Road vulnerable accessory facility detection and evaluation method based on mobile measurement and ai analysis

Abstract

The application discloses a road vulnerable accessory facility detection and evaluation method based on mobile measurement and AI analysis, and belongs to the technical field of road intelligent operation and maintenance. The method comprises the following steps: synchronously collecting point cloud data, image data and pose data through a vehicle-mounted mobile measurement system; constructing a dynamic elevation datum based on road fluctuation characteristics, and realizing separation of target point cloud and environment point cloud by adopting a hierarchical clustering and threshold self-adaptive adjustment strategy; establishing a typical accessory facility structure feature library, and extracting complete geometric parameters of vertical structures by density clustering and feature matching; constructing a digital surface model based on multi-temporal point clouds and performing differential operation, and determining a four-level state evaluation grade based on comparison of a change value and a maintenance standard.

Description

Technical Field

[0001] This invention belongs to the field of intelligent road operation and maintenance technology, specifically involving a method for detecting and evaluating vulnerable road ancillary facilities based on mobile measurement and AI analysis. It is applicable to asset statistics, geometric detection, change analysis, and condition evaluation of road ancillary facilities such as sound barriers, gantries, and crash barriers. Background Technology

[0002] As a crucial component of road infrastructure, expressway ancillary facilities are constantly exposed to the complex outdoor environment. Affected by factors such as climate change, vehicle vibration, and external loads, they are prone to safety risks such as column tilting, slab damage, and structural deformation, directly threatening road traffic safety. To promptly identify and eliminate these safety hazards, regular inspections and condition assessments of road ancillary facilities are necessary.

[0003] Existing road infrastructure inspection technologies mainly include manual inspection and vehicle-mounted measurement. Manual inspection requires significant manpower and resources, has low efficiency and limited coverage, and its results are heavily influenced by subjective factors, making high-precision quantitative evaluation difficult. Furthermore, it cannot promptly detect subtle deformations and hidden defects in the infrastructure. Traditional fixed-point measurement methods can only inspect specific key areas, failing to achieve large-scale continuous data collection, resulting in insufficient data integrity and making it difficult to support the needs of full lifecycle management of infrastructure.

[0004] Chinese invention patent application CN119599986A discloses a method and system for detecting defects in road ancillary facilities based on deep learning. This technical solution utilizes the PointRend deep convolutional network model to achieve image target detection, classification, and accurate segmentation of guardrails and anti-glare panels. It extracts the guardrail mask region and performs boundary traversal search, combining least-squares line fitting and the RANSAC algorithm to select the optimal guardrail boundary points. The distance from the candidate edge points to the fitted boundary is used to determine whether fractures or bends exist. This technical solution has the following shortcomings: First, this method mainly relies on two-dimensional images for defect identification, lacking the ability to accurately measure three-dimensional geometric information and failing to obtain key geometric parameters such as the verticality and flatness of the facilities. Second, this method can only identify the current state of the facilities, lacking the ability to compare and analyze multi-temporal data, and cannot quantify the deformation evolution trend of the facilities. Third, this method targets only guardrails and anti-glare panels, not covering more types of road ancillary facilities such as sound barriers and gantries. Fourth, the point cloud filtering algorithm of this method does not consider the influence of road undulations, limiting the filtering accuracy in complex road scenarios.

[0005] While existing vehicle-mounted measurement technologies attempt to collect data using vehicle-mounted LiDAR or photogrammetry, they still have significant limitations: point cloud data is highly discrete and large, susceptible to interference from road undulations and the surrounding environment, making it difficult for traditional filtering algorithms to adaptively separate effective data; geometric information extraction for vertical structures is incomplete, resulting in low classification accuracy; and there is a lack of clear change analysis indicators, making it difficult to accurately identify facility defects and deformations, and thus unable to achieve quantitative condition assessment. Furthermore, existing technologies have not formed a complete technical system from data acquisition to intelligent analysis, then to detection and evaluation, and finally to maintenance and disposal. The detection results are disconnected from operational practices, failing to meet the actual needs of digital management and maintenance of road facilities throughout their entire lifecycle. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for detecting and evaluating vulnerable road ancillary facilities based on mobile measurement and AI analysis. By constructing an adaptive point cloud filtering algorithm, an upright structure integrity extraction algorithm, and a change detection and evaluation model based on DSM differential, this method achieves high-efficiency data acquisition, high-precision geometric detection, and quantitative condition evaluation of road ancillary facilities.

[0007] The technical solution adopted in this invention is as follows: A method for detecting and evaluating vulnerable road ancillary facilities based on mobile measurement and AI analysis includes the following steps: Step S1, Multi-source sensor data synchronous acquisition: Data on road ancillary facilities is acquired by the vehicle-mounted mobile measurement system while the vehicle is in motion. The vehicle-mounted mobile measurement system includes a 3D laser scanner, a high-resolution linear camera and a POS system. The 3D laser scanner continuously scans the road ancillary facilities to obtain raw point cloud data, the high-resolution linear camera synchronously acquires panoramic image data, and the POS system records the vehicle pose data in real time and provides a spatiotemporal reference for the raw point cloud data and panoramic image data.

[0008] Step S2, adaptive filtering of point cloud data: coordinate transformation and noise reduction preprocessing are performed on the original point cloud data. A dynamic elevation reference surface is constructed based on the road undulation characteristics. Hierarchical clustering and threshold adaptive adjustment strategy are used to filter the preprocessed point cloud. The filtering threshold is automatically adjusted according to the elevation difference and spatial distribution density of adjacent point clouds to separate the target point cloud of road ancillary facilities from the environmental point cloud of vehicles and the surrounding environment.

