Airfield pavement management system based on AI and three-dimensional scanning

Abstract

The invention relates to the field of airfield pavement disease identification, and provides an airfield pavement management system based on AI and three-dimensional scanning, which comprises three-dimensional laser scanning equipment and an AI server, the three-dimensional laser scanning equipment is deployed on an airport pavement and is used for collecting point cloud data of a target area; the AI server comprises a filtering unit, a flattening processing unit and a crack identification unit; the filtering unit is used for performing multi-dimensional filtering on the point cloud data to generate a target point cloud without noise points; the flattening processing unit is used for generating a flatness cloud picture of the target area according to the target point cloud and positioning a standard exceeding area; wherein the flatness cloud picture comprises a two-dimensional cloud picture and a three-dimensional cloud picture; and the crack identification unit is used for inputting the exceeding area into the U-Net model, extracting a crack mask through an RGB space and an HSV space, and outputting a crack geometric parameter according to the crack mask.

Description

Technical Field

[0001] This invention relates to the field of airport pavement defect identification, and in particular to an airport pavement management system based on AI and 3D scanning. Background Technology

[0002] Airport pavements are a core component of civil aviation infrastructure, and their engineering quality and health status directly affect aircraft operational safety and airport operational efficiency. Currently, the inspection, acceptance, and maintenance of airport pavements both domestically and internationally still mainly rely on traditional manual inspections and sampling measurement methods.

[0003] For data acquisition, traditional surveying equipment such as levels, total stations, and pavement leveling instruments are commonly used to obtain key parameters through local sampling of the pavement. For defect identification, on-site visual inspection by technicians or assisted interpretation based on two-dimensional images is primarily relied upon to record and assess common defects such as cracks, spalling, and misalignment. For project acceptance and cost control, manual comparison of design drawings with on-site measured data is often used to calculate deviations in quantities and materials used, thereby estimating cost overruns.

[0004] In recent years, 3D laser scanning technology has been gradually introduced into the field of civil engineering. Its non-contact, high-precision, and full-view data acquisition capabilities have provided a new technical approach for pavement measurement. At the same time, artificial intelligence technology has made significant progress in areas such as image recognition and point cloud processing, and some studies have attempted to apply deep learning algorithms to road crack identification and classification.

[0005] However, existing technological solutions mostly remain at the stage of improving single technical aspects and have not yet formed an integrated solution covering the entire process of "data acquisition - intelligent analysis - decision support". Especially in application scenarios such as airport pavement, which have high precision and high reliability requirements, the existing technology system still has many technical bottlenecks and systemic defects.

[0006] Existing technical problems: Low detection efficiency and limited coverage: Traditional inspection methods rely on manual sampling, which makes it difficult to achieve comprehensive coverage of the entire pavement area and results in insufficient data representativeness. The low sampling frequency and long cycle cannot meet the needs of rapid acceptance and high-frequency inspections of large airports, seriously affecting project progress and operation and maintenance response speed.

[0007] Disease identification is highly subjective and lacks quantitative standards. The existing disease identification process relies heavily on the experience and judgment of technical personnel, lacks a unified quantitative standard and objective evaluation system, resulting in highly subjective and inconsistent identification results, making it difficult to trace history and analyze trends.

[0008] Cost control lags behind, making it difficult to detect material overconsumption in a timely manner: Deviations between material usage and design specifications during construction are often only discovered through manual calculations in the later stages of the project. The lack of real-time monitoring and early warning mechanisms leads to delayed cost control and makes it difficult to achieve refined management of the construction process.

[0009] Low degree of technological integration and weak system integration: Although 3D laser scanning and AI technology each have their own advantages, existing applications are mostly isolated systems that lack end-to-end integration from data collection, processing, and analysis to business decision-making. This fails to fully leverage the synergistic effects of the technologies and makes it difficult to support the overall digital transformation of airport pavement management. Summary of the Invention

[0010] This application proposes an airport pavement management system based on AI and 3D scanning to address the problems mentioned in the background art above: To achieve the above objectives, this application provides the following technical solution: In the first aspect, this application proposes an airport pavement management system based on AI and 3D scanning, the system including a 3D laser scanning device and an AI server; The three-dimensional laser scanning equipment is deployed on the airport pavement to collect point cloud data of the target area; The AI ​​server includes a filtering unit, a leveling unit, and a crack recognition unit. The filtering unit is used to perform multidimensional filtering on point cloud data to generate a target point cloud with noise removed; the multidimensional filtering includes: RANSAC model fitting filtering, DBSCAN clustering filtering, cubic radius filtering, statistical filtering, local contrast filtering and rasterized local threshold filtering. The leveling processing unit is used to generate a leveling cloud map of the target area based on the target point cloud and to locate the areas exceeding the standard; the leveling cloud map includes a two-dimensional cloud map and a three-dimensional cloud map. The crack identification unit is used to input the out-of-specification area into the U-Net model, extract the crack mask through RGB and HSV space, and output the crack geometric parameters based on the crack mask.

[0011] In conjunction with the first aspect, the three-dimensional laser scanning device includes: Acquire a regional image of the target area and generate a basic scan path; the basic scan path includes the airport runway, taxiway, and apron. Based on the basic scanning path, the acquisition accuracy of each scanning area is determined, and the first scanning strategy of the 3D laser scanning device is generated; wherein the acquisition accuracy is not less than the millimeter level. According to the first scanning strategy, the device attitude and motion trajectory data at each moment are determined; wherein, the device attitude is used to determine the local point density, and the motion trajectory data is used to determine the reflection intensity distribution. Determine whether there are abnormal point cloud data where the local point density is lower than a preset first threshold and the reflection intensity variance is higher than a preset second threshold; wherein, abnormal point cloud data is a region with degraded scan quality; wherein, the preset first threshold represents the minimum point cloud density and the preset second threshold represents the minimum reflection intensity threshold; Based on abnormal point cloud data, an attenuation model for laser scanning in heterogeneous pavement media is established, and the optimal compensation scanning angle and compensation scanning line spacing are calculated. A second scanning strategy is generated based on the compensated scanning angle and the compensated scanning line spacing.

[0012] In conjunction with the first aspect, the three-dimensional laser scanning device is also equipped with a synchronous multispectral environmental sensor; wherein, Multispectral environmental sensors are used to collect real-time environmental parameters of the area; these real-time environmental parameters include: ambient light intensity, surface temperature, and atmospheric visibility. Based on real-time environmental parameters, deploy high-speed vibration sensors and acoustic sensors that operate synchronously with the 3D laser scanning equipment; The characteristics of abnormal high-frequency vibrations were determined using high-speed vibration sensors. Based on acoustic sensors, determine the first acoustic characteristics of the three-dimensional laser scanning equipment in operation and the second acoustic characteristics of the environmental background; Based on real-time environmental parameters, abnormal high-frequency vibration characteristics, first acoustic characteristics, and second acoustic characteristics, a multi-dimensional compensation model is constructed to perform dynamic calibration when each frame of point cloud data is generated.

[0013] In conjunction with the first aspect, the filtering unit includes an index node, a first filtering decision node, and a second filtering decision node; The index node is used to divide local point cloud clusters based on point cloud data and calculate the feature vector of each local point cloud cluster; the feature vector includes curvature variance, normal vector consistency and neighborhood density gradient. The first filtering decision node is used to output a filtering strategy based on the feature vector and through a preset lightweight decision model. The filtering strategy includes the optimal filtering algorithm, the filtering execution order of the optimal filtering algorithm and the combination of initial parameters. The filtering strategy is used to remove point cloud noise from the point cloud data in order to determine the target point cloud.

