An asphalt pavement defect intelligent identification and positioning method and system

Abstract

The present application provides a kind of asphalt pavement defect intelligent identification and positioning method, comprising: using grating sensor array to collect pavement waveform data, and using image acquisition module to collect synchronous pavement image and identify feature area by image recognition strategy, according to the feature area of image recognition and corresponding pavement waveform data, obtain pre-training feature library, the pre-training feature library includes known feature and corresponding waveform template;Real-time waveform data is collected using grating sensor, and real-time waveform data is matched with the waveform template in pre-training feature library, and the similarity score of each waveband is obtained;The similarity score of each waveband is analyzed respectively, when the similarity score is higher than similarity threshold, then the corresponding area is known feature, when the similarity feature is lower than similarity threshold, then the corresponding area is defect feature;Obtain defect feature area and determine the type and severity of defect by defect analysis strategy.

Description

Technical Field

[0001] This invention relates to the field of road defect identification technology, and in particular to an intelligent identification and location method and system for asphalt pavement defects. Background Technology

[0002] Asphalt pavement is widely used in the construction of transportation infrastructure such as highways and urban roads due to its advantages such as driving comfort, convenient construction, and good noise reduction. However, under the influence of multiple factors such as long-term traffic load, natural environmental erosion, and deviations in construction techniques, asphalt pavement is prone to various defects such as depressions, bumps, cracks, loosening, and foreign object adhesion. Some defects may even extend into the deeper layers of the pavement, forming complex defects that link the surface and deeper layers. If these defects are not identified and located in a timely manner, they will gradually aggravate the pavement damage, shorten its service life, increase road maintenance costs, affect driving safety and comfort, and even cause traffic accidents.

[0003] Currently, the identification and location of defects in asphalt pavements mainly rely on manual inspections and traditional testing techniques. Manual inspections are affected by factors such as the experience, sense of responsibility, and workload of the inspectors, resulting in problems such as low identification efficiency, high rates of missed and false detections, insufficient location accuracy, high testing costs, and the inability to achieve large-scale continuous testing. Furthermore, they are difficult to effectively identify deep pavement defects and cannot record the development status of defects in real time. Traditional testing techniques often use single sensors to collect data, such as identifying surface defects through image acquisition modules or collecting pavement structure data through waveform sensors. These techniques suffer from drawbacks such as limited data, incomplete feature recognition, and weak anti-interference capabilities. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of this invention is to provide an intelligent identification and location method and system for asphalt pavement defects, so as to overcome the above-mentioned shortcomings of the existing technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for intelligent identification and localization of defects in asphalt pavement, characterized in that it includes: The pre-trained feature construction step involves using a grating sensor array to collect road surface waveform data and using an image acquisition module to simultaneously capture road surface images. Feature regions are then identified using an image recognition strategy. Based on the image-recognized feature regions and the corresponding road surface waveform data, a pre-trained feature library is obtained. The pre-trained feature library includes known features and corresponding waveform templates. The feature acquisition step involves using a grating sensor to collect real-time waveform data, and then performing adversarial matching between the real-time waveform data and waveform templates in a pre-trained feature library to obtain similarity scores for each band. The defect analysis step includes setting a similarity threshold and analyzing the similarity score of each band. When the similarity score is higher than the similarity threshold, the corresponding region is identified as a known feature and its feature type is marked. When the similarity score is lower than the similarity threshold, the corresponding region is identified as a defect feature. The defect identification step involves acquiring the defect feature area and determining the type and severity of the defect through defect analysis strategies.

[0007] Preferably, the grating sensor includes a laser grating emitter, and the defect analysis step includes a laser analysis strategy, which includes: The laser data acquisition step involves obtaining waveform data with similarity features below the similarity threshold as surface defect areas and constructing surface contour waveforms. The gradient change rate acquisition steps involve observing the fluctuations of the waveform segment by segment, quantifying the intensity of the waveform fluctuations, obtaining the gradient characterization value of each band, and extracting the largest gradient characterization value as the maximum gradient change rate. The depth acquisition step involves acquiring reference areas on both sides of the area to be analyzed, acquiring the height of the two reference areas respectively, acquiring the reference height representing the normal state of the road surface, comparing the deviation of the highest and lowest points of the waveform in the area to be analyzed from the reference height, and acquiring the larger offset as the absolute depth. The defect analysis step involves constructing dynamic thresholds based on historical defect data of the road segment. These dynamic thresholds include a gradient change rate threshold and an absolute depth threshold. If the absolute depth of the surface defect area is higher than the depth threshold and the gradient change rate is greater than the gradient change threshold, then when the highest point is higher than the reference surface, it is determined to be a convex defect; when the highest point is lower than the reference surface, it is determined to be a concave defect.

[0008] Preferably, the laser analysis strategy also includes a road surface foreign object identification step. When a protrusion-type defect is detected, defect features are acquired, including surface reflection intensity, surface curvature, and texture complexity. A foreign object feature identification library is constructed, and the extracted defect features are matched with the knowledge base. If the reflection intensity is higher than the protrusion range of the road surface itself, the number of curvature inflection points is greater than that of the road surface protrusion, and the texture is messy, it is determined to be a road foreign object. Otherwise, it is determined to be a road surface protrusion. Based on the extension length of the protrusion area along the driving direction and the vehicle speed, the passage time of the protrusion area is obtained. If the passage time of the protrusion area is lower than the time threshold, the confidence that the protrusion is a foreign object is increased.