[0009] Step S3, complete extraction of vertical structure facilities: extract the normal vector and elevation change rate feature parameters of the target point cloud, identify the feature point set of the vertical structure, use density-based clustering algorithm to segment the target area according to the preset facility structure size threshold, establish a structural feature library of typical auxiliary facilities, and achieve target classification through feature matching, and extract geometric parameters including height, verticality, spacing and flatness.

[0010] Step S4, Change Detection and Quantitative Evaluation Based on DSM Differential: The target point cloud collected in the current period and the point cloud collected in the past are rasterized to construct a digital surface model. Differential operation is performed on the current digital surface model and the historical digital surface model to obtain change detection results. The change detection results include the location, range and amount of change of the facility defect area and the deformation area. The condition evaluation level of the facility is determined based on the comparison results of the amount of change and the preset maintenance standard threshold.

[0011] Preferably, in step S1, the vehicle speed is 30km / h to 80km / h, the point cloud density of the original point cloud data is not lower than a preset point cloud density threshold, the field of view overlap rate between the 3D laser scanner and the high-resolution linear camera is not lower than 80%, and the sampling frequency of the POS system is not lower than 100Hz.

[0012] Preferably, in step S2, constructing the dynamic elevation datum includes: dividing the preprocessed point cloud into continuous segmented regions along the longitudinal direction of the road; performing planar fitting on the road surface point cloud in each segmented region to obtain a local datum; and smoothly connecting adjacent local datums to form the dynamic elevation datum. Automatically adjusting the filtering threshold includes: calculating the elevation difference between each point in the point cloud and the dynamic elevation datum; when the elevation difference exceeds a preset elevation difference threshold, the point is classified as the target point cloud; and the accuracy of the filtering process is not less than 95%.

[0013] Preferably, in step S3, typical auxiliary facilities include sound barriers, gantry frames, and crash barriers. The testing standards for geometric parameters include: the allowable deviation of the verticality of the sound barrier columns is no more than 0.5%, the allowable deviation of the flatness of the sound barrier panels is no more than 5 mm / m, the allowable deviation of the horizontality of the gantry frame beams is no more than 0.3%, and the allowable deviation of the flatness of the crash barrier surface is no more than 3 mm / m.

[0014] Preferably, in step S4, the raster resolution of the digital surface model is not greater than the preset DSM raster resolution threshold, and the differential operation is performed by pixel-by-pixel subtraction after aligning the coordinate references of the current digital surface model and the historical digital surface model using the hierarchical clustering difference method.

[0015] Preferably, in step S4, the status evaluation level includes four levels: when the change is not greater than the first change threshold, it is judged as a good level; when the change is greater than the first change threshold but not greater than the second change threshold, it is judged as a mild disease level; when the change is greater than the second change threshold but not greater than the third change threshold, it is judged as a moderate disease level; and when the change is greater than the third change threshold, it is judged as a severe disease level.

[0016] The beneficial effects of this invention are as follows: First, the detection efficiency and coverage are significantly improved. This invention adopts a vehicle-mounted mobile measurement mode, which does not require closed roads and can achieve large-scale continuous detection at normal driving speeds. The detection efficiency is more than 50 times higher than that of traditional manual inspection, and the coverage reaches 100%, solving the problems of low efficiency and incomplete coverage of traditional methods.

[0017] Second, the point cloud filtering algorithm has high accuracy and strong adaptability. The adaptive hierarchical clustering filtering algorithm proposed in this invention can dynamically adjust the filtering threshold according to the road undulation characteristics, effectively separating the target point cloud from the environmental point cloud, with a filtering accuracy of no less than 95%, overcoming the shortcomings of traditional filtering algorithms in terms of limited accuracy in complex road scenarios.

[0018] Third, the completeness of the extracted vertical structures is high. This invention establishes a structural feature library of typical auxiliary facilities and combines it with a density clustering algorithm to achieve accurate classification and complete geometric information extraction of vertical structural facilities such as sound barriers and gantry frames. The extraction completeness is no less than 98%, and the geometric measurement error is controlled within ±3mm.

[0019] Fourth, the change detection has a high degree of quantification. This invention establishes a change detection model based on multi-temporal DSM differential analysis, which can capture millimeter-level deformation, and realizes quantitative evaluation of facility service performance through a four-level condition evaluation model, with objective and reliable evaluation results.

[0020] Fifth, the technology system is complete and closed-loop. This invention constructs a complete technology system from data collection to intelligent analysis and then to detection and evaluation. The detection results can be directly connected to the digital management and maintenance platform to support scientific management and maintenance decisions for road ancillary facilities throughout their entire life cycle. Attached Figure Description

[0021] Figure 1 This is an overall flowchart of the method for detecting and evaluating vulnerable road ancillary facilities provided in this embodiment of the invention; Figure 2 This is a flowchart of the adaptive filtering process for point cloud data provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the complete extraction process of upright structural facilities provided in this embodiment of the invention; Figure 4 This is a flowchart of DSM differential change detection and quantitative evaluation provided in the embodiments of the present invention. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0023] This invention provides a method for detecting and evaluating vulnerable road ancillary facilities based on mobile measurement and AI analysis, such as... Figure 1As shown, the method includes four core steps: synchronous acquisition of multi-source sensor data, adaptive filtering of point cloud data, complete extraction of vertical structure facilities, and change detection and quantitative evaluation based on DSM differential. The following is a detailed description of each step.

