A road crack identification method, medium and system

Through linear structure light sensing and inertial navigation compensation technology, combined with Gaussian background modeling and GWO algorithm, convolutional neural networks are used to identify road surface cracks, solving the low-precision problems caused by environmental and attitude influence in the existing technology, achieving high Accurate crack detection.

CN118446976BActive Publication Date: 2024-11-15LAN SHEN (BEI JING) KE JI YOU XIAN GONG SI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410537878.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-30
Publication Date
2024-11-15
Estimated Expiration
2044-04-30

AI Technical Summary

Technical Problem

The existing two-dimensional and three-dimensional crack detection technologies are affected by factors such as light, shadows, stains, posture, vertical and horizontal slopes, and surface textures, making it difficult to meet the requirements of high-precision detection.

Method used

The linear structured light sensing device is used to perform continuous scanning of the road surface, combined with the inertial guide device to measure attitude data for compensation, filtering and Gaussian background modeling to screen potential crack feature points, optimize grouping using GWO algorithm, and input data into the convolutional neural network Make identification.

Benefits of technology

It realizes high-precision identification of road cracks, reduces misidentification, and improves detection accuracy and reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118446976B_ABST
    Figure CN118446976B_ABST
Patent Text Reader

Abstract

The present invention discloses a road crack identification method, medium and system, including: using a line structured light sensing device to continuously scan the road surface to obtain a laser point cloud elevation matrix; using an inertial navigation device to measure the real-time attitude data of a scanning device to compensate the laser point cloud elevation matrix; filtering the compensated laser point cloud elevation matrix to obtain a road depth matrix; converting the road depth matrix into a depth image and cutting it into a plurality of initial sub-image blocks; screening the potential crack feature points in the initial sub-image blocks by Gaussian background modeling; optimizing the grouping by a GWO algorithm, dividing the potential crack feature points constituting a connected area in the potential crack feature points into one group; eliminating the groups whose number is less than a number threshold, and eliminating the unrelated potential crack feature points in the group, to obtain the final sub-image block; inputting the final sub-image block into a convolutional neural network, and outputting the result of whether there is a crack. The present invention can accurately identify cracks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of road surface detection, and in particular to a road surface crack identification method, medium and system. Background Art

[0002] Cracks are one of the most common and important surface diseases of infrastructure. After the initial fine cracks appear, as surface rain or snow water infiltrates, the water content near the cracks increases, resulting in reduced strength. The infrastructure will be accelerated under the long-term effects of loads or environmental factors, and severe structural damage will occur. After the crack disease appears, the later the disease characteristics are detected, the greater the impact of the disease on the performance of the infrastructure, and the more funds are required for repair and maintenance.

[0003] Existing technologies mainly identify cracks in the following two ways:

[0004] (1) Crack feature extraction technology based on two-dimensional images

[0005] The crack feature extraction of two-dimensional images is mainly based on line scanning sensing technology. Through the line scanning camera, continuous one-dimensional images are collected, and a certain number of one-dimensional images are combined to obtain a two-dimensional surface image. On this basis, machine vision technology such as expansion, corrosion, opening operation, closing operation, edge detection algorithm, etc. are used to extract crack features. Since two-dimensional images are inevitably affected by environmental factors such as lighting conditions, shadows, stains, etc., even the most advanced algorithms currently have difficulty distinguishing crack features from the above-mentioned influencing factors, resulting in a large number of misidentifications and unable to meet the actual engineering accuracy requirements.

[0006] (2) Crack feature extraction technology based on 3D data

[0007] Three-dimensional data acquisition technologies mainly include three-dimensional imaging technologies such as laser radar, stereo vision, and line structured light. Among them, high-precision three-dimensional data acquisition mainly uses line structured light technology, and its principle is: using the principle of triangulation, using a camera to collect surface contour elevation, and imaging through three-dimensional reconstruction. This technology can better solve the shortcomings of two-dimensional image crack feature extraction that is easily affected by environmental interference. However, in engineering applications, high-precision crack detection based on three-dimensional images still faces huge challenges due to factors such as the posture of the three-dimensional scanning carrier, the vertical and horizontal slopes of the surface, the depth of the surface texture, and the surface markings. Summary of the invention

[0008] The embodiments of the present invention provide a pavement crack identification method, medium and system to solve the problem that existing two-dimensional and three-dimensional crack detection is affected by various factors and is difficult to meet high-precision detection requirements.

