Intelligent road marking quality evaluation system based on image analysis

Abstract

The invention provides a road marking quality intelligent evaluation system based on image analysis, and relates to the technical field of road facility monitoring, and the system comprises a detection unit, a marking quality evaluation unit, a marking wear prediction unit, a GPS positioning unit, a vehicle driving information unit, a control unit and a remote central control unit. The marking quality evaluation unit is constructed based on a differential geometry theory and comprises a curvature flow edge detection module, a marking geometric characteristic manifold representation module and a multi-scale differential invariant evaluation module, the system regards a marking as a two-dimensional manifold, and the marking quality is evaluated by calculating differential geometric quantities such as a Gaussian curvature, an average curvature and a shape index. And constructing geodesic distance measurement on the feature manifold, and evaluating the completeness, visibility and reflective performance of the marked line. And the marking wear prediction unit predicts the service life of the marking based on the traffic flow information and the environment characteristic mapping relation model, and generates a maintenance suggestion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of road facility monitoring, specifically to a highway marking quality intelligent evaluation system based on image analysis, which is used for automatic evaluation and prediction of quality parameters such as the integrity, visibility and retroreflective performance of highway markings. BACKGROUND

[0002] Highway markings are an important infrastructure for road traffic safety, and their quality directly affects the safe driving and normal operation of intelligent driving systems. Traditional highway marking quality evaluation mainly relies on manual inspection, which has the problems of low efficiency, strong subjectivity and inability to quantitatively evaluate. Existing automated evaluation methods mostly use simple image analysis based on color and retroreflectivity, which is difficult to adapt to complex road surface environments and has insufficient evaluation accuracy and stability. Especially in complex situations such as changes in light, road pollution and local wear of markings, traditional evaluation techniques often have difficulty in accurately identifying and evaluating marking quality.

[0003] With the development of differential geometry and computer vision technology, it is possible to use marking geometric characteristics for quality evaluation. However, there is currently a lack of a technical solution that deeply integrates differential geometry theory with marking quality evaluation, which can reveal the geometric nature of marking quality and achieve high-precision and high-adaptability marking quality evaluation. SUMMARY

[0004] The purpose of the present application is to provide a highway marking quality intelligent evaluation system based on image analysis, which realizes accurate evaluation and prediction of highway marking quality by integrating differential geometry theory and image analysis technology.

[0005] The present application proposes a highway marking quality intelligent evaluation system based on image analysis, which includes:

[0006] A detection unit for acquiring highway marking images and performing image preprocessing;

[0007] A marking quality evaluation unit in communication connection with the detection unit for receiving the preprocessed marking images sent by the detection unit, the marking quality evaluation unit including:

[0008] A curvature flow edge detection module for regarding the preprocessed marking images as two-dimensional manifolds, constructing a curvature flow equation, extracting marking edge features through multi-scale representation, and outputting an edge feature point set;

[0009] A marking geometric characteristic manifold representation module in communication connection with the curvature flow edge detection module for receiving the edge feature point set, constructing a marking manifold parameterized representation, calculating the first and second fundamental forms of the marking surface, extracting Gaussian curvature and mean curvature features, and generating a geometric characteristic data set;

[0010] a multi-scale differential invariant evaluation module, in communication connection with the reticle geometric property manifold representation module, configured to receive the geometric property dataset, compute multi-order differential invariants, fuse multi-scale features, construct geodesic distance metric on the feature manifold, evaluate the integrity, visibility and retroreflective performance of the reticle, and output quality evaluation results;

[0011] a reticle wear-out prediction unit, in communication connection with the reticle quality evaluation unit, configured to receive the quality evaluation results, predict the service life of the reticle, and generate maintenance recommendations.

[0012] Preferably, the curvature flow edge detection module comprises:

[0013] an image standardization sub-module configured to convert the input image to a standard resolution and brightness range, and establish a mapping relationship between pixels and physical dimensions;

[0014] a curvature calculation sub-module configured to construct an image surface, and calculate the principal curvature and mean curvature of each point on the surface;

[0015] an evolution control sub-module configured to control the evolution speed of the surface according to the curvature value, so that the high curvature area evolves slowly and the low curvature area evolves quickly and smoothly;

[0016] a multi-scale edge representation sub-module configured to generate image representations of different scales, and retain edges that stably exist at multiple scales by comparing the position offsets of edges at different scales;

[0017] an arc length optimization sub-module configured to define an arc length functional containing a curvature term, and optimize the edge curve by minimizing the functional while maintaining the topological structure of the edge curve.

[0018] Preferably, the reticle geometric property manifold representation module comprises:

[0019] a curve parameterization sub-module configured to parameterize the edge curve as an arc length parameter representation, establish a local coordinate system, and divide the reticle area into a regular grid;

[0020] a differential geometric quantity calculation sub-module configured to calculate the metric tensor, the first fundamental form and the second fundamental form on the discrete grid, and describe the intrinsic and extrinsic geometric properties of the reticle surface;

[0021] a geometric invariant extraction sub-module configured to calculate geometric invariants independent of the observation viewing angle, including the Gaussian curvature, the mean curvature and the shape index;

[0022] a geometric feature mapping sub-module configured to establish a multi-dimensional feature space, generate a geometric feature vector for each area of the reticle, and identify areas with geometric property anomalies;

[0023] A quality parameter space construction submodule is configured to define basic quality dimensions such as integrity, flatness, adhesion, and uniformity, and map the geometric characteristics to the quality parameter space.

[0024] Preferably, the multiscale differential invariant evaluation module comprises:

[0025] A differential invariant calculation submodule is configured to calculate first-order invariants, second-order invariants, and high-order invariants.

[0026] A multiscale feature fusion submodule is configured to construct multiple scale levels covering from details to the whole, decompose the features at each scale into basic components, and reconstruct a feature vector according to the importance.