[0024] Step S1: Synchronous acquisition of data from multiple sensor sources The multi-source sensor data synchronous acquisition process utilizes a vehicle-mounted mobile measurement system to achieve full coverage data collection of roadside facilities while the vehicle is in normal operation. The hardware components of the vehicle-mounted mobile measurement system include four main parts: a 3D laser scanner, a high-resolution linear camera, a POS system, and a data storage and transmission module. These sensors form an organic whole through precise calibration and time synchronization mechanisms, ensuring spatial consistency and temporal synchronization of the acquired data.

[0025] A 3D laser scanner serves as the core sensor for acquiring 3D point cloud data of roadside facilities. In a preferred embodiment of the invention, the 3D laser scanner employs a phase-detection laser scanner or a pulsed laser scanner, with a measurement distance range of 0.5m to 200m, a point cloud density of no less than 50 points / cm², and an angular resolution of no more than 0.001°. The 3D laser scanner is mounted on a bracket on the top of a vehicle, and its scanning range covers the area of ​​roadside facilities on both sides of the road. During the scanning process, the 3D laser scanner continuously emits laser pulses at a preset scanning frequency and receives reflected signals. It calculates the 3D coordinates of the target point based on the time of flight or phase difference, and simultaneously records the reflection intensity information of each point. The number of scan lines of the 3D laser scanner directly affects the vertical distribution density of the point cloud. In a preferred embodiment of the invention, a laser scanner with no less than 32 scan lines is selected to ensure complete coverage of vertical structures. Furthermore, the 3D laser scanner must also have the capability of multiple echo recording to extract effective point clouds of roadside facilities in scenarios with vegetation obstruction.

[0026] A high-resolution linear camera is used to simultaneously acquire panoramic image data, providing data support for subsequent point cloud-image registration and texture information supplementation. In a preferred embodiment of the invention, the high-resolution linear camera has a resolution of no less than 50 megapixels and uses a linear scanning imaging method to acquire high-resolution panoramic images of both sides of the road. The high-resolution linear camera and the 3D laser scanner are rigidly connected and fixed on the same bracket, and their relative position and orientation are obtained and maintained stably through calibration. The field-of-view overlap rate between the 3D laser scanner and the high-resolution linear camera is no less than 80% to ensure effective registration of the point cloud and image data. The exposure time and gain parameters of the high-resolution linear camera need to be adaptively adjusted according to the ambient lighting conditions to obtain clear texture images. In tunnels or areas with insufficient lighting, an onboard supplementary lighting system can be used to assist imaging.

[0027] The POS system consists of an inertial navigation unit and a GNSS receiver, used to record the vehicle's position, attitude, and time information in real time, providing a high-precision spatiotemporal reference for point cloud data and image data. In a preferred embodiment of the invention, the POS system has a positioning accuracy better than ±2cm, an attitude measurement accuracy better than 0.02°, and a sampling frequency of not less than 100Hz. The inertial navigation unit measures the vehicle's three-axis acceleration and three-axis angular velocity in real time, while the GNSS receiver synchronously acquires the vehicle's position coordinates and timing information, achieving high-precision pose calculation through a combined navigation algorithm. The output data of the POS system includes the vehicle's position coordinates in the geodetic coordinate system, attitude angles (heading angle, pitch angle, roll angle), and corresponding timestamps. To improve positioning accuracy, the POS system supports a real-time dynamic differential positioning mode, achieving centimeter-level positioning by receiving correction data from ground reference stations or network differential services. In areas where GNSS signals are blocked (such as under overpasses or at tunnel entrances), the POS system relies on the inertial navigation unit for short-term estimation and performs pose correction through a combined navigation algorithm after signal recovery.

[0028] The data storage and transmission module is responsible for storing the raw data collected by each sensor in real time, ensuring data integrity and time synchronization. In a preferred embodiment of the invention, the data storage module uses a high-speed solid-state drive with a storage capacity of no less than 2TB and a data write speed of no less than 500MB / s. Each sensor achieves time synchronization through a unified trigger signal, with a time synchronization error of no more than 1ms. The data storage module also needs to be equipped with a data verification mechanism to monitor the integrity of the stored data in real time. When data loss or corruption is detected, an alarm is triggered promptly and the abnormal location is recorded for subsequent retesting.

[0029] During data acquisition, the vehicle travels normally along the target road at a speed of 30 km / h to 80 km / h, simultaneously activating various sensors to collect data. A 3D laser scanner continuously scans roadside facilities and the surrounding environment to acquire high-density 3D point cloud data; a high-resolution linear camera simultaneously acquires panoramic images; and a POS system records vehicle pose data in real time. The data storage module stores the above data in a time-stamped manner, forming a raw dataset containing point cloud data, image data, and pose data. The vehicle speed range for data acquisition is determined comprehensively based on the point cloud density requirements and facility detection accuracy requirements. Lower speeds can acquire higher density point cloud data but have lower acquisition efficiency, while higher speeds have higher acquisition efficiency but may result in insufficient point cloud density. In a preferred embodiment of the invention, for sound barrier facilities requiring millimeter-level geometric detection, it is recommended to control the vehicle speed below 50 km / h; for gantry facilities requiring only macroscopic deformation detection, the vehicle speed can be increased to 80 km / h.