[0009] In a first aspect, a pavement crack identification method is provided, comprising:

[0010] When the inspection vehicle is driving on the road, the road surface is continuously scanned by a line structured light sensing device carried by the inspection vehicle to obtain the original laser point cloud elevation matrix of the road surface;

[0011] The inertial navigation device carried by the detection vehicle is used to measure the real-time posture data of the line structured light scanning device, and the real-time posture data is used to compensate the original laser point cloud elevation matrix to obtain a compensated laser point cloud elevation matrix;

[0012] Filtering each laser point cloud elevation in the compensated laser point cloud elevation matrix to obtain a road surface depth matrix;

[0013] Converting the road surface depth matrix into a depth image, and cutting the depth image into a plurality of initial sub-image blocks of the same size;

[0014] Screening and obtaining potential crack feature points in each of the initial sub-image blocks by Gaussian background modeling;

[0015] Optimizing the grouping by the GWO algorithm, dividing all the potential crack feature points that can constitute a connected area into one group;

[0016] Eliminating all potential crack feature points in the groups whose number is less than a threshold value, and eliminating irrelevant potential crack feature points in each group, to obtain a final sub-image block corresponding to a final group;

[0017] Each of the final sub-image blocks is input into a convolutional neural network, and a result of whether the final sub-image block has a crack is output.

[0018] In a second aspect, a computer-readable storage medium is provided, on which computer program instructions are stored; when the computer program instructions are executed by a processor, the pavement crack identification method as described in the embodiment of the first aspect is implemented.

[0019] In a third aspect, a pavement crack identification system is provided, comprising: a computer-readable storage medium as described in the embodiment of the second aspect.

[0020] In this way, the embodiment of the present invention can accurately identify road cracks with high recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0022] Figure 1 is a flow chart of a pavement crack identification method according to a preferred embodiment of the present invention;

[0023] Figure 2 is a flow chart of a pavement crack identification method according to another preferred embodiment of the present invention;

[0024] Figure 3 This is a filtered point cloud image of the embodiment of the present invention with a horizontal area of ​​3.5 m and a vertical area of ​​2 m;

[0025] Figure 4 This is a filtered point cloud image of 3.5 m in horizontal direction and 10 m in vertical direction according to an embodiment of the present invention;

[0026] Figure 5 is a depth image after cutting the initial sub-image block with a length of 3.5 m in horizontal direction and 2 m in vertical direction according to an embodiment of the present invention;

[0027] Figure 6 This is a schematic diagram of crack identification results of an embodiment of the present invention with a length of 3.5 m in the horizontal direction and 2 m in the vertical direction;

[0028] Figure 7 It is a schematic diagram of crack identification results of a 3.5 m horizontal and 10 m vertical crack according to an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] The embodiment of the present invention discloses a method for identifying pavement cracks. Figure 1 As shown, the method of the embodiment of the present invention includes the following steps:

[0031] Step S101: When the inspection vehicle is driving on the road, the road surface is continuously scanned by a line structured light sensing device carried by the inspection vehicle to obtain an original laser point cloud elevation matrix of the road surface.

[0032] In the embodiment of the present invention, the laser point cloud elevation matrix corresponds to a road surface with a size of 1mm*1mm.

[0033] The line structured light sensing device may be a three-dimensional line structured light sensor, which uses high-frequency (above 5000 Hz) laser scanning to obtain high-resolution three-dimensional point cloud data.

[0034] The 3D line structured light sensor is generally composed of a laser emission module, an image receiving module and related optical components. The laser emission module generates a high-frequency linear laser, which is repeatedly scanned by a scanning mirror on the target surface. The image receiving module uses a high-speed camera to capture the laser lines reflected by the surface, and can reconstruct the 3D point cloud data by combining the scanning information. By optimizing the laser power, exposure time, imaging parameters, etc., point cloud data with a high signal-to-noise ratio can be obtained, laying the foundation for subsequent feature extraction.