[0027] A feature manifold construction submodule is configured to define a similarity measure between features, construct a Riemannian metric of the feature manifold, and establish a geodesic line connection between feature points.

[0028] A quality score generation submodule is configured to define a metric function of quality difference, establish a mapping from the geodesic distance to the quality score, generate an integrity score, a visibility score, and a reflectivity performance score.

[0029] An abnormal pattern recognition submodule is configured to extract a main variation pattern based on principal component analysis on the manifold, and identify and classify abnormal regions.

[0030] Preferably, the system further comprises:

[0031] A GPS positioning unit is configured to obtain a real-time position and a driving direction of the vehicle.

[0032] A vehicle driving information unit is configured to obtain a real-time speed, an acceleration, and a steering angle of the vehicle.

[0033] A control unit is in communication connection with the GPS positioning unit and the vehicle driving information unit, and is configured to determine whether to start the marking line detection according to the vehicle position and driving information, and send a detection instruction to the detection unit.

[0034] The detection unit comprises a camera acquisition unit, an illumination detection unit, a color compensation unit, and an image preprocessing unit. The camera acquisition unit acquires a marking line image according to the detection instruction. The illumination detection unit obtains a real-time illumination value. The color compensation unit adjusts the color and intensity of the light source according to the real-time illumination value. The image preprocessing unit performs noise filtering and feature extraction on the image after compensation of the illumination.

[0035] Preferably, the marking line quality evaluation unit further comprises:

[0036] A region score generation module is configured to generate a quality score map of different regions of the marking line, and intuitively display the problem regions.

[0037] a whole grade evaluation module for calculating a whole quality grade according to the area score, and dividing the target line quality into five grades of excellent, good, medium, poor and bad;

[0038] an evaluation report generation module for summarizing the quality score and the abnormal information, and generating a detailed evaluation report containing charts and position markers;

[0039] a historical trend analysis module for comparing the evaluation results of the same position at different times, and analyzing the change trend of the target line quality over time.

[0040] Preferably, the system further comprises:

[0041] a remote central control unit in communication connection with the detection unit and the target line quality evaluation unit, for issuing an illumination compensation instruction to the detection unit through the Internet of Things, receiving the preprocessed feature image, and transmitting it to the target line quality evaluation unit;

[0042] Preferably, the remote central control unit comprises a mobile network communication module and a network control module, the mobile network communication module connects the remote data storage and analysis module through a mobile 4G or 5G network, and the network control module is used to control the illumination compensation scheme of the detection unit, adjust the image color, and optimize the image signal transmission.

[0043] Preferably, the target line wear prediction unit comprises:

[0044] a feature mapping sub-unit for constructing a target line period and environmental feature mapping relationship model based on target line positioning parameters, detected vehicle driving data and environmental data;

[0045] a traffic flow prediction sub-unit for predicting traffic flow by establishing a traffic flow time change model;

[0046] a target line service life prediction sub-unit for predicting the service life of the target line based on the traffic flow information, the target line period and the environmental feature mapping relationship model, and recommending a maintenance scheme according to the service life;

[0047] Preferably, the environmental data comprises illumination conditions and weather conditions.

[0048] Preferably, the system further comprises:

[0049] a parameter self-adaptive adjustment module for automatically fine-tuning the algorithm parameters according to the consistency of the evaluation results and the artificial verification;

[0050] a model updating module for periodically collecting verification data and optimizing the evaluation model;

[0051] An abnormal sample processing module is configured to record the evaluation abnormality of the target line sample, and add the target line sample to the training set after manual confirmation, thereby improving the system recognition accuracy.

[0052] The parameter adaptive adjustment module maintains different parameter configuration schemes for different road surface types, different weather conditions and different target line types.

[0053] Preferably, the system further comprises:

[0054] A cloud management module is configured to receive and store the target line image and the evaluation result.

[0055] A data analysis platform is in communication connection with the cloud management module, and is configured to provide historical data analysis and trend prediction.

[0056] A management interface is in communication connection with the data analysis platform, and is configured to support remote configuration and monitoring.

[0057] A third-party system interface is configured to integrate with external systems such as intelligent transportation systems and road maintenance management systems through a standard API, to realize data sharing and collaborative decision-making.

[0058] The present application has the following beneficial effects:

[0059] 1. By introducing the curvature flow theory to construct the edge detection technology, the accuracy of target line edge detection in complex road surface environment is significantly improved, and the problem that the traditional edge detection is easily disturbed by noise, light change and road cracks is solved;

[0060] 2. The target line is regarded as a two-dimensional manifold for geometric characteristic representation, which reveals the target line quality state from the geometric nature, breaks through the limitation of traditional pixel value analysis, and can capture small but key quality changes;

[0061] 3. A multi-scale differential invariant is used to construct an evaluation model, which realizes comprehensive quantization and accurate rating of the target line quality, and can simultaneously focus on the macroscopic integrity and microscopic details of the target line;

[0062] 4. A target line life prediction technology based on geometric characteristics is established, which can accurately predict the target line degradation trend and provide a scientific basis for road target line maintenance;

[0063] 5. The system has strong environmental adaptability and can maintain stable evaluation performance under different light conditions, different road surface types and different weather conditions;

[0064] 6. Modular design is adopted, the data flow between components is clear, and the system is easy to integrate and functionally extended, and has good industrial practicability. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1The overall architecture diagram of the system of the present application;

[0066] Figure 2 The functional module diagram of the marking line quality evaluation unit of the present application;

[0067] Figure 3 The workflow diagram of the curvature flow edge detection module of the present application;

[0068] Figure 4 The workflow diagram of the marking line geometric property manifold characterization module of the present application;

[0069] Figure 5 The workflow diagram of the multi-scale differential invariant evaluation module of the present application;

[0070] Figure 6 The workflow diagram of the marking line wear prediction unit of the present application. DETAILED DESCRIPTION

[0071] Reference will now be made to Figure 1 - Figure 6 The present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0072] As shown in Figure 1 , the present application provides a highway marking line quality intelligent evaluation system based on image analysis, which comprises a detection unit 1, a marking line quality evaluation unit 2, a marking line wear prediction unit 3, a GPS positioning unit 4, a vehicle driving information unit 5, a control unit 6 and a remote central control unit 7.

