Pavement health information extraction system and method based on large model analysis

Abstract

The invention relates to the technical field of pavement analysis, in particular to a pavement health information extraction system and method based on large model analysis. A pavement health information extraction system based on large model analysis comprises an unmanned aerial vehicle information acquisition device, an unmanned aerial vehicle control device, a three-dimensional modeling device and a model information collection and analysis device. According to the technical scheme, the three unmanned aerial vehicles synchronously collect the three-dimensional video information of the road surface in parallel, and the high-precision three-dimensional modeling and dynamic analysis system of the road surface is constructed in combination with the relative position data updated in real time. Compared with a traditional mode of reciprocating acquisition of a single unmanned aerial vehicle, the scheme realizes three-dimensional coverage and synchronous acquisition of road surface information by utilizing multi-unmanned aerial vehicle cooperative operation, and completes rapid fusion modeling of a three-dimensional scene based on accurate relative position information, so that the generation efficiency of a road surface three-dimensional model is improved by more than three times, and the spatial positioning accuracy reaches a centimeter level.

Description

Technical Field

[0001] The present application relates to the technical field of pavement analysis, and in particular to a pavement health information extraction system and method based on large model analysis. Background Art

[0002] The contents of this section merely provide background information related to this application and may not constitute prior art.

[0003] During highway construction, the route is typically planned first, with construction then proceeding step by step along the route. The construction process typically includes road grading, tunnel excavation, bridge erection, followed by foundation construction and road paving. Due to the long construction period, continuous route inspections are required after the initial road grading. By acquiring and analyzing the road's three-dimensional structural information, construction teams can promptly adjust subsequent construction schedules and plans. If, during construction, rain causes landslides and falling rocks to cover the already-graded road surface, the construction process and plan must be reassessed, and the adjusted costs calculated. Alternatively, if new buildings appear along the construction route, their legality must be determined, and subsequent construction plans must be revised based on the new surface conditions.

[0004] To this end, regular inspections along planned routes are required during highway construction. The current mainstream inspection method uses drones: drones fly along planned routes, collecting image data along the way. Using image fusion technology, they generate three-dimensional road surface information, which is then used to assess road health. Drone inspections typically use oblique photogrammetry to obtain data. However, practice has shown that the information obtained from a single shooting angle is limited. In particular, for tall objects on the road, two or three different angles are often required to fully capture their surface information. Existing technical solutions typically require drones to repeatedly fly back and forth along the route, continuously collecting data and gradually refining the three-dimensional scene to ultimately construct a complete three-dimensional model. This method of extracting three-dimensional information is inefficient in practice. Even with multiple round-trip drone flights, complex road conditions often struggle to obtain accurate three-dimensional information. This results in inaccurate three-dimensional road surface modeling, which in turn affects the accuracy and efficiency of road health assessments. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a road health information extraction system and extraction method based on large model analysis to solve the technical problems mentioned in the background technology.

[0006] The purpose of this application is achieved through the following technical solutions: A road health information extraction system based on large model analysis includes: The drone information collection device includes three drones for aerial photography, which are deployed parallel to the road to be measured and fly synchronously along the road to be measured to obtain three-dimensional video information. The three-dimensional video information includes: video data synchronously collected by the three drones; relative position information corresponding to each frame of the video data; the relative position information includes at least the real-time relative position between the three drones; UAV control device: used to control the flight direction and attitude of each aerial photography UAV in real time, guiding it to fly along the predetermined route; 3D modeling device: receives 3D video information in real time and generates a 3D road surface model based on the video data and its corresponding relative position information; Model information collection and analysis device: collects pavement 3D models generated in different time periods, compares and analyzes the current pavement 3D model with the historical pavement 3D model, and extracts and generates pavement health information that characterizes changes in pavement status.

[0007] This technical solution uses three drones to synchronously collect three-dimensional road surface video information in parallel, combining it with real-time updated relative position data to build a high-precision three-dimensional road surface modeling and dynamic analysis system. Compared to the traditional model of reciprocating acquisition by a single drone, this solution utilizes multiple drones working together to achieve three-dimensional coverage and synchronous acquisition of road surface information. Based on precise relative position information, it rapidly integrates and models three-dimensional scenes, increasing the efficiency of generating three-dimensional road surface models by more than three times and achieving centimeter-level spatial positioning accuracy. Through time-series comparative analysis of three-dimensional models from different time periods, it can automatically identify the development trajectories of road surface defects such as cracks, subsidence, and rutting, accurately extracting quantitative indicators such as crack width change and subsidence depth, and increasing the accuracy of extracting road surface state change information to over 95%. This effectively addresses the low efficiency and high subjective error of traditional manual inspections, provides precise data support with temporal and spatial continuity for road maintenance decisions, and significantly enhances the automation level and engineering application value of road health assessments.