[0030] The output data for this step are raw point cloud data D01, panoramic image data D02, and pose data D03. The raw point cloud data is in the format of an N×4 matrix, where N is the number of points, and each point contains four attributes: three-dimensional coordinates (X, Y, Z) and reflection intensity I. The panoramic image data is in the format of an RGB image sequence, with each image bearing an acquisition timestamp. The pose data is in the format of a time series containing position coordinates, attitude angles, and timestamps. These three types of data are correlated through timestamps, laying the foundation for subsequent data registration and fusion processing.

[0031] Step S2, adaptive filtering processing of point cloud data The adaptive filtering process for point cloud data preprocesses and filters the raw point cloud data to achieve efficient separation of the target point cloud of road ancillary facilities from the environmental point cloud. For example... Figure 2 As shown, this step includes three sub-processes: coordinate transformation and denoising preprocessing, dynamic elevation datum construction, and hierarchical clustering filtering.

[0032] The coordinate transformation and denoising preprocessing sub-process first transforms the raw point cloud data from the vehicle coordinate system to the geodetic coordinate system, and then removes isolated and noisy points from the point cloud. The core of the coordinate transformation is to utilize the pose data provided by the POS system to transform each laser point step-by-step from the scanner's own coordinate system to the carrier coordinate system, the navigation coordinate system, and finally to the geodetic coordinate system. (Point setting...) The coordinates in the scanner coordinate system are: The mounting matrix of the scanner relative to the carrier is as follows: Translation vector The attitude matrix of the vehicle in the navigation coordinate system is (Calculated from heading angle, pitch angle, and roll angle), the transformation matrix from the navigation coordinate system to the geodetic coordinate system is: The position of the carrier in the geodetic coordinate system is Then point coordinates in the geodetic coordinate system The calculation is performed as follows: First, the point coordinates in the scanner coordinate system are transformed to the carrier coordinate system. Then, the point coordinates in the carrier coordinate system are transformed to the navigation coordinate system. Next, the point coordinates in the navigation coordinate system are transformed to the geodetic coordinate system. Finally, the carrier position is superimposed to obtain the final coordinates.

[0033] The denoising preprocessing employs statistical filtering to remove isolated and noisy points. For each point, the average distance to all points in its neighborhood is calculated. If this average distance exceeds the global average distance plus a preset standard deviation, the point is identified as a noise point and removed. In a preferred embodiment of the invention, the number of neighborhood points is set to 50, and the determination factor is set to 2.0.

[0034] The dynamic elevation datum construction sub-process builds a dynamic reference plane that varies longitudinally with the road's undulation characteristics, providing an adaptive elevation judgment benchmark for subsequent filtering. Traditional filtering methods typically use fixed elevation thresholds, which are prone to misjudgment when the road has variations in longitudinal and transverse slopes. The dynamic elevation datum construction algorithm proposed in this invention is as follows: First, the preprocessed point cloud is divided into continuous segmented regions along the longitudinal direction of the road. In a preferred embodiment of the invention, the segment length is set to 10m to 20m, and an overlap region of 2m to 5m is set between adjacent segments to ensure smooth connection of the reference plane. During segmentation, the road centerline is determined based on the vehicle trajectory information provided by the POS system, and the point cloud is segmented longitudinally using the centerline as a reference.

[0035] Secondly, the road surface point cloud is identified and its plane is fitted within each segmented region. The identification of the road surface point cloud is based on elevation statistical features. Within each segmented region, the point cloud is sorted by elevation value, and points with elevation values ​​within the lowest 5% to 15% range are selected as the candidate set of road surface point clouds. Then, the RANSAC algorithm is used to fit the road surface plane from the candidate set to obtain the plane equation of the local reference plane. Let the plane equation of the local reference plane be: , in, For the first The plane normal vectors of each segment. The term is a plane constant and satisfies the normalization condition. .

[0036] Finally, adjacent local reference surfaces are smoothly connected to form a dynamic elevation reference surface. In the overlapping areas of adjacent segments, a weighted average method is used to achieve a smooth transition of the reference surface. (Points are set.) Located in the The segment and the first Within the overlapping region of the segment, its distance to the segment... The distance between the centers of each segment is , to the The distance between the centers of each segment is Then point Corresponding dynamic reference surface elevation value Calculated according to the following adaptive elevation datum fusion formula: , in, Point plane coordinates Substitute the first The elevation values ​​obtained from the segmented datum surface equations are calculated using the following method: ; Point Substituting the plane coordinates into the first Elevation values ​​obtained by calculating the equations of segmented datum surfaces; and Points Distance weighting factors to the centers of the two segments.

[0037] The hierarchical clustering filtering subprocess employs a two-level clustering strategy to separate the target point cloud from the environment point cloud. The first-level clustering divides the point cloud into a ground layer and a non-ground layer based on the point cloud's elevation attributes. For each point, the elevation difference between it and the dynamic elevation datum is calculated. : , in, This is the actual elevation value of that point. This is the dynamic reference surface elevation value corresponding to that point. When Less than the preset ground layer threshold When, this point is classified as a ground-level point cloud; when Greater than or equal to At that time, the point is classified as a non-ground layer point cloud. In a preferred embodiment of the present invention, The value ranges from 0.3m to 0.5m.