[0014] In conjunction with the first aspect, the optimal filtering algorithm includes: RANSAC model fitting filter is used to fit the ground plane in point cloud data and remove noise points that deviate too far from the plane. DBSCAN clustering filter is used to cluster candidate points above the ground in point cloud data and remove small clusters with a size smaller than a preset height threshold. Three-stage radius filtering includes: the first stage retains points with ≥10 neighboring points with a radius of 0.18m; the second stage retains points with ≥10 neighboring points with a radius of 0.14m; and the third stage retains points with ≥15 neighboring points with a radius of 0.30m. Statistical filtering is used to iteratively calculate the mean and standard deviation of ground point heights and remove outliers whose heights exceed a preset maximum value. Local contrast filtering is used to determine the local ground height by using a preset quantile of the neighborhood height, and to delete isolated high points with height differences exceeding a preset value; Rasterized local threshold filtering is used to divide point cloud data into grids, dynamically generate local thresholds, and remove noise points that exceed the local threshold.

[0015] In conjunction with the first aspect, the leveling processing unit includes: Projection nodes: used to project the target point cloud onto a two-dimensional horizontal plane and construct a Delaunay triangulation; Calculation node: Used to calculate the residual between the elevation of the three vertices of each triangular facet in the Delaunay triangulation and the elevation of the local microplane where the corresponding triangular facet is located, as the initial value of the flatness of the core region of the triangular facet; Weighted nodes are used to collect the initial flatness values ​​of all adjacent triangle faces for each vertex of the triangulation network, and calculate the weighted flatness value of the corresponding vertex using a weight allocation function based on kernel density estimation; wherein, the weight function is related to the cosine of the angle between the triangle area and its normal vector and the average normal vector at the vertex of the triangulation network; 2D rendering node: Used to generate a continuous, high-precision 2D flatness field covering the entire target area by using bicubic spline interpolation to calculate the weighted flatness values ​​of all vertices, and then render a 2D cloud map based on this field. 3D rendering node: Used to superimpose the 2D flatness field as a height offset onto the 2D projection coordinates of the target point cloud to generate a 3D deformable cloud map that characterizes the spatial distribution features of flatness.

[0016] In conjunction with the first aspect, the leveling processing unit also includes a labeling interface that is synchronously output by the crack identification unit: wherein, The annotation interface is used to simultaneously display the 3D deformation cloud map and the corresponding real scene orthophoto; When any out-of-range area is selected by the user via a location command, an independent window is generated, and the original 3D point cloud corresponding to the selected area is highlighted and rendered through the independent window.

[0017] In conjunction with the first aspect, the crack identification unit includes a first channel and a second channel; wherein, The first channel is a grayscale image of the reflection intensity generated by the point cloud projection of the area exceeding the standard. The second channel is a multi-scale normal rate of change feature image obtained based on cloud computing of points in the over-standard area.

[0018] In conjunction with the first aspect, the U-Net model includes parallel encoding branch channels and decoding branch channels; wherein, The encoded branch channel is used to extract spatial context features of the reflectance intensity image and geometric texture features of the normal change rate image; The decoding branch channel is used to adaptively weight and fuse spatial context features and geometric texture features through a cross-attention mechanism, and outputs a crack pixel-level prediction mask that represents reflection properties and three-dimensional geometric characteristics.

[0019] In conjunction with the first aspect, the calculation of the crack geometry parameters includes width calculation, depth calculation, and volume calculation; wherein, When calculating the geometric parameters of the crack, all three-dimensional point clouds belonging to the crack are extracted from the point cloud data in advance based on the crack pixel-level prediction mask; When calculating the width, the crack skeleton line is segmented according to the three-dimensional point set. In the normal plane of each segment point, the crack point cloud is projected onto the corresponding plane. The distribution range of the projected point cloud in the direction perpendicular to the skeleton line is calculated as the local width of the projected point cloud. The weighted average of the local widths of all segment points is taken, and the weight is the reciprocal of the curvature of the projected point cloud. During depth calculation, based on the 3D point cloud, a reference pavement point cloud without cracks is selected around the crack point cloud. The local pavement surfaces of the crack area and the reference area are fitted by the moving least squares method, and the signed distance from the crack point to the reference surface is calculated as the depth. When calculating the volume, the depth value of each crack point is multiplied by the equivalent area it occupies in the local triangular mesh based on the three-dimensional point set, and the sum is used to determine the estimated volume of the crack.

[0020] The beneficial effects of this invention are as follows: This invention acquires 3D point cloud data of the pavement using a 3D laser scanning device. An AI server then removes noise through multi-dimensional filtering to obtain a high-quality target point cloud, generates a smoothness cloud map to locate areas exceeding the standard, and finally extracts cracks from the areas exceeding the standard using a U-Net model and multi-spatial features, outputting quantitative parameters. This achieves intelligent management of the entire process from data acquisition to defect assessment, aiming to improve the reliability and accuracy of data processing in complex and noisy environments such as airport pavements.

[0021] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0023] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0024] In the attached diagram: Figure 1 This is a diagram illustrating the execution process of an airport pavement management system based on AI and 3D scanning, as described in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the execution steps of the three-dimensional laser scanning device in an embodiment of the present invention. Detailed Implementation

[0025] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0026] This application proposes an intelligent system for airport pavement condition detection, assessment and management. The system acquires three-dimensional spatial data of the pavement through three-dimensional laser scanning, and an AI server performs automated preprocessing, analysis and defect identification on the data, and finally outputs the corresponding defect report.

[0027] Example 1: See Figure 1 This application proposes an airport pavement management system based on AI and 3D scanning, the system including a 3D laser scanning device and an AI server; Three-dimensional laser scanning equipment is deployed on the airport pavement to collect point cloud data of the target area; The AI ​​server includes a filtering unit, a leveling unit, and a crack recognition unit; The filtering unit is used to perform multidimensional filtering on point cloud data to generate a target point cloud with noise removed; the multidimensional filtering includes: RANSAC model fitting filtering, DBSCAN clustering filtering, cubic radius filtering, statistical filtering, local contrast filtering and rasterized local threshold filtering. The leveling processing unit is used to generate a leveling cloud map of the target area based on the target point cloud and to locate the areas exceeding the standard; the leveling cloud map includes a two-dimensional cloud map and a three-dimensional cloud map. The crack identification unit is used to input the out-of-specification area into the U-Net model, extract the crack mask through RGB and HSV space, and output the crack geometric parameters based on the crack mask.

[0028] In the specific implementation process of this application, firstly, vehicle-mounted, rack-mounted, or handheld 3D laser scanning equipment is used to collect point cloud data on the airport road surface. The 3D laser scanning equipment emits laser beams and receives reflected signals, calculates the 3D coordinates of the target points, and generates point cloud data. It can quickly and extensively collect high-precision 3D geometric information of the pavement. Compared with 2D detection technology, the data dimensions are richer and can completely preserve details such as the elevation and texture of the pavement.

[0029] An AI server is an intelligent computing device that integrates a filtering unit, a flatness processing unit, and a crack identification unit. It is responsible for point cloud data preprocessing, flatness analysis, crack identification, and quantification.