[0073] In an embodiment of the present application, the control unit 6 is connected to the GPS positioning unit 4 and the vehicle driving information unit 5, and determines whether the marking line detection needs to be started according to the acquired vehicle position and driving information, and sends a detection instruction to the detection unit 1. The detection unit 1 acquires the highway marking line image and performs image preprocessing, and transmits the preprocessed marking line image to the marking line quality evaluation unit 2. The marking line quality evaluation unit 2 performs quality evaluation on the marking line image, and outputs the quality evaluation result to the marking line wear prediction unit 3. The marking line wear prediction unit 3 predicts the service life of the marking line according to the quality evaluation result, and generates a maintenance suggestion.

[0074] Preferably, the remote central control unit 7 is in communication connection with the detection unit 1 and the marking line quality evaluation unit 2, and issues an illumination compensation instruction to the detection unit 1 through the Internet of Things, receives the preprocessed feature image, and transmits it to the marking line quality evaluation unit 2. This architecture realizes centralized management and sharing of data, and improves the flexibility and scalability of the system.

[0075] As shown in Figure 1 , the detection unit 1 comprises a camera acquisition unit 11, an illumination detection unit 12, a color compensation unit 13 and an image preprocessing unit 14.

[0076] In one embodiment of the present application, the camera acquisition unit 11 adopts an industrial-grade camera with a resolution of 1920x1080, a frame rate of 30 fps, and a field of view angle of 75°, which can cover the full width of the lane. The camera is installed on the front bumper of the vehicle, about 60 cm above the ground, and the imaging angle is adjustable.

[0077] The light detection unit 12 includes a photosensitive sensor with a dynamic range of up to 120 dB, which can obtain the ambient light value in real time. In actual application, the light value usually varies within the range of 0-100000 lux. The system sets processing strategies under different light conditions, for example, when the light value is below 100 lux (night or tunnel environment), the system will increase the exposure time and turn on the auxiliary light source; when the light value is between 100-10000 lux (overcast or evening), the system adopts standard exposure parameters; when the light value is higher than 10000 lux (sunny noon), the system will reduce the exposure time and start strong light suppression.

[0078] The color compensation unit 13 adjusts the color and intensity of the light source according to the real-time light value obtained by the light detection unit 12, to realize the light compensation of the alignment image. In one specific embodiment, the color compensation unit 13 adopts the following light compensation formula: .

[0079] Wherein: is the compensated light intensity, with the unit of lux; is the initial intensity of the light source, with the unit of lux; k is the compensation proportion coefficient, dimensionless, usually taking the value range of 0.8-1.2; is the color RGB value under standard light conditions, which is a three-dimensional vector; is the color RGB value of the current ambient light, which is a three-dimensional vector. Each component of the RGB value ranges from 0 to 255, representing the intensity of the red, green, and blue color channels respectively.

[0080] The color RGB value of the light source is calculated by the following formula: .

[0081] Wherein: is the output color RGB value of the light source, which is a three-dimensional vector; is the initial value of the light source color, which is a three-dimensional vector; is the adjustment function, dimensionless; W is the weather information (sunny, overcast, rainy, snowy, etc.); T is the time information (morning, noon, afternoon, dusk, night, etc.). The adjustment function adjusts the basic color value according to the weather and time conditions, so that the compensated light is closer to the natural light condition.

[0082] The image preprocessing unit 14 performs noise filtering and feature extraction on the image compensated for illumination. Specifically, first, Gaussian filtering is used for noise reduction processing, and the filter kernel size is usually 5x5 or 7x7, and the sigma value is automatically adjusted between 1.0-2.5 according to the image noise level; then color space conversion is performed to convert the RGB image to HSV or Lab color space, which facilitates subsequent scale bar extraction; and finally histogram equalization is performed to enhance the image contrast. The preprocessed image is transmitted to the scale bar quality assessment unit 2 for further analysis.

[0083] As shown in Figure 2 , the scale bar quality assessment unit 2 is the core innovative module of the present application, including a curvature flow edge detection module 21, a scale bar geometric feature manifold representation module 22, a multi-scale differential invariant evaluation module 23, a region score generation module 24, a overall grade evaluation module 25, an evaluation report generation module 26 and a historical trend analysis module 27.

[0084] As shown in Figure 3 , the curvature flow edge detection module 21 includes an image standardization sub-module 211, a curvature calculation sub-module 212, an evolution control sub-module 213, a multi-scale edge representation sub-module 214 and an arc length optimization sub-module 215.

[0085] The image standardization sub-module 211 converts the input image to a standard resolution and brightness range, establishing a mapping relationship between pixels and physical dimensions. In the preferred embodiment of the present application, the standard resolution is set to 1280x720 pixels, the brightness range is normalized to 0-1, and the conversion relationship between pixels and physical dimensions is calculated according to the camera installation height and angle, with a typical value of 1 pixel corresponding to 0.5-2 millimeters.

[0086] The curvature calculation sub-module 212 treats the grayscale image as a height field, constructs a continuous image surface, and calculates the principal curvature and average curvature of each point on the surface. In actual implementation, first, the gradient field of the image is calculated: .

[0087] wherein: is the gradient vector at point , which is a two-dimensional vector; is the grayscale value of the image at point , dimensionless, ranging from 0 to 1; and are the partial derivatives of the grayscale value with respect to and , respectively, representing the rate of change of the grayscale value in the and directions.