[0009] Preferably, the grating sensor further includes an infrared grating emitter, and the defect analysis step further includes an infrared analysis strategy. The infrared analysis strategy includes a defect feature acquisition step, which acquires waveform data with similarity features lower than a similarity threshold as abnormal regions, identifies abnormal features, records the spatial location and intensity level of the abnormal features to form an abnormal feature set, and simultaneously acquires surface defect regions to form a surface defect region set. Defect features are then formed based on the surface defect set and the abnormal feature set.

[0010] Preferably, the infrared analysis strategy further includes a deep defect acquisition step, which involves spatially registering the surface defect set with the abnormal feature set, filtering out overlapping or adjacent regions to form candidate association pairs, screening the candidate association pairs for deep defects, and determining that if the intensity level of the abnormal feature associated with the surface defect region is higher than the conventional abnormal threshold corresponding to the surface defect feature, then a deep defect exists below the surface defect region. Furthermore, by combining the waveform distortion degree corresponding to the abnormal feature and the size of the surface defect, the type and risk level of the deep defect are determined.

[0011] Preferably, the pre-trained feature construction step further includes a dynamic update strategy, which includes: The feature acquisition and labeling step uses an image recognition strategy to determine whether the feature region is a complete feature. If the recognition result is a partial feature, the image region corresponding to the partial feature and the road waveform data collected at the same time are extracted, and the vehicle's position information is obtained as partial feature information and stored in the cache. In the data accumulation and calibration steps, when the vehicle travels to the same or adjacent area again, the feature acquisition and labeling steps are repeated to obtain new partial feature information. Based on the vehicle location information, the partial features and the partial features in the cache are spatially registered, and the waveform similarity and image feature point matching degree of the overlapping area are calculated. A confidence score is generated. When the confidence score is higher than the matching threshold, it is determined that the new partial feature information and the partial feature information in the cache belong to the same physical feature. The feature fusion step fuses the road surface waveform data in overlapping areas based on similarity to generate continuous and complete road surface waveform data. The complete feature region is then obtained by stitching together the road surface images.

[0012] Preferably, the data accumulation and calibration step also includes a time-space consistency verification sub-step, which analyzes the timing of the occurrence of distortion peaks or valleys in the newly acquired partial feature information of the road surface waveform data, and compares the spatial occurrence order of feature edges in the image region corresponding to the same partial feature in the direction of vehicle travel. If the waveform distortion timing is logically consistent with the spatial occurrence order of image feature edges, the original confidence level generated by waveform similarity and image feature point matching degree is enhanced. If there is a logical contradiction between the waveform distortion timing and the spatial progression order of image feature edges, the confidence level is corrected.

[0013] Preferably, the adversarial matching includes constructing known feature templates and real-time data acquisition channels, mapping each known feature template and its corresponding real-time road surface data to the same high-dimensional feature space. The mapping process causes similar features to form clusters in the feature space, while dissimilar features form separate distributions. The distribution characteristics of the feature space are continuously optimized through an adversarial network mechanism, so that the distribution area of ​​the known feature template in the feature space and the distribution area of ​​the real-time data are adversarially distinguished. In the adversarially optimized feature space, for any real-time data, the distribution proximity between it and each template in the set of known feature templates is calculated, and this proximity is used as the similarity score between the two. The similarity score reflects the probability or closeness of the real-time data falling into the feature distribution area defined by the known feature template. Based on the calculated similarity score, the real-time data is determined to be the feature type corresponding to the known feature template with the highest similarity, or when all similarity scores are lower than a preset threshold, the real-time data is determined to be an unknown feature or a defect feature.

[0014] An intelligent identification and location system for asphalt pavement defects includes: The pre-trained feature construction module uses a grating sensor array to collect road surface waveform data and uses an image acquisition module to collect and capture road surface images simultaneously. It also uses an image recognition strategy to identify feature regions. Based on the image-recognized feature regions and the corresponding road surface waveform data, a pre-trained feature library is obtained. The pre-trained feature library includes known features and corresponding waveform templates. The feature acquisition module uses a grating sensor to collect real-time waveform data, and performs adversarial matching between the real-time waveform data and waveform templates in the pre-trained feature library to obtain similarity scores for each band. The defect analysis module has a similarity threshold. It analyzes the similarity score of each band. When the similarity score is higher than the similarity threshold, the corresponding area is identified as a known feature and the corresponding feature type is obtained. When the similarity score is lower than the similarity threshold, the corresponding area is identified as a defect feature.