[0035] Step S102: using the inertial navigation device carried by the inspection vehicle to measure the real-time attitude data of the line structured light scanning device, and using the real-time attitude data to compensate the original laser point cloud elevation matrix to obtain a compensated laser point cloud elevation matrix.

[0036] It should be understood that the inertial navigation unit (IMU) should be installed near the line structured light scanning device to collect the real-time attitude data of the line structured light scanning device. The inertial navigation unit includes: a three-axis accelerometer and a three-axis gyroscope, which are used to measure the angle and displacement changes of the line structured light scanning device in real time, and have the advantage of high precision.

[0037] Specifically, this step includes the following process:

[0038] 1. Perform Kalman filtering on the real-time posture data of the line structured light scanning device to obtain the six-axis posture data of the line structured light scanning device.

[0039] The six-axis attitude data includes: roll angle roll, pitch angle pitch, yaw angle yaw, x, y, z. The filtering process is an existing technology and will not be described in detail here.

[0040] 2. The six-axis posture data of the line structured light scanning device is used to perform coordinate system conversion and displacement compensation on each laser point cloud elevation in the original laser point cloud elevation matrix to obtain a compensated laser point cloud elevation matrix.

[0041] The above-mentioned coordinate system conversion and displacement compensation are prior arts and will not be described in detail here.

[0042] This step can eliminate the influence of the posture change of the line structured light scanning device on the point cloud elevation measurement, reduce the error, and obtain a more accurate point cloud elevation matrix.

[0043] In addition, through experimental analysis, it was determined that the IMU's angle measurement accuracy must be within 0.1 degrees and the position measurement accuracy must be within 1 mm in order to effectively eliminate the impact of posture on point cloud elevation.

[0044] Step S103: filtering each laser point cloud elevation in the compensated laser point cloud elevation matrix to obtain a road surface depth matrix.

[0045] Specifically, this step includes the following steps:

[0046] 1. Perform a one-dimensional moving average filter on each laser point cloud elevation Z(x,y) in the compensated laser point cloud elevation matrix along the x direction to obtain a filtering result Z x (x,y).

[0047] Specifically, the first filtering result Z x The calculation formula for (x,y) is as follows:

[0048]

[0049] Among them, w is the half width of the filter window, which can generally be 3 to 5 pixels.

[0050] 2. Filter the result Z along the y direction x (x, y) is filtered again with one-dimensional moving average to obtain the road depth Z xy (x,y) road depth matrix.

[0051] Specifically, the road surface depth Z xy The calculation formula for (x,y) is as follows:

[0052]

[0053] Through this step, bidirectional high-pass filtering can be achieved, effectively eliminating the influence of the horizontal and vertical slopes of the surface on the crack depth and highlighting the local undulation characteristics of the crack. Figure 3 and 4 shown.

[0054] Step S104: converting the road surface depth matrix into a depth image, and cutting the depth image into a plurality of initial sub-image blocks of the same size.

[0055] Specifically, this step includes the following process:

[0056] 1. Each road depth Z in the road depth matrix xy (x, y) is normalized to the grayscale range of [0, 255], and the normalized road depth Z is obtained. img (x,y) composed of depth images.

[0057] Specifically, the calculation formula for the normalized road surface depth is as follows:

[0058]

[0059] Among them, max(Z) and min(Z) are the maximum and minimum values ​​of the road surface depth respectively.

[0060] 2. The depth image is evenly cut into initial sub-image blocks of a preset size and a preset resolution along the X direction and the Y direction.

[0061] The best cutting scale can be determined through experimental analysis. For example, the preset size can be 10cm×10cm, and the preset resolution can be 1mm×1mm, which can capture the local features of most cracks. The X direction and Y direction are the X direction and Y direction of the image coordinate system. The effect diagram is as follows Figure 5 shown.

[0062] By cutting the depth image, the local crack features can be better highlighted, which facilitates subsequent feature analysis and identification.

[0063] Step S105: Screening and obtaining potential crack feature points in each initial sub-image block by Gaussian background modeling.