[0088] Then the Hessian matrix of the image is calculated: .

[0089] wherein: is the Hessian matrix of the image , which is a 2x2 matrix; , , and are the second-order partial derivatives of the gray value , representing the second-order change rate of the gray value in each direction. Since the mixed partial derivatives are equal, i.e. , the Hessian matrix is a symmetric matrix.

[0090] Based on the gradient field and the Hessian matrix, the principal curvatures and of the surface are calculated:

[0091] .

[0092] wherein: and are the principal curvatures of the surface, with the unit of pixel⁻¹; |∇I| is the module length of the gradient vector, and the calculation formula is |∇I|=√((∂I / ∂x)²+(∂I / ∂y)²); eigenvalues represent the eigenvalues of the calculation matrix, and the results are two real numbers, which are the maximum principal curvature κ1 and the minimum principal curvature κ2. The principal curvature describes the bending degree of the surface in different directions, and its sign indicates the bending direction, with a positive value indicating a convex and a negative value indicating a concave.

[0093] The mean curvature H and the Gaussian curvature K are respectively:

[0094] .

[0095] .

[0096] wherein: H is the mean curvature, with the unit of pixel⁻¹, representing the average value of the bending degree of the surface; K is the Gaussian curvature, with the unit of pixel⁻², representing the intrinsic bending degree of the surface, which is closely related to the topological properties of the surface. The mean curvature and the Gaussian curvature are important geometric properties of the surface, which are related to the shape and deformation characteristics of the surface.

[0097] The evolution control submodule 213 controls the evolution speed of the surface according to the curvature value, so that the high curvature area (potential edge) evolves slowly and the low curvature area is quickly smoothed. Specifically, the following curvature flow equation is used to control the surface evolution:

[0098] .

[0099] wherein: represents the change rate of the image with respect to time t; is the edge stopping function, dimensionless; div denotes the divergence operator, which calculates the divergence of a vector field; is the normalized gradient vector, which indicates the direction of the gradient. This equation describes the evolution of the image under the effect of curvature flow, where high curvature regions (e.g., edges) evolve slowly, preserving edge information, while low curvature regions quickly smooth, suppressing noise.

[0100] Edge stopping function is commonly defined as:

[0101] .

[0102] where: is a control parameter, dimensionless, typically taking a value of 1.5-2 times the average of the image gradient magnitude. This function approaches 0 in regions with large gradients (e.g., edges), slowing down the evolution speed; it approaches 1 in regions with small gradients (e.g., flat regions), speeding up the evolution speed.

[0103] In practical applications, the evolution step size is set to 0.05-0.2, automatically adjusted according to the image complexity. The number of iterations is usually 20-50, with the convergence condition being that the edge changes less than 0.5 pixels. For straight-line type scales, the curvature threshold is set to 0.01-0.05 pixels-1; for curved type scales, the curvature threshold is set to 0.05-0.15 pixels-1.

[0104] The multi-scale edge representation submodule 214 generates 5-8 different scale image representations, with a scale factor set to 1.5-2.0. Small scales (1-2 pixels) are used to capture details, and large scales (10-15 pixels) are used to suppress noise. By comparing the position shifts of edges at different scales, edges that are stable at multiple scales are retained.

[0105] The arc length optimization submodule 215 defines an arc length functional that includes a curvature term, and optimizes the edge curve by minimizing the functional while maintaining the topological structure of the edge curve. The arc length functional is defined as:

[0106] .

[0107] where: is the arc length functional, representing the energy of the curve ; denotes the edge curve; is the curvature of the curve, with units of pixels-1; is a weight coefficient, dimensionless, typically taking a value of 0.5-2.0; is the curve element, representing a small length along the curve; denotes the derivative of the curve line integral, which is the sum of the function values at all points on the curve. The functional contains two parts: the first part 1 corresponds to the length of the curve, and the second part corresponds to the smoothness of the curve. By minimizing the functional, a curve that is both short and smooth can be obtained, which is suitable for representing the edge of the reticle.

[0108] The arc length functional is minimized by the gradient descent method, and the edge point is moved no more than 0.5 pixels at each iteration, while the topological preservation constraint is applied to prevent the curve from intersecting or breaking.

[0109] The optimized edge curve is output to the reticle geometric feature manifold representation module 22 as an edge feature point set.

[0110] As shown in Figure 4 , the reticle geometric feature manifold representation module 22 includes a curve parameterization submodule 221, a differential geometric quantity calculation submodule 222, a geometric invariant extraction submodule 223, a geometric feature mapping submodule 224, and a quality parameter space construction submodule 225.

[0111] The curve parameterization submodule 221 parameterizes the edge curve as an arc length parameter representation, establishes a local coordinate system, and divides the reticle region into a regular grid. In a specific implementation, a feature point is selected every 5-10 centimeters along the edge curve, and the sampling is densified at the curvature change. For each feature point, a Frenet frame is established, and tangent vector , normal vector , and binormal vector are defined:

[0112] .

[0113] .

[0114] .

[0115] wherein: represents the position vector of the point on the curve with parameter s, which is a three-dimensional vector with a unit of centimeters; s is the arc length parameter along the curve, with a unit of centimeters; is the unit tangent vector of the point, which is dimensionless, indicating the tangent direction of the curve at the point; is the unit normal vector of the point, which is dimensionless, perpendicular to the tangent vector, pointing to the direction of the curve bending; is the unit binormal vector of the point, which is dimensionless, perpendicular to the tangent vector and the normal vector; represents the derivative of the position vector with respect to the arc length parameter s; represents the derivative of the tangent vector with respect to the arc length parameter s; represents the modulus of the vector . Represents the vector cross product operation. The Frenet frame is an important tool for describing the local geometric properties of a curve and can be used to analyze the bending and torsion properties of a curve.