[0015] The beneficial effects of this invention are as follows: By coordinating the grating sensor array and the image acquisition module, road surface waveform data and road surface image data are acquired synchronously. Combined with image recognition strategies, feature regions are located, and a pre-trained feature library containing known features and corresponding waveform templates is constructed. This achieves multi-dimensional feature complementarity and avoids feature omissions caused by single data acquisition. Simultaneously, a dynamic update strategy for the pre-trained feature library is designed. Through the acquisition, accumulation, spatial registration, fusion, and spatiotemporal consistency verification of partial features, incomplete partial features can be effectively identified and fused to generate continuous and complete feature information. This dynamically enriches the feature library content, adapts to the diversity and dynamic changes of road surface defects, significantly reduces the rate of missed and false detections, and improves the comprehensiveness and accuracy of defect identification. The dynamic update strategy incorporates vehicle location information to perform spatial registration on some features, ensuring the spatial accuracy of feature fusion. A spatiotemporal consistency verification sub-step is added, which corrects the confidence level by comparing the waveform distortion time sequence with the spatial progression order of image feature edges, ensuring the spatiotemporal consistency of feature recognition and positioning. This further improves the accuracy of defect positioning, enabling precise location of defects and providing accurate location references for subsequent maintenance work, thus reducing maintenance costs. An adversarial matching approach is adopted to match real-time waveform data with waveform templates in a pre-trained feature library. By constructing a high-dimensional feature space adversarial network mechanism, the feature space distribution is optimized, enabling similar features to cluster and dissimilar features to separate, thereby improving the sensitivity and anti-interference ability of feature matching. It can quickly and accurately distinguish between known features and defect features. At the same time, the accuracy of feature matching is further improved by using similarity scores for quantitative judgment, thus improving the efficiency of defect identification. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a flowchart of the dynamic update process of the pre-trained feature library of the present invention; Figure 3 This is a flowchart of the laser analysis strategy of the present invention; Figure 4 This is a flowchart of the infrared analysis strategy of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] It should be noted that when a component is described as "fixed to" another component, it can be directly on the other component or may have a component in between. When a component is considered "connected to" another component, it can be directly connected to the other component or may have a component in between. When a component is considered "set on" another component, it can be directly set on the other component or may have a component in between. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0020] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings: like Figures 1-4 As shown, the present invention provides an intelligent identification and location method for asphalt pavement defects, including: The pre-training feature construction step involves acquiring road surface waveform data using a grating sensor array and simultaneously capturing road surface images using an image acquisition module. Feature regions are identified using image recognition strategies. Based on the identified feature regions and corresponding road surface waveform data, a pre-training feature library is obtained, including known features and corresponding waveform templates. Through multi-source data collaborative acquisition and analysis, a comprehensive and accurate feature reference system is established. The grating sensor array, as the core data acquisition component, consists of a laser grating emitter and an infrared grating emitter. It can acquire longitudinal elevation change information of the road surface at a high-density sampling frequency, forming continuous road surface waveform data. This waveform data accurately reflects the physical characteristics of the road surface, such as its undulations and structural changes, providing a quantitative basis for subsequent feature matching. The image acquisition module works synchronously with the grating sensor array, capturing road surface visual information in real time using high-definition imaging equipment. Its acquisition frequency is strictly aligned with the sensor data sampling frequency, ensuring that each set of waveform data corresponds to a road surface image at the same spatiotemporal location, achieving spatiotemporal consistency in data correlation.

[0021] In the feature region identification process, the image recognition strategy employs multi-scale feature extraction and semantic segmentation techniques. First, the acquired road surface images are preprocessed, including noise reduction, contrast enhancement, and distortion correction, to eliminate the influence of factors such as lighting variations, shadow interference, and shooting angle deviations on the recognition results. Then, shallow feature extraction captures basic information such as road surface texture and edges, followed by deep feature extraction to mine semantic information from the image. Combined with prior knowledge of road surface features, various regions in the image are classified and identified, filtering out feature regions containing known features or potential defects. These feature regions include both normal road surface structural features and potentially abnormal areas, providing clear visual references for subsequent waveform template construction.

[0022] The construction of the pre-trained feature library involves associating and mapping the feature regions obtained from image recognition with the corresponding road surface waveform data. For each identified feature region, key feature parameters are extracted from its corresponding waveform data, including peak value, valley value, period, and amplitude. Combined with the visual attributes of the feature region, a one-to-one correspondence between the feature and the waveform template is formed. Known features cover the regular structural features formed during the normal use of asphalt pavement, as well as defect features of clearly defined types from historical inspections. After manual annotation and verification, these features, along with the corresponding waveform templates, are stored in the pre-trained feature library to form a standardized reference benchmark.

[0023] The pre-training feature construction step also includes a dynamic update strategy, which includes: The feature acquisition and labeling step uses an image recognition strategy to determine whether the feature region is a complete feature. If the recognition result is a partial feature, the image region corresponding to the partial feature and the road waveform data collected at the same time are extracted, and the vehicle's position information is obtained as partial feature information and stored in the cache. The image recognition strategy performs a completeness check on the identified feature regions. When the recognition result is incomplete due to limitations in the shooting angle, occlusion, or insufficient sensor sampling range, the system automatically extracts the image segment corresponding to that feature, simultaneously extracts the waveform data of the corresponding road segment collected by the grating sensor at the same time, and obtains the current location information through the vehicle's onboard positioning module. This partial feature information is categorized and stored in a cache, while auxiliary information such as the collection time and vehicle direction is recorded to provide a basis for subsequent data association.

