CN110197157B - Pavement crack growth detection method based on historical crack data - Google Patents

Abstract

The invention discloses a road surface crack growth detection method based on historical crack data, which comprises the following steps: GPS initial positioning, namely extracting a plurality of similar images close to the current road surface image by comparing the current positioning information with position information in historical map data; obtaining precisely matched image data in a plurality of similar images close to the current road surface image after the image-level positioning is matched with the characteristic points; the pixel level positions image characteristic points matched through the ORB, calculates an H matrix, and maps historical mark crack pixels to a current crack image by using the H matrix; historical crack pixel distribution mapped in the current crack image is analyzed based on the historical cracks of the RGM, the intensity distribution of the pixels of the mapped cracks is represented by a Gaussian model, and finally the pixel values meeting the conditions are divided into cracks. The method of the invention provides an effective and reliable strategy for researching the change of the crack state along with time by referring to historical crack data, thereby greatly simplifying and improving the detection and identification of the crack.

Description

A pavement crack growth detection method based on historical crack data

Technical Field

The invention relates to image processing technology, and in particular to a pavement crack growth detection method based on historical crack data.

Background Art

At present, there have been some achievements in the detection of pavement cracks, such as the patent CN106548182, which was applied for on November 2, 2016, and the patent name is "Pavement crack detection method and device based on deep learning and main cause analysis", which discloses a pavement crack detection method based on convolutional neural network and main cause analysis. The patent CN106324084, which was applied for on August 30, 2016, and the patent name is "Crack depth detection method", discloses a method and instrument for detecting crack depth value.

In recent years, with the rapid development of pavement crack detection technology, the demand for pavement crack detection methods, especially crack propagation analysis of pavement, has become increasingly urgent. At present, the supervised method based on machine learning is the main means of pavement crack detection. The premise of the application of this method is to train a large amount of labeled data, which is difficult to be widely used in practice. This patent proposes a method of mapping and comparing historical image data with existing crack images at a certain time interval, which can improve the detection accuracy while reducing the amount of training data.

Summary of the invention

The technical problem to be solved by the present invention is to provide a pavement crack growth detection method based on historical crack data in view of the defects in the prior art.

The technical solution adopted by the present invention to solve the technical problem is: a pavement crack growth detection method based on historical crack data, comprising the following steps:

1) Real-time acquisition of the current pavement crack image and synchronous acquisition of the positioning information corresponding to the current pavement crack image;

2) Initial positioning: Based on the current positioning information and the location information in the historical map data, extract the historical image data within the threshold distance from the current positioning information from the historical road image data, and select multiple historical images similar to the current road image;

3) Image-level positioning: Find an image closest to the road surface among the similar images roughly matched in step 2), as follows:

3.1) grayscale processing of the image;

3.2) Calculate the Hamming distance of all feature descriptors of the current grayscale image and similar images, compare the Hamming distance to achieve rough matching of feature points, and thus find the same feature points of the current grayscale image and similar images. These matched feature points are used as road surface fingerprint information;

3.3) By comparing the number of identical feature points, find an image closest to the road surface in the roughly matched similar images;

4) Pixel-level localization: Map the historical cracks marked in the closest pavement image to the current pavement crack image;

5) Based on the RGM (Region Growth Method) mapping crack analysis of the current pavement crack image, all pixels belonging to newly grown cracks in the collected crack image are detected.

According to the above scheme, the method for obtaining the feature descriptor in step 3.2) is as follows:

3.2.1) Detect image feature points: Detect corner points as image feature points using the Harris algorithm;

3.2.2) Adding direction information: Adding direction information to the extracted feature points so that the direction of the extracted feature points remains unchanged. The direction is obtained by calculating the entropy value of the pixel block through the moment of the image in the circular window;

3.2.3) ORB feature point matching:

Select n pairs of features on the feature points to form a mapping matrix s. The size of the matrix s is 2×2n, and its elements are the coordinates of each feature pair on the X and Y axes. Then use the direction from the feature point to the centroid to get the affine transformation matrix R. Use the matrix R to calculate the new description matrix s _θ , and combine it with the BRIEF descriptor to get the ORB feature descriptor.