[0030] The filtering unit, an AI server responsible for noise removal from point cloud data, employs a multi-dimensional noise reduction strategy combining RANSAC model fitting filtering, DBSCAN clustering filtering, cubic radius filtering, statistical filtering, local contrast filtering, and rasterized local threshold filtering. This strategy eliminates noise from different dimensions: RANSAC model fitting removes outliers, DBSCAN density clustering removes isolated points, cubic radius filtering removes locally sparse points, statistical filtering removes points deviating from the mean, local contrast filtering removes abnormal height points, and rasterized threshold filtering removes noise within the raster. This comprehensively covers various noise scenarios, resulting in more thorough noise reduction and more complete preservation of valid data compared to single filtering methods. For example, addressing the complexity of airport pavement point cloud data and the presence of noise sources including vehicles, shadows, pavement debris, and insufficient scanner systems, this application utilizes multi-dimensional filtering to maximize the removal of various noise points while preserving the true pavement features.

[0031] The pavement smoothness processing unit is a functional unit within the AI ​​server responsible for pavement smoothness analysis. It outputs two types of visual cloud maps and locates areas that do not meet smoothness standards. The smoothness processing unit calculates the elevation deviation of each point in the target point cloud and maps the deviation to color / grayscale to form cloud maps, including 3D and 2D cloud maps. The 2D cloud map displays the distribution of locations exceeding the standard in a 2D planar view; the 3D cloud map displays elevation changes in a 3D stereoscopic view. By comparing the cloud maps with preset smoothness thresholds, the exceeding areas can be located. The visualized output of pavement smoothness distribution, with the 2D cloud map used for rapid screening of planar exceeding locations and the 3D cloud map clearly displaying elevation anomalies and accurately locating areas requiring repair, combined with the 2D cloud map output, allows inspectors to identify exceeding areas from a planar perspective and also determine detailed information about exceeding standards from the 3D perspective of pavement undulations.

[0032] The crack identification unit is the functional unit in the AI ​​server responsible for the detection and quantification of pavement cracks. It employs deep learning and multi-spatial feature extraction, inputting data of the affected area into the U-Net segmentation model. This model combines multi-dimensional features from RGB (color) and HSV (hue / saturation / brightness) spaces to extract crack masks, and then calculates geometric parameters such as crack length, width, and area. The U-Net model improves crack segmentation accuracy, and the multi-spatial features enhance its resistance to interference from lighting, stains, and other factors, outputting high-precision crack detection results. This achieves high-precision, pixel-level crack identification and geometric parameterization of crack states. The U-Net segmentation model excels at small target segmentation, and the RGB and HSV spaces can be combined with scene data such as lighting and stains to improve the accuracy of crack identification, enabling the output of geometric parameters to move from qualitative to specific quantitative results.

[0033] The beneficial effects of this application are as follows: This application acquires 3D point cloud data of the pavement using a 3D laser scanning device. The AI ​​server then removes noise through multi-dimensional filtering to obtain a high-quality target point cloud, generates a smoothness cloud map to locate areas exceeding the standard, and finally extracts cracks from the areas exceeding the standard using the U-Net model and multi-spatial features, and outputs quantitative parameters. This achieves intelligent management of the entire process from data acquisition to defect assessment, with the aim of improving the reliability and accuracy of data processing in complex noise environments such as airport pavements.

[0034] Example 2: See Figure 2 3D laser scanning equipment includes: Acquire a regional image of the target area and generate a basic scan path; the basic scan path includes the airport runway, taxiway, and apron. Based on the basic scanning path, the acquisition accuracy of each scanning area is determined, and the first scanning strategy of the 3D laser scanning device is generated; wherein the acquisition accuracy is not less than the millimeter level. According to the first scanning strategy, the device attitude and motion trajectory data at each moment are determined; wherein, the device attitude is used to determine the local point density, and the motion trajectory data is used to determine the reflection intensity distribution. Determine whether there are abnormal point cloud data where the local point density is lower than a preset first threshold and the reflection intensity variance is higher than a preset second threshold; wherein, abnormal point cloud data is a region with degraded scan quality; wherein, the preset first threshold represents the minimum point cloud density and the preset second threshold represents the minimum reflection intensity threshold; Based on abnormal point cloud data, an attenuation model for laser scanning in heterogeneous pavement media is established, and the optimal compensation scanning angle and compensation scanning line spacing are calculated. A second scanning strategy is generated based on the compensated scanning angle and the compensated scanning line spacing.

[0035] In the specific implementation of this application, the 3D laser scanning equipment collects image information of the core area of ​​the airport pavement and plans an initial scanning path framework covering the airport runway, taxiway, and apron. Based on the functional area division of the airport pavement, the scanning range is determined using image recognition or preset area information, and a basic path is generated; ensuring that the scan covers all key functional areas of the airport pavement and avoiding omission of important parts.

[0036] For different pavement areas covered by the basic scanning path, a data acquisition accuracy standard of no less than millimeter level is set, and an initial scanning execution plan is formulated. Furthermore, based on the importance of the airport pavement areas—for example, since runway accuracy requirements are higher than apron accuracy—corresponding accuracy is allocated to form the first scanning strategy. This ensures high accuracy in critical areas while optimizing scanning resource allocation and avoiding over-scanning.

[0037] It is understandable that by setting the first scanning strategy corresponding to the basic scanning path and acquisition accuracy, it is possible to determine that there is a set quality baseline in the key area.

[0038] Based on the first scanning strategy, the spatial attitude of the scanning device is calculated in real time, including its angle and height. The motion trajectory is the device's movement path; the device attitude affects the local point cloud density, and the motion trajectory affects the spatial distribution of laser reflection intensity. The device attitude determines the laser beam coverage and point cloud density, while the motion trajectory determines the scanning area sequence and the distribution pattern of reflection intensity data. By precisely controlling the attitude and trajectory, the local point cloud density is ensured to meet the standards, and the reflection intensity distribution is uniform, thereby improving the quality and consistency of the original data.

[0039] During the judgment process, the scanned point cloud data is examined. If the point density in a local area is lower than the preset minimum point cloud density threshold, and the variance of the reflection intensity is higher than the preset minimum reflection intensity threshold, then the area is determined to be an abnormal point cloud area with degraded scan quality. By comparing the point cloud density and the variance of the reflection intensity with preset thresholds, areas where scan quality has deteriorated due to uneven pavement media and equipment deviations are identified; thus, low-quality data areas are promptly discovered, and compensation areas are determined.

[0040] Understandably, by performing real-time and quantitative diagnostics on scanning quality and dynamically linking the motion parameters of the equipment with the point cloud quality indicators, it is possible to accurately locate areas of data quality degradation caused by uneven pavement materials, water accumulation and oil stains, or equipment vibration. Under the criteria that low local point density indicates poor scanning geometry or severe absorption, resulting in insufficient geometric coverage, and high reflection intensity variance indicates uneven pavement medium reflection characteristics, it is possible to accurately determine whether there is a heterogeneous medium causing signal attenuation.

[0041] For anomalous point cloud regions, a mathematical model of laser energy attenuation in non-uniform surface media (such as surfaces of different materials and roughness) is constructed. Based on this model, the optimal compensation angle and line spacing for improving scanning quality are calculated. The law of laser reflection attenuation on non-uniform surfaces is analyzed, and scanning parameters are optimized through the model. This addresses the scanning quality degradation caused by non-uniform surfaces and improves the point cloud quality in anomalous regions.

[0042] The second scanning strategy is based on the calculated optimal compensation angle and line spacing, and formulates a supplementary scanning execution plan for abnormal point cloud areas. The compensation parameters are converted into an executable scanning strategy to guide the equipment to perform a second scan of abnormal areas; accurately repair low-quality data areas and ensure the consistency of point cloud data quality across the entire pavement.