[0024] The data accumulation and calibration steps involve repeating the feature acquisition and labeling steps when the vehicle travels to the same or adjacent area again to obtain new partial feature information. Based on the vehicle's location information, spatial registration is performed on partial features and features in the cache. The waveform similarity and image feature point matching degree of the overlapping area are calculated, and a confidence score is generated. If the confidence score is higher than the matching threshold, the new partial feature information and the partial feature information in the cache are determined to belong to the same physical feature. When the vehicle subsequently passes through the same or adjacent road segment, the system repeats the feature acquisition and labeling process to obtain new partial feature information. Based on the vehicle's positioning information, the system spatially registers the newly acquired partial features with historical partial features stored in the cache, determining their overlap in physical space using coordinate alignment technology. The similarity of waveform data within the overlapping area is then calculated, quantifying the matching degree by comparing the waveform morphology and the degree of difference in key parameters. Simultaneously, feature points in the feature regions of the two images are extracted, and the number and accuracy of matching feature points are calculated. A confidence score is generated by combining the waveform similarity and image feature point matching degree. When the confidence level reaches the preset matching threshold, it is determined that the new and old feature information originates from the same physical feature, providing a basis for subsequent fusion operations.

[0025] The feature fusion step fuses overlapping areas of road surface waveform data based on similarity, generating continuous and complete road surface waveform data. Complete feature regions are then obtained through road surface image stitching. For multiple partial feature data identified as the same physical feature, the system performs weighted fusion based on similarity in overlapping areas of the waveform data. Regions with higher similarity have greater weight, ensuring that the fused waveform data accurately reflects the complete shape of the feature and generates a continuous, unbroken waveform curve. Simultaneously, corresponding image segments are stitched together, and image fusion technology eliminates stitching artifacts, restoring the complete visual shape of the feature region. The fused complete feature information is added to a pre-trained feature library for supplementation and updating. If a complete feature is still not formed after fusion, the updated partial feature information continues to be stored in a cache, awaiting further supplementation with subsequent data collection.

[0026] The data accumulation and calibration steps also include a temporal and spatial consistency verification sub-step. This involves analyzing the timing of distortion peaks or valleys in the newly acquired feature information, and comparing the spatial order of feature edges in the image region corresponding to the same feature along the vehicle's direction of travel. If the waveform distortion timing and the spatial order of image feature edges are logically consistent, the original confidence level generated by waveform similarity and image feature point matching is enhanced. If there is a logical contradiction between the waveform distortion timing and the spatial progression order of image feature edges, the confidence level is corrected. The analysis focuses on the timing of distortion peaks or valleys in the waveform data corresponding to the newly acquired features, while simultaneously observing the spatial order of feature edges in the corresponding image along the vehicle's direction of travel. Since the operation of the sensor and image acquisition module during vehicle movement has a fixed temporal relationship, under normal circumstances, the waveform distortion timing and the spatial progression order of image feature edges should remain logically consistent. When the order of the two matches, it indicates that there is no abnormality in the data collection process, and the system will automatically enhance the previously generated original confidence level. If a logical contradiction occurs, it indicates that there may be data collection errors or feature mismatches. The system will correct the confidence level accordingly based on the degree of contradiction to reduce the risk of misfusion.

[0027] The feature acquisition step utilizes a grating sensor to collect real-time waveform data. This real-time waveform data is then adversarially matched against waveform templates in a pre-trained feature library to obtain similarity scores for each band. Through high-precision data acquisition and intelligent matching algorithms, rapid comparison between real-time road features and the pre-trained feature library is achieved. During real-time waveform data acquisition, the laser grating emitter and infrared grating emitter operate synchronously. The laser grating focuses on the road surface, capturing road elevation changes with micron-level precision, forming high-frequency waveform data reflecting surface undulations. The infrared grating penetrates the road surface, acquiring thermal radiation signals from the road's internal structure, converting them into low-frequency waveform data reflecting deeper conditions. These two types of waveform data are integrated to form multi-band real-time waveform data, comprehensively covering the physical characteristics of both the road surface and deeper layers, providing rich quantitative evidence for subsequent matching analysis. During data acquisition, the sensor employs an adaptive sampling frequency adjustment mechanism, dynamically adjusting the sampling interval according to vehicle speed to ensure sufficient data density under different driving conditions, avoiding feature information loss due to speed changes.

[0028] Adversarial matching involves constructing known feature templates and real-time data acquisition channels. Each known feature template and its corresponding real-time road surface data are mapped to the same high-dimensional feature space. This mapping process causes similar features to cluster together, while dissimilar features are distributed separately. An adversarial network mechanism continuously optimizes the distribution characteristics of the feature space, creating an adversarial distinction between the distribution areas of known feature templates and the distribution areas of real-time data. In the adversarially optimized feature space, for any real-time data, the distribution proximity between it and each template in the known feature template set is calculated. This proximity is used as the similarity score, reflecting the probability or closeness of the real-time data falling into the feature distribution area defined by the known feature templates. Based on the calculated similarity score, the real-time data is determined to be the feature type corresponding to the known feature template with the highest similarity. Alternatively, if all similarity scores are below a preset threshold, the real-time data is determined to be an unknown feature or a defective feature. By constructing a high-dimensional feature space and an adversarial optimization mechanism, accurate distinction between real-time data and templates is achieved. The construction of known feature templates and real-time data acquisition channels is a fundamental prerequisite for adversarial matching. The known feature templates are derived from validated standardized features in the pre-trained feature library. Each template contains complete waveform feature parameters and corresponding feature attribute information. The real-time data acquisition channel is responsible for preprocessing the multi-band waveform data acquired by the grating sensor, including data noise reduction, outlier removal, and data standardization, to eliminate the impact of sensor noise, environmental interference, and other factors on data quality, ensuring the consistency and reliability of the data input to the matching system.

