An automatic identification and repair method for asphalt pavement cracks on highways

By combining multi-scale image analysis and sub-pixel line structure detection with Riesz wavelet transform and Steger local detection, the accuracy and continuity issues of automatic identification and repair of asphalt pavement cracks under complex road conditions are solved, achieving high-precision crack identification and repair control.

CN123109037APending Publication Date: 2026-11-17CHINA RAILWAY 19TH BUREAU GRP 3RD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611381736.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-08
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing automatic identification and repair technologies for asphalt pavement cracks suffer from low detection accuracy, false positives and false negatives, and inaccurate repair control under complex pavement textures, crack width and direction variations, and localized intermittent conditions.

Method used

By employing multi-scale image analysis, sub-pixel line structure detection, joint matching of crack features, and continuous trajectory recovery methods, and through Riesz wavelet transform and Steger local detection, candidate crack regions are extracted and continuous repair control sequences are generated, achieving high crack recognition accuracy, good trajectory continuity, and accurate repair positioning.

Benefits of technology

It improves the accuracy of crack identification and repair precision under complex asphalt pavement conditions, reduces false detections and center position deviation, and provides highly automated repair control.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123109037A_ABST
    Figure CN123109037A_ABST
Patent Text Reader

Abstract

This invention discloses an automatic identification and repair method for asphalt pavement cracks on highways, comprising the following steps: acquiring asphalt pavement images and establishing a correspondence between image locations and road locations; performing multi-scale analysis on the images to determine candidate crack regions based on different scales, directions, and response intensities; extracting high-precision crack center points by combining crack width and extension direction, and filtering effective center points through consistency of direction, scale, and intensity; constructing crack center trajectories based on effective center points, and connecting and restoring fracture trajectories that meet continuity conditions; converting the position, direction, and width of the crack trajectory into repair location, travel direction, and effective range, generating a repair control sequence, and controlling the repair mechanism to complete the repair along the crack trajectory. This invention achieves continuous processing of crack identification, trajectory restoration, and repair control.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of road maintenance and pavement defect detection technology, and in particular to an automatic identification and repair method for cracks in asphalt pavement on highways. Background Technology

[0002] With the increase in the mileage of expressways and the improvement of the level of intelligent road maintenance, the automatic detection and repair of asphalt pavement cracks has become an important application scenario in road inspection and maintenance operations. Currently, expressway crack detection usually uses vehicle-mounted camera equipment to continuously collect pavement images, and identifies crack areas through image stitching, edge detection, line structure extraction or multi-scale image analysis methods. Then, based on the identified crack locations, the movement path and operation range of the repair equipment are planned to realize the detection and repair of asphalt pavement cracks.

[0003] Existing automatic crack identification and repair technologies for asphalt pavements still have significant limitations under conditions of complex pavement textures, crack width and direction variations, and localized crack discontinuities. On one hand, traditional crack detection methods typically use fixed detection scales, fixed judgment thresholds, or single image responses to determine the crack center. When the same crack undergoes width changes, directional shifts, or localized grayscale weakening along its extension direction, the detection scale struggles to consistently match the actual crack width. Furthermore, linear responses formed by pavement joints, texture edges, and changes in brightness are easily identified as cracks, leading to crack center location misalignment, false detections, or missed detections. On the other hand, existing methods often rely on direct connections between adjacent locations to identify crack trajectory breaks caused by local occlusion, wear, and weak responses. It is difficult to simultaneously determine the continuity of the fractured portion based on crack direction changes, width changes, and trajectory extension trends, easily resulting in incorrect connections or missing trajectories. Simultaneously, the lack of point-to-point correspondence between crack identification results and subsequent repair control in terms of location, direction of travel, and repair range makes it difficult to directly convert the identified crack trajectory into continuous repair control information, affecting the correspondence between the automatic repair process and the actual crack morphology.

[0004] Therefore, how to provide an automatic identification and repair method for cracks in asphalt pavement on highways is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an automatic identification and repair method for asphalt pavement cracks on highways. This invention fully utilizes multi-scale image analysis, sub-pixel line structure detection, joint matching of crack features, and continuous trajectory recovery methods. It details the implementation process from asphalt pavement image acquisition, crack candidate region identification, accurate crack center extraction, effective center point screening, fracture trajectory recovery to repair control sequence generation and automatic repair. It has the advantages of high crack identification accuracy, good trajectory continuity, accurate repair positioning, and high degree of automation.

[0006] An automatic identification and repair method for asphalt pavement cracks on highways according to an embodiment of the present invention includes the following steps: S1. Acquire asphalt pavement images and complete spatial registration, establish the mapping relationship between image pixel coordinates and road position coordinates, and generate the image to be detected; S2. Perform multi-scale Riesz wavelet transform on the image to be detected, determine the crack candidate region based on the scale response extrema and direction response extrema at each location, and generate a Riesz feature set containing the main response scale, main response direction and response intensity. S3. The Steger local detection scale is controlled by the main response scale in the Riesz feature set, the center search direction is constrained by the main response direction, and the line structure judgment threshold is adaptively determined by the response intensity to extract the candidate sub-pixel center points of the crack. S4. Associate the candidate sub-pixel center points of the crack with the corresponding Riesz features, and retain only the center points that simultaneously satisfy the main response direction and local line direction, the main response scale and the line response scale, and the dual response intensity matching to form an effective center point set. S5. Construct the crack center trajectory based on the effective center point set, and jointly compare the changes in the main response direction, the main response scale, and the center point extension trend on both sides of the trajectory breakpoint. When the three meet the continuity condition, restore the missing trajectory segment and generate a continuous sub-pixel crack trajectory. S6. Convert the sub-pixel coordinates, main response direction, and main response scale of each trajectory point in the continuous sub-pixel crack trajectory into the repair position, travel direction, and effective range, respectively, to form a point-by-point associated repair control sequence, and control the repair execution mechanism to perform crack repair along the sequence.

[0007] Optionally, S1 includes: S11. During the detection of the vehicle traveling along the highway, asphalt pavement images are continuously collected, and the longitudinal mileage, lateral position and vehicle posture of the road corresponding to the collection time of each frame image are simultaneously acquired. The collection time is used as the association index to form image-road position synchronization data. S12. Based on the camera calibration parameters, installation height and vehicle attitude, project each frame image onto the road plane, and calculate the predicted displacement of the current frame relative to the previous frame based on the longitudinal mileage change and lateral position change of the road at adjacent acquisition times, thereby limiting the matching search range of the overlapping area of ​​adjacent frames. S13. Extract the road surface texture points that are common to the previous and next frames within the matching search range, calculate the image registration residual based on the positional correspondence of the texture points, and limit the correction direction and correction range of the registration residual with the predicted displacement to obtain the correction transformation relationship of the current frame relative to the previous frame. S14. According to the image acquisition order, the cumulative correction transformation relationship is carried out, and each frame of image is transformed to a unified road plane coordinate system in turn. The cumulative result is corrected frame by frame using the road position of each frame. S15. Divide the unified road plane coordinate system into equally spaced road coordinate grids, and project each frame pixel to the corresponding grid position; for the same grid position covered by multiple frames, compare the distance from the corresponding pixel to the projection center of their respective image optical axis, retain the pixel with the smallest distance, and delete the duplicate covered pixels; S16. Connect the road coordinate grids that have completed pixel filtering according to the longitudinal mileage of the road, and establish a fixed index between the road coordinates of each grid position and the corresponding pixel coordinates to form a continuous image to be detected with a road position mapping relationship.