[0043] This application's 3D laser scanning device generates a basic path and a first scanning strategy covering key areas, controlling the device's attitude and trajectory to acquire data. By detecting abnormal point cloud areas, it establishes a non-homogeneous pavement attenuation model to calculate compensation parameters, generating a second scanning strategy for supplementary scanning, thus achieving a dynamically optimized scanning process throughout the entire workflow. This dynamic detection and compensation mechanism solves the pain points of uneven quality and the inability to repair degraded areas in traditional scanning.

[0044] Understandably, the attenuation model in the homogeneous pavement medium is built by reverse engineering based on anomalous point cloud data. It is used to calculate the optimal compensation scanning angle and linear spacing, enabling the scanning device to integrate the current environment for scanning.

[0045] The beneficial effects of the above scheme are: To prevent scanning quality degradation in complex noisy environments, this application employs intelligent scanning, utilizing real-time quality feedback acquisition and adaptive compensation via a physical model to proactively process passive data. Unlike simple filtering, interpolation, and enhancement processes, this approach builds a physical model based on anomalous data and calculates the theoretically optimal compensation scanning geometry parameters, thereby improving the accuracy of recognition results while saving computational resources.

[0046] Example 3: The three-dimensional laser scanning equipment is also equipped with a synchronous multispectral environmental sensor; among which, Multispectral environmental sensors are used to collect real-time environmental parameters of the area; these real-time environmental parameters include: ambient light intensity, surface temperature, and atmospheric visibility. Based on real-time environmental parameters, deploy high-speed vibration sensors and acoustic sensors that operate synchronously with the 3D laser scanning equipment; The characteristics of abnormal high-frequency vibrations were determined using high-speed vibration sensors. Based on acoustic sensors, determine the first acoustic characteristics of the three-dimensional laser scanning equipment in operation and the second acoustic characteristics of the environmental background; Based on real-time environmental parameters, abnormal high-frequency vibration characteristics, first acoustic characteristics, and second acoustic characteristics, a multi-dimensional compensation model is constructed to perform dynamic calibration when each frame of point cloud data is generated.

[0047] In the specific implementation of this application, the three-dimensional laser scanning equipment is equipped with a multispectral environmental sensor that is synchronized with the scanning process in time, in order to collect environmental parameters that affect the scanning accuracy.

[0048] Multispectral sensors detect environmental information in different spectral bands (such as visible light and infrared) and align it with the working sequence of the scanning equipment to ensure temporal consistency between environmental and scanning data. Multispectral sensors are used to capture external environmental disturbances. Strong light affects the signal-to-noise ratio, temperature changes cause thermal drift and material deformation in the equipment, and visibility (humidity, particulate matter) directly affects laser transmission and echo intensity. High-speed vibration sensors are used to capture abnormal high-frequency vibrations from external machinery and the sensor's own condition; for example, passing vehicles and wind loads can generate high-frequency vibrations that can easily lead to scanning geometric distortion. Acoustic sensors can reflect the health status of motors and scanning mirrors; the acoustic characteristics of the environmental background can be correlated with specific external activities, such as the severe vibration interference caused by aircraft takeoff and landing.

[0049] The multidimensional compensation model is used to convert the perceived environmental parameters, abnormal high-frequency vibration parameters, acoustic noise and other interference quantities into the correction quantities of point cloud data, so as to determine the comprehensive interference error mapping model. When generating each frame of data, it can suppress environmental noise from all the root causes of interference.

[0050] The beneficial effects of the above scheme are: This application addresses the issue of decreased point cloud data accuracy in complex noisy environments due to varying environmental conditions and internal / external mechanical disturbances. By adding heterogeneous sensors to the physical sensing stage of data acquisition, it achieves source noise acquisition and compensation, enabling hardware sensing and model optimization. Furthermore, this application employs a multispectral sensing device combining high-speed vibration and acoustic sensing to construct a multi-dimensional compensation model. This is a specific design for airport pavement scanning, a particular source of full-spectrum interference, and addresses the randomness and ambiguity of environmental interference. Through hardware isolation, this application acquires key environmental variables affecting laser scanning quality in real time, resolving the accuracy fluctuations caused by neglecting environmental factors in traditional scanning methods.

[0051] Example 4: The filtering unit includes an index node, a first filtering decision node, and a second filtering decision node; The index node is used to divide local point cloud clusters based on point cloud data and calculate the feature vector of each local point cloud cluster; the feature vector includes curvature variance, normal vector consistency and neighborhood density gradient. The first filtering decision node is used to output a filtering strategy based on the feature vector and through a preset lightweight decision model. The filtering strategy includes the optimal filtering algorithm, the filtering execution order of the optimal filtering algorithm and the combination of initial parameters. The filtering strategy is used to remove point cloud noise from the point cloud data in order to determine the target point cloud.

[0052] In an optimizable embodiment, a second filtering decision node is deployed. The second filtering decision node is used to extract local feature vectors from the initial target point cloud again and calculate the filtering accuracy index through a preset verification model. If the filtering accuracy index does not reach the preset threshold, the parameter combination of the optimal filtering algorithm is adjusted and fed back to the first filtering decision node to re-execute the filtering. If the filtering accuracy index reaches the preset threshold, the final target point cloud is output. The filtering accuracy index includes noise retention rate, effective feature retention rate, and point cloud density uniformity.

[0053] In the specific implementation of this application, the first filtering decision node in the filtering unit, based on the feature vector provided by the index node, uses a lightweight decision model to generate a filtering strategy that includes the optimal filtering algorithm, execution order, and initial parameters, thereby achieving noise removal and determining the target point cloud. In this process, a lightweight decision model, such as a simplified decision tree or a small neural network, adaptively selects the most suitable filtering scheme, algorithm, order, and parameters based on the feature vectors of local clusters, specifically addressing different types of noise. This solves the problems of poor performance and low efficiency caused by traditional filtering algorithms being singular or having fixed parameters, improving the accuracy and computational efficiency of noise removal.

[0054] In this application, the index node is used to extract features from the scanned raw point cloud data. The raw point cloud data is divided into blocks by partitioning local point cloud clusters. Feature vectors are calculated for each local point cloud cluster after partitioning. These feature vectors include three types of data: curvature variance, normal vector consistency, and neighborhood density gradient. Curvature variance reflects the severity of surface undulations; normal vector consistency reflects the smoothness of the surface; and neighborhood density gradient reflects the uniformity of point distribution. The first filtering decision node maps the feature vectors to specific filtering strategies based on a pre-defined lightweight decision model. The lightweight decision model uses decision trees, random forests, or small neural networks. The filtering strategy includes the corresponding algorithm, execution order, and initial parameter combinations.

[0055] In this application, the lightweight decision model employs a random forest algorithm. In a practical training implementation, the number of decision trees is set to 100, the maximum depth to 10, and a feature selection method based on information gain is used. During model training, a dataset containing at least 5000 airport pavement crack samples is used. The training data is divided into at least 3000 small crack samples, 1500 medium crack samples, and 500 large crack samples. 80% of the samples are used for training, and the remaining 20% ​​are used for validation. Under these training conditions, the method in this application achieves an accuracy of 92.5% in the airport pavement crack identification task, which is 15.3% higher than traditional methods.

[0056] The beneficial effects of this application are: The filtering unit divides local point cloud clusters and extracts feature vectors through index nodes. The first filtering decision node uses a lightweight decision model to output an adaptive filtering strategy, achieving intelligent and efficient noise removal. For the complex scenario of airport pavements, the filtering unit is used to generate the most efficient filtering and purification scheme, maximizing noise removal while preventing over-smoothing and erroneous deletion of real pavement details. Index nodes are used to introduce the point cloud data required for decision-making into the lightweight decision model, allowing the model to consider different feature scenarios and synergistic effects, thus improving parameter sensitivity during decision-making.