[0116] The marking area is divided into a regular grid. The grid size is usually set to 1×1 cm. The number of pixels corresponding to the image is calculated based on the camera installation height and angle. Special boundary conditions are applied to the marking edge and end points to ensure calculation stability.

[0117] The differential geometry calculation submodule 222 calculates the metric tensor, the first fundamental form and the second fundamental form on the discrete grid to describe the intrinsic and extrinsic geometric properties of the reticle surface.

[0118] The first basic form is calculated as follows:

[0119] .

[0120] in: It is the first fundamental form, representing the intrinsic measure of the surface; , , is the measurement coefficient, dimensionless; and Represent the parameterized representation of the marking surface Parameters and The partial derivative of Represents vector inner product operation; and The first fundamental form describes intrinsic geometric quantities such as distance, angle, and area on a surface, independent of how the surface is embedded in three-dimensional space.

[0121] The second basic form is calculated as follows:

[0122] .

[0123] in: It is the second fundamental form, which represents the external geometric characteristics of the surface; , , is the coefficient of the second fundamental form in centimeters ; is the surface normal vector, dimensionless; 、 and Respectively The second fundamental form describes the curvature and curvature direction of the surface in three-dimensional space, reflecting the external geometric characteristics of the surface.

[0124] The geometric invariant extraction submodule 223 calculates geometric invariants independent of the observation angle, including Gaussian curvature, mean curvature, and shape index. The Gaussian curvature K and the mean curvature H are respectively:

[0125]

[0126]

[0127] wherein: is the Gaussian curvature, with a unit of centimeter , representing the degree of intrinsic bending of the surface; is the mean curvature, with a unit of centimeter , representing the average value of the degree of extrinsic bending of the surface; is the coefficient of the first fundamental form; is the coefficient of the second fundamental form. The Gaussian curvature is an intrinsic invariant of the surface, which does not change with how the surface is bent, and only changes when the surface is torn or glued, and is an important index for describing the topological properties of the surface.

[0128] The shape index is defined as:

[0129]

[0130] wherein: is the shape index, dimensionless, with a value range of and is the principal curvature, with a unit of centimeter-1, and , is the inverse tangent function. Different values of the shape index represent different local shapes: represents a pit, represents a valley, represents a saddle point, represents a ridge, represents a convex peak. The shape index provides an intuitive description of the local open-line shape of the surface, and is not affected by the scaling of the surface.

[0131] The geometric feature mapping submodule 224 establishes a multi-dimensional feature space, generates a geometric feature vector for each region of the reticle, and identifies regions with abnormal geometric characteristics. In an embodiment of the present application, the geometric feature vector contains 12 geometric descriptors, including Gaussian curvature, mean curvature, shape index, principal curvature ratio, curvature derivative, etc.

[0132] In the reticle quality evaluation, different geometric characteristics have the following mapping relationship with the reticle quality problems:​​​​​​​

[0133] Gauss curvature anomaly (out of ±20% normal range) corresponds to material peeling or bulging;

[0134] Average curvature mutation (change of adjacent regions more than 30%) represents edge wear;

[0135] Shape index change (deviation from normal range > 25%) indicates abnormal surface morphology of the reticle.

[0136] The quality parameter space construction submodule 225 defines basic quality dimensions such as integrity, flatness, adhesion, and uniformity, and maps the geometric characteristics to the quality parameter space. The value range of each quality parameter is 0-100, and the threshold is set as follows: excellent: 90-100 points; good: 75-89 points; medium: 60-74 points; poor: 40-59 points; and bad: 0-39 points.

[0137] The geometric characteristic data set is output to the multi-scale differential invariant evaluation module 23 for further analysis.

[0138] As shown in Figure 5 , the multi-scale differential invariant evaluation module 23 includes a differential invariant calculation submodule 231, a multi-scale feature fusion submodule 232, a feature manifold construction submodule 233, a quality score generation submodule 234, and an abnormal pattern recognition submodule 235.

[0139] The differential invariant calculation submodule 231 calculates first-order invariants, second-order invariants, and high-order invariants. The first-order invariants include the trace of the metric tensor (local area change of the surface) and the determinant (surface stretching ratio), with 4-6 points sampled per square centimeter. The second-order invariants include the Gaussian curvature, the average curvature, and the principal curvature ratio, with 2-3 points sampled per square centimeter. The high-order invariants include the eigenvalue spectrum of the shape operator and the curvature derivative, with 1-2 points sampled per square centimeter in key areas.

[0140] The multi-scale feature fusion submodule 232 constructs 3-5 scale levels, covering features from details (centimeter level) to the whole (meter level). The features at each scale are decomposed into basic components, the most discriminative components are selected for retention, and the feature vectors are reconstructed according to importance. In different application scenarios, the scale weights can be automatically adjusted: wear detection: fine scale weight increases to 60%-70%; overall evaluation: medium scale weight dominates (50%-60%); trend analysis: large scale weight increases to 40%-50%.

[0141] The feature manifold construction submodule 233 defines the similarity measure between features, constructs the Riemannian metric of the feature manifold, and establishes the geodesic line connection between feature points. The geodesic distance between feature points is defined as:

[0142] .

[0143] where: is the geodesic distance between feature points and , unit depends on the metric of feature vector; represents the minimum value in all paths connecting and ; represents the integral on the parameter interval [0,1]; represents the path connecting feature points and ; , , is the Riemann metric, which defines the inner product of vectors in the tangent space; is the tangent vector of the path, representing the tangent vector of the path at parameter ; represents the square length of the tangent vector at point under the metric . The geodesic distance measures the shortest path length between two points on the feature manifold, which is a natural measure of feature similarity.