[0008] When obtaining road surface information through high-altitude aerial photography, the flight altitude must be increased to ensure the safety of drone flight and expand the field of view, but this results in significant measurement errors. On the one hand, the increase in flight altitude causes the slight angular deviations in the relative positions of drones to be magnified, seriously affecting the matching accuracy of feature points in the image, making it difficult to accurately align the road surface three-dimensional model when constructing it; on the other hand, traditional wireless positioning technology can only obtain distance information between drones, and cannot effectively handle the height errors between drones. It is difficult to accurately determine the relative position relationship between drones, resulting in a significant reduction in the accuracy and reliability of road surface information collection. To address this issue, this application provides the following technical solutions: Aerial photography drones include: The drone body is used to provide an aerial flight platform; A high-speed camera is mounted on the drone body at a fixed angle, tilted downward; The wireless positioning module generates initial relative position parameters with other aerial photography drones through millimeter wave positioning; Laser transmitter, fixedly installed on the drone; The laser receiver forms two laser receiving planes on both sides of the drone body; An information processor is connected to the wireless positioning module and the laser receiver signal to generate relative position information; Among them, the laser transmitter of the aerial photography drone sends a laser signal to the laser receiver of the adjacent aerial photography drone, and the information processor corrects the relative position information based on the position where the laser irradiates the laser receiving plane.

[0009] This solution effectively overcomes the positioning challenges of high-altitude aerial photography by constructing a "wireless positioning initial measurement + laser refinement" system for obtaining drone relative positions. The wireless positioning module provides a preliminary position reference for the drone, quickly establishing a relative position baseline. The laser transmitter, in conjunction with two laser receiving surfaces, accurately measures subtle differences in the relative position of the drones based on the laser's impact point, achieving high-precision corrections at the centimeter or even millimeter level. This significantly reduces feature point matching errors caused by relative positioning errors between drones and significantly improves the accuracy and reliability of the 3D road model, effectively enhancing the efficiency and quality of road inspections.

[0010] When multiple drones collaborate to collect road surface information, the accumulated errors from each drone's independent positioning and the lack of a unified positional reference make it difficult to establish accurate positional relationships between them. If each drone relies solely on its own positioning system to obtain positional information, not only will it be unable to eliminate positioning errors caused by factors such as flight altitude and environmental interference, but it will also cause inaccurate positional relationships during image matching, leading to feature point mismatches and 3D model dislocations. This will seriously affect the accuracy and completeness of 3D road surface modeling, and in turn, reduce the reliability of road health information extraction.

[0011] Furthermore, the aerial photography drones on both sides obtain the relative positions of the aerial photography drone in the middle, and the aerial photography drone in the middle obtains its own current geographical position.

[0012] This solution effectively addresses the positioning coordination challenges faced by multi-UAV collaborative operations by establishing a "center positioning + side coordination" position sensing mechanism. The center UAV acquires its precise geographic location as a global benchmark, while the two side UAVs use the center UAV as a reference for relative position measurement, forming a three-in-one positional relationship network. This design unifies the positioning errors of each UAV into a common coordinate system, improving relative position accuracy to the decimeter level, significantly reducing the mismatch rate of feature points in image matching, and improving the integrity and accuracy of 3D model assembly.

[0013] When drones perform road information collection tasks at high altitudes, they are easily affected by factors such as atmospheric turbulence and airflow disturbances, resulting in unstable flight posture, jitter, tilt, etc. The images collected during this period have serious geometric distortions, and due to the lack of stable lens pointing parameters, it is impossible to accurately establish the spatial correspondence between images. This makes it easy for feature points to be mismatched during subsequent image matching, and it is difficult to achieve effective fusion in the three-dimensional modeling process, resulting in distortion, dislocation and other problems in the three-dimensional model of the road surface, which seriously reduces the accuracy and reliability of road health information extraction and cannot meet the needs of high-precision road detection. To address this problem, this application provides the following technical solutions: The drone information collection device also includes: The video information screening module is connected to the aerial photography drone signal and is used to obtain the video data collected synchronously by each aerial photography drone, and to screen out the image frames when the aerial photography drone's flight posture is stable to generate three-dimensional video information.

[0014] This solution effectively mitigates the negative impact of unstable drone flight on data collection by introducing a video information screening module and establishing an intelligent image quality control mechanism. Based on real-time drone flight attitude data, this module accurately identifies periods of stable flight and automatically selects high-quality image frames with no noticeable jitter and controlled geometric distortion, ensuring consistent 3D video quality. The 3D modeling process achieves sub-pixel model stitching accuracy, significantly enhancing the integrity and accuracy of the 3D road model.

[0015] When drones collaborate to collect road information, they face two technical bottlenecks: First, the attitude data of the built-in gyroscope of the flight control system is difficult to read directly due to permission restrictions, resulting in the inability to obtain the drone's flight attitude in real time; second, relying solely on the attitude information of a single drone, it is impossible to determine whether the collection array composed of three drones is in the predetermined shooting position as a whole. As a result, video frames under disturbances such as turbulence lack spatial position correlation, making it difficult to determine their modeling value. This, in turn, causes confusion in feature point matching and misalignment of model splicing during three-dimensional information fusion, seriously affecting the accuracy of road surface detection. To this end, this application provides the following technical solutions: The video information screening module includes: A video information acquisition unit is used to acquire the video data collected synchronously by each aerial photography drone, and synchronously arrange the video data acquired by the three aerial photography drones based on the time tags to generate a video frame matrix; The video information screening unit obtains the time period when the laser receiver of each aerial photography UAV receives the laser signal, and defines the time period when all laser receivers receive the laser signal as the ideal time period; The video information cropping unit takes the portion corresponding to the ideal time period in the video frame matrix as the three-dimensional video information.