[0057] Example 5: The optimal filtering algorithm includes: RANSAC model fitting filter is used to fit the ground plane in point cloud data and remove noise points that deviate too far from the plane. DBSCAN clustering filter is used to cluster candidate points above the ground in point cloud data and remove small clusters with a size smaller than a preset height threshold. Three-stage radius filtering includes: the first stage retains points with ≥10 neighboring points with a radius of 0.18m; the second stage retains points with ≥10 neighboring points with a radius of 0.14m; and the third stage retains points with ≥15 neighboring points with a radius of 0.30m. Statistical filtering is used to iteratively calculate the mean and standard deviation of ground point heights and remove outliers whose heights exceed a preset maximum value. Local contrast filtering is used to determine the local ground height by using a preset quantile of the neighborhood height, and to delete isolated high points with height differences exceeding a preset value; Rasterized local threshold filtering is used to divide point cloud data into grids, dynamically generate local thresholds, and remove noise points that exceed the local threshold.

[0058] In the specific implementation of this application, RANSAC model fitting filtering is applied to the process of fitting point cloud data using a mathematical model to determine the terrain or remove points on the ground that deviate from the plane. Specifically, the RANSAC algorithm is used to fit the ground plane, calculating the specific height of each point from the plane, thereby filtering out points that deviate too far from the plane; these points are considered noise. The ground is preserved while high points are removed; significantly high points exceeding a preset threshold are deleted, and points near the ground that conform to the processed ground plane are retained.

[0059] In the process of refining sparse high-value points, cluster filtering uses DBSCAN clustering on candidate points above the ground to remove small clusters with a cluster size smaller than a preset clustering threshold.

[0060] It is understandable that model fitting filtering and clustering filtering are used for coarse filtering, RANSAC is used to separate ground and non-ground points, i.e., vehicles, equipment, etc., and DBSCAN performs semantic clustering on non-ground points to remove small-scale floating noise.

[0061] Radius filtering (based on neighborhood density) involves repeatedly calculating the number of neighboring points within a specified radius using cKDTree and deleting points with insufficient neighbor numbers (sparse outliers). cKDTree is a KD-tree spatial index data structure implemented in C in Python's scipy library. It is used to efficiently perform nearest neighbor search and range search operations in multidimensional space. In this application, it is used to perform nearest neighbor search and range search based on noisy points to determine the points that need filtering.

[0062] The first radius is 0.18m, and points with ≥10 adjacent points are retained, corresponding to medium-sized cracks; The second radius is 0.14m, and points with ≥10 neighboring points are further screened, corresponding to small cracks; The third method retains points with ≥15 adjacent points at a radius of 0.30m, corresponding to large cracks; The samples of the three radii mentioned above were all extracted from the 5,000 airport pavement crack samples.

[0063] In the process of constrained optimization, in order to achieve iterative restriction on the refitting of certain values ​​after each iteration, this application restricts the points that have already been filtered by radius. The residual high points are filtered with a radius of 0.30m, and points with ≥15 neighboring points are retained.

[0064] Understandably, radius filtering is a multi-scale filtering method used to pass through an increasing sequence with a specific radius. The purpose of two smaller radius filters is to gradually remove small particle noise attached to the real surface; the third larger radius filter is to recover and retain the real pavement features with large curvature or edges that may have been damaged by the first two filters, i.e. crack edges or joints.

[0065] Statistical filtering, based on mean and standard deviation, iteratively calculates the mean (mu) and standard deviation (sigma) of ground point height, removes outliers whose height exceeds a predefined threshold and the maximum of the three standard deviations, and dynamically adjusts the threshold through statistical characteristics.

[0066] Local contrast filtering, in the application of local contrast enhancement, aims to improve the contrast of different regions in an image and reduce or remove high-brightness points in the image. It calculates the local neighborhood of each point and estimates the local ground height using the 20th quantile of the neighborhood height, deleting isolated high points whose height difference exceeds the height difference interval.

[0067] Rasterized local thresholding filtering cleans up anomalous features in data and images by dividing the point cloud into grids, calculating local height statistics (quantiles and standard deviation) within each grid, dynamically generating local thresholds, and deleting locally sparse points with heights exceeding the threshold to avoid accidental deletion of edge regions.

[0068] Example 6: The leveling unit includes: Projection nodes: used to project the target point cloud onto a two-dimensional horizontal plane and construct a Delaunay triangulation; Calculation node: Used to calculate the residual between the elevation of the three vertices of each triangular facet in the Delaunay triangulation and the elevation of the local microplane where the corresponding triangular facet is located, as the initial value of the flatness of the core region of the triangular facet; Weighted nodes are used to collect the initial flatness values ​​of all adjacent triangle faces for each vertex of the triangulation network, and calculate the weighted flatness value of the corresponding vertex using a weight allocation function based on kernel density estimation; wherein, the weight function is related to the cosine of the angle between the triangle area and its normal vector and the average normal vector at the vertex of the triangulation network; 2D rendering node: Used to generate a continuous, high-precision 2D flatness field covering the entire target area by using bicubic spline interpolation to calculate the weighted flatness values ​​of all vertices, and then render a 2D cloud map based on this field. 3D rendering node: Used to superimpose the 2D flatness field as a height offset onto the 2D projection coordinates of the target point cloud to generate a 3D deformable cloud map that characterizes the spatial distribution features of flatness.

[0069] In this application, the projection nodes in the leveling processing unit project the filtered target point cloud data from three-dimensional space onto a two-dimensional horizontal plane, ignoring the Z-axis elevation information, and construct a Delaunay triangulation structure based on the projected two-dimensional point set. Utilizing the Delaunay triangulation's property of maximizing the minimum angle, slender triangles are avoided, ensuring mesh uniformity and transforming the discrete point cloud into a structured two-dimensional triangular mesh. In leveling calculations and local region partitioning, the geometric rationality of the pavement area covered by each triangular facet is ensured, reducing sources of error in subsequent calculations.

[0070] For each triangular facet in the triangular mesh, the computation node fits the local microplane containing that facet, preferably using the least squares method to fit the plane. The residual is calculated by comparing the actual elevation of the three vertices with the elevation of the fitted plane. This residual provides a preliminary value for the flatness of the core region of the triangular facet. Local microplane fitting reflects the true undulation of the ground in the area covered by the triangular facet, and the magnitude of the residual directly quantifies the flatness of that area. It accurately captures the flatness details of each local region, avoiding the loss of local information caused by global plane fitting. For each vertex in the triangular mesh, the weighted node collects the initial flatness values ​​of all its adjacent triangular faces. Using a weighted function estimated by kernel density, which includes weight factors positively correlated with the area of ​​the triangular facet and the cosine of the angle between the facet's normal vector and the average normal vector at the vertex, the weighted sum is used to obtain the final flatness value for that vertex. The weight allocation of kernel density estimation can be based on the geometric importance of triangular facets, i.e., facets with larger areas have a greater impact on vertices and directional consistency, i.e., a smaller angle between normal vectors indicates a flatter region and a higher weight to adjust the contribution of the initial value; by integrating the flatness information of multiple local regions in the neighborhood, a more accurate and robust vertex flatness value can be obtained, eliminating the random error of calculating a single triangular facet.