[0038] The second-layer clustering divides the non-ground layer point cloud into clusters of auxiliary facilities and discrete noise based on spatial distribution density. The adaptive density clustering filtering formula proposed in this invention is as follows: , in, For point The local density value, This represents the total number of points in the non-ground layer point cloud. This is an indicator function (it takes the value 1 when the condition is met, and 0 otherwise). For point With point The Euclidean distance between them This is the adaptive neighborhood search radius.

[0039] Adaptive neighborhood search radius Dynamic adjustments are made based on the road longitudinal slope and cross slope change rates, calculated using the following formula: , in, The reference neighborhood radius is 0.3m in a preferred embodiment of the present invention; The longitudinal slope angle of the road; The cross slope angle of the road; and To adjust the coefficients, in a preferred embodiment of the present invention, the values ​​are 0.5 and 0.3, respectively.

[0040] On point Local density value Density greater than the preset threshold If the point belongs to the auxiliary facility target cluster, it is determined that the point belongs to the discrete noise cluster and is discarded. In a preferred embodiment of the present invention, The value range is from 10 to 30.

[0041] The output data for this step is the target point cloud D06. After adaptive filtering, the separation accuracy between the target point cloud and the environment point cloud is no less than 95%.

[0042] Step S3: Complete extraction of the upright structural facility. The complete extraction process for vertical structure facilities involves identifying and extracting road ancillary facilities from the target point cloud and measuring their geometric parameters. For example... Figure 3 As shown, this step includes four sub-processes: feature parameter extraction, density clustering segmentation, structural feature matching, and geometric parameter calculation.

[0043] The feature parameter extraction sub-process calculates the local geometric features of each point in the target point cloud for subsequent upright structure identification. The feature parameters extracted in this invention include normal vector, rate of change of elevation, linearity, flatness, and sphericity.

[0044] The calculation of the normal vector is based on the analysis of the local neighborhood covariance matrix of the point. For a point... First, obtain its radius as The set of all points within the spherical neighborhood Then calculate the covariance matrix of the neighborhood point set: , in, For the number of neighboring points, Let be the centroid coordinates of the neighborhood point set. For the covariance matrix... Eigenvalue decomposition yields three eigenvalues. and the corresponding feature vectors Minimum eigenvalue corresponding feature vector That is, a point The estimated value of the normal vector.

[0045] The rate of change of elevation is used to identify vertical structures perpendicular to the ground. The elevation change rate is defined as the ratio of the vertical distribution range to the horizontal distribution range of the point cloud within its neighborhood. The formula for calculating the elevation change rate of a vertical structure proposed in this invention is as follows: , in, and These are the maximum and minimum elevation values ​​of the neighboring point cloud, respectively; , , , Neighboring point clouds in direction and Maximum and minimum coordinate values ​​of the direction; To prevent small positive numbers from being divided by zero, in a preferred embodiment of the invention, the value is set to 0.001. When the elevation change rate... When the value exceeds a preset threshold (2.0 in a preferred embodiment of the invention), the determination point is... Located on an upright structure.

[0046] Linearity, flatness, and sphericity are calculated based on the eigenvalues ​​of the covariance matrix: , in, Linearity reflects the degree to which a point cloud extends along a certain direction; Flatness reflects the degree to which a point cloud is distributed on a certain plane; Sphericity reflects the degree of dispersion of point cloud distribution. Vertical structures (such as columns) typically have high linearity, while plate-like structures (such as sound barrier panels) typically have high flatness.

[0047] The density clustering segmentation subprocess uses the DBSCAN algorithm to segment the target point cloud into independent facility target clusters. The core parameters of the DBSCAN algorithm include the neighborhood search radius. and minimum points In a preferred embodiment of the present invention, The value is set adaptively based on the point cloud density, ranging from 0.05m to 0.15m. The value ranges from 20 to 50.

[0048] The DBSCAN algorithm is executed as follows: First, calculate its value for each point. The number of points in the neighborhood, if the number of points is not less than These are then marked as core points; subsequently, the core points that are density-reachable from each other and their... All points within a neighborhood are clustered into the same target cluster; finally, points that do not belong to any cluster are marked as noise points. After clustering, target clusters whose sizes do not meet the requirements are filtered out according to a preset facility structure size threshold. In a preferred embodiment of the present invention, the diameter of the sound barrier column ranges from 5cm to 15cm and the height ranges from 2m to 5m; the diameter of the gantry column ranges from 20cm to 50cm and the height ranges from 5m to 10m; the height of the crash barrier ranges from 0.5m to 1.5m.

[0049] The structural feature matching subprocess establishes a structural feature library of typical ancillary facilities and achieves target classification through feature matching. The structural feature library stores standard geometric features of various ancillary facilities, including size ranges, shape features, and spatial distribution patterns. The structural feature matching similarity calculation formula proposed in this invention is as follows: , in, For target cluster With the first feature in the feature library Type of facility template Matching similarity; The feature dimension; For target cluster The One eigenvalue; template The The mean of each feature; template The The standard deviation of each feature is used to classify the target cluster into the facility type with the highest matching similarity. If the highest similarity is lower than a preset threshold (0.6 in a preferred embodiment of the invention), it is marked as an unknown type.

[0050] The geometric parameter calculation subprocess performs precise geometric parameter measurements on the classified clusters of facilities. The measurement items are determined based on the facility type: for sound barriers, the measurements include column height, column verticality, panel flatness, and spacing between adjacent panels; for gantry frames, the measurements include column height, column verticality, beam horizontality, and overall deformation; and for crash barriers, the measurements include wall height, wall surface flatness, and height consistency.

