm per area. Cracks with a total length greater than three times the diagonal length of the slab and divided into more than three independent areas were classified as broken slabs, and four broken slabs were identified. The remaining cracks were classified as other cracks, totaling 18. The distribution characteristics of different types of cracks are shown in Table 1.

[0095] Table 1. Statistical Table of Different Types of Cracks

[0096] The classified cracks were divided into several crack segments with a length of 10cm, and sampling points were selected at 1cm intervals on each crack segment. A cross-sectional area perpendicular to the crack's direction was extracted, centered on the sampling point. Auxiliary laser scanning lines with tilt angles of ±15° and -15° along the scanning carrier's travel direction were set within the cross-sectional area to form a cross-scan coverage. Weak echo signals with an intensity 30% lower than the normal echo intensity within the cross-sectional area were amplified using an adaptive gain algorithm. The gain coefficient was dynamically calculated based on the ratio of the weak echo signal intensity to a preset echo intensity threshold, with an upper limit of 5 times. For pixels still lacking depth data, a neighborhood depth interpolation algorithm was used for completion. Pixels with valid depth data within an 8-neighborhood of the pixel lacking depth data were selected, and the average depth was calculated as the interpolation depth.

[0097] Within the cross-sectional area, a pixel-by-pixel scan is performed along a direction perpendicular to the crack's orientation. Pixels with a depth gradient greater than 0.3 mm are identified as candidate crack edge points. Two candidate crack edge points located on opposite sides of the crack with opposite depth gradient directions are selected as the final crack edge points. The Euclidean distance between the two crack edge points on the horizontal plane is calculated as the crack width, with an accuracy of 0.1 mm. Figure 9 As shown, the width calculation results are displayed in pseudo-color, with purple representing the widest crack area, transitioning sequentially to red, orange, blue, and green, representing different widths from coarse to fine, clearly reflecting the variation of crack width along its length. Point cloud data of normal road surface within a 5mm radius on both sides of the crack are selected within the cross-sectional area. Least squares are used for plane fitting, and the fitted plane serves as the reference plane for the cross-sectional area. The bottom point of the crack with the largest depth value within the cross-sectional area is identified, and the vertical distance from the bottom point of the crack to the reference plane is calculated as the crack depth. The accuracy of the crack depth is 0.1mm. Figure 10 As shown, the depth calculation results are also displayed in pseudo-color, with purple representing the deepest crack area, transitioning to red, orange, blue, and green in sequence, representing different depths from deep to shallow, intuitively presenting the spatial distribution characteristics of crack depth.

[0098] A typical transverse crack was selected for detailed analysis. The crack, with a total length of 2.3m, was divided into 23 crack segments, each containing 10 sampling points. For example... Figure 11 As shown, the crack lines reveal the complete shape and centerline of the transverse crack, with both width and depth reaching their maximum in the central region. After crack treatment, as... Figure 12 As shown, the main crack and branch cracks were successfully identified and marked, noise interference was filtered out, and the effective crack topology was preserved. Additionally, Figure 13 and Figure 14This diagram shows the calculated width and depth of the alligator crack. The arithmetic mean of the crack width and depth at all sampling points within each crack segment was calculated to obtain the segmental average crack width and depth, accurate to 0.1 mm. Measurement data shows that the maximum width of the transverse crack is 5.2 mm, occurring at a crack length of 1.2 m, and the maximum depth is 15.8 mm, occurring at a crack length of 1.1 m. The peak locations of the width and depth are basically consistent, conforming to the general pattern of crack development. A total of 468 m of cracks were quantitatively analyzed across the entire test section, with an average crack width of 3.6 mm and an average crack depth of 9.2 mm. A complete crack damage assessment report was generated, providing accurate data support for pavement maintenance decisions.

[0099] This invention represents a significant technological advancement over traditional crack detection methods. Traditional threshold-based edge detection methods rely on manually set fixed thresholds when segmenting cracks from the background, making them susceptible to interference from road surface texture and lighting variations, resulting in significant noise and boundary breaks in the segmentation results. This invention employs a minimum cut-maximum flow-based crack-background segmentation algorithm. By constructing a pixel-level undirected graph and defining edge weights as a weighted combination of the absolute values ​​of depth and grayscale differences, it optimizes the segmentation results globally. Minimizing the sum of edge weights corresponding to the segmentation boundary indicates that the segmentation contour is located at the position of most dramatic depth and grayscale changes, thus accurately capturing the crack contour and exhibiting stronger robustness against noise and local blurring. Traditional skeleton extraction methods based on morphological refinement rely on fixed structural elements, exhibiting poor adaptability to changes in crack width and topological branches, and are prone to producing breakpoints and burrs. This invention employs an ant colony optimization-based crack skeleton extraction and refinement algorithm. Pixels within the crack segmentation contour are considered as the foraging space of an ant colony. The global search capability and positive feedback mechanism of the ant colony algorithm are used to extract the crack skeleton centerline. By setting pheromone concentration inversely proportional to the distance from the pixel to the segmentation contour boundary, virtual ants are guided to preferentially roam in the crack center region. Combined with a depth gradient heuristic function, the virtual ants move along a direction of continuously varying depth. The extracted crack skeleton centerline maintains a single-pixel width and completely preserves the crack's topological connectivity. Traditional crack depth measurement methods struggle to accurately obtain crack depth data in deep and narrow crack scenarios due to the inner wall obstructing the vertical laser beam after it enters the crack, resulting in missing bottom echo signals. This invention adds auxiliary laser scanning lines with ±15° tilt angles to the vertical laser scanning to form cross-scan coverage. The tilted laser beam enters the crack from different angles, increasing the sampling probability of the crack bottom point cloud. Combined with adaptive gain amplification and neighborhood depth interpolation algorithms to complete the missing depth data, the completeness and accuracy of cross-sectional crack depth measurement are significantly improved.