[0008] Optionally, S2 includes: S21. Read the pixel grayscale values ​​corresponding to each road coordinate grid in the image to be detected, expand the image boundary in a mirror manner, and convert the expanded image to the frequency domain; using the same radial bandpass mother wavelet as a reference, change the radial support range of the mother wavelet in order from high frequency to low frequency to construct a multi-scale wavelet frequency domain window with progressively increasing scale, so that each scale corresponds to a road surface linear structure width interval; S22. For each scale, combine the corresponding wavelet frequency domain window with the horizontal Riesz operator and the vertical Riesz operator respectively, perform two orthogonal Riesz component transformations of the same scale on the image to be detected, and then convert the transformation results to the spatial domain to obtain the first Riesz component map and the second Riesz component map with consistent positions at each scale, where the first Riesz component map is the first Riesz component map. The two Riesz components of the scale are obtained as follows: ; in, The image to be detected. For the first Wavelet frequency domain window of scale, For horizontal and vertical frequencies, The first The first and second Riesz components of the scale. S23. Taking each pixel position in the image to be detected as the center position, in the... Within the wavelet space support range corresponding to the scale, multiple candidate normal directions passing through the center position are set; for each candidate normal direction, the distance corresponding to the current scale wavelet half support width is moved from the center position along both sides of the normal direction to obtain the first sampling position and the second sampling position on both sides of the center position, and the Riesz components of the two sampling positions are projected onto the current candidate normal direction respectively. S24. Compare the signs of the normal projections at the first and second sampling positions. Only when the signs of the two normal projections are opposite will a two-sided antiphase response be formed for the candidate normal direction. The intensity of the two-sided antiphase response is determined according to the negative value of the product of the two normal projections. ; in, As the central location, As candidate normal directions, and These are the projections of the Riesz components on both sides of the center position onto the candidate normal direction; the bilateral antiphase response intensities corresponding to each candidate normal direction at the same center position are compared, and the candidate normal direction that obtains the local maximum value of the direction is determined as the crack normal, and the direction perpendicular to the crack normal is recorded as the crack response direction at the current scale; S25. Search for the nearest Riesz projection extrema points on both sides of the center position along the crack normal. Form an inverse extrema pair between the two projection extrema points and take the midpoint of the inverse extrema pair as the center reprojection point of the current scale. In order of scale from small to large, take the inverse extrema pair of the previous scale as the starting position of the search on both sides of the next scale. Search for the nearest inverse projection extrema point of the next scale outward along the corresponding crack normal. In this way, inverse extrema pairs and center reprojection points are generated scale by scale. S26. Map the center projection points generated at adjacent scales to the same road coordinate grid in the image to be detected. Connect only the scale responses where the center projection points continuously fall into the same pixel neighborhood, the two opposite extreme points are always located on both sides of the center projection point, and the crack response directions maintain the adjacent direction order, to form a cross-scale Riesz response chain with the same road surface linear structure as the object. Compare the bilateral opposite response intensities of each scale in each cross-scale Riesz response chain. Determine the scale with a response intensity greater than the response intensity of the adjacent scale in the chain as the main response scale, and determine the crack response direction corresponding to this scale as the main response direction. Determine the corresponding bilateral opposite response intensity as the response intensity. S27. Retain the center projection points that simultaneously possess cross-scale Riesz response chains, principal response scale, and directional local maxima. Connect spatially adjacent center projection points sequentially according to the principal response direction. When the connection direction of adjacent center projection points belongs to the same directional order as the corresponding principal response direction, classify the corresponding center projection points into the same crack candidate region. Record the principal response scale, principal response direction, and response intensity according to the pixel position of each center projection point to generate a Riesz feature set.

[0009] Optionally, S3 includes: S31. Read the main response scale, main response direction and response intensity corresponding to each center projection point in the Riesz feature set, and obtain the pixel distance between the inverse extreme point pairs under the main response scale. Use this pixel distance as the local line structure width corresponding to the current center projection point. S32. Based on the width intervals of the road surface linear structure corresponding to each Riesz scale, establish a corresponding Steger Gaussian detection scale interval for each width interval; map the positional proportion of the local linear structure width in the width interval corresponding to the current master response scale to the corresponding Steger Gaussian detection scale interval to determine the local detection scale of the current center projection point: ; in, Center return point The width of the local line structure, For the first The width range corresponding to the main response scale S33. Read adjacent spatial center projection points sequentially along the main response direction. When adjacent center projection points belong to the same cross-scale Riesz response chain, use the local detection scale of the previous center projection point as the scale search starting point of the next center projection point. Only correct the detection scale within the Steger Gaussian detection scale interval corresponding to the main response scale of the next center projection point to form a local detection scale sequence that is continuously arranged along the crack extension direction. S34. Perform Gaussian convolution on the neighborhood of each center re-projection point according to the local detection scale sequence and calculate the Hessian matrix. Take the normal of the main response direction as the center search reference direction and select the feature vector with the smaller angle with the normal from the two feature vectors of the Hessian matrix as the actual center search direction. S35. Calculate the first-order and second-order grayscale changes along the actual center search direction, and determine the sub-pixel offset distance from the center re-projection point to the center of the line structure based on the two; when the offset position is within the neighborhood of the current pixel and the corresponding line structure response reaches the judgment threshold determined by the Riesz response intensity, retain the offset position. S36. Convert the retained offset position into sub-pixel coordinates in the image to be detected, and associate and record the corresponding main response scale, main response direction, response intensity and local detection scale to form the candidate sub-pixel center point of the crack.

[0010] Optionally, establishing the corresponding Steger Gaussian detection scale interval includes: for center re-projection points with the same principal response scale in the Riesz feature set, reading pairs of inverse extreme points, and determining the local line structure width range corresponding to the principal response scale based on the distance between the pairs of inverse extreme points; for center re-projection points within this width range, convolving the neighborhood of the center re-projection point with Gaussian scales arranged in ascending order and calculating the Hessian matrix, extracting the normal eigenvalues ​​representing the line structure in the Hessian matrix along the crack normal, and recording the Gaussian scale corresponding to the change from increasing to decreasing of the absolute value of the normal eigenvalue as the response peak scale of the center re-projection point; and dividing each center re-projection point within the same principal response scale into... The response peak scales obtained from the center re-projection points are arranged in ascending order according to the width of the corresponding local line structure. The response peak scale corresponding to the smallest local line structure width is used as the lower limit of the Steger Gaussian detection scale for the main response scale, and the response peak scale corresponding to the largest local line structure width is used as the upper limit of the Steger Gaussian detection scale, thus forming the Steger Gaussian detection scale interval corresponding to the Riesz main response scale. Subsequently, for any center re-projection point, based on the position of the spacing between its antiphase extreme points within the corresponding local line structure width range, the local detection scale used for calculating the Hessian matrix of the center re-projection point is determined from the Steger Gaussian detection scale interval.

[0011] Optionally, S4 includes: S41. According to the center re-projection point index retained when the crack candidate sub-pixel center point is generated, establish a one-to-one correspondence between each crack candidate sub-pixel center point and the corresponding center re-projection point, read the main response direction, main response scale, response intensity, anti-phase extreme point pair and local detection scale of the center re-projection point, and at the same time read the Hessian matrix eigenvector and normal eigenvalue corresponding to the crack candidate sub-pixel center point. S42. Select the eigenvector of the Hessian matrix with a smaller angle to the normal of the main response direction as the Steger normal direction. Rotate the Steger normal direction by 90 degrees to obtain the Steger local line direction. According to the candidate direction interval used by the multi-scale Riesz wavelet transform, assign the main response direction and the Steger local line direction to the corresponding direction positions respectively. When they are assigned to the same direction position, retain the candidate sub-pixel center point of the crack; otherwise, delete the point. S43. For the center point of the candidate sub-pixel of the crack retained by the direction check, take its local detection scale as the center, and take the adjacent smaller and larger scales in the corresponding Steger Gaussian detection scale interval respectively. For the same sub-pixel position, recalculate the Hessian matrix at the three scales. Obtain the absolute value of the normal eigenvalue along the Steger normal direction in the three Hessian matrices respectively. When the absolute value of the normal eigenvalue corresponding to the local detection scale is greater than the corresponding values ​​of the smaller and larger scales at the same time, record the local detection scale as the Steger line response scale. S44. Based on the width-scale correspondence order used when establishing the Riesz line structure width interval and the Steger Gaussian detection scale interval, determine the Gaussian detection scale interval to which the Steger line response scale belongs, and obtain the Riesz line structure width interval corresponding to this interval; check the Riesz line structure width interval with the line structure width interval corresponding to the main response scale of the center projection point item by item. If the two correspond to the same scale number, determine that the main response scale and the Steger line response scale match; otherwise, delete the current center point. S45. Using the current candidate sub-pixel center point of the crack as the intermediate sampling position, move one pixel spacing to each side along the Steger normal direction to establish the first lateral sampling position and the second lateral sampling position; using the three sampling positions as centers, obtain the projection of the Riesz components on both sides along the corresponding Riesz crack normal and calculate the bilateral antiphase response, while reading the absolute value of the Hessian normal eigenvalue of the three sampling positions; when the bilateral antiphase response of the intermediate sampling position is greater than the bilateral antiphase response of the two lateral sampling positions, and the absolute value of the Hessian normal eigenvalue of the intermediate sampling position is greater than the corresponding value of the two lateral sampling positions, determine that the peak position of the Riesz response coincides with the peak position of the Steger line response, and complete the dual response intensity matching; S46. Retain the candidate sub-pixel center points of the crack that sequentially satisfy the consistency of direction order, scale number, and the coincidence of the double response peak positions. Write the corresponding main response direction, main response scale, Riesz response intensity, Steger line response scale, and Hessian normal feature value according to the sub-pixel coordinates of each point. Form an effective center point set according to the position of the sub-pixel coordinates in the image to be detected.