[0051] The method for measuring the verticality of a column is as follows: Fit a cylindrical surface to the column point cloud to obtain the equation of the cylinder axis, and calculate the angle between the axis direction vector and the vertical direction vector. Verticality deviation is defined as The method for measuring the flatness of the plate is as follows: perform plane fitting on the point cloud of the plate, calculate the distance from each point to the fitted plane, and define the flatness deviation as the maximum distance deviation per unit length.

[0052] The output data for this step is the classification label D08 and the geometric parameter set D09. The completeness of the extracted upright structure is no less than 98%, and the measurement error of the geometric parameters is no greater than ±3mm.

[0053] Step S4: Change detection and quantitative evaluation based on DSM differential analysis The DSM-based change detection and quantitative evaluation process identifies defective and deformed areas of facilities by comparing multi-temporal point cloud data to construct a digital surface model, and then performs a quantitative condition assessment. For example... Figure 4 As shown, this step includes four sub-processes: DSM model construction, differential calculation, identification of changed areas, and condition evaluation. This step is the core link in the present invention to achieve quantitative assessment of facility service performance. By establishing a multi-temporal data comparison and analysis mechanism, the deformation evolution trend of the facility over time can be effectively captured.

[0054] The DSM model construction sub-process rasterizes the target point cloud to generate a digital surface model. First, a regular raster grid covering the facility area is established. The raster resolution is set according to the detection accuracy requirements, and in a preferred embodiment of this invention, it ranges from 2cm to 5cm. The selection of the raster resolution needs to comprehensively consider both detection accuracy requirements and computational efficiency. A smaller raster resolution can acquire more detailed surface morphology information, which is beneficial for capturing minute deformations, but it will significantly increase data storage and computation time; a larger raster resolution can improve processing efficiency, but may miss local detail changes. For sound barrier facilities requiring millimeter-level deformation monitoring, a raster resolution of 2cm to 3cm is recommended; for overall deformation monitoring of large facilities such as gantry cranes, a raster resolution of 4cm to 5cm can be used.

[0055] Then, for each grid cell, all point clouds falling within that cell are statistically analyzed, and the elevation statistics of these points are calculated as the DSM value of that grid. This invention uses the maximum elevation value as the grid DSM value to accurately reflect the elevation information of the facility surface. , in, For the first DSM value of each grid cell; For the first cell to fall within this grid cell Elevation values ​​of each point; Indicates the first The spatial range of each grid cell. For empty grid cells where no point cloud falls, a neighborhood interpolation method is used to fill them. In a preferred embodiment of the invention, an inverse distance weighted interpolation method is used for empty grid cells. The interpolation result of the empty grid cell is obtained by weighted averaging of the DSM values ​​of the effective grid cells within the surrounding 8 neighborhoods, with the weights inversely proportional to the square of the distance. In addition, to eliminate the influence of random noise during the point cloud acquisition process, median filtering or mean filtering can be used to preprocess the point cloud elevation values ​​falling within the grid cell when calculating the grid DSM values.

[0056] The differential operation subprocess performs coordinate alignment between the current DSM and historical DSMs, followed by pixel-by-pixel subtraction. Coordinate alignment employs a stable region registration method, selecting a stable reference region (such as fixed buildings, ground markers, etc.) around the facility as the registration datum. The ICP algorithm is used to accurately align the coordinate datums of the two DSMs. The selection of a stable reference region should meet the following conditions: First, the reference region has not undergone displacement or deformation between the two acquisition periods; second, the reference region has a sufficient number of feature points to support reliable registration calculations; third, the reference region has good coverage and clear geometric features in the point cloud. In practical applications, concrete piers, metal pole bases, or permanent ground markers on both sides of roads can be selected as stable reference regions.

[0057] After coordinate alignment, calculate the difference DSM: , in, The current DSM Raster values; For the first time in the history of DSM Raster values; The values ​​represent differences, with positive values ​​indicating an increase in elevation (such as deformation or bulging) and negative values ​​indicating a decrease in elevation (such as defects or depressions). The calculation results of the differential DSM can intuitively reflect the morphological changes of the facility between two periods, and serve as the basic data for subsequent identification of change areas and condition assessment.

[0058] The change region identification sub-process extracts significantly changed regions from the differential DSM. First, a change detection threshold is set. ,when The grid is determined to have changed at that time. In a preferred embodiment of the present invention, The value ranges from 5mm to 10mm. Setting the change detection threshold requires comprehensive consideration of the measurement system's accuracy and actual detection needs. If the threshold is set too low, measurement noise may be misjudged as a real change; if the threshold is set too high, minute but important deformation signals may be missed. In a preferred embodiment of the invention, the change detection threshold is set to 2 to 3 times the nominal accuracy of the measurement system to ensure the reliability of the detection results.

[0059] Then, connected component analysis is performed on the changing grid, aggregating adjacent changing grids into changing regions, and calculating the location, extent, and change statistics for each changing region. The change statistics include the maximum change, average change, and area within the changing region. The connected component analysis uses the eight-neighbor connectivity criterion, meaning two changing grids are considered connected if they are adjacent in any of the eight directions (up, down, left, right, upper left, upper right, lower left, lower right). After connected component analysis, isolated changing regions with excessively small areas are filtered out to eliminate noise interference. In a preferred embodiment of the invention, isolated changing regions with an area less than nine times the square of the grid resolution are filtered out.