[0071] The 2D rendering node utilizes bicubic spline interpolation to interpolate the weighted smoothness values ​​of the discrete triangulation vertices into a continuous 2D smoothness field covering the entire target area. This field is then transformed into a visualized 2D contour map, where color depth represents smoothness quality. Bicubic spline interpolation possesses second-order continuous differentiability, generating smooth and high-precision continuous surfaces. The effect is to transform discrete vertex data into a continuous smoothness distribution field. The 2D contour map visually displays the spatial variations in pavement smoothness, facilitating the rapid identification of high-risk areas, such as regions with poor smoothness.

[0072] The 3D rendering node uses the values ​​in the 2D smoothness field as height offsets, with larger offsets in areas of poor smoothness. These offsets are superimposed onto the 2D projection coordinates of the target point cloud, generating a 3D deformation cloud map that intuitively reflects the spatial distribution of pavement smoothness. The height offset directly maps to the quantified smoothness value, and the 3D visualization enhances spatial perception. The effect is to transform abstract smoothness data into intuitive deformation effects in 3D space, helping users quickly understand the 3D deformation characteristics of the pavement, which can be understood as local depressions and bulges, thereby improving the efficiency and accuracy of pavement condition analysis.

[0073] The flattening processing unit of this application constructs a structured triangular network through projection nodes, calculates the initial value of local flatness through computing nodes, obtains accurate vertex values ​​by integrating neighborhood information through weighted nodes, generates a continuous flatness field through two-dimensional rendering nodes, and realizes spatial visualization through three-dimensional rendering nodes, forming a complete process from discrete point cloud to high-precision visualized flatness distribution.

[0074] Example 7: The leveling unit also includes a labeling interface that is synchronously output with the crack identification unit: wherein, The annotation interface is used to simultaneously display the 3D deformation cloud map and the corresponding real scene orthophoto; When any out-of-range area is selected by the user via a location command, an independent window is generated, and the original 3D point cloud corresponding to the selected area is highlighted and rendered through the independent window.

[0075] In this application, the pavement leveling unit is equipped with an interactive interface that can maintain temporal or spatial synchronization with the output of the crack identification unit, used to integrate and display the pavement leveling analysis results and crack identification results. This breaks down the information barriers between the two units, allowing users to simultaneously obtain the correlation information between pavement leveling and cracks on the same interface, significantly improving the efficiency of comprehensive pavement problem analysis.

[0076] The annotation interface simultaneously displays a 3D deformation cloud map from the pavement smoothing unit, reflecting the pavement smoothness distribution and the corresponding orthophoto of the real scene in that area; for example, aerial or high-precision images of the pavement. Through a spatial coordinate alignment algorithm, the analysis results of the 3D deformation cloud map are precisely matched with the pixel positions of the real orthophoto, achieving synchronous display. Users can intuitively compare the analysis conclusions of the 3D data with the actual physical condition of the pavement, quickly verifying the accuracy of the analysis results and reducing misjudgments caused by purely data-driven approaches.

[0077] Users typically use location commands, such as clicking or selecting areas exceeding the standard in the annotation interface. Once an area exceeds the standard (e.g., unevenness exceeding the standard or cracks), the system automatically generates an independent interactive window to highlight and display the corresponding original 3D point cloud. The location command triggers spatial coordinate matching of the area, the independent window isolates the target area, and the highlight rendering uses color or transparency differences to emphasize the details of the original point cloud. This allows users to focus on viewing the original data details of the area exceeding the standard and to analyze the causes of the problem in depth, such as the specific shape of the pavement depression or the 3D characteristics of the cracks.

[0078] The annotation interface of this application integrates the output of the flattening processing and crack identification unit through a synchronization mechanism, synchronously displays the three-dimensional deformation cloud map and the real orthophoto, and supports user interaction to select the out-of-standard area, triggering the display of the highlighted details of the original three-dimensional point cloud, realizing the correlation visualization and interactive analysis of multi-source data.

[0079] Example 8: The crack identification unit includes a first channel and a second channel; wherein, The first channel is a grayscale image of the reflection intensity generated by the point cloud projection of the area exceeding the standard. The second channel is a multi-scale normal rate of change feature image obtained based on cloud computing of points in the over-standard area.

[0080] In this application, the crack identification unit adopts a dual-channel parallel structure, including a first channel for extracting material reflection features and a second channel for extracting geometric edge features. The dual-channel design captures different dimensional features of the crack, such as material differences and geometric abrupt changes, avoiding the limitations of single features. It provides multi-dimensional complementary input features for the AI ​​model, significantly improving the comprehensiveness and accuracy of crack identification. The input to the first channel is a grayscale image generated by projecting the 3D point cloud of the out-of-standard area (i.e., the area with excessive flatness) onto a 2D plane, using the reflection intensity value of the point cloud as the grayscale value. The reflection intensity of the point cloud is closely related to the pavement material; for example, changes in material or surface condition at the crack will cause abnormal reflection intensity. Projecting as a grayscale image allows for feature extraction using mature image analysis techniques. This accurately captures the material reflection differences in the crack area, providing an intuitive visual feature basis for crack identification and reducing the blind spot of the AI ​​model's reliance on geometric features. The input to the second channel is the rate of change of the normal vector of each point in the out-of-standard area point cloud at multiple scales, and the rate of change value is mapped to a feature image. The normal vector of a 3D point at the edge of a crack undergoes a sudden change, meaning the normal vector has a high rate of change. Multi-scale calculations can adapt to cracks of different widths and depths. It effectively captures the geometric edge features of cracks, supplementing crack scenarios with geometric abrupt changes but no material differences that cannot be covered by reflection intensity features, and significantly improving recognition robustness.

[0081] The crack recognition unit of this application uses a dual-channel structure to extract features of the out-of-standard area from two dimensions: material reflection (first channel) and geometric edge (second channel), providing multi-source complementary input data for the subsequent AI recognition module.

[0082] Example 9: The U-Net model includes parallel encoding branch channels and decoding branch channels; wherein, The encoded branch channel is used to extract spatial context features of the reflectance intensity image and geometric texture features of the normal change rate image; The decoding branch channel is used to adaptively weight and fuse spatial context features and geometric texture features through a cross-attention mechanism, and outputs a crack pixel-level prediction mask that represents reflection properties and three-dimensional geometric characteristics.

[0083] In this application, the U-Net model employs a parallel encoding and decoding branch structure. The encoding branch focuses on feature extraction, while the decoding branch focuses on feature fusion and crack prediction mask output. This parallel branch structure allows for the simultaneous processing of two heterogeneous features: reflection intensity images and normal change rate images. The clear division of labor between the encoding and decoding branches avoids mutual interference between features, improving feature processing efficiency and providing an independent and high-quality feature foundation for subsequent accurate fusion. This reduces information loss caused by processing multiple features in a single branch. The encoding branch is used to extract spatial context features from the reflection intensity image and geometric texture features from the normal change rate image, respectively. Through typical U-Net encoding operations such as convolution and downsampling, the spatial context features of the reflection intensity image are captured hierarchically, such as the global distribution relationship between the crack region and the background. It also captures the geometric texture features of the normal change rate image, such as the details of abrupt normal changes at the crack edges. These two key feature types are selectively extracted, covering the global spatial information and local geometric details required for crack recognition, providing comprehensive feature support for the fusion stage. The decoding branch channel utilizes a cross-attention mechanism to adaptively weight and fuse spatial context features and geometric texture features, outputting a pixel-level prediction mask for cracks that represents reflection properties and 3D geometric characteristics. The cross-attention mechanism dynamically assigns weights by calculating the correlation between the two types of features; for example, geometric texture features in the crack edge region have higher weights, while spatial context features in the background region have higher weights, achieving adaptive feature fusion. The pixel-level prediction mask restores spatial resolution through upsampling, accurately marking each pixel as a crack. This achieves complementary advantages between the two types of features, significantly improving the accuracy and robustness of crack recognition. The pixel-level output meets the fine-grained localization requirements for airport pavement crack detection.