[0012] Optionally, S5 includes: S51. Establish spatial adjacency relationship according to the sub-pixel coordinates of the effective center point. Starting from any effective center point, search for the next effective center point in the forward neighborhood corresponding to its main response direction. Select the effective center points with the same direction order as the main response direction and the smallest lateral offset in turn and connect them until there are no effective center points in the forward neighborhood that meet the conditions, thus forming the crack center trajectory segment. S52. Mark the end center point of each crack center trajectory segment that has only a one-sided connection relationship as the trajectory breakpoint. Read each trajectory breakpoint and its center points that are continuously arranged inward along the trajectory. Determine the extension direction of the breakpoint according to the connection direction of the continuous center points, and obtain the main response direction, main response scale and anti-phase extreme point pair corresponding to the trajectory breakpoint. S53. For the trajectory breakpoints of different crack center trajectory segments, pair them up in pairs. Compare the connection direction of one trajectory breakpoint to another trajectory breakpoint with the extension direction of the two end breakpoints respectively. Only retain the breakpoint pairs where the extension direction of both ends points to the area between breakpoints and the connection direction is in the same direction sequence or adjacent direction sequence as the main response direction of both ends. S54. For the retained breakpoint pairs, compare the scale order of the main response scales at both ends; when both ends are at the same scale order or adjacent scale order, determine the search bandwidth between breakpoints by the larger value of the distance between the opposite extreme point pairs at both ends, and establish a strip-shaped search region extending in opposite directions along the breakpoint extension direction between the two trajectory breakpoints. S55. Proceed pixel by pixel from the two trajectory breakpoints into the strip search area. At each advancement position, search for the maximum position of the Riesz two-sided antiphase response along the normal of the current main response direction and take this position as the recovery center point. Read the Riesz response according to the main response scale corresponding to the previous recovery center point and its adjacent scales. Retain the recovery center points whose response maximum scale is in the same scale order or adjacent scale order as the previous recovery center point scale, and update the subsequent advancement direction with the connection direction of two consecutive recovery center points. S56. When the restored center point sequences generated at both ends enter the same pixel neighborhood, and the connection direction of the end ends is in the same direction sequence as the connection direction of the restored center point sequence on the other side, connect the two restored center point sequences and insert the corresponding crack center trajectory segment according to the sub-pixel coordinates; keep the breakpoints that do not meet the conditions of direction continuity, scale continuity and opposite extension disconnected, process all trajectory breakpoints in sequence, and generate continuous sub-pixel crack trajectories.

[0013] The beneficial effects of this invention are: This invention spatially registers asphalt pavement images and uses multi-scale Riesz wavelet transform to extract crack response information at different scales. Crack candidate regions are determined based on scale response extrema, directional response extrema, and response intensity. Simultaneously, the main response scale, main response direction, and response intensity are used to constrain the subsequent crack center detection process. This allows the detection scale to be adjusted according to changes in crack line width, and the center search direction to match the actual crack extension direction. This reduces the impact of fixed detection scales and fixed search directions on crack identification with varying widths and directions, improving the accuracy of crack candidate region identification and center localization under complex asphalt pavement conditions.

[0014] This invention associates Riesz features with Steger subpixel line structure detection results, and sequentially checks the consistency of the main response direction and local line direction, matches the main response scale and line response scale, and matches the peak positions of the dual responses for the candidate subpixel center points of cracks. Only center points that simultaneously meet the multiple feature consistency conditions are retained. This invention can use the mutual verification between different feature extraction mechanisms to eliminate non-crack linear responses formed by road texture, joints, gray-scale abrupt changes, and local noise, reducing false detections and center position shifts caused by single response judgments. At the same time, it obtains effective center points with subpixel positioning accuracy, providing a more accurate spatial position basis for crack trajectory construction and subsequent repair control. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an automatic identification and repair method for asphalt pavement cracks on highways proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figure 1 An automatic identification and repair method for cracks in asphalt pavement of highways includes the following steps: S1. Acquire asphalt pavement images and complete spatial registration, establish the mapping relationship between image pixel coordinates and road position coordinates, and generate the image to be detected; S2. Perform multi-scale Riesz wavelet transform on the image to be detected, determine the crack candidate region based on the scale response extrema and direction response extrema at each location, and generate a Riesz feature set containing the main response scale, main response direction and response intensity. S3. The Steger local detection scale is controlled by the main response scale in the Riesz feature set, the center search direction is constrained by the main response direction, and the line structure judgment threshold is adaptively determined by the response intensity to extract the candidate sub-pixel center points of the crack. S4. Associate the candidate sub-pixel center points of the crack with the corresponding Riesz features, and retain only the center points that simultaneously satisfy the main response direction and local line direction, the main response scale and the line response scale, and the dual response intensity matching to form an effective center point set. S5. Construct the crack center trajectory based on the effective center point set, and jointly compare the changes in the main response direction, the main response scale, and the center point extension trend on both sides of the trajectory breakpoint. When the three meet the continuity condition, restore the missing trajectory segment and generate a continuous sub-pixel crack trajectory. S6. Convert the sub-pixel coordinates, main response direction, and main response scale of each trajectory point in the continuous sub-pixel crack trajectory into the repair position, travel direction, and effective range, respectively, to form a point-by-point associated repair control sequence, and control the repair execution mechanism to perform crack repair along the sequence.

[0018] In this embodiment, S1 includes: S11. During the detection of the vehicle traveling along the highway, asphalt pavement images are continuously collected, and the longitudinal mileage, lateral position and vehicle posture of the road corresponding to the collection time of each frame image are simultaneously acquired. The collection time is used as the association index to form image-road position synchronization data. S12. Based on the camera calibration parameters, installation height and vehicle attitude, project each frame image onto the road plane, and calculate the predicted displacement of the current frame relative to the previous frame based on the longitudinal mileage change and lateral position change of the road at adjacent acquisition times, thereby limiting the matching search range of the overlapping area of ​​adjacent frames. S13. Extract the road surface texture points that are common to the previous and next frames within the matching search range, calculate the image registration residual based on the positional correspondence of the texture points, and limit the correction direction and correction range of the registration residual with the predicted displacement to obtain the correction transformation relationship of the current frame relative to the previous frame. S14. According to the image acquisition order, the cumulative correction transformation relationship is carried out, and each frame image is sequentially transformed to a unified road plane coordinate system. The cumulative result is corrected frame by frame using the road position of each frame, so that the projected pixels of the same road position in different frames remain in the same position. S15. Divide the unified road plane coordinate system into equally spaced road coordinate grids, and project each frame pixel to the corresponding grid position; for the same grid position covered by multiple frames, compare the distance from the corresponding pixel to the projection center of their respective image optical axis, retain the pixel with the smallest distance, and delete the duplicate covered pixels; S16. Connect the road coordinate grids that have completed pixel filtering according to the longitudinal mileage of the road, and establish a fixed index between the road coordinates of each grid position and the corresponding pixel coordinates to form a continuous image to be detected with a road position mapping relationship.

[0019] In this embodiment, S2 includes: S21. Read the pixel grayscale values ​​corresponding to each road coordinate grid in the image to be detected, expand the image boundary in a mirror manner, and convert the expanded image to the frequency domain; using the same radial bandpass mother wavelet as a reference, change the radial support range of the mother wavelet in order from high frequency to low frequency to construct a multi-scale wavelet frequency domain window with progressively increasing scale, so that each scale corresponds to a road surface linear structure width interval; S22. For each scale, combine the corresponding wavelet frequency domain window with the horizontal Riesz operator and the vertical Riesz operator respectively, perform two orthogonal Riesz component transformations of the same scale on the image to be detected, and then convert the transformation results to the spatial domain to obtain the first Riesz component map and the second Riesz component map with consistent positions at each scale, where the first Riesz component map is the first Riesz component map. The two Riesz components of the scale are obtained as follows: ; in, The image to be detected. For the first Wavelet frequency domain window of scale, For horizontal and vertical frequencies, The first The first and second Riesz components of the scale. S23. Taking each pixel position in the image to be detected as the center position, in the... Within the wavelet space support range corresponding to the scale, multiple candidate normal directions passing through the center position are set; for each candidate normal direction, the distance corresponding to the current scale wavelet half support width is moved from the center position along both sides of the normal direction to obtain the first sampling position and the second sampling position on both sides of the center position, and the Riesz components of the two sampling positions are projected onto the current candidate normal direction respectively. S24. Compare the signs of the normal projections at the first and second sampling positions. Only when the signs of the two normal projections are opposite will a two-sided antiphase response be formed for the candidate normal direction. The intensity of the two-sided antiphase response is determined according to the negative value of the product of the two normal projections. ; in, As the central location, As candidate normal directions, and These are the projections of the Riesz components on both sides of the center position onto the candidate normal direction; the bilateral antiphase response intensities corresponding to each candidate normal direction at the same center position are compared, and the candidate normal direction that obtains the local maximum value of the direction is determined as the crack normal, and the direction perpendicular to the crack normal is recorded as the crack response direction at the current scale; S25. Search for the nearest Riesz projection extrema points on both sides of the center position along the crack normal determined in S24. Form an antiphase extrema pair between the two projection extrema points and take the midpoint of the antiphase extrema pair as the center reprojection point of the current scale. In order of scale from small to large, take the antiphase extrema pair of the previous scale as the starting position of the search on both sides of the next scale. Search for the nearest antiphase projection extrema point of the next scale outward along the corresponding crack normal. In this way, antiphase extrema pairs and center reprojection points are generated scale by scale. S26. Map the center projection points generated at adjacent scales to the same road coordinate grid in the image to be detected. Connect only the scale responses where the center projection points continuously fall into the same pixel neighborhood, the two opposite extreme points are always located on both sides of the center projection point, and the crack response directions maintain the adjacent direction order, to form a cross-scale Riesz response chain with the same road surface linear structure as the object. Compare the bilateral opposite response intensities of each scale in each cross-scale Riesz response chain. Determine the scale with a response intensity greater than the response intensity of the adjacent scale in the chain as the main response scale, and determine the crack response direction corresponding to this scale as the main response direction. Determine the corresponding bilateral opposite response intensity as the response intensity. S27. Retain the center projection points that simultaneously possess cross-scale Riesz response chains, principal response scale, and directional local maxima. Connect spatially adjacent center projection points sequentially according to the principal response direction. When the connection direction of adjacent center projection points belongs to the same directional order as the corresponding principal response direction, classify the corresponding center projection points into the same crack candidate region. Record the principal response scale, principal response direction, and response intensity according to the pixel position of each center projection point to generate a Riesz feature set.