[0064] Specifically, this step includes the following process:

[0065] 1. For each initial sub-image block, use the normalized road depth Z img (x, y) statistics are used to obtain the depth histogram of each initial sub-image block.

[0066] Obtaining a depth histogram through depth is a prior art and will not be described in detail here.

[0067] 2. Assuming that there are no cracks in each initial sub-image block and the road depth distribution of each feature point in each initial sub-image block approximately obeys the Gaussian distribution, the expectation maximization algorithm is used to fit the final depth mean and final depth standard deviation of the depth histogram of each initial sub-image block.

[0068] The road depth distribution approximately follows the Gaussian distribution and can be expressed as follows:

[0069]

[0070] The Expectation-Maximization (EM) algorithm is an iterative optimization algorithm that converges to the optimal solution by alternately calculating the expected value of latent variables and model parameters. Specifically, the EM algorithm is as follows:

[0071] (1) For each initial sub-image block, the depth mean and depth standard deviation of the depth histogram of the initial sub-image block are calculated as the depth mean and depth standard deviation of the initialized histogram.

[0072] (2) Calculate the probability that each feature point in the initial sub-image block belongs to the Gaussian background for the i-th time.

[0073] The calculation formula for the current probability obtained by the i-th calculation is as follows:

[0074] p (i) (Z img (x,y)|μ (i-1) ,σ (i-1) ).

[0075] Among them, μ (i-1) and σ (i-1) are the depth mean and depth standard deviation calculated for the i-1th time respectively.

[0076] (3) The depth mean and depth standard deviation of the depth histogram of the initial sub-image block are calculated using the probability for the i-th iteration.

[0077] The calculation formulas for the depth mean and depth standard deviation are as follows:

[0078]

[0079]

[0080] Among them, μ (i) and σ (i) are the depth mean and depth standard deviation calculated for the i-th time respectively.

[0081] (4) Repeat the above steps (2) and (3) until the iteration termination condition is met, and use the depth mean and depth standard deviation obtained when the iteration terminates as the final depth mean and final depth standard deviation.

[0082] Specifically, the iteration termination condition is: |μ (i) -μ (i-1) |<ε and |σ (i) -σ (i-1) |<ε, where ε is the convergence accuracy threshold.

[0083] 3. Calculate the normalized road depth Z of each initial sub-image block img The absolute value of the difference between (x, y) and the final depth mean of the initial sub-image block.

[0084] The absolute value of the difference can be expressed as follows:

[0085] ΔZ(x,y)=|Z img (x,y)-μ.

[0086] Where ΔZ(x,y) is the normalized road depth Z in the initial sub-image block. img The absolute value of the difference between (x, y) and the final depth mean of the initial sub-image block, μ is the final depth mean of the initial sub-image block.

[0087] 4. Using the final depth mean and final depth standard deviation of the initial sub-image block, calculate the adaptive threshold of the initial sub-image block.

[0088] Specifically, the calculation formula of the adaptive threshold is as follows:

[0089] D=μ+kσ.

[0090] Wherein, D is the adaptive threshold, k is the adaptive coefficient, which can be 2 to 3, and σ is the final depth standard deviation of the initial sub-image block.

[0091] By designing the calculation method of the above-mentioned adaptive threshold, it is automatically adjusted according to the actual depth distribution of different areas, so as to better adapt to the different surface undulations and improve the detection accuracy of crack feature points.

[0092] For flat areas, the depth distribution is more concentrated, σ is smaller, and the adaptive threshold D is also smaller, which is conducive to detecting fine cracks; for rough areas, the depth distribution is more dispersed, σ is larger, and the adaptive threshold D is also larger, which can filter out the noise points caused by surface undulations.

[0093] This step effectively extracts potential crack feature points from the three-dimensional point cloud data through the design of an adaptive threshold D, laying the foundation for the subsequent extraction of crack lines and geometric parameters.

[0094] 5. The normalized road depth Z in the initial sub-image block img The feature point whose absolute value of the difference corresponding to (x, y) is greater than the adaptive threshold is regarded as the potential crack feature point.