[0060] The condition assessment sub-process determines the condition assessment level of the facility based on a comparison between changes and maintenance standards. This invention establishes a four-level condition assessment model, with the following assessment formula: , in, The status evaluation level; The maximum change in the area of ​​change; , , This is a maintenance standard threshold. In a preferred embodiment of the invention, for sound barrier facilities, Set as the facility's allowable deviation value. Set to twice the allowable deviation value. Set to 3 times the allowable deviation value. For example, when the allowable deviation for the flatness of the sound barrier panel is 5 mm / m, , , The design of the four-level condition assessment model follows the grading principle of highway maintenance technical specifications. A good level indicates that the facility is in normal service and does not require immediate action; a minor level indicates that the facility has minor defects and routine maintenance is recommended; a moderate level indicates that the facility has obvious defects and special repairs are required; and a severe level indicates that the facility has serious safety hazards and emergency measures need to be taken immediately.

[0061] The output data for this step are change detection result D13 and condition assessment level D14. Change detection can capture millimeter-level deformation, and the condition assessment results provide a quantitative basis for subsequent maintenance and treatment decisions. The data format of the change detection results includes attributes such as the boundary coordinates, centroid coordinates, maximum change, average change, change area, and change type (defect or deformation) of the changed area; the data format of the condition assessment level is the four-level classification result and the corresponding confidence score.

[0062] Step S5: Output and application of detection and evaluation results The output and application of testing and evaluation results will summarize and integrate the processing results of the above steps to generate a structured testing and evaluation report, and connect it to the digital management platform through a standard data interface.

[0063] The inspection and evaluation report includes: asset statistics, including the quantity, location coordinates, and basic attributes of various ancillary facilities within the inspection scope; geometric inspection results, including the geometric parameter measurements of each facility and their deviations from design standards; change inspection results, including the location, range, amount of change, and type of change of each area; condition evaluation results, including the condition evaluation level and disease statistics of each facility; and maintenance recommendations, providing differentiated maintenance and treatment suggestions based on the level and type of disease.

[0064] The inspection and evaluation reports are stored in structured data formats (such as JSON and XML) and connected to the digital management platform through standard data interfaces. The digital management platform can visualize, query, retrieve, and dynamically update the inspection and evaluation data, enabling digital management of road ancillary facilities assets and supporting scientific management and maintenance decisions throughout the entire lifecycle.

[0065] The following example, using a sound barrier testing project on an elevated road as an example, details the specific application process of this invention.

[0066] The elevated road is 12km long, and the metal plate sound barriers have been in service for 8 years. They have hidden dangers such as cracking of the plates and tilting of the columns, and need to be fully inspected and evaluated.

[0067] For equipment selection and integration, the core equipment selected includes a RIEGL VQ-1560i 3D laser scanner (point cloud density 60 points / cm², measurement distance 0.5m to 150m), a Canon EOS R5 high-resolution camera (50 megapixels), and a POS system consisting of a Trimble R10 GNSS and IMU inertial navigation (positioning accuracy ±1cm). The 3D laser scanner and camera are fixed to a vehicle-mounted roof bracket, and the POS system is installed in the vehicle's cab, connected to the data storage module via a data cable.

[0068] For data acquisition, the vehicle traveled at 60 km / h along the elevated road, simultaneously activating all sensors. A 3D laser scanner continuously scanned the sound barrier and its surrounding environment, collecting approximately 120 GB of point cloud data; cameras simultaneously captured panoramic images, acquiring a total of 24,000 images; and the POS system recorded the vehicle's pose data. The acquisition process took 2 hours, achieving full coverage data collection of the sound barrier along the entire road section.

[0069] In terms of data processing and analysis, CloudCompare software was used to denoise the original point cloud during the preprocessing and filtering stages. An adaptive filtering algorithm from this invention was used to construct a dynamic reference surface (elevation adjustment step size of 0.1m) to separate the sound barrier point cloud from the environmental interference point cloud. After filtering, the purity of the target point cloud reached 96.8%. In the vertical structure extraction stage, a sound barrier structural feature library was established (column diameter 8cm, height 3m, panel thickness 10mm, width 1.2m). The DBSCAN algorithm was used for clustering and segmentation, extracting 326 columns and 1630 panels, achieving an extraction completeness of 98.5%. In the geometric detection stage, the verticality of the columns was calculated through point cloud plane fitting, revealing that 12 columns had a verticality deviation of 0.6% to 0.8% (exceeding the standard). Through panel point cloud flatness analysis, 35 panels were found to have a flatness deviation of 6mm / m to 10mm / m (exceeding the standard). During the change analysis and evaluation phase, point cloud data collected 3 years ago were compared to construct two DSM models (5cm resolution). Differential calculations revealed 8 board defects (area 0.1m² to 0.3m²) and 5 column deformations (deformation amount 12mm to 18mm). Based on the evaluation model, 23 cases of mild disease, 15 cases of moderate disease, and 7 cases of severe disease were identified.

[0070] Regarding testing results and repair procedures, testing data, evaluation results, and repair records are integrated into a digital platform to achieve asset statistics for sound barriers (326 columns and 1630 panels), visual labeling of defect locations, and tracking of repair progress, supporting subsequent maintenance decisions.