[0084] The U-Net model in this application adopts a parallel encoder-decoder branch structure. The encoder branch extracts the spatial context features of the reflection intensity image and the geometric texture features of the normal change rate image, respectively. The decoder branch adaptively fuses the two types of features through a cross-attention mechanism, and finally outputs a pixel-level crack prediction mask.

[0085] Example 10: The calculation of crack geometry parameters includes width calculation, depth calculation, and volume calculation; among which, When calculating the geometric parameters of the crack, all three-dimensional point clouds belonging to the crack are extracted from the point cloud data in advance based on the crack pixel-level prediction mask; When calculating the width, the crack skeleton line is segmented according to the three-dimensional point set. In the normal plane of each segment point, the crack point cloud is projected onto the corresponding plane. The distribution range of the projected point cloud in the direction perpendicular to the skeleton line is calculated as the local width of the projected point cloud. The weighted average of the local widths of all segment points is taken, and the weight is the reciprocal of the curvature of the projected point cloud. During depth calculation, based on the 3D point cloud, a reference pavement point cloud without cracks is selected around the crack point cloud. The local pavement surfaces of the crack area and the reference area are fitted by the moving least squares method, and the signed distance from the crack point to the reference surface is calculated as the depth. When calculating the volume, the depth value of each crack point is multiplied by the equivalent area it occupies in the local triangular mesh based on the three-dimensional point set, and the sum is used to determine the estimated volume of the crack.

[0086] In this application, the quantification process of crack geometric parameters encompasses the calculation of three core physical parameters: crack width, crack depth, and crack volume. These three parameters are key indicators for assessing the severity of airport pavement cracks (width reflects the lateral expansion range, depth reflects the longitudinal damage level, and volume reflects the overall scale of damage). This comprehensive approach covers the geometric characteristics of cracks, providing multi-dimensional data support for pavement maintenance priority determination and repair plan formulation. Utilizing the crack pixel-level prediction mask output by the U-Net model in claim 9, the pixel positions marked as cracks in the two-dimensional image are mapped back to the original three-dimensional point cloud data, filtering out the three-dimensional point set belonging only to the crack region. The prediction mask accurately marks the two-dimensional spatial location of the crack, which can be associated with the three-dimensional point cloud through coordinate mapping. Focusing on the crack region for parameter calculation eliminates interference from irrelevant background point clouds, reducing computational load and improving the accuracy of parameter calculation.

[0087] In specific calculations, Width is calculated by measuring the horizontal distance of the point cloud data; depth is calculated by measuring the vertical distance of the point cloud data; volume is calculated based on both width and depth.

[0088] The width calculation steps are as follows: Divide the crack into several segments along the crack skeleton line; construct a normal plane for each segment point, i.e., perpendicular to the tangent direction of the skeleton line; project the crack point cloud within the segment onto this normal plane; take the range of coordinate distribution of the projected points in the direction perpendicular to the skeleton line as the local width; use the inverse of the curvature of each segment point cloud as the weight to perform a weighted average of the local widths to obtain the overall crack width; the crack skeleton line represents the central axis of the crack, and the normal plane ensures that the width measurement direction is perpendicular to the crack extension direction; the inverse curvature weight makes the curvature smaller, that is: the straighter segments of the crack contribute more, and the width is more stable; accurately calculate the crack width, adapt to crack bending changes, and the weighted average reduces the influence of local anomalies on the overall width.

[0089] The depth calculation steps are as follows: Select a pavement point cloud without cracks around the crack point cloud as a reference; fit the local surface of the crack area and the normal pavement surface of the reference area respectively by moving least squares method; calculate the signed distance from each crack point to the reference surface (negative value indicates depression, positive value indicates convexity) as the crack depth. The moving least squares method can fit local non-uniform surfaces, with the reference surface reflecting the undulation of the normal pavement; signed distances distinguish between depressions and bulges; it accurately calculates crack depth, eliminates the interference of overall pavement undulation on depth measurement, and accurately reflects the actual damage degree of cracks.

[0090] The volume calculation steps are as follows: construct a local triangular mesh from the 3D point cloud of the crack region; calculate the equivalent area occupied by each crack point in the triangular mesh; multiply the depth value of each crack point by its equivalent area; sum the products of all points to obtain the estimated volume of the crack; the volume is the integral of the depth over the crack area, and the equivalent area represents the spatial contribution of the point cloud in the mesh; quantify the overall failure volume of the crack and assess its impact on the pavement structural strength.

[0091] This application first extracts a precise 3D point set of cracks by using a crack pixel-level prediction mask. Then, for the three key parameters of crack width, depth, and volume, targeted methods are used to perform quantitative calculations, namely, weighted segmentation of skeleton lines, moving least squares surface fitting, and accumulation of triangular mesh equivalent surfaces.

[0092] In practical implementation, this application uses the classic U-Net model as its basic framework. It progressively extracts image features through convolution and pooling operations to capture abstract features of cracks, such as texture and shape. Then, it gradually restores the image resolution through transposed convolution and skip connections, mapping the abstract features back to pixel-level crack segmentation results. The model output is logits without sigmoid activation, and end-to-end training is performed using BCEWithLogitsLoss to improve numerical stability.

[0093] The refined preprocessing of crack masks addresses the common use of red markers in crack masks by combining multiple color spaces for accurate mask extraction: RGB space (red channel threshold) and HSV space (red hue range) are used simultaneously to detect red regions, and the results are merged using bitwise_or to ensure the integrity and accuracy of the crack mask. This solves the problem of potential missed or false detections in a single color space, providing a more reliable supervision signal for the model.

[0094] For loss functions designed for imbalanced data, since cracks typically account for a very small percentage of images (few positive samples), F1OptimizedLoss is employed: based on BCEWithLogitsLoss, it sets extremely high weights for positive samples (pos_weight=100.0), forcing the model to focus more on the small number of cracked regions. This avoids the model from being biased towards outputting crack-free results due to an excessively high proportion of negative samples (non-cracked regions).

[0095] Based on the crack analysis results, during actual implementation, a structured report template will be automatically invoked to generate a comprehensive analysis report integrating data, charts, and a 3D visualization model with a single click. Simultaneously, based on the analysis results and preset rules, decision support information such as maintenance priority suggestions and risk area identification will be automatically output. The raw data collected during this cycle, the processed clean point cloud, all intermediate analysis results, and the final report will be systematically stored in the central database along with project metadata, forming a traceable and reusable digital archive of pavement data.

[0096] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An airport pavement management system based on AI and 3D scanning, characterized in that, The system includes a 3D laser scanning device and an AI server; The three-dimensional laser scanning equipment is deployed on the airport pavement to collect point cloud data of the target area; The AI ​​server includes a filtering unit, a leveling unit, and a crack recognition unit. The filtering unit is used to perform multidimensional filtering on point cloud data to generate a target point cloud with noise removed; the multidimensional filtering includes: RANSAC model fitting filtering, DBSCAN clustering filtering, cubic radius filtering, statistical filtering, local contrast filtering and rasterized local threshold filtering. The leveling processing unit is used to generate a leveling cloud map of the target area based on the target point cloud and to locate the areas exceeding the standard; the leveling cloud map includes a two-dimensional cloud map and a three-dimensional cloud map. The crack identification unit is used to input the out-of-specification area into the U-Net model, extract the crack mask through RGB and HSV space, and output the crack geometric parameters based on the crack mask.