According to the above scheme, the method for determining the image feature point in step 3.2.1) is: comparing the difference between a pixel point P in the image and the pixel points on a circular window composed of a number of surrounding points, the sum of the differences of the several points is N, and when N is greater than the determination criterion, the point is determined to be an image feature point;

Among them, I(x) is the gray value of the current pixel, I(p) is the gray value of the pixel P, ε is the set threshold, circle(p) is the radiation range on the circular window around the pixel P, and the formula is substituted for the pixels that meet the range requirements.

According to the above scheme, the pixel-level positioning in step 4) is to map the marked historical crack pixels into the current crack image.

According to the above scheme, the pixel-level positioning in step 4) is to obtain the H matrix of the two images by matching the image feature points through ORB, and then use the H matrix to map the marked historical crack pixels to the current crack image.

According to the above scheme, the specific steps of crack mapping based on multi-scale positioning in step 4) are:

Assuming that the road is a plane, the basic geometric shape can be described by the homography matrix under the pinhole camera model. Then, the two linear constraints on the homography matrix show that the historical crack labels are mapped to the query image through the following relationship:

Where n is the number of historical crack label pixels, [u′ _i y′ _i ] ^T is the coordinate of the mapped crack data in the historical data, and [u _i y _i ] ^T is the coordinate of the historical crack label pixel.

According to the above scheme, the specific steps of historical crack analysis based on RGM in step 5) are:

5.1) Mapped crack grayscale value distribution analysis: The query crack labels mapped in the above step are used as "ideal" initial seed points. Starting from these seed points, the region is expanded by finding adjacent points with similar attributes to the seed points, including color and intensity. By mapping the labels, the pixel values are calculated from the mapping correspondence on the query image, and the image intensity distribution histogram of the mapped crack data is plotted using statistical principles.

5.2) Use Gaussian model to represent the intensity distribution of mapped crack pixels:

A Gaussian model is established to represent this distribution, and its distribution characteristics are used as operators. The corresponding mean and standard deviation calculation formulas are:

Where N is the number of pixels mapped to crack labels.

5.3) Growth crack analysis: When the intensity of a point in the image meets the following conditions, it is classified as a crack:

I(p _μ ,p _v )∈[0,ω+λσ]

Where λ is a constant value determined according to the actual application situation.

According to the above solution, the positioning information in step 1) is GPS positioning information.

The beneficial effects of the present invention are as follows: crack detection is carried out through three steps: GPS initial positioning, image-level positioning, and historical crack analysis based on RGM; GPS initial positioning extracts multiple similar images close to the current road surface image by comparing the current positioning information with the location information in the historical map data. Image-level positioning detects corner points through the Harris algorithm, and adds direction information to the extracted feature points so that the direction of the extracted feature points remains unchanged, and obtains precisely matched image data after ORB feature point matching. Historical crack analysis based on RGM maps the matched historical crack labels to the query image, and uses a Gaussian model to represent the intensity distribution of the mapped crack pixels, and finally classifies the pixel values that meet the conditions as cracks. This method proposes an effective and reliable strategy for studying the changes in crack status over time by referring to historical crack data, which greatly simplifies and improves the detection and identification of cracks.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

FIG1 is a flow chart of a method according to an embodiment of the present invention;

FIG2 is a flow chart of a method according to an embodiment of the present invention;

FIG3 is a schematic diagram of initial positioning of coordinate matching according to an embodiment of the present invention;

FIG4 is a schematic diagram of a seed point and a growth area according to an embodiment of the present invention;

FIG. 5 is an intensity histogram of a mapped crack data image according to an embodiment of the present invention.

DETAILED DESCRIPTION

In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

As shown in Figures 1 and 2, the crack growth detection method based on historical crack data of the embodiment of the present invention includes: multi-scale positioning; crack mapping based on multi-scale positioning and mapping crack analysis based on RGM. Among them, multi-scale positioning includes GPS initial positioning, image-level positioning and pixel-level positioning; mapping crack analysis based on RGM includes historical crack gray value distribution analysis, Gaussian model construction and growth crack analysis.