[0071] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features, but these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for detecting and evaluating vulnerable road ancillary facilities based on mobile measurement and AI analysis, characterized in that, Includes the following steps: Step S1, Multi-source sensor data synchronous acquisition: Data on road ancillary facilities is acquired by a vehicle-mounted mobile measurement system while the vehicle is in motion. The vehicle-mounted mobile measurement system includes a 3D laser scanner, a high-resolution linear camera, and a POS system. The 3D laser scanner continuously scans the road ancillary facilities to obtain raw point cloud data. The high-resolution linear camera synchronously acquires panoramic image data. The POS system records vehicle pose data in real time and provides a spatiotemporal reference for the raw point cloud data and panoramic image data. Step S2, adaptive filtering of point cloud data: The original point cloud data is subjected to coordinate transformation and denoising preprocessing. A dynamic elevation reference surface is constructed based on the road undulation characteristics. The preprocessed point cloud is filtered using a hierarchical clustering and threshold adaptive adjustment strategy. The hierarchical clustering and threshold adaptive adjustment strategy includes: the first layer clustering divides the point cloud into a ground layer and a non-ground layer based on the elevation attribute of the point cloud; the second layer clustering divides the non-ground layer point cloud into ancillary facility target clusters and discrete noise clusters based on spatial distribution density, and separates the target point cloud of road ancillary facilities from the environmental point cloud of vehicles and the surrounding environment. Step S3, complete extraction of upright structure facilities: Extract the normal vector and elevation change rate feature parameters of the target point cloud, identify the feature point set of the upright structure, use a density-based clustering algorithm to segment the target area according to the preset facility structure size threshold, establish a structural feature library of typical auxiliary facilities, and achieve target classification through feature matching, extracting geometric parameters including height, verticality, spacing and flatness; use the pose data provided by the POS system to register and fuse the target point cloud with the panoramic image data, and use the texture information of the panoramic image to supplement the area information missing in the target point cloud due to occlusion or reflection; Step S4, Change Detection and Quantitative Evaluation Based on DSM Differential: The target point cloud collected in the current period and the point cloud collected in the past are rasterized to construct a digital surface model. The current digital surface model and the historical digital surface model are differentially calculated to obtain the change detection results. The change detection results include the location, range and amount of change of the facility defect area and the deformation area. The facility status evaluation result is determined based on the comparison result of the amount of change with the preset maintenance standard threshold. Step S5, Output and Application of Inspection and Evaluation Results: Integrate asset statistics, geometric inspection results, change inspection results, and condition evaluation results to generate an inspection and evaluation report. The inspection and evaluation report is connected to the digital management and maintenance platform through a standard data interface to support the full life-cycle management and maintenance decisions of road ancillary facilities. The asset statistics include the quantity, location coordinates, and basic attributes of various ancillary facilities within the inspection range; the geometric inspection results include the geometric parameter measurements of each facility and their deviations from design standards; the change inspection results include the location, range, amount of change, and type of change of each change area; and the condition evaluation results include the condition evaluation results and defect statistics of each facility.

2. The method according to claim 1, characterized in that, In step S1, the vehicle travels at a speed of 30 km / h to 80 km / h, the point cloud density of the original point cloud data is not lower than a preset point cloud density threshold, the field-of-view overlap rate between the 3D laser scanner and the high-resolution linear camera is not lower than 80%, and the sampling frequency of the POS system is not lower than 100 Hz.

3. The method according to claim 1, characterized in that, In step S2, constructing the dynamic elevation reference surface includes: dividing the preprocessed point cloud into continuous segmented regions along the longitudinal direction of the road; performing planar fitting on the road surface point cloud in each segmented region to obtain a local reference surface; and smoothly connecting adjacent local reference surfaces to form the dynamic elevation reference surface. Automatically adjusting the filtering threshold includes: calculating the elevation difference between each point in the point cloud and the dynamic elevation reference surface; when the elevation difference exceeds a preset elevation difference threshold, the point is classified as the target point cloud; and the accuracy of the filtering process is not less than 95%.

4. The method according to claim 1, characterized in that, In step S3, the typical auxiliary facilities include sound barriers, gantry frames, and crash barriers. The testing standards for the geometric parameters include: the allowable deviation of the verticality of the sound barrier columns is no more than 0.5%, the allowable deviation of the flatness of the sound barrier panels is no more than 5 mm / m, the allowable deviation of the horizontality of the gantry frame beams is no more than 0.3%, and the allowable deviation of the flatness of the crash barrier surface is no more than 3 mm / m.

5. The method according to claim 1, characterized in that, In step S4, the raster resolution of the digital surface model is not greater than the preset DSM raster resolution threshold. The difference operation uses the hierarchical clustering difference method to align the coordinate references of the current digital surface model and the historical digital surface model before performing pixel-by-pixel subtraction.

6. The method according to claim 1, characterized in that, In step S4, the status evaluation result includes four levels: when the change is not greater than the first change threshold, it is judged as a good level; when the change is greater than the first change threshold but not greater than the second change threshold, it is judged as a mild disease level; when the change is greater than the second change threshold but not greater than the third change threshold, it is judged as a moderate disease level; and when the change is greater than the third change threshold, it is judged as a severe disease level.

7. The method according to claim 1, characterized in that, In step S3, the segmentation of the target area using a density-based clustering algorithm based on a preset facility structure size threshold includes: setting a neighborhood search radius parameter and a minimum number of points parameter; determining whether a point is a core point based on the number of points in the neighborhood of each point; clustering mutually density-reachable core points and points in their neighborhoods into the same target cluster; and filtering target clusters whose sizes do not conform to the preset range based on the facility structure size threshold.

Images