[0095] The traditional method of directly using the depth threshold is easily affected by complex surface morphology and produces more false detections. The embodiment of the present invention can effectively suppress interference such as surface undulations and shadows through the above-mentioned Gaussian background modeling method, accurately determine the final depth mean and final depth standard deviation of each local area, and use the final depth mean as the depth standard. It can more accurately identify points that deviate from the Gaussian distribution as potential crack feature points, further improving the robustness of the algorithm.

[0096] By fitting the depth features of each area with a Gaussian distribution model, a reasonable depth threshold can be adaptively determined to better adapt to different surface conditions. This adaptive threshold design improves the reliability of crack feature point detection. Gaussian background modeling utilizes the statistical distribution characteristics of depth data, without manually specifying the depth threshold, and can more objectively determine the boundary between cracks and non-crack areas. This automated feature extraction without manual intervention improves the applicability and versatility of the entire system.

[0097] Step S106: optimizing the grouping by the GWO algorithm, and dividing all the potential crack feature points that can constitute a connected area into one group.

[0098] Specifically, this step includes the following process:

[0099] 1. For each potential crack feature point, calculate the distance between the potential crack feature point and each other potential crack feature point, and divide the two potential crack feature points whose distance is less than the initial radius parameter into the same group.

[0100] The distance calculation can be calculated using the Euclidean distance calculation formula, as follows:

[0101]

[0102] Among them, (x 1 ,y 1 ) and (x 2 ,y 2 ) are the coordinates of two potential crack feature points. The coordinates are the coordinates of the crack feature points in the initial sub-image block in the image coordinate system.

[0103] When clustering, the initial radius parameter r is set, and the potential crack feature points with a distance less than the initial radius parameter r are grouped into the same group. The initial radius parameter r is generally 5-10 mm and can be adjusted according to actual conditions.

[0104] 2. For each group, the initial radius parameter is optimized by the gray wolf hunting algorithm to obtain the final radius parameter of the group, and the feature points whose distance from the final alpha wolf in the gray wolf hunting algorithm is greater than the final radius parameter of the group are removed from the group, and the remaining feature points in the group are retained as potential crack feature points that can constitute connected areas.

[0105] The Grey Wolf Optimizer (GWO) is used for optimization.

[0106] The GWO algorithm is a swarm intelligence optimization algorithm that simulates the social behavior of gray wolves, including the relationship between the leadership of alpha wolves, beta wolves, and delta wolves and the masses. Through continuous iterative optimization, a better potential crack feature point grouping scheme can be obtained. The GWO algorithm can be used for optimization to better discover the topological relationship between potential crack feature points, organize discrete potential crack feature points into connected areas with topological relationships, provide a basis for subsequent crack connection, and improve the accuracy of crack connection.

[0107] Specifically, this step includes the following process:

[0108] (1) Initialize the positions of n gray wolves.

[0109] by represents the individual position of the gray wolf, then Where n is the number of potential crack feature points in the group.

[0110] 2. Calculate the fitness function of each gray wolf individual The value of .

[0111] The fitness function can be defined as the average Euclidean distance from the gray wolf individual to each potential crack feature point in the group.

[0112] 3. Identify the alpha wolf, beta wolf, and delta wolf.

[0113] Among them, the alpha wolf The beta wolf is the individual with the highest fitness. The second most fit individual, delta wolf It is the individual with the third highest fitness.

[0114] 4. Update the positions of other individuals (i.e. mega wolves) based on the positions of alpha wolves, beta wolves, and delta wolves:

[0115]

[0116]

[0117]

[0118] in, The calculation method is the same as only The values ​​are different. and The calculation method is the same as only The values ​​are different. and is a vector of random numbers between [0,1]. is the radius parameter, it should be understood that It will change with the iteration update, and the average Euclidean distance from the current alpha wolf position to each potential crack feature point in the group is the current

[0119] 5. Repeat the above steps until the termination condition is met, and the average Euclidean distance from the current alpha wolf position to each potential crack feature point in the group is used as the final radius parameter.

[0120] The termination condition is reaching the number of iterations or the convergence of the fitness function value.