Among them, the specific steps of GPS initial positioning are:

S1. The image sensor module installed on the vehicle to be positioned collects the current road crack image in real time, and the positioning module simultaneously collects the current positioning information;

S2. By comparing the current positioning information with the location information in the historical map data, multiple similar images close to the current road image are extracted from the historical map data, and finally the historical image data within the threshold distance is obtained. The following formula is satisfied:

d _ji = dist(G _j ,G _i )

Pos＝{G _i |d _ij ≤k}

Among them, _Gj is the GPS coordinate of the j-th query crack data, _Gi is the GPS coordinate of the i-th historical crack data, d is the GPS distance between the current crack data and the historical crack data, k is the threshold distance of the candidate historical crack data, and Pos is a set of historical crack candidate images.

At the same time, the positioning accuracy of GPS in practical applications is 10 meters. In this step, the positioning accuracy is used as a threshold to perform a preliminary screening of historical cracks, thereby completing the GPS initial positioning and serving as the first step of multi-scale positioning.

The specific steps of image-level positioning are:

S1. Detect image feature points: Detect corner points using the Harris algorithm, that is, compare the difference between a pixel point P in the image and the pixel points on the circle composed of several surrounding points. The formula is:

Among them, I(x) is the gray value of the current pixel, I(p) is the gray value of the pixel P, ε is the set threshold, and the sum of the response function values of several points is N. When N is greater than the judgment standard, the point is an image feature point;

Generally, we take the integer part of 3/4 of the N value as the judgment standard;

S2. Adding direction information: Adding direction information to the extracted feature points makes the direction of the extracted feature points unchanged. The direction is calculated by the moment of the image in the circular window. The moment of the image is defined as follows:

Therefore, the entropy value of the image is calculated at each moment as follows:

The orientation of a pixel is defined as follows:

θ＝arctan(m ₁₀ ,m ₀₁ )

Where, m _pq , (p, q∈(0,1)) represents the moment of the image block, I(x, y) represents the gray value at the coordinate (x, y), and θ is the angle between the centroid and the feature point;

S3, ORB feature point matching:

Select n pairs of features at the feature points to form a matrix s:

The size of the matrix s is 2×2n, where x _i , y _i (i∈(1,n)) represents the coordinates of the i-th feature pair on the X and Y axes. The direction from the feature point to the centroid is used to obtain the affine transformation matrix R. The new description matrix s _θ is calculated using the matrix R. The formula is:

Combined with the BRIEF descriptor, we get the ORB feature descriptor:

g _n (P, θ)＝f _n (P)|(x _i ,y _i )∈S _θ

At the same time, the selected value of n is 256, and based on the obtained 256-dimensional ORB feature descriptor, the Hamming distance of all feature descriptors of the current grayscale image to be located and similar images is calculated, and the Hamming distance is compared to achieve coarse matching of feature points, and the RANSAC algorithm is used to remove incorrect matches, so as to find the same feature points of the current grayscale image to be located and similar images, and these matched feature points are used as road surface fingerprint information; by comparing the number of identical feature points, an image closest to the road surface is found in the coarsely matched similar images, and the image has the largest number of feature points matching the image to be located, so as to achieve precise matching of the image to be located using the fingerprint information.

Pixel-level positioning is used to find the H matrix of the two images, and then use the H matrix to map the marked historical cracks to the current crack image. The specific steps are:

Assuming that the road is a plane, under the pinhole camera model, the basic geometric shape can be described by the homography matrix. The specific formula is as follows:

_hi represents the components of the matrix H.

Rewrite the above formula as follows:

From the above formula, we can get two linear constraints on the homography matrix:

Using the calculated homography matrix, the historical crack labels are mapped to the query image through the following relationship:

Among them, n is the historical crack label pixel, [u′ _i y′ _i ] ^T is the coordinate of the mapped crack data in the historical data, and [u _i y _i ] ^T is the coordinate of the historical crack label pixel.

Therefore, the mapped crack labels on the query image can be represented by a set of 2D image coordinates as follows:

Q＝{[u′ _i y′ _i ] ^T }(i＝1,2,…n)

The specific steps of historical crack analysis based on RGM are:

S1. Analysis of historical crack grayscale value distribution: The query crack labels mapped in the above steps are used as "ideal" initial seed points. Starting from these seed points, the region is expanded by finding adjacent points with similar properties to the seed points (e.g., similar color, intensity, etc.). By mapping the labels, the pixel values can be calculated from the mapped correspondence on the query image. Using statistical principles, the image intensity distribution histogram of the mapped crack data is plotted.