[0029] The mapping process to a high-dimensional feature space aims to transform low-dimensional waveform data into feature vectors in a high-dimensional space, achieving a deep representation of features. This mapping operation is accomplished through a feature extraction algorithm. This algorithm extracts key features from the waveform data across multiple dimensions, including temporal features such as peak-to-valley distribution, duration, and periodicity, and frequency domain features such as frequency components, amplitude distribution, and phase characteristics. It also integrates auxiliary information such as sensor type and band attributes to form a multi-dimensional feature set. Each known feature template and real-time waveform data are transformed into corresponding feature vectors through this mapping process. Since the template feature vectors and real-time feature vectors reside in the same high-dimensional feature space, the differences between different features are amplified in this high-dimensional space.

[0030] Adversarial networks (ANNs) achieve continuous optimization of the feature space distribution through a dynamic game between a generator and a discriminator. The generator learns the distribution patterns of known feature templates and simulates the generation of feature vectors similar to those known features. The discriminator distinguishes whether the input feature vector comes from known feature templates or real-time data. During the game, both continuously adjust their parameters; the generator attempts to generate feature vectors that are harder for the discriminator to distinguish, while the discriminator continuously improves its discriminative ability. This adversarial process causes similar features to gradually cluster in the high-dimensional space, while features of different categories disperse, forming a clear separation. This makes the boundaries of the feature space more distinct, significantly improving the accuracy of subsequent matching.

[0031] Calculating distributional proximity is a crucial step in obtaining similarity scores. In the adversarially optimized high-dimensional feature space, each feature vector corresponds to a specific spatial location. For any set of real-time feature vectors, the system calculates the spatial distance between it and the feature vector corresponding to each known feature template in the pre-trained feature library. The distance directly reflects the degree of similarity between the two. Distributional proximity is derived from the relative relationship of spatial distance; the closer the spatial distance, the higher the proximity and the higher the corresponding similarity score; the farther the spatial distance, the lower the proximity and the lower the similarity score.

[0032] The essence of similarity score is a quantitative expression of the probability that real-time data belongs to a certain known feature template. The higher the score, the greater the probability that the real-time data falls within the feature distribution area defined by the known feature template. When the similarity score of real-time data is higher than a preset threshold, the system determines that the road surface area corresponding to the real-time data is the feature type corresponding to the known feature template. If the similarity scores corresponding to all known feature templates are lower than the preset threshold, it indicates that the road surface feature corresponding to the real-time data does not appear in the pre-trained feature library and is determined to be an unknown feature or a defect feature, providing a clear analysis object for subsequent defect analysis steps.

[0033] The defect analysis process involves setting similarity thresholds and analyzing the similarity scores of each band. When the similarity score is higher than the threshold, the corresponding area is identified as a known feature and its feature type is marked. When the similarity score is lower than the threshold, the corresponding area is identified as a defect feature. Through multi-dimensional threshold judgment and classification strategy analysis, accurate classification and defect identification of road features are achieved. The similarity threshold is set using a dynamic adaptive mechanism, not a fixed value. It is determined by combining the distribution range of similar features in the pre-trained feature library, the false judgment rate statistics of historical detection data, and the differences in road conditions of different road sections. During the analysis, the system independently evaluates the similarity scores of the laser band and the infrared band to ensure that the feature judgment of different types of data does not interfere with each other. When the similarity score of a certain band is higher than the similarity threshold of the corresponding band, it indicates that the feature of that area highly matches the known features in the pre-trained feature library, and the system automatically marks the specific type of the feature. When the score is lower than the threshold, it means that the feature of that area is not covered by the feature library and is directly identified as a defect feature, entering the subsequent specialized analysis process.

[0034] The grating sensor includes a laser grating emitter, and the defect analysis steps include a laser analysis strategy, which includes: The laser data acquisition process involves acquiring waveform data with similarity scores below a similarity threshold as surface defect areas and constructing a surface contour waveform. From all waveband data, waveform data with similarity scores below the threshold are selected; the corresponding road surface areas are then identified as suspected surface defect areas. When constructing the surface contour waveform based on this data, the original data is smoothed to eliminate waveform jitter caused by instantaneous sensor noise, while retaining key feature points that reflect the true undulations of the road surface, ensuring the accuracy and reliability of the contour waveform.

[0035] The gradient change rate acquisition process involves observing the waveform fluctuations segment by segment, quantifying the severity of these fluctuations, obtaining the gradient characterization value for each band, and extracting the maximum gradient characterization value as the maximum gradient change rate. This quantification of waveform fluctuation severity helps determine abrupt changes in the road surface layer. The system divides the surface profile waveform into several continuous bands of fixed length, calculating the ratio of the elevation change to the horizontal distance between adjacent data points for each segment. This ratio represents the gradient characterization value for the corresponding band. Statistical analysis of the gradient characterization values ​​across all bands extracts the maximum value as the maximum gradient change rate. This value effectively reflects the steepness of the defect area's edge, providing crucial information for subsequent defect type identification.