[0020] In this embodiment, S3 includes: S31. Read the main response scale, main response direction and response intensity corresponding to each center projection point in the Riesz feature set, and obtain the pixel distance between the inverse extreme point pairs under the main response scale. Use this pixel distance as the local line structure width corresponding to the current center projection point. S32. Establish the width intervals of the road linear structure corresponding to each Riesz scale, and establish a corresponding Steger Gaussian detection scale interval for each width interval; map the position ratio of the local linear structure width in the width interval corresponding to the current main response scale to the corresponding Steger Gaussian detection scale interval to determine the local detection scale of the current center projection point: ; in, Center return point The width of the local line structure, For the first The width range corresponding to the main response scale S33. Read adjacent spatial center projection points sequentially along the main response direction. When adjacent center projection points belong to the same cross-scale Riesz response chain, use the local detection scale of the previous center projection point as the scale search starting point of the next center projection point. Only correct the detection scale within the Steger Gaussian detection scale interval corresponding to the main response scale of the next center projection point to form a local detection scale sequence that is continuously arranged along the crack extension direction. S34. Perform Gaussian convolution on the neighborhood of each center re-projection point according to the local detection scale sequence and calculate the Hessian matrix. Take the normal of the main response direction as the center search reference direction and select the feature vector with the smaller angle with the normal from the two feature vectors of the Hessian matrix as the actual center search direction. S35. Calculate the first-order and second-order grayscale changes along the actual center search direction, and determine the sub-pixel offset distance from the center re-projection point to the center of the line structure based on the two; when the offset position is within the neighborhood of the current pixel and the corresponding line structure response reaches the judgment threshold determined by the Riesz response intensity, retain the offset position. S36. Convert the retained offset position into sub-pixel coordinates in the image to be detected, and associate and record the corresponding main response scale, main response direction, response intensity and local detection scale to form the candidate sub-pixel center point of the crack.

[0021] In this embodiment, establishing the corresponding Steger Gaussian detection scale interval includes: for center re-projection points with the same main response scale in the Riesz feature set, reading pairs of inverse extreme points, and determining the local line structure width range corresponding to the main response scale based on the distance between the pairs of inverse extreme points; for center re-projection points within this width range, convolving the neighborhood of the center re-projection point with Gaussian scales arranged in ascending order and calculating the Hessian matrix, extracting the normal eigenvalues ​​representing the line structure in the Hessian matrix along the crack normal, and recording the Gaussian scale corresponding to when the absolute value of the normal eigenvalue changes from increasing to decreasing as the response peak scale of the center re-projection point; and then, for each center re-projection point within the same main response scale, [the following steps are taken]. The response peak scales obtained from the center projection point are arranged in ascending order according to the width of the corresponding local line structure. The response peak scale corresponding to the smallest local line structure width is used as the lower limit of the Steger Gaussian detection scale for the main response scale, and the response peak scale corresponding to the largest local line structure width is used as the upper limit of the Steger Gaussian detection scale, thus forming the Steger Gaussian detection scale interval corresponding to the Riesz main response scale. Subsequently, for any center projection point, based on the position of the spacing between its antiphase extreme points within the corresponding local line structure width range, the local detection scale used for calculating the Hessian matrix of the center projection point is determined from the Steger Gaussian detection scale interval.

[0022] In this embodiment, S4 includes: S41. According to the center re-projection point index retained when the crack candidate sub-pixel center point is generated, establish a one-to-one correspondence between each crack candidate sub-pixel center point and the corresponding center re-projection point, read the main response direction, main response scale, response intensity, anti-phase extreme point pair and local detection scale of the center re-projection point, and at the same time read the Hessian matrix eigenvector and normal eigenvalue corresponding to the crack candidate sub-pixel center point. S42. Select the eigenvector of the Hessian matrix with a smaller angle to the normal of the main response direction as the Steger normal direction. Rotate the Steger normal direction by 90 degrees to obtain the Steger local line direction. According to the candidate direction interval used by the multi-scale Riesz wavelet transform, assign the main response direction and the Steger local line direction to the corresponding direction positions respectively. When they are assigned to the same direction position, retain the candidate sub-pixel center point of the crack; otherwise, delete the point. S43. For the center point of the candidate sub-pixel of the crack retained by the direction check, take its local detection scale as the center, and take the adjacent smaller and larger scales in the corresponding Steger Gaussian detection scale interval respectively. For the same sub-pixel position, recalculate the Hessian matrix at the three scales. Obtain the absolute value of the normal eigenvalue along the Steger normal direction in the three Hessian matrices respectively. When the absolute value of the normal eigenvalue corresponding to the local detection scale is greater than the corresponding values ​​of the smaller and larger scales at the same time, record the local detection scale as the Steger line response scale. S44. Based on the width-scale correspondence order used when establishing the Riesz line structure width interval and the Steger Gaussian detection scale interval, determine the Gaussian detection scale interval to which the Steger line response scale belongs, and obtain the Riesz line structure width interval corresponding to this interval; check the Riesz line structure width interval with the line structure width interval corresponding to the main response scale of the center projection point item by item. If the two correspond to the same scale number, determine that the main response scale and the Steger line response scale match; otherwise, delete the current center point. S45. Using the current candidate sub-pixel center point of the crack as the intermediate sampling position, move one pixel spacing to each side along the Steger normal direction to establish the first lateral sampling position and the second lateral sampling position; using the three sampling positions as centers, obtain the projection of the Riesz components on both sides along the corresponding Riesz crack normal and calculate the bilateral antiphase response, while reading the absolute value of the Hessian normal eigenvalue of the three sampling positions; when the bilateral antiphase response of the intermediate sampling position is greater than the bilateral antiphase response of the two lateral sampling positions, and the absolute value of the Hessian normal eigenvalue of the intermediate sampling position is greater than the corresponding value of the two lateral sampling positions, determine that the peak position of the Riesz response coincides with the peak position of the Steger line response, and complete the dual response intensity matching; S46. Retain the candidate sub-pixel center points of the crack that sequentially satisfy the consistency of direction order, scale number, and the coincidence of the double response peak positions. Write the corresponding main response direction, main response scale, Riesz response intensity, Steger line response scale, and Hessian normal feature value according to the sub-pixel coordinates of each point. Form an effective center point set according to the position of the sub-pixel coordinates in the image to be detected.

[0023] In this embodiment, S5 includes: S51. Establish spatial adjacency relationship according to the sub-pixel coordinates of the effective center point. Starting from any effective center point, search for the next effective center point in the forward neighborhood corresponding to its main response direction. Select the effective center points with the same direction order as the main response direction and the smallest lateral offset in turn and connect them until there are no effective center points in the forward neighborhood that meet the conditions, thus forming the crack center trajectory segment. S52. Mark the end center point of each crack center trajectory segment that has only a one-sided connection relationship as the trajectory breakpoint. Read each trajectory breakpoint and its center points that are continuously arranged inward along the trajectory. Determine the extension direction of the breakpoint according to the connection direction of the continuous center points, and obtain the main response direction, main response scale and anti-phase extreme point pair corresponding to the trajectory breakpoint. S53. For the trajectory breakpoints of different crack center trajectory segments, pair them up in pairs. Compare the connection direction of one trajectory breakpoint to another trajectory breakpoint with the extension direction of the two end breakpoints respectively. Only retain the breakpoint pairs where the extension direction of both ends points to the area between breakpoints and the connection direction is in the same direction sequence or adjacent direction sequence as the main response direction of both ends. S54. For the retained breakpoint pairs, compare the scale order of the main response scales at both ends; when both ends are at the same scale order or adjacent scale order, determine the search bandwidth between breakpoints by the larger value of the distance between the opposite extreme point pairs at both ends, and establish a strip-shaped search region extending in opposite directions along the breakpoint extension direction between the two trajectory breakpoints. S55. Proceed pixel by pixel from the two trajectory breakpoints into the strip search area. At each advancement position, search for the maximum position of the Riesz two-sided antiphase response along the normal of the current main response direction and take this position as the recovery center point. Read the Riesz response according to the main response scale corresponding to the previous recovery center point and its adjacent scales. Retain the recovery center points whose response maximum scale is in the same scale order or adjacent scale order as the previous recovery center point scale, and update the subsequent advancement direction with the connection direction of two consecutive recovery center points. S56. When the restored center point sequences generated at both ends enter the same pixel neighborhood, and the connection direction of the end ends is in the same direction sequence as the connection direction of the restored center point sequence on the other side, connect the two restored center point sequences and insert the corresponding crack center trajectory segment according to the sub-pixel coordinates; keep the breakpoints that do not meet the conditions of direction continuity, scale continuity and opposite extension disconnected, process all trajectory breakpoints in sequence, and generate continuous sub-pixel crack trajectories.