[0100] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A three-dimensional quantitative analysis method for millimeter-level cracks on infrastructure surfaces, characterized in that, A high-frequency line structured light sensor is used to continuously scan the surface of infrastructure to acquire 1mm resolution laser point cloud elevation matrix, 3D depth image, and 2D grayscale image. An inertial navigation system is used to collect the attitude parameters of the scanning carrier to establish a timestamp-to-spatial position mapping function for motion compensation registration and correction. A pixel-level undirected graph is constructed, and a crack-background segmentation algorithm based on minimum cut-maximum flow is used to identify the crack and background segmentation contours. The crack skeleton centerline is extracted using an ant colony optimization algorithm. Crack network connection and branch identification are performed by outputting connection decision parameters and branch identification parameters through a crack topology connection optimization model. Cracks are classified according to crack direction angle and crack length characteristics. A 15-degree tilt angle of ±15 degrees is set in the crack cross-section area to assist the laser scanning line to form cross-scan coverage. For weak echo signals, adaptive gain amplification and neighborhood depth interpolation are used to complete the missing depth data. The crack width and crack depth of the cross-section are calculated.

2. The method according to claim 1, characterized in that, The steps for establishing the timestamp-to-spatial-position mapping function are as follows: the inertial navigation system collects the position coordinates, velocity vectors, attitude angles, and angular velocity data of the scanning carrier at a frequency of 1000Hz, and performs position and attitude interpolation calculations on the timestamps of the 1mm resolution laser point cloud elevation matrix sampling time using the cubic spline interpolation method.

3. The method according to claim 2, characterized in that, The motion compensation registration correction steps are as follows: the position offset and attitude rotation of the scanning carrier are obtained according to the timestamp to spatial position mapping function; the attitude rotation matrix transformation is performed on the 1mm resolution laser point cloud elevation matrix coordinates to compensate for the attitude change of the scanning carrier; and the nearest neighbor matching method is used for spatial registration.

4. The method according to claim 3, characterized in that, The steps for constructing a pixel-level undirected graph are as follows: each pixel of the 3D depth image and the 2D grayscale image is treated as a graph node, and an undirected edge is established between each graph node and its 8 neighboring pixels. The edge weight is the sum of the absolute values ​​of the depth difference and the absolute values ​​of the grayscale difference, weighted by weight coefficients of 0.6 and 0.4, respectively.

5. The method according to claim 4, characterized in that, The steps for selecting crack seed points specifically involve identifying depth abrupt change points in the 3D depth image. Depth abrupt change points are pixels whose absolute depth difference with their 8 neighboring pixels is greater than 0.5 mm. Connectivity analysis is then performed on these depth abrupt change points to select those with a connected component area greater than [missing value]. The set of deep mutation points is used as the seed point set of the crack.

6. The method according to claim 5, characterized in that, The steps for selecting background seed points are as follows: First, identify uniform grayscale road surface regions in the 2D grayscale image. A uniform grayscale road surface region is a region where the pixel grayscale variance is less than 10. Randomly select pixels within the uniform grayscale road surface region as the set of background seed points.

7. The method according to claim 6, characterized in that, The steps of the crack-background segmentation algorithm based on minimum cut-maximum flow are as follows: connect the crack seed point set to the source point and the background seed point set to the sink point. Set the weight of the connecting edge to infinity. Use the Boykov-Kolmogorov algorithm to solve for the maximum flow from the source point to the sink point. Obtain the pixel-level undirected graph minimum cut set according to the maximum flow minimum cut theorem. The boundary of the minimum cut set forms the crack and background segmentation contour.

8. The method according to claim 7, characterized in that, Pheromones concentration is defined as the initial pheromone concentration of each pixel being inversely proportional to the shortest distance from the pixel to the crack and background segmentation contour. The shortest distance is obtained by calculating the Euclidean distance from the pixel to the boundary of the crack and background segmentation contour. The initial pheromone concentration value is dimensionless by dividing the reciprocal of the shortest distance by the maximum value of the reciprocal of the shortest distance.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions that, when executed in a computer, are used to perform a three-dimensional quantitative analysis method for millimeter-level cracks on the surface of infrastructure, as described in any one of claims 1-8.

10. A three-dimensional quantitative analysis system for millimeter-level cracks on infrastructure surfaces, characterized in that, The system comprises the computer-readable storage medium of claim 9, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

Images