[0144] The quality score generation submodule 234 defines a metric function of quality difference, establishes a mapping from geodesic distance to quality score, and generates integrity score, visibility score and reflectivity performance score. The quality score and the geodesic distance are related as follows:

[0145] .

[0146] where: is the quality score, dimensionless, with a value range of 0-100; is the natural exponential function; is the proportionality coefficient, with a unit of the reciprocal of geodesic distance, adjusted according to the evaluation standard, usually taking a value of 0.1-0.5; is the geodesic distance. This mapping function converts the geodesic distance into the quality score, the smaller the distance, the higher the score, the larger the distance, the lower the score, which conforms to the intuition of quality evaluation.

[0147] The abnormal pattern recognition submodule 235 extracts the main variation pattern based on the principal component analysis on the manifold, identifies and classifies the abnormal area. By comparing the actual marking line features with the standard pattern, the system can identify different types of abnormalities such as wear, cracking, peeling, and pollution. In an embodiment of the present application, the determination criteria for abnormal patterns are as follows: wear: the curvature change of the marking line edge exceeds the threshold value (usually 30%), and the surface flatness decreases; cracking: local extreme value of Gaussian curvature appears, and the shape index jumps; peeling: adhesion degree parameter is lower than 60 points, and the surface uniformity is poor; pollution: the retroreflective performance score is more than 50% lower than the normal value, but the geometric characteristics are basically normal.

[0148] The quality evaluation result is output to the area score generation module 24, the overall grade evaluation module 25, the evaluation report generation module 26, and the historical trend analysis module 27, and is transmitted to the marking line wear prediction unit 3 for further analysis.

[0149] The area score generation module 24 generates a quality score map of different areas of the marking line to visually display the problem areas. Specifically, the marking line is divided into several areas (usually 10x10 cm or 20x20 cm), the quality score of each area is calculated, and the quality condition is visually displayed using pseudo-color mapping. In a preferred embodiment of the present application, the red-yellow-green color system is used to represent the quality from poor to good, which facilitates the quick identification of problem areas by management personnel.

[0150] The overall grade evaluation module 25 calculates the overall quality grade according to the area score, and divides the marking line quality into five grades: excellent, good, medium, poor, and bad. The overall score calculation uses the weighted average method:

[0151] .

[0152] wherein: is the overall score, dimensionless, with a value range of 0-100; is the score of the i-th area, dimensionless, with a value range of 0-100; is the weight coefficient, dimensionless, satisfying ; is the total number of areas; represents the weighted sum of all areas. The weight coefficient can be set according to the importance of the area, for example, the weight of the marking line at the turning place and near the intersection can be appropriately increased.

[0153] The evaluation report generation module 26 summarizes the quality score and abnormal information to generate a detailed evaluation report containing charts and location markers. The report content includes: marking line location information (GPS coordinates, road name, stake number, etc.), overall quality grade, area score map, abnormal area list and its type, severity and location, and maintenance suggestions.

[0154] The historical trend analysis module 27 compares the evaluation results of the same location at different times to analyze the trend of the change of the marking line quality over time. By fitting the degradation curve, the trend of the change of the marking line quality can be predicted to provide a reference for maintenance decision-making. The degradation curve usually adopts an exponential model:

[0155] .

[0156] wherein: represents the quality score at time , dimensionless, with a value range of 0-100; is the initial quality score, dimensionless, with a value range of 0-100, usually the quality score of a new marking line, close to 100; is a natural exponential function; is a degradation coefficient, with a unit of time , related to factors such as traffic flow, climate conditions, and marking line materials; is time, with a unit of month or year. The model describes the process of exponential decay of the marking line quality over time, which is consistent with the actual observed marking line degradation law.

[0157] As shown in Figure 6 , the marking line wear prediction unit 3 includes a feature mapping sub-unit 31, a traffic flow prediction sub-unit 32, and a marking line life prediction sub-unit 33.

[0158] The feature mapping sub-unit 31 constructs a marking line period and environmental feature mapping relationship model based on the marking line positioning parameters, detected vehicle driving data, and environmental data. The environmental data includes light conditions (sunny, cloudy, rainy, snowy, etc.) and weather conditions (temperature, humidity, precipitation, etc.). The mapping relationship model adopts a weighted fusion method:

[0159] .

[0160] wherein: is the output of the mapping relationship model, which is a feature vector; is the th feature mapping function that maps the input parameters to feature values; is the marking line positioning parameter, including position coordinates, road type, etc.; is the detection period parameter, including date, time, etc.; is the environmental parameter, including temperature, humidity, light, etc.; is the weight coefficient, dimensionless, satisfying is the number of feature mapping functions; represents the weighted sum of all feature mapping functions. The model fuses multi-dimensional heterogeneous data into consistent feature representation, which is convenient for subsequent analysis.

[0161] The traffic flow prediction subunit 32 predicts the traffic flow by establishing a traffic flow time variation model. The time variation model considers the intra-day variation, intra-week variation and seasonal variation:

[0162] .

[0163] wherein: denotes the traffic flow at time , with the unit of vehicle / hour; denotes the reference traffic flow, with the unit of vehicle / hour, representing the traffic flow under standard conditions; , and denote the intra-day, intra-week and seasonal variation functions, respectively, with no dimension; , , and denote the hour, day, month and year time variables, respectively; denotes the annual growth rate, with no dimension, usually taking the value of 2%-5%. The model comprehensively considers the multi-scale time variation of the traffic flow, and can accurately predict the traffic conditions at different time periods.

[0164] The marking life prediction subunit 33 predicts the marking service life based on the traffic flow information, marking time period and environmental characteristic mapping relationship model, and recommends a maintenance scheme according to the service life. The marking service life prediction formula is:

[0165] .