S2. Use the Gaussian model to represent the intensity distribution of the mapped crack pixels:

All of the above pixel values I(μ',ν') associated with the mapped crack labels follow a certain pattern, namely Gaussian distribution, so Gaussian distribution can be used as a constraint for crack extension. To this end, a Gaussian model is established to represent this distribution, and its distribution characteristics are used as operators. The corresponding mean and standard deviation can be calculated as:

Where N is the number of pixels mapped to crack labels.

It should be noted that each query crack image has a unique developed Gaussian model, so the image used is robust and reliable when performing region growing calculations. Through the constructed Gaussian model, it is possible to quickly determine whether the adjacent image has similar properties to the seed point. At the same time, when performing region growing calculations, since the crack area in the image usually has a lower image intensity, the range of image intensity, such as the mean and standard deviation, can be set from the Gaussian model parameters.

S3. Growth crack analysis: When the intensity of a point in the image meets the following conditions, it is classified as a crack:

I(p _μ ,p _v )∈[0,ω+λσ]

Where λ is a constant determined according to the actual application situation.

In another specific embodiment of the present invention:

The reference-based crack analysis method of the embodiment of the present invention consists of three main modules: 1) multi-scale positioning; 2) mapping the historical crack image to the query crack image; 3) crack post-processing and analysis. Each query crack data and historical crack data contains GPS information and crack images. In addition, the historical crack labels in the historical crack data, whether manually extracted or automatically extracted, are represented as a set of pixels belonging to the crack in the pavement image (pixel level). Therefore, a point set is used to represent each historical crack data as follows:

m _i ={G _i ,I _i ,L _i } i∈(1,n)

Where n is the number of historical crack data, _Gi is GPS information, _Ii is the pavement crack image, and _Li is the historical crack label.

The specific process is shown in Figure 1.

The main purpose of the multi-scale positioning module is to establish the image correspondence between the current fracture data and the historical fracture data. This method adopts a coarse-to-fine strategy to achieve multi-scale positioning through three steps: GPS initial positioning, image-level positioning, and pixel-level positioning.

GPS initial positioning:

The initial positioning of GPS data is shown in Figure 3. Let _Gj be the GPS coordinates of the jth query crack data, and _Gi be the GPS coordinates of the i-th historical crack data. The distance between the two can be calculated as:

d _ji = dist(G _j ,G _i )

Using GPS data matching, a set of candidate historical crack images is obtained whose associated GPS coordinates and query GPS coordinates are within a threshold distance. GPS-based initial positioning allows a limited number of candidate images to be extracted from a large number of collected historical crack images. The mathematical description of the task is as follows:

Pos＝{G _i |d _ij ≤k}

Among them, k is the threshold distance for selecting candidate historical crack data. It determines whether the i-th historical crack data is close enough to the j-th query crack data. In practical applications, according to the error of GPS positioning, the threshold distance is 10 meters.

Based on the initial positioning results, image matching is applied to achieve image-level positioning. The purpose of image-level positioning is to find the "nearest" historical crack image in the candidate historical crack data. "Nearest" means that the distance between the camera position associated with the query crack image and the matched historical crack image is the smallest among all historical crack images. This method uses the matching of local image feature point pairs to achieve image-level positioning.

The matching of local image feature point pairs is implemented using the ORB algorithm. ORB is an image matching algorithm that combines oFAST (FAST with orientation) and rBRIEF (rotated BRIEF). All images are represented by some local feature points. The local feature point pairs of two images represent the same area in the two images. Among them, oFAST is used for feature point detection and rBRIEF is used for feature descriptor calculation.

oFAST first uses the Harris angle measure to select different feature points. Secondly, the direction information is added so that the extracted feature points are direction-invariant. The direction is calculated by the moment of the image within the circular window. The moment of the image is defined as follows:

Therefore, the entropy value of the image block is calculated at each moment as follows:

The orientation of the image patch is defined as follows:

θ＝atan2(m ₀₁ ,m ₁₀ )

Among them, atan2 is a variant of the tangent function tan, and the BRIEF descriptor is a bit string description of an image patch constructed by a set of binary intensity tests. Consider a smooth image pixel P, two binary test indicators x and y at arbitrary positions, and the logical result of comparing their image intensities:

where P(x) is the intensity of the image at a point x. Therefore, the BRIEF descriptor is defined as a vector of n binary test indicators x and y:

In the literature, there are many solutions on how to select n binary tests. In the embodiment of the present invention, a Gaussian distribution around the center of the image block is used, and the vector length n=256 is selected. Therefore, the ORB feature descriptor is represented by a 256-bit string. In order to make the BRIEF descriptor invariant to rotation, the BRIEF description is performed according to the direction of the key point. For any binary test of feature set n, define the following 2×n matrix

From the direction θ of the image patch, the corresponding rotation matrix can be calculated as follows:

Then a controllable S _θ can be constructed through the rotation matrix as shown below:

S _θ ＝R _θ S

Therefore, the rotation invariant descriptor, also called ORB descriptor, can be calculated as follows:

g _n (Ρ, θ)＝f _n (Ρ)|(x _i ,y _i )∈S _θ

Map historical crack labels to query images:

Once the “recent” historical crack image is obtained, the underlying geometric relationship between the query and historical images can be obtained. Assuming the sidewalk is a plane, under the pinhole camera model, the underlying geometry can be described by the homography matrix as follows:

_hi represents the components of the matrix H.

The above two equations can be rewritten as follows:

From the above formula, we can get two linear constraints on the homography matrix:

Since the homography matrix can be determined on a scale, it can be calculated from at least 4 point pairs. In practical applications, the direct linear transformation (DLT) can be applied to calculate the homography matrix, and the result can be optimized using the Levenberg-Marquardt (LM) method.

Using the calculated homography matrix, the historical crack labels can be mapped to the query image as shown below:

Where n is the historical crack label pixel. [u′ _i y′ _i ] ^T is the coordinate of the mapped crack data in the historical data. [u _i y _i ] ^T is the coordinate of the historical crack label pixel.

Therefore, the mapped crack labels on the query image can be represented by a set of 2D image coordinates as shown below:

Q＝{[u′ _i y′ _i ] ^T }(i＝1,2,…n)

Crack post-processing and regional growth analysis:

In practical applications, query crack labels can be quickly implemented by mapping historical crack labels into query crack images. Since the crack conditions may deteriorate during the time interval between the query and historical crack data collection, there are still some pixels belonging to newly generated cracks that need to be detected. Since RGM relies on good crack seed points, the query crack labels mapped in the above steps can be used as "ideal" initial seed points. Starting from these seed points, the region is expanded by finding neighboring points with similar properties to the seed points (e.g., similar color, intensity, etc.), as shown in Figure 4. The red point in the middle is the initial seed point, and the blue point is the growth area.

By mapping the labels, the pixel values can be calculated from the mapped correspondence on the query image. All of these pixel values associated with the mapped crack labels follow a certain pattern, as shown in Figure 5. Figure 5 shows a typical image intensity distribution similar to a Gaussian model mapping crack pixels, which can be used as a constraint for crack extension. Therefore, the embodiment of the present invention develops a Gaussian model to accomplish this task. The corresponding mean and standard deviation can be calculated as:

Where N is the number of pixels mapped to crack labels.

It is important to note that each query crack image has a unique developed Gaussian model, making it robust and reliable for region growing. From the computed Gaussian model, it is quickly possible to determine whether neighboring images have similar properties to the seed point. Since crack regions in an image usually have lower image intensity, the range of image intensity can be set from the Gaussian model parameters such as mean and standard deviation. Therefore, a point is classified as a crack when its intensity satisfies the following conditions:

I(p _μ ,p _v )∈[0,ω+λσ]

Where λ is the determination ratio, which is set based on experience in practical applications.

Therefore, RGM can apply the above formula to detect all adjacent image points near the seed point. Using RGM, all pixels belonging to newly grown cracks in the query crack image can be detected.