[0121] The algorithm automatically optimizes the grouping of feature points by defining a reasonable fitness function, making the points within the group as compact as possible and the points between groups as separated as possible. This adaptive grouping strategy can better adapt to different crack morphologies and distribution characteristics. Compared with the traditional simple clustering method based on distance threshold, the GWO algorithm can capture the complex relationship between feature points, effectively aggregate adjacent feature points, and maintain the separation between different crack areas, obtaining more reasonable grouping results, which can lay a good foundation for the accurate extraction of crack geometric parameters, thereby improving the accuracy of subsequent crack connection and parameter extraction.

[0122] Step S107: Eliminate all potential crack feature points in groups whose number is less than the number threshold, and eliminate irrelevant potential crack feature points in each group, to obtain a final sub-image block corresponding to the final group.

[0123] Specifically, all potential crack feature points in groups with a number less than a threshold value may be removed first. Groups with too few numbers may not represent real crack features. In this way, the removal of irrelevant potential crack feature points in each group may only be targeted at the remaining groups to improve efficiency.

[0124] The quantity threshold can be 5-10 and adjusted according to actual conditions.

[0125] Specifically, the removal of irrelevant potential crack feature points in each group can be determined in the following way:

[0126] 1. For each group, determine the distance between each potential crack feature point in the group and other potential crack feature points in the group, and determine the shortest distance corresponding to each potential crack feature point in the group.

[0127] The distance can be calculated using the Euclidean distance calculation formula.

[0128] 2. If the shortest distance corresponding to a potential crack feature point in the group is greater than the final radius parameter, the potential crack feature point is eliminated.

[0129] The feature point groups optimized in step S106 are analyzed to eliminate isolated feature points without any association, which are likely to be false features caused by noise or other interference.

[0130] Through the two elimination operations in this step, the reliability of crack feature points can be further improved, noise and irrelevant interference features can be filtered out, and the foundation can be laid for the subsequent extraction of crack lines and geometric parameters.

[0131] Step S108: Input each final sub-image block into a convolutional neural network, and output a result of whether the final sub-image block has a crack.

[0132] As a typical deep learning model, the convolutional neural network (CNN) model performs well in image classification tasks. The size of the final sub-image block can be 10cm×10cm. The input is a depth image, and the output is a binary classification result of whether the depth image is a crack.

[0133] The convolutional neural network model is pre-trained to make its output more accurate. The training samples include a large number of depth images of actual crack areas and non-crack areas (such as logos, markings, etc.), and the parameters of the CNN model are trained through supervised learning. Assume that there are M samples in the training data set, and the i-th sample is Among them, y i ∈{0,1}, indicating whether the sample is a crack area (0 means non-crack, 1 means crack). The objective function of the CNN model is:

[0134]

[0135] Among them, Θ is the parameter of the CNN model, is the output function of the CNN model, is the cross entropy loss function.

[0136] Using the trained CNN model, all the final sub-image blocks are classified and inferred, the parts identified as non-crack areas are eliminated, and the areas that may contain cracks are retained. The recognition results are as follows: Figure 6 and 7 shown.

[0137] This step uses deep learning technology to accurately separate crack features from 3D point cloud data and eliminate the influence of other interfering areas (such as logos, markings, etc.). The CNN model can effectively capture the deep features of the crack area and improve the accuracy and reliability of crack detection.

[0138] Preferably, after step S108, the method of the embodiment of the present invention further includes:

[0139] Step S109: the calculation result is the characteristic value of the crack in the final sub-image block having the crack.

[0140] Specifically, this step includes the following process:

[0141] 1. For a final sub-image block having a crack, extract the coordinates of each feature point on the contour line of the crack in the final sub-image block.

[0142] Extracting the coordinates of the feature points on the contour line of the crack is an existing technology. Specifically, according to the spatial distribution and topological relationship of the feature points on the contour line, image refinement, skeleton extraction and other methods can be used to complete the extraction of the feature points on the contour line of the crack to obtain the contour connection line.

[0143] Assume that the coordinates of the jth feature point on the crack contour line are (x j ,y j ,Z img (x j ,y j )).

[0144] 2. A curve model of the crack in the final sub-image block is obtained by fitting the coordinates of each feature point on the contour line of the crack in the final sub-image block.