[0166] wherein: denotes the predicted marking service life, with the unit of month; denotes the reference life under standard conditions, with the unit of month, determined according to the marking material and construction quality; denotes the correction coefficient, with no dimension, related to environmental factors such as temperature and pollution index, usually between 0.6-1.2; denotes the marking environmental attribute parameter, with no dimension, related to the environmental characteristics of the location where the marking is located, usually between 0.8-1.5; denotes the traffic flow correction coefficient, with no dimension, usually proportional to the-0.6 power of the traffic flow, i.e. , wherein denotes the daily average traffic flow. The model considers various factors affecting the marking life, and can accurately predict the service life of the marking.

[0167] In the preferred embodiment of the present application, according to the predicted service life, the system gives the following maintenance suggestions: service life < 3 months: prefer to replace; service life 3-6 months: plan to replace; service life 6-12 months: regular inspection; service life > 12 months: normal monitoring.

[0168] The remote central control unit 7 includes a mobile network communication module and a network control module.

[0169] The mobile network communication module connects the remote data storage and analysis module, the target line quality evaluation unit 2 and the target line wear prediction unit 3 through the mobile 4G or 5G network. In an embodiment of the present application, the communication module adopts 4G / 5G dual-mode communication, supports a maximum upload speed of 100 Mbps, and ensures that image data can be transmitted in real time.

[0170] The network control module is used to control the light compensation scheme of the detection unit 1, control the image noise reduction processing scheme through camera data processing, adjust the image color, and optimize the image signal transmission. In actual application, the network control module automatically adjusts the data transmission strategy according to the current network condition, such as reducing the image resolution or using a higher compression rate when the network is congested, to ensure the real-time performance of the system.

[0171] The system of the present application also includes a parameter adaptive adjustment module, a model updating module, an abnormal sample processing module, a cloud management module, a data analysis platform, a management interface and a third party system interface.

[0172] The parameter adaptive adjustment module automatically fine-tunes the algorithm parameters according to the consistency of the evaluation results and manual verification. In actual application, the system regularly collects the target line quality evaluation results of manual verification, compares them with the system evaluation results, and calculates the consistency index. When the consistency is lower than the threshold value (usually 85%), the system automatically adjusts the related algorithm parameters to improve the evaluation accuracy.

[0173] The model updating module regularly collects verification data to optimize the evaluation model. In a preferred embodiment of the present application, the system updates the model once a quarter, using the incremental learning method to gradually improve the evaluation model.

[0174] The abnormal sample processing module records the target line samples with evaluation abnormalities, adds them to the training set after manual confirmation, and improves the system recognition accuracy. In particular, for samples with large differences between the system evaluation results and manual evaluation, the system will automatically mark and submit them for manual review. The confirmed samples will be used as training data for model optimization.

[0175] The cloud management module is used to receive and store target line images and evaluation results. In an embodiment of the present application, a distributed storage architecture is adopted to support PB-level data storage and provide efficient data retrieval and backup functions.

[0176] The data analysis platform provides historical data analysis and trend prediction functions, supports multi-dimensional data query and visual display. Through the platform, management personnel can understand the overall situation and change trend of the target line quality in the region, and provide a basis for macro decision-making.

[0177] The management interface supports remote configuration and monitoring, provides web and mobile access methods, and facilitates management personnel to understand the system running state and evaluation results at any time.

[0178] The third-party system interface is integrated with external systems such as intelligent transportation systems and road maintenance management systems through standard APIs, realizing data sharing and collaborative decision-making.

[0179] The image analysis-based highway marking quality intelligent evaluation system of the present application is suitable for marking quality evaluation and maintenance management of various roads such as expressways and urban roads. The system can be installed on a special detection vehicle or a regular road inspection vehicle to realize normalized monitoring of marking quality.

[0180] In the expressway scenario, the system focuses on the retroreflective performance and integrity of the marking, increases the acquisition frequency, reduces the single processing area, and ensures the evaluation accuracy under high-speed driving conditions. In the urban road scenario, the system enhances the anti-interference ability, focuses on the visibility and wear state of the marking, and adapts to marking recognition under complex backgrounds.

[0181] For special road sections such as tunnel entrances and exits, bridges and viaducts, construction areas, etc., the system adopts targeted processing strategies to ensure the accuracy and reliability of the evaluation.

[0182] Through the system of the present application, the road management department can timely find out the marking quality problems, scientifically formulate the maintenance plan, improve the road safety, and reduce the maintenance cost. At the same time, the marking quality information provided by the system also provides important support for the automatic driving and advanced auxiliary driving systems.

[0183] The above only describes the preferred embodiments of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. The intelligent highway marking quality assessment system based on image analysis is characterized by: include: A detection unit, used to acquire highway marking images and perform image preprocessing; A marking quality assessment unit is communicatively connected to the detection unit and is configured to receive the pre-processed marking image sent by the detection unit. The marking quality assessment unit includes: a curvature flow edge detection module, configured to treat the preprocessed marking image as a two-dimensional manifold, construct a curvature flow equation, extract marking edge features through multi-scale representation, and output an edge feature point set; a marking line geometric property manifold characterization module, communicatively connected to the curvature flow edge detection module, configured to receive the edge feature point set, construct a parameterized representation of the marking line manifold, calculate the first basic form and the second basic form of the marking line surface, extract Gaussian curvature and mean curvature features, and generate a geometric property data set; a multi-scale differential invariant evaluation module, communicatively connected to the marking geometric characteristic manifold characterization module, configured to receive the geometric characteristic dataset, calculate multi-order differential invariants, fuse multi-scale features, construct a geodesic distance metric on the feature manifold, evaluate the integrity, visibility, and reflective performance of the marking, and output a quality assessment result; The road marking wear prediction unit is in communication with the road marking quality assessment unit and is configured to receive the quality assessment result, predict the service life of the road marking, and generate maintenance recommendations.