In another specific embodiment of the present invention:

The present invention provides a crack growth detection method based on historical crack data, comprising the following steps:

Historical map collection stage:

S1. The image sensor module installed on the map collection vehicle collects road images in real time, and the positioning module and the inertial navigation module synchronously collect positioning information and inertial navigation information;

S2, preprocessing the collected road surface image to obtain a corresponding grayscale image; preprocessing the positioning information and inertial navigation information to obtain position information corresponding to the grayscale image;

S3, matching each grayscale image with the position information one by one, combining the actual vehicle speed and the acquisition frequency of the image sensor module, making each grayscale image have a fixed interval, and obtaining an atlas;

Positioning phase:

S4, collecting the current road image in real time through the image sensor module set on the vehicle to be positioned, and the positioning module synchronously collects the current positioning information, and extracts multiple similar images close to the current road image from the map set by comparing the current positioning information with the position information in the map set;

S5, preprocessing the current road surface image to obtain a current grayscale image; extracting image features from the current grayscale image and similar images to obtain feature points that meet feature uniqueness, time invariance, and feature translation and rotation invariance as road surface fingerprint information, and obtaining an image that is closest to the road surface based on the road surface fingerprint information;

S6. According to the relative position relationship between the fingerprint information in the closest road surface image and the current road surface image, the accurate positioning result of the current road surface image is calculated.

Furthermore, the method for extracting image features in step S5 of the present invention is:

S51, detecting image feature points: detecting corner points by Harris algorithm, that is, comparing the difference between a pixel point P in the image and the pixel points on the circle composed of several surrounding points, the formula is:

Among them, I(x) is the gray value of the current pixel, I(p) is the gray value of the pixel P, ε is the set threshold, and the sum of the response function values of several points is N. When N is greater than the judgment standard, the point is an image feature point;

Determine the main direction of the feature point based on the direction from the feature point to the centroid:

θ＝arctan(m ₁₀ ,m ₀₁ )

Where, m _pq , (p, q∈(0,1)) represents the moment of the image block, I(x, y) represents the gray value at the coordinate (x, y), and θ is the angle between the centroid and the feature point;

S52, describe the feature points:

Select n pairs of features at the feature points to form a matrix s:

The size of the matrix s is 2×2n, where x _i , y _i (i∈(1,n)) represents the coordinates of the i-th feature pair on the X and Y axes. The direction from the feature point to the centroid is used to obtain the affine transformation matrix R. The new description matrix s _θ is calculated using the matrix R. The formula is:

Combined with the BRIEF descriptor, we get the ORB feature descriptor:

g _n (P, θ)＝f _n (P)|(x _i ,y _i )∈M _θ

Among them, n is 256;

S53, feature point matching: Based on the obtained 256-dimensional ORB feature descriptor, calculate the Hamming distance of all feature descriptors of the current grayscale image to be located and similar images, compare the Hamming distance to achieve rough matching of feature points, and use the RANSAC algorithm to remove incorrect matches, so as to find the same feature points between the current grayscale image to be located and similar images, and these matched feature points are used as road surface fingerprint information; by comparing the number of identical feature points, find an image closest to the road surface in the roughly matched similar images, which has the largest number of feature points matching with the image to be located, so as to use the fingerprint information to achieve precise matching of the image to be located.

Furthermore, the method for calculating the accurate positioning result of the current road surface image in step S6 of the present invention is:

The relative position relationship between the two images is obtained by using the fingerprint information of the current road surface image to be located and the similar images. The position information of the current road surface image to be located is calculated by using the position and relative position relationship of the similar images, thereby realizing the metric-level positioning of the vehicle.

Furthermore, in step S2 of the present invention, the positioning information and the inertial navigation information are preprocessed to obtain the position information corresponding to the grayscale image by converting the GPS data and the inertial navigation data into the position information represented by longitude and latitude.

Furthermore, in step S3 of the present invention, the fixed interval of each image in the atlas is 0.5 m.

It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims (9)

1. A pavement crack growth detection method based on historical crack data is characterized by comprising the following steps:

1) Acquiring a current pavement crack image in real time, and synchronously acquiring positioning information corresponding to the current pavement crack image;

2) Initial positioning: according to the current positioning information and the position information in the historical map data, extracting historical image data within a threshold distance from the current positioning information from the historical road surface image data, and selecting a plurality of similar images of the current road surface image from the historical road surface image data;

3) Image level localization: finding a closest road surface image from the roughly matched similar images in the step 2), wherein the method specifically comprises the following steps:

3.1 Gray-scale processing the image;

3.2 Calculating the Hamming distances of all feature descriptors of the current gray image and the similar image, comparing the Hamming distances to realize coarse matching of feature points, and then finding out the same feature points of the current gray image and the similar image, wherein the matched feature points are used as pavement fingerprint information;

3.3 By comparing the number of the same feature points, a closest road surface image is found in the rough matching similar images;

4) Pixel level positioning: mapping the historical cracks which are closest to the marks in the pavement image into the current pavement crack image;

5) And (3) carrying out RGM-based mapping crack analysis on the current pavement crack image, and detecting all pixels belonging to newly grown cracks in the collected crack image.