[0145] The fitting can be performed using existing fitting methods, for example, using geometric models such as B-spline curves and polynomial curves for fitting.

[0146] Take the B-spline curve as an example. The B-spline curve is a commonly used smooth interpolation curve, and its expression is:

[0147]

[0148] in, is the control point, is the B-spline basis function, d is the order of the spline curve, usually 3 or 4. The B-spline curve has the characteristics of local control and smoothness, and can better fit the geometric shape of the crack.

[0149] 3. Calculate the characteristic value of the crack in the final sub-image block according to the curve model of the crack in the final sub-image block.

[0150] The characteristic values ​​of the crack include: length, width, depth and depth-to-length ratio of the crack.

[0151] By fitting the curve model obtained, quantitative parameters such as the length, width, and depth of the crack can be calculated. The crack depth can be directly obtained from the Z coordinate of the fitting curve, the length can be calculated from the projected length of the curve, and the width can be calculated from the height difference on both sides of the crack. The specific calculation formula is as follows:

[0152] Length of crack:

[0153] The width of the crack: W = max{Z img (x j ,y j )}-min{Z img (x j ,y j )}.

[0154] Crack depth: D = min{Z img (x j ,y j )}.

[0155] Crack Depth-Length Ratio: D / L

[0156] The depth-to-length ratio of a crack can be used to characterize the extent of crack development.

[0157] The above quantitative parameters are helpful to further analyze the severity of cracks.

[0158] Through this step, the remaining crack feature points are connected into crack lines according to the topological relationship, and the crack lines are accurately represented using geometric models such as curve fitting. The crack depth, length, ratio and other indicators are calculated to further complete crack identification. The accurate extraction of crack geometric parameters from three-dimensional point cloud data is achieved, providing a key basis for subsequent structural health monitoring and assessment.

[0159] In addition, an embodiment of the present invention further provides a computer-readable storage medium, on which computer program instructions are stored; when the computer program instructions are executed by a processor, the pavement crack identification method as described in the above embodiment is implemented.

[0160] In addition, an embodiment of the present invention further provides a road surface crack identification system, including: a computer-readable storage medium as described in the above embodiment.

[0161] In summary, the embodiments of the present invention can accurately identify road cracks with high recognition accuracy and can accurately extract geometric parameters of the cracks.

[0162] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A road crack identification method, It is characterized in that include: When the inspection vehicle is driving on the road, the road surface is continuously scanned by a line structured light scanning device carried by the inspection vehicle to obtain the original laser point cloud elevation matrix of the road surface; The inertial navigation device carried by the inspection vehicle is used to measure the real-time posture data of the line structured light scanning device, and the real-time posture data is used to compensate the original laser point cloud elevation matrix to obtain a compensated laser point cloud elevation matrix; Filtering each laser point cloud elevation in the compensated laser point cloud elevation matrix to obtain a road surface depth matrix; Converting the road surface depth matrix into a depth image, and cutting the depth image into a plurality of initial sub-image blocks of the same size; Screening and obtaining potential crack feature points in each of the initial sub-image blocks by Gaussian background modeling; Optimizing the grouping by the GWO algorithm, dividing all the potential crack feature points that can constitute a connected area into one group; Eliminating all potential crack feature points in the groups whose number is less than a threshold value, and eliminating irrelevant potential crack feature points in each group, to obtain a final sub-image block corresponding to a final group; Each of the final sub-image blocks is input into a convolutional neural network, and a result of whether the final sub-image block has a crack is output.

2. The pavement crack identification method according to claim 1, It is characterized in that After the step of outputting the result of whether the final sub-image block has cracks, the method further includes: The result is calculated as a characteristic value of a crack in the final sub-image block having a crack, wherein the characteristic value of the crack includes: a length, a width, a depth, and a depth-to-length ratio of the crack.

3. The pavement crack identification method according to claim 1, It is characterized in that The step of obtaining the compensated laser point cloud elevation matrix comprises: Performing Kalman filtering on the real-time posture data of the line structured light scanning device to obtain six-axis posture data of the line structured light scanning device; The six-axis posture data of the line structured light scanning device are used to perform coordinate system conversion and displacement compensation on each laser point cloud elevation in the original laser point cloud elevation matrix to obtain a compensated laser point cloud elevation matrix.