[0024] In this embodiment, S6 includes: S61. Read the sub-pixel coordinates of each trajectory point according to the trajectory order of the continuous sub-pixel crack trajectory, and obtain the main response direction and main response scale associated with each trajectory point; for the recovery center point, read the main response direction and main response scale determined during its recovery process, and form trajectory point data units arranged in trajectory order. S62. Based on the mapping relationship between image pixel coordinates and road position coordinates, the sub-pixel coordinates of each trajectory point are converted into the longitudinal and lateral positions of the road. Then, based on the installation pose of the repair execution mechanism relative to the image acquisition device, the road position is converted to the coordinate system of the repair execution mechanism to obtain the repair position corresponding to each trajectory point. S63. Transform the main response direction of each trajectory point to the coordinate system of the repair actuator, and determine the positive direction according to the sequential position of adjacent trajectory points, so that the transformed main response direction is consistent with the trajectory advancement direction, and use the corresponding direction angle as the travel direction of each repair position; S64. Based on the width range of the road surface linear structure corresponding to the main response scale of each trajectory point, read the distance between the opposite extreme point pairs of the trajectory point, and convert the distance into the actual crack width according to the road coordinate grid spacing of the image to be detected. Use the actual crack width to determine the range of action of the repair material perpendicular to the direction of travel. S65. Combine the repair position, direction of travel, and range of action corresponding to the same trajectory point into a single repair control unit, and arrange each repair control unit sequentially according to the trajectory order of the continuous sub-pixel crack trajectory; for adjacent control units, form a moving segment based on the adjacent repair position, and write the direction of travel and range of action corresponding to the next control unit at the end of the moving segment to form a point-to-point associated repair control sequence; S66. Control the repair execution mechanism to read the repair control sequence sequentially, so that the execution end moves to the corresponding repair position and advances along the corresponding direction of travel, while adjusting the repair width to the corresponding range of action; after completing the current control unit, read the next control unit, until all repair control units corresponding to the continuous sub-pixel crack trajectory are executed. Example

[0025] To verify the feasibility of the proposed automatic identification and repair method for asphalt pavement cracks in highways in actual road maintenance environments, this invention was applied to continuous inspection and crack repair operations on asphalt concrete pavements of operational highways. The selected road area simultaneously included longitudinal cracks, transverse cracks, diagonal cracks, bending cracks, and locally discontinuous cracks. Some cracks exhibited a change in width from thin to wide or vice versa along their extension direction. Other cracks were located near construction joints, wheel track edges, and coarse aggregate outlines. The road surface also exhibited localized shadows, reflections, surface contamination, and grayscale variations caused by wear. These road conditions highlight the limitations of existing fixed detection scales in adapting to crack width variations, the tendency of single-line structure detection to mistake non-crack textures for cracks, the ease with which weak-response areas can disrupt the trajectory, and the difficulty in directly converting detection results into control parameters for repair equipment.

[0026] In practical applications, an image acquisition device facing the road surface is installed on the underside of the inspection vehicle, ensuring that the image acquisition field of view covers the main road surface area of ​​the lane to be inspected. Asphalt pavement images are continuously acquired during vehicle movement, while the vehicle positioning unit and attitude detection unit simultaneously output the longitudinal road mileage, lateral vehicle position, and vehicle attitude corresponding to the acquisition time. Image, road position data, and vehicle attitude data are correlated using the same acquisition time, ensuring that each frame of the road surface image has corresponding vehicle road position and attitude information, thus forming continuous image-road position synchronized data.

[0027] Before image stitching, the calibration parameters of the image acquisition equipment, the equipment installation height, and the current vehicle posture are read, and each frame of road surface image is projected onto the road plane. For two adjacent frames, instead of performing unconstrained matching directly across the entire image, the longitudinal mileage change and lateral position change of the road between the two acquisition times are first read. From this, the predicted displacement of the subsequent frame relative to the previous frame on the road plane is calculated. This predicted displacement is used to limit the overlapping area of ​​adjacent images and the texture matching search range, thereby restricting image matching to an area that corresponds to the actual direction and distance of vehicle movement.

[0028] Within the defined overlapping area, common road surface texture points between consecutive frames are extracted. These texture points can originate from aggregate boundaries, fine surface textures, and locally stable grayscale regions. Image registration residuals are calculated based on the positional correspondence between these common texture points. Simultaneously, the vehicle prediction displacement is used to constrain the residual correction direction and magnitude, yielding the correction transformation relationship between the current frame and the previous frame. The correction transformation relationships are continuously accumulated frame by frame according to the vehicle image acquisition sequence, transforming the continuous road surface images frame by frame to a unified road plane coordinate system. The longitudinal mileage and lateral position of each frame are then used again to correct and accumulate the transformation results, reducing road position drift caused by the gradual accumulation of errors during continuous stitching.

[0029] The unified road plane is further divided into continuous and equally spaced road coordinate grids. Pixels in each frame of the road surface image are projected onto the corresponding road grid according to the aforementioned spatial transformation relationship. When the same road grid is repeatedly covered by multiple consecutive frames, the distances from multiple candidate pixels to the projection center of their respective image optical axes are calculated. Pixels with smaller distances from the projection center are retained as the valid pixels corresponding to that road grid, while other overlapping pixels are deleted. This pixel filtering method prioritizes the use of information from the central imaging region of the image for the same road location, reducing the interference caused by perspective changes at image edges and overlapping textures on crack detection. After filtering, the road coordinate grids are connected sequentially along the longitudinal direction of the road, and a fixed index between the road coordinates and image pixel coordinates is established for each road grid. This results in a detection image that possesses both continuous road surface image information and a mapping relationship to the actual road location.

[0030] After the image to be detected is formed, the image boundary is mirrored and expanded, and the expanded image is converted to the frequency domain. Using the same radial bandpass wavelet as a reference, multiple frequency domain windows with gradually increasing scales are formed by varying the radial support range from high to low frequency. Different frequency domain windows correspond to different widths of road surface linear structures, allowing fine cracks, medium-width cracks, and wider cracks to form corresponding image responses at a scale suitable for their own widths, rather than treating all cracks uniformly at a fixed detection scale.

[0031] For each scale, the corresponding wavelet frequency domain window is combined with the horizontal and vertical Riesz operators respectively, and the image to be detected is subjected to Riesz component transformation in two orthogonal directions. Then, it is transformed back to the spatial domain to obtain a first and second Riesz component map with consistent positions. For any center position in the image to be detected, multiple candidate normal directions passing through that center position are set within the wavelet spatial support range corresponding to the current scale. Along each candidate normal direction, the image is moved from the center position to both sides by a distance corresponding to the half-support width of the current scale, obtaining two symmetrical sampling positions. The Riesz components at the two sampling positions are then projected onto the current candidate normal direction.

[0032] When the projection signs of two sampling locations along the current candidate normal direction are opposite, it indicates that there are local gray-level changes in opposite directions on both sides of the center position, forming a two-sided antiphase response in the candidate normal direction. For multiple candidate normal directions at the same center position, the intensity of the two-sided antiphase response corresponding to each direction is compared. The candidate direction that obtains the directional local maximum response is determined as the crack normal at the current scale, and the direction perpendicular to the crack normal is taken as the crack response direction at the current scale. This process enables crack detection to not only rely on gray-level magnitude, but also to simultaneously utilize the antiphase change relationship on both sides of the crack and the directional local maximum relationship to determine the possible linear crack structure.

[0033] After the crack normal is determined, the nearest Riesz projection extrema points are searched along this normal to both sides of the center position. These two extrema points are paired as inverse extrema points, and the midpoint between them is taken as the center reprojection point for the current scale. When processing from a smaller scale to a larger scale, the two inverse extrema points obtained in the previous scale are used as the starting positions for searching on both sides of the next scale. New inverse extrema positions are searched from the inside out at the next scale, ensuring a continuous correspondence between the detection of the same pavement line structure at different scales.

[0034] The center projection points obtained at adjacent scales are then mapped to the same road coordinate grid. Only when the center projection points at different scales are continuously located in the same local pixel neighborhood, the two out-of-phase extreme points are always located on either side of the center, and the crack response directions maintain a continuous relationship between adjacent directions, are the responses at the corresponding scales connected to form a cross-scale Riesz response chain for the same pavement linear structure. For each cross-scale Riesz response chain, the bilateral out-of-phase response intensities at different scales are compared. The scale with a response intensity greater than that of the adjacent scale is taken as the dominant response scale, and the crack response direction corresponding to that scale is taken as the dominant response direction. Simultaneously, the response intensity at that scale is recorded.