[0036] The depth acquisition step involves acquiring reference areas on both sides of the area to be analyzed, and obtaining the height of each reference area to establish a baseline height representing the normal state of the road surface. The deviations of the highest and lowest points of the waveform within the analysis area from this baseline height are compared, and the larger offset is taken as the absolute depth. The vertical dimension of the defect is quantified by establishing this baseline height. The system automatically selects intact road surface areas unaffected by defects on both sides of the analysis area as reference areas, typically selecting continuous waveform data within a certain distance before and after the analysis area. The average elevation of all data points within the two reference areas is calculated, and the midpoint is taken as the baseline height representing the normal state of the road surface. Subsequently, the elevation differences between the highest and lowest points of the waveform within the analysis area relative to the baseline height are compared, and the larger absolute value is selected as the absolute depth. This value directly reflects the severity of the defect in the vertical direction.

[0037] The defect analysis process involves constructing dynamic thresholds based on historical defect data for road segments. These dynamic thresholds include gradient rate of change thresholds and absolute depth thresholds. If the absolute depth of a surface defect area exceeds the depth threshold and the gradient rate of change exceeds the gradient change threshold, the defect is classified as a convex defect if the highest point is above the reference surface, and as a concave defect if the highest point is below the reference surface. This dynamic threshold system, built based on historical defect data, can adapt to the differences in pavement characteristics across different road segments. The system retrieves historical defect records for the currently detected road segment, statistically analyzes the gradient rate of change and absolute depth data for similar defects at different times, and combines this with factors such as pavement design standards, service life, and traffic flow to determine the gradient rate of change threshold and absolute depth threshold through statistical analysis. When the absolute depth of a surface defect area exceeds the depth threshold and the maximum gradient rate of change exceeds the gradient change threshold, the system further determines the defect based on the position of the highest point of the waveform relative to the reference surface: if the highest point is above the reference surface, it is classified as a convex defect; if the highest point is below the reference surface, it is classified as a concave defect, achieving accurate classification of surface defects.

[0038] The laser analysis strategy also includes a road surface foreign object (GMO) identification step. When a protrusion-like defect is detected, defect features are acquired, including surface reflection intensity, surface curvature, and texture complexity. A GMO feature identification database is constructed, and the extracted defect features are matched with the knowledge base. If the reflection intensity is higher than the protrusion range of the road surface itself, the number of curvature inflection points is greater than that of the road surface protrusion, and the texture is messy, it is identified as a road GMO; otherwise, it is identified as a road surface protrusion. Based on the extension length of the protrusion area along the driving direction and the vehicle speed, the transit time of the protrusion area is obtained. If the transit time of the protrusion area is lower than a time threshold, the confidence that the protrusion is a GMO is increased. The road surface GMO identification step is a further subdivision of protrusion-like defects, avoiding misclassification of road surface foreign objects as road surface structural protrusions. The system extracts three core features from protruding defect areas: surface reflection intensity is calculated from the amplitude of the laser reflection signal; the material difference between the road surface protrusion and the foreign object leads to a significant difference in reflection intensity. Surface curvature is calculated using the second derivative of continuous waveform data, reflecting the smoothness of the protruding area's contour. Texture complexity is determined by analyzing the pixel grayscale change frequency of the corresponding road surface image. A foreign object feature recognition library stores feature data of common road surface foreign objects such as stones, debris, and spilled materials. The extracted defect features are compared with the data in the library. When the reflection intensity exceeds the normal range of a road surface protrusion, the number of curvature inflection points is greater than the typical value for a road surface structure protrusion, and the texture appears chaotic, it is initially identified as a road surface foreign object; otherwise, it is identified as a road surface protrusion. Simultaneously, the system calculates the time required for a vehicle to pass through the protruding area based on the length of the protruding area along the vehicle's travel direction and the current vehicle speed. If this time is lower than a preset time threshold, it indicates that the protruding area is small, further increasing the confidence that it is a foreign object and ensuring the accuracy of the recognition results.

[0039] The grating sensor also includes an infrared grating emitter. The defect analysis step further includes an infrared analysis strategy, which includes a defect feature acquisition step. This involves acquiring waveform data with similarity scores below a similarity threshold as anomalous regions, identifying anomalous features, recording their spatial location and intensity level to form an anomalous feature set, and simultaneously acquiring surface defect regions to form a surface defect region set. Defect features are then formed based on the surface defect set and the anomalous feature set. Infrared waveform data with similarity scores below a threshold are filtered out, and their corresponding regions are marked as anomalous regions. When identifying anomalous features, the focus is on analyzing the intensity changes and spatial distribution of the infrared signal, recording the specific spatial coordinates and intensity level of the anomalous features to form an anomalous feature set. Simultaneously, surface defect region information obtained from the laser analysis strategy is integrated to form a surface defect region set. The combination of these two approaches constructs a more comprehensive defect feature system.