2. The system according to claim 1, wherein: The curvature flow edge detection module includes: Image normalization submodule, used to convert the input image to a standard resolution and brightness range and establish a mapping relationship from pixels to physical size; The curvature calculation submodule is used to construct the image surface and calculate the principal curvature and mean curvature of each point on the surface; The evolution control submodule is used to control the evolution speed of the surface according to the curvature value, so that the high curvature area evolves slowly and the low curvature area evolves quickly and smoothly; The multi-scale edge representation submodule is used to generate image representations at different scales and retain edges that are stable at multiple scales by comparing the position offsets of edges at different scales; The arc length optimization submodule is used to define an arc length functional including a curvature term and optimize the edge curve by minimizing the functional while maintaining the topological structure of the edge curve.

3. The system according to claim 1, wherein: The marking line geometric characteristic manifold characterization module includes: The curve parameterization submodule is used to parameterize the edge curve into arc length parameter representation, establish a local coordinate system, and divide the marking area into a regular grid; The differential geometry calculation submodule is used to calculate the metric tensor, the first fundamental form and the second fundamental form on the discrete grid to describe the intrinsic and extrinsic geometric properties of the reticulated surface; The geometric invariant extraction submodule is used to calculate the geometric invariants that are independent of the observation angle, including Gaussian curvature, mean curvature and shape index; The geometric feature mapping submodule is used to establish a multidimensional feature space, generate geometric feature vectors for each area of ​​the marking line, and identify areas with abnormal geometric characteristics; The quality parameter space construction submodule is used to define the basic quality dimensions of integrity, flatness, adhesion and uniformity, and map the geometric characteristics to the quality parameter space.

4. The system according to claim 1, wherein: The multi-scale differential invariant evaluation module includes: Differential invariant calculation submodule, used to calculate first-order invariants, second-order invariants and higher-order invariants; The multi-scale feature fusion submodule is used to construct multiple scale levels covering details to the whole, decompose the features at each scale into basic components, and reconstruct the feature vector according to the weighted importance; The feature manifold construction submodule is used to define the similarity measure between features, construct the Riemannian metric of the feature manifold, and establish geodesic connections between feature points; The quality score generation submodule is used to define the measurement function of quality difference, establish the mapping from geodesic distance to quality score, and generate the integrity score, visibility score and reflective performance score; The abnormal pattern recognition submodule is used to extract the main variation patterns based on principal component analysis on the manifold and identify and classify abnormal areas.

5. The system according to claim 1, wherein: Also includes: GPS positioning unit, used to obtain the vehicle's real-time location and driving direction; Vehicle driving information unit, used to obtain the vehicle's real-time speed, acceleration and steering angle; a control unit, in communication with the GPS positioning unit and the vehicle driving information unit, for determining whether to start road marking detection based on the vehicle position and driving information, and sending a detection instruction to the detection unit; Among them, the detection unit includes a camera acquisition unit, a light detection unit, a color compensation unit and an image preprocessing unit. The camera acquisition unit collects the marking line image according to the detection instruction, the light detection unit obtains the real-time light value, the color compensation unit adjusts the light source color and intensity according to the real-time light value, and the image preprocessing unit performs noise filtering and feature extraction on the image after light compensation.

6. The system according to claim 1, wherein: The marking quality assessment unit further includes: The regional scoring generation module is used to generate quality score maps for different areas of the marking line, visually displaying problem areas; The overall grade assessment module is used to calculate the overall quality grade based on the regional scores, and classify the marking quality into five grades: excellent, good, medium, poor, and inferior; An evaluation report generation module, which summarizes quality scores and exception information and generates a detailed evaluation report containing charts and location marks; The historical trend analysis module is used to compare the evaluation results of the same location at different times and analyze the changing trend of road marking quality over time.

7. The system according to claim 1, wherein: The system further comprises: a remote central control unit, in communication with the detection unit and the marking quality assessment unit, for issuing illumination compensation instructions to the detection unit via the Internet of Things, receiving pre-processed feature images, and transmitting them to the marking quality assessment unit; Among them, the remote central control unit includes a mobile network communication module and a network control module. The mobile network communication module is connected to the remote data storage and analysis module through a mobile 4G or 5G network. The network control module is used to control the illumination compensation scheme of the detection unit, adjust the image color, and optimize the image signal transmission.

8. The system according to claim 1, wherein: The marking wear prediction unit includes: The feature mapping subunit is used to construct a mapping relationship model between road marking period and environmental features based on road marking positioning parameters, vehicle driving data and environmental data; Traffic flow prediction subunit, used to predict traffic flow by establishing a traffic flow time variation model; The road marking life prediction subunit is used to predict the service life of road markings based on the relationship model between traffic flow information, road marking period and environmental characteristics, and recommend maintenance plans based on the service life; The environmental data includes lighting conditions and weather conditions.

9. The system according to claim 1, wherein: The system further comprises: Parameter adaptive adjustment module, used to automatically fine-tune algorithm parameters based on the consistency of evaluation results and manual verification; Model update module, used to regularly collect verification data and optimize the evaluation model; The abnormal sample processing module is used to record and evaluate abnormal marking samples, and add them to the training set after manual confirmation to improve the system's recognition accuracy; The parameter adaptive adjustment module maintains different parameter configuration schemes for different road types, different weather conditions and different road marking types.

10. The system according to claim 1, wherein: The system further comprises: A cloud management module for receiving and storing marking images and evaluation results; A data analysis platform, in communication with the cloud management module, for providing historical data analysis and trend prediction; A management interface, communicating with the data analysis platform to support remote configuration and monitoring; The third-party system interface is used to integrate with external systems such as intelligent transportation systems and road maintenance management systems through standard APIs to achieve data sharing and collaborative decision-making.

Images