[0035] The resulting principal response scale represents the main width level corresponding to the current crack location, the principal response direction represents the local extension direction of the crack, and the response intensity represents the degree of linear structural response of the crack at that location. Spatially adjacent center projection points are further connected according to the principal response direction. Only when the connection direction of adjacent center projection points is within the same directional range as the corresponding principal response direction are these center projection points classified into the same crack candidate region. The center projection points in each candidate region are associated and stored with their principal response scale, principal response direction, and response intensity, thus forming the Riesz feature set used for subsequent sub-pixel detection.

[0036] For each central projection point in the Riesz feature set, its principal response scale, principal response direction, and response intensity are read. The pixel distance between two out-of-phase extrema at that principal response scale is also read, and this distance is used as the local line structure width at the current location. Corresponding Steger Gaussian detection scale intervals are established for the line structure width intervals corresponding to different Riesz scales, thus establishing a width-scale correspondence between the two different crack detection mechanisms.

[0037] When establishing the Steger Gaussian detection scale interval, center projection points with the same Riesz principal response scale are grouped together, and the spacing between the antiphase extreme point pairs corresponding to each center projection point is read. Multiple spacings yield the local line structure width range corresponding to that Riesz principal response scale. For center projection points within this width range, the neighborhood of each center point is processed using a Gaussian scale that increases from small to large, and the Hessian matrix is ​​calculated. Normal eigenvalues ​​representing the line structure response are read along the Riesz crack normal. When the absolute value of the normal eigenvalue changes from increasing to decreasing with increasing Gaussian scale, the corresponding Gaussian scale is recorded as the peak response scale of the current line structure width.

[0038] Among multiple center projection points corresponding to the same Riesz principal response scale, the local line structure widths are arranged from smallest to largest, along with the corresponding response peak scales. The response peak scale corresponding to the smallest line structure width is taken as the lower limit of the Steger detection scale corresponding to that Riesz scale, and the response peak scale corresponding to the largest line structure width is taken as the upper limit of the detection scale, thus establishing a correspondence between the Riesz line structure width interval and the Steger Gaussian detection scale interval. When processing any center projection point subsequently, the actual local detection scale used for that point is determined based on the position of the distance between the inverse extreme points of that point within the corresponding width interval.

[0039] For adjacent center projection points along the same crack extension direction, when the preceding and following center points belong to the same cross-scale Riesz response chain, the already determined local detection scale of the preceding center point is used as the starting point for the scale search of the following center point. Adjustment is only allowed within the Steger detection scale interval corresponding to the main response scale of the following center point, thus forming a continuously varying local detection scale sequence along the crack extension direction. In this way, as the crack gradually changes from narrow to wide, the detection scale can be gradually increased; as the crack changes from wide to narrow, the detection scale can be continuously decreased, avoiding the inability of a fixed detection scale to simultaneously accommodate different crack widths.

[0040] Gaussian convolutions are performed on the neighborhood of each center re-projection point according to the determined local detection scale, and the Hessian matrix is ​​calculated. Using the normal direction of the Riesz principal response as the reference direction for center search, the eigenvector with the smaller angle to this reference normal is selected from the two eigenvectors of the Hessian matrix as the actual center search direction. The first-order and second-order changes in local grayscale are calculated along the actual search direction, and the sub-pixel offset distance from the current center re-projection point to the center of the line structure is determined based on these two changes. When the offset sub-pixel position is still within the current reasonable pixel neighborhood, and the corresponding line structure response reaches the judgment threshold determined by the Riesz response intensity, the sub-pixel position is retained, and the principal response scale, principal response direction, Riesz response intensity, and local detection scale are simultaneously associated with this position, thus forming the candidate sub-pixel center point for the crack.

[0041] Since construction joints, coarse aggregate edges, and local shadows in actual asphalt pavements may also form strong line structures, joint verification is carried out after the candidate sub-pixel center points are generated. According to the center re-projection point index saved when the candidate sub-pixel center points were generated, each candidate sub-pixel center point is matched one-to-one with the original Riesz center re-projection point. At the same time, the main response direction, main response scale, Riesz response intensity, inverse extreme point pairs, local detection scale, and the Hessian matrix eigenvector and normal eigenvalue of the current sub-pixel position are read.

[0042] From the eigenvectors of the Hessian matrix, the direction with the smallest angle to the normal of the Riesz principal response direction is selected as the Steger normal. This normal is then rotated to obtain the Steger local line direction. Following the candidate direction intervals set in the Riesz analysis phase, the Riesz principal response direction and the Steger local line direction are assigned to their corresponding direction sequences. If both fall into the same direction sequence, the candidate center point is retained; otherwise, it is deleted. This process first eliminates candidate structures where the local orientation judgments of Riesz analysis and Steger line structure analysis are significantly inconsistent.

[0043] For a center point that passes the orientation verification, taking the current local detection scale as the center, adjacent smaller and larger scales are selected within the corresponding Steger Gaussian detection scale interval, and the Hessian matrix at the same sub-pixel position is recalculated at each of the three scales. When the absolute value of the normal eigenvalue corresponding to the current local detection scale is greater than the corresponding values ​​at both the smaller and larger scales, the current detection scale is identified as the Steger line response scale. Then, according to the pre-established width-scale correspondence order, the Gaussian scale interval to which the Steger line response scale belongs is determined, and the Riesz line structure width interval corresponding to this interval is obtained. When this width interval has the same scale order as the Riesz width interval corresponding to the current center point's main response scale, the Riesz main response scale and the Steger line response scale are determined to be a match; otherwise, the current candidate center point is deleted.

[0044] After the orientation and scale have been verified, the candidate sub-pixel center point is used as the intermediate sampling position. The sampling positions are then shifted one pixel interval to each side along the Steger normal, forming two lateral sampling positions. The Riesz two-sided inverse response and the absolute value of the Hessian normal eigenvalue are read at both the intermediate and lateral positions. Only when the Riesz two-sided inverse response at the intermediate sub-pixel position is simultaneously higher than that at both lateral positions, and the absolute value of the Hessian normal eigenvalue at the intermediate position is also simultaneously higher than that at both lateral positions, is the Riesz response peak and the Steger line response peak spatially considered to coincide. Finally, only candidate sub-pixel center points with consistent orientation, matching scale, and consistent dual response peak positions are retained, forming the effective center point set.

[0045] After the set of effective center points is formed, spatial adjacency relationships are established based on the sub-pixel coordinates of each center point. Starting from any effective center point, the next effective center point is searched in the forward neighborhood corresponding to its main response direction. Center points with connection directions within the same range as the current main response direction and with small lateral offsets are preferentially selected for connection. This search continues forward in the same manner until no effective center point satisfying the conditions exists in the forward neighborhood, thus obtaining several crack center trajectory segments.

[0046] For trajectory ends with only a unilateral connection, they are identified as trajectory breakpoints. The breakpoints and multiple center points continuously arranged inwards along the trajectory are read. The outward extension trend of the trajectory breakpoint is calculated based on the connection direction between these center points. Simultaneously, the main response direction, main response scale, and inverse extreme point pair information corresponding to the breakpoint are read. Candidate pairings are performed for breakpoints in different trajectory segments. Only when the extension directions of both breakpoints point to the region between them, and the connection direction between the breakpoints is within the same or adjacent direction range as the main response directions at both ends, is the breakpoint pair retained as the object to be restored.

[0047] Further comparing the principal response scales at both ends of the breakpoint to be recovered, when the two breakpoints are at the same or adjacent scale positions, the distance between the two inverse extreme point pairs is read, and the search bandwidth between the breakpoints is determined by the larger distance. A strip-shaped search region extending towards each other along their respective extension directions is established between the two breakpoints. The search proceeds pixel-by-pixel from the two breakpoints into the strip-shaped region, searching for the maximum position of the Riesz bilateral inverse response along the normal of the current principal response direction at each advancing position, and this position is used as the recovery center point.

[0048] After obtaining a new recovery center point, the Riesz response of the previous recovery center point's main response scale and its adjacent scales is read. Only the positions where the response maxima scale is in the same or adjacent scale order as the previous recovery point are retained, and the next advancement direction is updated according to the connection direction of two consecutive recovery center points. When the recovery center point sequences generated from two breakpoints enter the same pixel neighborhood and the connection directions at both ends remain continuous, the two recovery sequences are connected and inserted into the original crack trajectory segment. Breakpoint pairs with discontinuous directions, significant scale jumps, or opposite extension trends at both ends are not connected, thereby avoiding the incorrect splicing of two independent cracks simply because the breakpoints are close together.

[0049] After the continuous subpixel crack trajectory is generated, the subpixel coordinates, main response direction, and main response scale of each trajectory point are read in the order of the trajectory. For the center point that is restored in the fracture area, the main response direction and main response scale determined during the restoration process are also read, so that the normal detection points and restored points form trajectory point data with the same format.