4. The pavement crack identification method according to claim 1, It is characterized in that The step of obtaining the road surface depth matrix comprises: Perform a one-dimensional moving average filter on each laser point cloud elevation Z(x,y) in the compensated laser point cloud elevation matrix along the x direction to obtain a filtering result Z x (x,y); The result of a single filter Z along the y direction x (x, y) is filtered again with one-dimensional moving average to obtain the road depth Z xy The road depth matrix composed of (x, y); Among them, the first filtering result Z x The calculation formula for (x,y) includes: Wherein, the road surface depth Z xy The calculation formula for (x,y) includes: Where w is the half width of the filter window.

5. The pavement crack identification method according to claim 3, It is characterized in that The step of cutting the depth image into a plurality of initial sub-image blocks of the same size comprises: Each road depth Z in the road depth matrix xy (x, y) is normalized to the grayscale range of [0, 255], and the normalized road depth Z is obtained. img The depth image composed of (x, y); Uniformly cutting the depth image into initial sub-image blocks of a preset size and a preset resolution along the X direction and the Y direction; The calculation formula of the normalized road surface depth includes: Among them, max(Z) and min(Z) are the maximum and minimum values ​​of the road surface depth respectively.

6. The pavement crack identification method according to claim 5, It is characterized in that The step of screening and obtaining potential crack feature points in each of the initial sub-image blocks by Gaussian background modeling includes: For each of the initial sub-image blocks, the normalized road depth Z is used. img (x, y) obtains a depth histogram of each of the initial sub-image blocks; Assuming that there is no crack in each of the initial sub-image blocks, and that the road depth distribution of each feature point in each of the initial sub-image blocks approximately obeys a Gaussian distribution, an expectation maximization algorithm is used to fit the final depth mean and final depth standard deviation of the depth histogram of each of the initial sub-image blocks; Calculate the normalized road depth Z of each of the initial sub-image blocks img The absolute value of the difference between (x, y) and the final depth mean of the initial sub-image block; Calculating an adaptive threshold of the initial sub-image block by using a final depth mean and a final depth standard deviation of the initial sub-image block; The normalized road depth Z in the initial sub-image block img The feature point whose absolute value of the difference corresponding to (x, y) is greater than the adaptive threshold is taken as a potential crack feature point; The calculation formula of the adaptive threshold is as follows: D = μ + kσ; Where D is the adaptive threshold, k is the adaptive coefficient, μ is the final depth mean of the initial sub-image block, and σ is the final depth standard deviation of the initial sub-image block.

7. The pavement crack identification method according to claim 1, It is characterized in that The step of dividing the potential crack feature points that can constitute a connected area among all the potential crack feature points into one group comprises: For each of the potential crack feature points, calculate the distance between the potential crack feature point and each of the other potential crack feature points, and divide the two potential crack feature points whose distance is smaller than the initial radius parameter into the same group; For each group, the initial radius parameter is optimized by the gray wolf hunting algorithm to obtain the final radius parameter of the group, and the feature points whose distance from the final alpha wolf in the gray wolf hunting algorithm is greater than the final radius parameter of the group are removed from the group, and the remaining feature points in the group are retained as the potential crack feature points that can constitute the connected area.

8. The road surface crack identification method according to claim 7, It is characterized in that The step of removing the irrelevant potential crack feature points in each group includes: For each of the groups, determining the distance between each potential crack feature point in the group and other potential crack feature points in the group, and determining the shortest distance corresponding to each potential crack feature point in the group; If the shortest distance corresponding to a potential crack feature point in the group is greater than the final radius parameter, the potential crack feature point is removed.

9. A computer-readable storage medium, Features: The computer-readable storage medium stores computer program instructions; when the computer program instructions are executed by a processor, the pavement crack identification method according to any one of claims 1 to 8 is implemented.

10. A road crack identification system, It is characterized in that include: The computer readable storage medium of claim 9.