2. The airport pavement management system based on AI and 3D scanning as described in claim 1, characterized in that, The three-dimensional laser scanning device includes: Acquire a regional image of the target area and generate a basic scan path; the basic scan path includes the airport runway, taxiway, and apron. Based on the basic scanning path, the acquisition accuracy of each scanning area is determined, and the first scanning strategy of the 3D laser scanning device is generated; wherein the acquisition accuracy is not less than the millimeter level. According to the first scanning strategy, the device attitude and motion trajectory data at each moment are determined; wherein, the device attitude is used to determine the local point density, and the motion trajectory data is used to determine the reflection intensity distribution. Determine whether there are abnormal point cloud data where the local point density is lower than a preset first threshold and the reflection intensity variance is higher than a preset second threshold; wherein, abnormal point cloud data is a region with degraded scan quality; wherein, the preset first threshold represents the minimum point cloud density and the preset second threshold represents the minimum reflection intensity threshold; Based on abnormal point cloud data, an attenuation model for laser scanning in heterogeneous pavement media is established, and the optimal compensation scanning angle and compensation scanning line spacing are calculated. A second scanning strategy is generated based on the compensated scanning angle and the compensated scanning line spacing.

3. The airport pavement management system based on AI and 3D scanning as described in claim 1, characterized in that, The three-dimensional laser scanning equipment is also equipped with synchronous multispectral environmental sensors; among which... Multispectral environmental sensors are used to collect real-time environmental parameters of the area; these real-time environmental parameters include: ambient light intensity, surface temperature, and atmospheric visibility. Based on real-time environmental parameters, deploy high-speed vibration sensors and acoustic sensors that operate synchronously with the 3D laser scanning equipment; The characteristics of abnormal high-frequency vibrations were determined using high-speed vibration sensors. Based on acoustic sensors, determine the first acoustic characteristics of the three-dimensional laser scanning equipment in operation and the second acoustic characteristics of the environmental background; Based on real-time environmental parameters, abnormal high-frequency vibration characteristics, first acoustic characteristics, and second acoustic characteristics, a multi-dimensional compensation model is constructed to perform dynamic calibration when each frame of point cloud data is generated.

4. An airport pavement management system based on AI and 3D scanning as described in claim 1, characterized in that, The filtering unit includes an index node and a first filtering decision node; The index node is used to divide local point cloud clusters based on point cloud data and calculate the feature vector of each local point cloud cluster; the feature vector includes curvature variance, normal vector consistency and neighborhood density gradient. The first filtering decision node is used to output a filtering strategy based on the feature vector and through a preset lightweight decision model. The filtering strategy includes the optimal filtering algorithm, the filtering execution order of the optimal filtering algorithm and the combination of initial parameters. The filtering strategy is used to remove point cloud noise from the point cloud data in order to determine the target point cloud.

5. An airport pavement management system based on AI and 3D scanning as described in claim 4, characterized in that, The optimal filtering algorithm includes: RANSAC model fitting filter is used to fit the ground plane in point cloud data and remove noise points that deviate too far from the plane. DBSCAN clustering filter is used to cluster candidate points above the ground in point cloud data and remove small clusters with a size smaller than a preset height threshold. Three-stage radius filtering includes: the first stage retains points with ≥10 neighboring points with a radius of 0.18m; the second stage retains points with ≥10 neighboring points with a radius of 0.14m; and the third stage retains points with ≥15 neighboring points with a radius of 0.30m. Statistical filtering is used to iteratively calculate the mean and standard deviation of ground point heights and remove outliers whose heights exceed a preset maximum value. Local contrast filtering is used to determine the local ground height by using a preset quantile of the neighborhood height, and to delete isolated high points with height differences exceeding a preset value; Rasterized local threshold filtering is used to divide point cloud data into grids, dynamically generate local thresholds, and remove noise points that exceed the local threshold.

6. An airport pavement management system based on AI and 3D scanning as described in claim 1, characterized in that, The leveling unit includes: Projection nodes: used to project the target point cloud onto a two-dimensional horizontal plane and construct a Delaunay triangulation; Calculation node: Used to calculate the residual between the elevation of the three vertices of each triangular facet in the Delaunay triangulation and the elevation of the local microplane where the corresponding triangular facet is located, as the initial value of the flatness of the core region of the triangular facet; Weighted nodes are used to collect the initial flatness values ​​of all adjacent triangle faces for each vertex of the triangulation network, and calculate the weighted flatness value of the corresponding vertex using a weight allocation function based on kernel density estimation; wherein, the weight function is related to the cosine of the angle between the triangle area and its normal vector and the average normal vector at the vertex of the triangulation network; 2D rendering node: Used to generate a continuous, high-precision 2D flatness field covering the entire target area by using bicubic spline interpolation to calculate the weighted flatness values ​​of all vertices, and then render a 2D cloud map based on this field. 3D rendering node: Used to superimpose the 2D flatness field as a height offset onto the 2D projection coordinates of the target point cloud to generate a 3D deformable cloud map that characterizes the spatial distribution features of flatness.

7. An airport pavement management system based on AI and 3D scanning as described in claim 6, characterized in that, The leveling unit also includes a labeling interface that is synchronously output by the crack identification unit: wherein, The annotation interface is used to simultaneously display the 3D deformation cloud map and the corresponding real scene orthophoto; When any out-of-range area is selected by the user via a location command, an independent window is generated, and the original 3D point cloud corresponding to the selected area is highlighted and rendered through the independent window.

8. An airport pavement management system based on AI and 3D scanning as described in claim 1, characterized in that, The crack identification unit includes a first channel and a second channel; wherein... The first channel is a grayscale image of the reflection intensity generated by the point cloud projection of the area exceeding the standard. The second channel is a multi-scale normal rate of change feature image obtained based on cloud computing of points in the over-standard area.

9. An airport pavement management system based on AI and 3D scanning as described in claim 8, characterized in that, The U-Net model includes parallel encoding branch channels and decoding branch channels; wherein, The encoded branch channel is used to extract spatial context features of the reflectance intensity image and geometric texture features of the normal change rate image; The decoding branch channel is used to adaptively weight and fuse spatial context features and geometric texture features through a cross-attention mechanism, and outputs a crack pixel-level prediction mask that represents reflection properties and three-dimensional geometric characteristics.

10. An airport pavement management system based on AI and 3D scanning as described in claim 9, characterized in that, The calculation of the crack geometry parameters includes width calculation, depth calculation, and volume calculation; among which, When calculating the geometric parameters of the crack, all three-dimensional point clouds belonging to the crack are extracted from the point cloud data in advance based on the crack pixel-level prediction mask; When calculating the width, the crack skeleton line is segmented according to the three-dimensional point set. In the normal plane of each segment point, the crack point cloud is projected onto the corresponding plane. The distribution range of the projected point cloud in the direction perpendicular to the skeleton line is calculated as the local width of the projected point cloud. The weighted average of the local widths of all segment points is taken, and the weight is the reciprocal of the curvature of the projected point cloud. During depth calculation, based on the 3D point cloud, a reference pavement point cloud without cracks is selected around the crack point cloud. The local pavement surfaces of the crack area and the reference area are fitted by the moving least squares method, and the signed distance from the crack point to the reference surface is calculated as the depth. When calculating the volume, the depth value of each crack point is multiplied by the equivalent area it occupies in the local triangular mesh based on the three-dimensional point set, and the sum is used to determine the estimated volume of the crack.

Images