[0040] The infrared analysis strategy also includes a deep defect acquisition step. This involves spatially registering the surface defect set with the anomalous feature set, filtering out overlapping or adjacent regions to form candidate association pairs, and then screening these pairs for deep defects. If the intensity level of the anomalous feature associated with the surface defect region is higher than the conventional anomalous threshold corresponding to the surface defect feature, a deep defect is determined to exist beneath the surface defect region. Further analysis, combined with the waveform distortion degree corresponding to the anomalous feature and the size of the surface defect, determines the type and risk level of the deep defect. The core of the deep defect acquisition step is establishing the correlation between surface defects and deep anomalies. The system first spatially registers the surface defect region set with the anomalous feature set, mapping the two types of data to the same physical space coordinate system using coordinate alignment technology. Regions that overlap or are close together are then filtered out to form candidate association pairs. When screening these candidate association pairs for deep defects, the conventional anomalous threshold corresponding to the surface defect region is used as the judgment standard. This threshold is based on statistical analysis of the infrared signal intensity of normal road surfaces. When the intensity level of the anomalous feature in a candidate association pair exceeds the conventional anomalous threshold, it indicates an anomaly in the road structure beneath the surface defect region, and the presence of a deep defect is determined. Subsequently, by combining the degree of waveform distortion corresponding to the abnormal characteristics (the more severe the waveform distortion, the more obvious the damage to the deep structure), and the size of the surface defects, the specific type of deep defects, such as crack extension and internal loosening, is comprehensively judged, and the potential road damage risk is assessed to determine the risk level.

[0041] The defect identification process involves acquiring the defect feature area and determining the defect type and severity using a defect analysis strategy. All output data from both laser and infrared analysis strategies are compiled, including defect types such as depressions, protrusions, foreign objects, and deep loosening, as well as key parameters such as location coordinates, absolute depth, gradient change rate, and anomaly intensity level. By comprehensively evaluating the deviation of these parameters from dynamic thresholds and considering the potential impact of the defect on driving safety and road surface lifespan, the severity level of the defect is ultimately determined, providing comprehensive and accurate technical support for subsequent road maintenance decisions.

[0042] An intelligent identification and location system for asphalt pavement defects includes: The pre-trained feature construction module uses a grating sensor array to collect road surface waveform data and an image acquisition module to simultaneously capture road surface images. It then uses an image recognition strategy to identify feature regions. Based on the image-recognized feature regions and the corresponding road surface waveform data, a pre-trained feature library is obtained. The pre-trained feature library includes known features and corresponding waveform templates. The feature acquisition module uses a grating sensor to collect real-time waveform data, and performs adversarial matching between the real-time waveform data and waveform templates in the pre-trained feature library to obtain similarity scores for each band. The defect analysis module has a similarity threshold. It analyzes the similarity score of each band. When the similarity score is higher than the similarity threshold, the corresponding area is identified as a known feature and the corresponding feature type is obtained. When the similarity score is lower than the similarity threshold, the corresponding area is identified as a defect feature.

[0043] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent identification and location of defects in asphalt pavement, characterized in that, include: The pre-trained feature construction step involves using a grating sensor array to collect road surface waveform data and using an image acquisition module to simultaneously capture road surface images. Feature regions are then identified using an image recognition strategy. Based on the image-recognized feature regions and the corresponding road surface waveform data, a pre-trained feature library is obtained. The pre-trained feature library includes known features and corresponding waveform templates. The feature acquisition step involves using a grating sensor to collect real-time waveform data, and then performing adversarial matching between the real-time waveform data and waveform templates in a pre-trained feature library to obtain similarity scores for each band. The defect analysis step includes setting a similarity threshold and analyzing the similarity score of each band. When the similarity score is higher than the similarity threshold, the corresponding region is identified as a known feature and its feature type is marked. When the similarity score is lower than the similarity threshold, the corresponding region is identified as a defect feature. The defect identification step involves acquiring the defect feature area and determining the type and severity of the defect through defect analysis strategies.

2. The intelligent identification and location method for asphalt pavement defects according to claim 1, characterized in that, The grating sensor includes a laser grating emitter, and the defect analysis step includes a laser analysis strategy, which includes: The laser data acquisition step involves obtaining waveform data with similarity features below the similarity threshold as surface defect areas and constructing surface contour waveforms. The gradient change rate acquisition steps involve observing the fluctuations of the waveform segment by segment, quantifying the intensity of the waveform fluctuations, obtaining the gradient characterization value of each band, and extracting the largest gradient characterization value as the maximum gradient change rate. The depth acquisition step involves acquiring reference areas on both sides of the area to be analyzed, acquiring the height of the two reference areas respectively, acquiring the reference height representing the normal state of the road surface, comparing the deviation of the highest and lowest points of the waveform in the area to be analyzed from the reference height, and acquiring the larger offset as the absolute depth. The defect analysis step involves constructing dynamic thresholds based on historical defect data of the road segment. These dynamic thresholds include a gradient change rate threshold and an absolute depth threshold. If the absolute depth of the surface defect area is higher than the depth threshold and the gradient change rate is greater than the gradient change threshold, then when the highest point is higher than the reference surface, it is determined to be a convex defect; when the highest point is lower than the reference surface, it is determined to be a concave defect.