[0050] Using the initially established mapping relationship between image pixel coordinates and road position coordinates, the sub-pixel coordinates of each trajectory point are converted into the corresponding longitudinal and lateral positions of the road. Then, based on the fixed installation position of the repair actuator relative to the image acquisition device, the road coordinates are transformed into the coordinate system of the repair actuator to obtain the actual repair position corresponding to each trajectory point.

[0051] The main response direction of each trajectory point is synchronously transformed to the coordinate system of the repair actuator, and the actual forward direction is determined based on the sequential position of adjacent trajectory points, ensuring that the transformed main response direction is consistent with the crack trajectory advancement direction, thus obtaining the forward direction corresponding to each repair position. Simultaneously, the pixel distance between the road surface structure width range corresponding to the main response scale of the trajectory point and the inverse extreme point of that trajectory point is read. Then, the pixel distance is converted into the actual crack width based on the actual spacing of the road coordinate grid, and the actual crack width is used to determine the effective range of the repair material perpendicular to the forward direction.

[0052] Thus, each crack trajectory point forms a single repair control unit consisting of the repair location, direction of travel, and range of action. All repair control units are arranged sequentially according to the crack trajectory. The repair locations of adjacent control units form the actuator moving segments, and the direction of travel and range of action of the next control unit are written at the end of each moving segment, ultimately forming a continuous repair control sequence that is interconnected point by point.

[0053] After the repair actuator begins operation, it sequentially reads the repair control sequence, moving the actuator to the current repair position and advancing along the current direction of travel, adjusting the repair width to the current effective range. After completing the current control unit, it continues reading the next control unit. When the crack direction undergoes continuous bending, the actuator continuously changes its direction of advancement according to the directions of adjacent trajectory points; when the crack locally widens, the repair effective range is synchronously expanded based on the actual crack width obtained from the main response scale and the spacing between inverse extreme points; when the crack narrows, the effective range is correspondingly reduced. By continuously executing all repair control units, the actual repair movement trajectory maintains a correspondence with the aforementioned continuous sub-pixel crack center trajectory.

[0054] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for automatic identification and repair of cracks in asphalt pavement of highways, characterized in that, include: S1. Acquire asphalt pavement images and complete spatial registration, establish the mapping relationship between image pixel coordinates and road position coordinates, and generate the image to be detected; S2. Perform multi-scale Riesz wavelet transform on the image to be detected, determine the crack candidate region based on the scale response extrema and direction response extrema at each location, and generate a Riesz feature set containing the main response scale, main response direction and response intensity. S3. The Steger local detection scale is controlled by the main response scale in the Riesz feature set, the center search direction is constrained by the main response direction, and the line structure judgment threshold is adaptively determined by the response intensity to extract the candidate sub-pixel center points of the crack. S4. Associate the candidate sub-pixel center points of the crack with the corresponding Riesz features, and retain only the center points that simultaneously satisfy the main response direction and local line direction, the main response scale and the line response scale, and the dual response intensity matching to form an effective center point set. S5. Construct the crack center trajectory based on the effective center point set, and jointly compare the changes in the main response direction, the main response scale, and the center point extension trend on both sides of the trajectory breakpoint. When the three meet the continuity condition, restore the missing trajectory segment and generate a continuous sub-pixel crack trajectory. S6. Convert the sub-pixel coordinates, main response direction, and main response scale of each trajectory point in the continuous sub-pixel crack trajectory into the repair position, travel direction, and effective range, respectively, to form a point-by-point associated repair control sequence, and control the repair execution mechanism to perform crack repair along the sequence.

2. The method for automatic identification and repair of cracks in asphalt pavement of highways according to claim 1, characterized in that, S1 includes: S11. During the detection of the vehicle traveling along the highway, asphalt pavement images are continuously collected, and the longitudinal mileage, lateral position and vehicle posture of the road corresponding to the collection time of each frame image are simultaneously acquired. The collection time is used as the association index to form image-road position synchronization data. S12. Based on the camera calibration parameters, installation height and vehicle attitude, project each frame image onto the road plane, and calculate the predicted displacement of the current frame relative to the previous frame based on the longitudinal mileage change and lateral position change of the road at adjacent acquisition times, thereby limiting the matching search range of the overlapping area of ​​adjacent frames. S13. Extract the road surface texture points that are common to the previous and next frames within the matching search range, calculate the image registration residual based on the positional correspondence of the texture points, and limit the correction direction and correction range of the registration residual with the predicted displacement to obtain the correction transformation relationship of the current frame relative to the previous frame. S14. According to the image acquisition order, the cumulative correction transformation relationship is carried out, and each frame of the image is sequentially converted to a unified road plane coordinate system. The cumulative result is then corrected frame by frame using the road position of each frame. S15. Divide the unified road plane coordinate system into equally spaced road coordinate grids, and project each frame pixel to the corresponding grid position; for the same grid position covered by multiple frames, compare the distance from the corresponding pixel to the projection center of their respective image optical axis, retain the pixel with the smallest distance, and delete the duplicate covered pixels; S16. Connect the road coordinate grids that have completed pixel filtering according to the longitudinal mileage of the road, and establish a fixed index between the road coordinates of each grid position and the corresponding pixel coordinates to form a continuous image to be detected with a road position mapping relationship.

3. The method for automatic identification and repair of cracks in asphalt pavement of highways according to claim 1, characterized in that, S2 includes: S21. Read the pixel gray values ​​corresponding to each road coordinate grid in the image to be detected, expand the image boundary in a mirror manner, and convert the expanded image to the frequency domain. Using the same radial bandpass mother wavelet as a reference, change the radial support range of the mother wavelet in order from high frequency to low frequency to construct a multi-scale wavelet frequency domain window with progressively increasing scale, so that each scale corresponds to a road surface linear structure width range. S22. For each scale, the corresponding wavelet frequency domain window is combined with the horizontal Riesz operator and the vertical Riesz operator respectively, and two orthogonal Riesz component transformations of the same scale are performed on the image to be detected. Then, the transformation results are converted to the spatial domain to obtain the first Riesz component map and the second Riesz component map with consistent positions at each scale, where the first Riesz component map is the first Riesz component map. The two Riesz components of the scale are obtained as follows: ; in, The image to be detected. For the first Wavelet frequency domain window of scale, For horizontal and vertical frequencies, The first The first and second Riesz components of the scale. S23. Taking each pixel position in the image to be detected as the center position, in the... Within the wavelet space support range corresponding to the scale, multiple candidate normal directions passing through the center position are set; for each candidate normal direction, the distance corresponding to the current scale wavelet half support width is moved from the center position along both sides of the normal direction to obtain the first sampling position and the second sampling position on both sides of the center position, and the Riesz components of the two sampling positions are projected onto the current candidate normal direction respectively. S24. Compare the signs of the normal projections at the first and second sampling positions. Only when the signs of the two normal projections are opposite will a two-sided antiphase response be formed for the candidate normal direction. The intensity of the two-sided antiphase response is determined according to the negative value of the product of the two normal projections. ; in, As the central location, As candidate normal directions, and These are the projections of the Riesz components on both sides of the center position onto the candidate normal direction; the bilateral antiphase response intensities corresponding to each candidate normal direction at the same center position are compared, and the candidate normal direction that obtains the local maximum value of the direction is determined as the crack normal, and the direction perpendicular to the crack normal is recorded as the crack response direction at the current scale; S25. Search for the nearest Riesz projection extrema points on both sides of the center position along the crack normal. Form an inverse extrema pair between the two projection extrema points and take the midpoint of the inverse extrema pair as the center reprojection point of the current scale. In order of scale from small to large, take the inverse extrema pair of the previous scale as the starting position of the search on both sides of the next scale. Search for the nearest inverse projection extrema point of the next scale outward along the corresponding crack normal. In this way, inverse extrema pairs and center reprojection points are generated scale by scale. S26. Map the center projection points generated at adjacent scales to the same road coordinate grid in the image to be detected. Connect only the scale responses where the center projection points continuously fall into the same pixel neighborhood, the two opposite extreme points are always located on both sides of the center projection point, and the crack response directions maintain the adjacent direction order, to form a cross-scale Riesz response chain with the same road surface linear structure as the object. Compare the bilateral opposite response intensities of each scale in each cross-scale Riesz response chain. Determine the scale with a response intensity greater than the response intensity of the adjacent scale in the chain as the main response scale, and determine the crack response direction corresponding to this scale as the main response direction. Determine the corresponding bilateral opposite response intensity as the response intensity. S27. Retain the center projection points that simultaneously possess cross-scale Riesz response chains, principal response scale, and directional local maxima. Connect spatially adjacent center projection points sequentially according to the principal response direction. When the connection direction of adjacent center projection points belongs to the same directional order as the corresponding principal response direction, classify the corresponding center projection points into the same crack candidate region. Record the principal response scale, principal response direction, and response intensity according to the pixel position of each center projection point to generate a Riesz feature set.