2. The method for detecting the crack growth of the road surface based on the historical crack data as claimed in claim 1, wherein the method for acquiring the feature descriptors in the step 3.2) is as follows:

3.2.1 Detect image feature points: detecting angular points as image characteristic points through a Harris algorithm;

3.2.2 Add direction information: adding direction information to the extracted characteristic points to enable the directions of the extracted characteristic points to be unchanged, wherein the directions are obtained by calculating entropy values of pixel blocks through moments of images in a circular window;

3.2.3 ORB feature point matching:

selecting n pairs of features on the feature points to form a mapping matrix s, wherein the size of the matrix s is 2X 2n, the elements of the matrix s are the coordinates of each feature pair on the X axis and the Y axis, then obtaining an affine transformation matrix R by utilizing the directions from the feature points to the centroid, and obtaining a new description matrix s by utilizing the matrix R for calculation _θ And combining the BRIEF descriptor to obtain the ORB feature descriptor.

3. The road surface crack growth detection method based on historical crack data of claim 2, wherein the method for judging the image characteristic points in the step 3.2.1) is as follows: comparing the difference values of pixel points on a circular window formed by one pixel point P in the image and a plurality of surrounding points, wherein the sum of the difference values of the plurality of points is N, and when N is greater than a judgment standard, judging the point as an image feature point;

wherein, I (x) is the gray value of the current pixel point, I (P) is the gray value of the pixel point P, epsilon is the set threshold, circle (P) is the radiation range on the circular window around the pixel point P.

4. The method for detecting road surface crack growth based on historical crack data of claim 1, wherein the pixel-level positioning in the step 4) is mapping marked historical crack pixels into the current crack image.

5. The method for detecting the crack growth on the road surface based on the historical crack data as claimed in claim 2, wherein the pixel-level positioning in the step 4) is to find an H matrix of two images through image feature points matched by an ORB, and then map the marked historical crack pixels into the current crack image by using the H matrix.

6. The method for detecting the crack growth on the road surface based on the historical crack data according to claim 5, wherein the step 4) of the crack mapping based on the multi-scale positioning comprises the following specific steps:

assuming that the road is a plane, under the pinhole camera model, the homography matrix can be used to describe the basic geometric shape, and then two linear constraints on the homography matrix are known, and the historical crack label is mapped to the query image through the following relationship:

wherein n is the number of historical crack label pixels, [ u' _i y′ _i ] ^T Coordinates for mapping fracture data in historical data, [ u ] u _i y _i ] ^T The coordinates of the historical crack label pixels.

7. The method for detecting the crack growth of the road surface based on the historical crack data as claimed in claim 1, wherein the step 5) of analyzing the historical crack based on the RGM comprises the following specific steps:

5.1 Map crack gray value distribution analysis: taking the query crack labels mapped in the above steps as 'ideal' initial seed points, and starting from the seed points, expanding the region by searching for adjacent points with similar attributes to the seed points, wherein the attributes comprise color and intensity; calculating pixel values from the mapping corresponding relation on the query image through the mapping labels, and drawing an image intensity distribution histogram of the mapping crack data by utilizing a statistical principle;

5.2 The intensity distribution of the mapped crack pixels is represented by a gaussian model:

a Gaussian model is established to represent the distribution, the distribution characteristics are used as operators, and the corresponding calculation formula of the mean value and the standard deviation is as follows:

wherein N is the number of the pixels of the mapping crack label;

5.3 Growth crack analysis: when the intensity of a certain point in the image satisfies the following condition, the point is divided into cracks:

I(p _μ ,p _v )∈[0,ω+λσ]

in the formula, λ is a constant value determined according to the actual application.

8. The road surface crack growth detection method based on historical crack data of claim 1, wherein the positioning information in the step 1) is GPS positioning information.

9. The road surface crack growth detection method based on historical crack data of claim 3, wherein the number of the pixel points on the circular window around the pixel point P is 12 or 16.

Images