3. The intelligent identification and location method for asphalt pavement defects according to claim 2, characterized in that, The laser analysis strategy also includes a road surface foreign object identification step. When a protrusion-type defect is detected, defect features are acquired, including surface reflection intensity, surface curvature, and texture complexity. A foreign object feature identification library is constructed, and the extracted defect features are matched with the knowledge base. If the reflection intensity is higher than the protrusion range of the road surface itself, the number of curvature inflection points is greater than that of the road surface protrusion, and the texture is messy, it is determined to be a road foreign object. Otherwise, it is determined to be a road surface protrusion. Based on the extension length of the protrusion area along the driving direction and the vehicle speed, the passage time of the protrusion area is obtained. If the passage time of the protrusion area is lower than the time threshold, the confidence that the protrusion is a foreign object is increased.

4. The intelligent identification and location method for asphalt pavement defects according to claim 2, characterized in that, The grating sensor also includes an infrared grating emitter, and the defect analysis step also includes an infrared analysis strategy. The infrared analysis strategy includes a defect feature acquisition step, which acquires waveform data with similarity features below a similarity threshold as abnormal regions, identifies abnormal features, records the spatial location and intensity level of abnormal features to form an abnormal feature set, and simultaneously acquires surface defect regions to form a surface defect region set. Based on the surface defect set and the abnormal feature set, defect features are formed.

5. The intelligent identification and location method for asphalt pavement defects according to claim 4, characterized in that, The infrared analysis strategy also includes a deep defect acquisition step, which involves spatially registering the surface defect set with the abnormal feature set, filtering out overlapping or adjacent regions to form candidate association pairs, screening the candidate association pairs for deep defects, and determining that if the intensity level of the abnormal feature associated with the surface defect region is higher than the conventional abnormal threshold corresponding to the surface defect feature, then a deep defect exists below the surface defect region. Furthermore, by combining the waveform distortion degree corresponding to the abnormal feature and the size of the surface defect, the type and risk level of the deep defect are determined.

6. The intelligent identification and location method for asphalt pavement defects according to claim 1, characterized in that, The pre-trained feature construction step also includes a dynamic update strategy, which includes: The feature acquisition and labeling step uses an image recognition strategy to determine whether the feature region is a complete feature. If the recognition result is a partial feature, the image region corresponding to the partial feature and the road waveform data collected at the same time are extracted, and the vehicle's position information is obtained as partial feature information and stored in the cache. In the data accumulation and calibration steps, when the vehicle travels to the same or adjacent area again, the feature acquisition and labeling steps are repeated to obtain new partial feature information. Based on the vehicle location information, the partial features and the partial features in the cache are spatially registered, and the waveform similarity and image feature point matching degree of the overlapping area are calculated. A confidence score is generated. When the confidence score is higher than the matching threshold, it is determined that the new partial feature information and the partial feature information in the cache belong to the same physical feature. The feature fusion step fuses the road surface waveform data in overlapping areas based on similarity to generate continuous and complete road surface waveform data. The complete feature region is then obtained by stitching together the road surface images.

7. The intelligent identification and location method for asphalt pavement defects according to claim 6, characterized in that, The data accumulation and calibration steps also include a time-space consistency verification sub-step, which analyzes the timing of the occurrence of distortion peaks or valleys in the newly acquired partial feature information of the road surface waveform data, and compares the spatial occurrence order of feature edges in the image region corresponding to the same partial feature in the direction of vehicle travel. If the waveform distortion timing is logically consistent with the spatial occurrence order of image feature edges, the original confidence level generated by waveform similarity and image feature point matching degree is enhanced. If there is a logical contradiction between the waveform distortion timing and the spatial progression order of image feature edges, the confidence level is corrected.

8. The intelligent identification and location method for asphalt pavement defects according to claim 1, characterized in that, The adversarial matching involves constructing known feature templates and real-time data acquisition channels, mapping each known feature template and its corresponding real-time road surface data to the same high-dimensional feature space. This mapping process causes similar features to form clusters in the feature space, while dissimilar features form separate distributions. The distribution characteristics of the feature space are continuously optimized through an adversarial network mechanism, enabling an adversarial distinction between the distribution areas of known feature templates and the distribution areas of real-time data. In the adversarially optimized feature space, for any real-time data, the distribution proximity between it and each template in the set of known feature templates is calculated. This proximity is used as the similarity score between the two. The similarity score reflects the probability or proximity of the real-time data to the feature distribution area defined by the known feature templates. Based on the calculated similarity score, the real-time data is determined to be the feature type corresponding to the known feature template with the highest similarity. Alternatively, when all similarity scores are below a preset threshold, the real-time data is determined to be an unknown feature or a defect feature.

9. An intelligent identification and location system for asphalt pavement defects, characterized in that, include: The pre-trained feature construction module uses a grating sensor array to collect road surface waveform data and uses an image acquisition module to collect and capture road surface images simultaneously. It also uses an image recognition strategy to identify feature regions. Based on the image-recognized feature regions and the corresponding road surface waveform data, a pre-trained feature library is obtained. The pre-trained feature library includes known features and corresponding waveform templates. The feature acquisition module uses a grating sensor to collect real-time waveform data, and performs adversarial matching between the real-time waveform data and waveform templates in the pre-trained feature library to obtain similarity scores for each band. The defect analysis module has a similarity threshold. It analyzes the similarity score of each band. When the similarity score is higher than the similarity threshold, the corresponding area is identified as a known feature and the corresponding feature type is obtained. When the similarity score is lower than the similarity threshold, the corresponding area is identified as a defect feature.

Images