4. The method for automatic identification and repair of cracks in asphalt pavement of highways according to claim 1, characterized in that, S3 includes: S31. Read the main response scale, main response direction and response intensity corresponding to each center projection point in the Riesz feature set, and obtain the pixel distance between the inverse extreme point pairs under the main response scale. Use this pixel distance as the local line structure width corresponding to the current center projection point. S32. Based on the width intervals of the road surface linear structure corresponding to each Riesz scale, establish a corresponding Steger Gaussian detection scale interval for each width interval; map the positional proportion of the local linear structure width in the width interval corresponding to the current master response scale to the corresponding Steger Gaussian detection scale interval to determine the local detection scale of the current center projection point: ; in, Center return point The width of the local line structure, For the first The width range corresponding to the main response scale This corresponds to the Steger Gaussian detection scale range; S33. Read adjacent spatial center projection points sequentially along the main response direction. When adjacent center projection points belong to the same cross-scale Riesz response chain, use the local detection scale of the previous center projection point as the scale search starting point of the next center projection point. Only correct the detection scale within the Steger Gaussian detection scale interval corresponding to the main response scale of the next center projection point to form a local detection scale sequence that is continuously arranged along the crack extension direction. S34. Perform Gaussian convolution on the neighborhood of each center re-projection point according to the local detection scale sequence and calculate the Hessian matrix. Take the normal of the main response direction as the center search reference direction and select the feature vector with the smaller angle with the normal from the two feature vectors of the Hessian matrix as the actual center search direction. S35. Calculate the first-order and second-order grayscale changes along the actual center search direction, and determine the sub-pixel offset distance from the center re-projection point to the center of the line structure based on the two; when the offset position is within the neighborhood of the current pixel and the corresponding line structure response reaches the judgment threshold determined by the Riesz response intensity, retain the offset position. S36. Convert the retained offset position into sub-pixel coordinates in the image to be detected, and associate and record the corresponding main response scale, main response direction, response intensity and local detection scale to form the candidate sub-pixel center point of the crack.

5. The method for automatic identification and repair of cracks in asphalt pavement of highways according to claim 4, characterized in that, Establishing the corresponding Steger Gaussian detection scale interval includes: for center re-projection points with the same principal response scale in the Riesz feature set, reading pairs of inverse extreme points, and determining the local line structure width range corresponding to the principal response scale based on the distance between the pairs of inverse extreme points; for center re-projection points within this width range, convolving the neighborhood of the center re-projection point with Gaussian scales arranged in ascending order and calculating the Hessian matrix, extracting the normal eigenvalues ​​representing the line structure in the Hessian matrix along the crack normal, and recording the Gaussian scale corresponding to the change from increasing to decreasing of the absolute value of the normal eigenvalue as the response peak scale of the center re-projection point; and then, for each center re-projection point under the same principal response scale,... The response peak scales obtained from the projection points are arranged in ascending order according to the width of the corresponding local line structure. The response peak scale corresponding to the smallest local line structure width is used as the lower limit of the Steger Gaussian detection scale for the main response scale, and the response peak scale corresponding to the largest local line structure width is used as the upper limit of the Steger Gaussian detection scale, thus forming the Steger Gaussian detection scale interval corresponding to the Riesz main response scale. Subsequently, for any center projection point, based on the position of the spacing between its antiphase extreme points within the corresponding local line structure width range, the local detection scale used for calculating the Hessian matrix of the center projection point is determined from the Steger Gaussian detection scale interval.

6. The method for automatic identification and repair of asphalt pavement cracks on highways according to claim 1, characterized in that, S4 includes: S41. According to the center re-projection point index retained when the crack candidate sub-pixel center point is generated, establish a one-to-one correspondence between each crack candidate sub-pixel center point and the corresponding center re-projection point, read the main response direction, main response scale, response intensity, anti-phase extreme point pair and local detection scale of the center re-projection point, and at the same time read the Hessian matrix eigenvector and normal eigenvalue corresponding to the crack candidate sub-pixel center point. S42. Select the eigenvector of the Hessian matrix with a smaller angle to the normal of the main response direction as the Steger normal direction. Rotate the Steger normal direction by 90 degrees to obtain the Steger local line direction. According to the candidate direction interval used by the multi-scale Riesz wavelet transform, assign the main response direction and the Steger local line direction to the corresponding direction positions respectively. When they are assigned to the same direction position, retain the candidate sub-pixel center point of the crack; otherwise, delete the point. S43. For the center point of the candidate sub-pixel of the crack retained by the direction check, take its local detection scale as the center, and take the adjacent smaller and larger scales in the corresponding Steger Gaussian detection scale interval respectively. For the same sub-pixel position, recalculate the Hessian matrix at the three scales. Obtain the absolute value of the normal eigenvalue along the Steger normal direction in the three Hessian matrices respectively. When the absolute value of the normal eigenvalue corresponding to the local detection scale is greater than the corresponding values ​​of the smaller and larger scales at the same time, record the local detection scale as the Steger line response scale. S44. Based on the width-scale correspondence order used when establishing the Riesz line structure width interval and the Steger Gaussian detection scale interval, determine the Gaussian detection scale interval to which the Steger line response scale belongs, and obtain the Riesz line structure width interval corresponding to this interval; check the Riesz line structure width interval with the line structure width interval corresponding to the main response scale of the center projection point item by item. If the two correspond to the same scale number, determine that the main response scale and the Steger line response scale match; otherwise, delete the current center point. S45. Using the current candidate sub-pixel center point of the crack as the intermediate sampling position, move one pixel spacing to each side along the Steger normal direction to establish the first lateral sampling position and the second lateral sampling position; using the three sampling positions as centers, obtain the projection of the Riesz components on both sides along the corresponding Riesz crack normal and calculate the bilateral antiphase response, while reading the absolute value of the Hessian normal eigenvalue of the three sampling positions; when the bilateral antiphase response of the intermediate sampling position is greater than the bilateral antiphase response of the two lateral sampling positions, and the absolute value of the Hessian normal eigenvalue of the intermediate sampling position is greater than the corresponding value of the two lateral sampling positions, determine that the peak position of the Riesz response coincides with the peak position of the Steger line response, and complete the dual response intensity matching; S46. Retain the candidate sub-pixel center points of the crack that sequentially satisfy the consistency of direction order, scale number, and the coincidence of the double response peak positions. Write the corresponding main response direction, main response scale, Riesz response intensity, Steger line response scale, and Hessian normal feature value according to the sub-pixel coordinates of each point. Form an effective center point set according to the position of the sub-pixel coordinates in the image to be detected.

7. The method for automatic identification and repair of asphalt pavement cracks on highways according to claim 1, characterized in that, S5 includes: S51. Establish spatial adjacency relationship according to the sub-pixel coordinates of the effective center point. Starting from any effective center point, search for the next effective center point in the forward neighborhood corresponding to its main response direction. Select the effective center points with the same direction order as the main response direction and the smallest lateral offset in turn and connect them until there are no effective center points in the forward neighborhood that meet the conditions, thus forming the crack center trajectory segment. S52. Mark the end center point of each crack center trajectory segment that has only a one-sided connection relationship as the trajectory breakpoint. Read each trajectory breakpoint and its center points that are continuously arranged inward along the trajectory. Determine the extension direction of the breakpoint according to the connection direction of the continuous center points, and obtain the main response direction, main response scale and anti-phase extreme point pair corresponding to the trajectory breakpoint. S53. For the trajectory breakpoints of different crack center trajectory segments, pair them up in pairs. Compare the connection direction of one trajectory breakpoint to another trajectory breakpoint with the extension direction of the two end breakpoints respectively. Only retain the breakpoint pairs where the extension direction of both ends points to the area between breakpoints and the connection direction is in the same direction sequence or adjacent direction sequence as the main response direction of both ends. S54. For the retained breakpoint pairs, compare the scale order of the main response scales at both ends; when both ends are at the same scale order or adjacent scale order, determine the search bandwidth between breakpoints by the larger value of the distance between the opposite extreme point pairs at both ends, and establish a strip-shaped search region extending in opposite directions along the breakpoint extension direction between the two trajectory breakpoints. S55. Proceed pixel by pixel from the two trajectory breakpoints into the strip search area. At each advancement position, search for the maximum position of the Riesz two-sided antiphase response along the normal of the current main response direction and take this position as the recovery center point. Read the Riesz response according to the main response scale corresponding to the previous recovery center point and its adjacent scales. Retain the recovery center points whose response maximum scale is in the same scale order or adjacent scale order as the previous recovery center point scale, and update the subsequent advancement direction with the connection direction of two consecutive recovery center points. S56. When the restored center point sequences generated at both ends enter the same pixel neighborhood, and the connection direction of the end ends is in the same direction sequence as the connection direction of the restored center point sequence on the other side, connect the two restored center point sequences and insert the corresponding crack center trajectory segment according to the sub-pixel coordinates; keep the breakpoints that do not meet the conditions of direction continuity, scale continuity and opposite extension disconnected, process all trajectory breakpoints in sequence, and generate continuous sub-pixel crack trajectories.