[0016] This solution addresses the challenges of flight control data access restrictions and array position determination by establishing a dual "laser signal-time tag" screening mechanism. The video information screening module synchronizes the video data of the three drones based on time tags, forming a time-sequenced video frame matrix. Simultaneously, through signal interaction with the laser receivers, it accurately identifies the "ideal time period" when all three drones are in their predetermined positions. This mechanism eliminates the need to directly read internal flight control data, determining the spatial consistency of the array solely through laser signals.

[0017] The road health information extraction system based on large model analysis also includes: an image preprocessing device for preprocessing three-dimensional video information; the preprocessing includes: noise denoising and lens correction.

[0018] In the technical solution provided in this application, the image preprocessing device can effectively improve the image quality and increase the construction accuracy of the three-dimensional model.

[0019] There are two technical pain points in the traditional 3D modeling process: first, feature point matching is easily affected by factors such as fluctuations in the drone's flight attitude and changes in image perspective, resulting in pixel mapping deviations and the formation of a set of feature point mismatches, which in turn causes geometric distortion of the 3D model; second, when building a model framework based on a single matching algorithm, there is a lack of global adjustment constraints, making it difficult to eliminate the accumulated errors between multi-view images and unable to meet high-precision detection requirements.

[0020] The 3D modeling device includes: A pixel matching module is used to extract feature points from 3D video information and map the feature points to generate a feature point mapping set; 3D model building module, which matches feature points based on bundle network adjustment to build the model framework; The model rendering module extracts rendering materials from the three-dimensional image to render the model frame and generate a three-dimensional road model.

[0021] This solution achieves high-precision reconstruction of pavement 3D models by constructing a three-level modeling system: feature point mapping, beam network adjustment, and texture rendering. The pixel matching module combines 3D video information collected synchronously by multiple drones with the relative position reference provided by laser positioning to form a high-density feature point set. The 3D model construction module incorporates a beam network adjustment algorithm. Based on the collinearity equation constraints of multi-view images, it globally optimizes the ground coordinates of feature points, achieving centimeter-level spatial positioning accuracy for the model framework and effectively eliminating cumulative errors. The model rendering module extracts spectrally consistent texture materials from the original video and achieves precise alignment of the texture with the model framework through coordinate mapping. This module can intuitively present the geometric morphology and textural characteristics of pavement defects such as cracks and rutting, providing a high-fidelity data foundation for the automated extraction of pavement health information and significantly enhancing the engineering application value of 3D modeling.

[0022] When building a 3D model, it is necessary to extract enough feature points from the 3D video information and match the feature points to provide enough feature points when building the 3D model. However, feature point extraction and feature point matching require a lot of computing power. If a neural network model is used to extract feature points, the accuracy is not high when the image feature changes in the video information are complex, which will cause a lot of distortion in the model framework. Based on this, the application provides the following technical solutions: The pixel matching module uses the following steps to extract feature points from a 3D image and map them: Step 1: Extract the flight speed of the drone from the 3D video information and divide the 3D video information into several video groups based on the flight speed. Each video group includes video streams captured by three drones in the same time period. Step 2: Extract several sample images with equal spacing from the video stream, extract feature points from the sample images based on the SIFT algorithm, and perform one-to-one matching of the feature points based on cosine distance to generate a sample matching set; Step 3: Use the sample matching set to train the convolutional network model and adjust the weight parameters in the convolutional network model; Step 4: Input the remaining images in the video stream into the trained convolutional network model to extract feature points and the mapping relationship between feature points; Step 5: The mapping relationship between the feature points extracted in step 2 and step 4 is used as a feature point mapping set.

[0023] In the technical solution provided by this application, the SIFT algorithm is used to extract features from some images in the video stream, which can greatly ensure the accuracy of feature point extraction and feature point mapping. This is then used as training data to train the convolutional network model, which can change its internal weight parameters, thereby having good feature recognition and feature comparison capabilities in the video group, and can quickly extract feature points and the mapping relationship between feature points. In this way, the convolutional network and the feature extraction algorithm are combined in this application, which not only ensures the efficiency of feature point extraction, but also ensures the accuracy of feature point extraction.

[0024] During the three-dimensional model construction process, feature point processing faces two technical bottlenecks: on the one hand, although traditional feature point extraction and matching algorithms (such as SIFT and SURF) have high accuracy, they are computationally complex. When faced with massive video data, there is a problem of excessive computing power consumption, which makes it difficult to meet real-time modeling requirements; on the other hand, when simply using a neural network model to extract feature points, due to the complex changes in road texture, lighting and perspective in video images, the model's generalization ability is insufficient, and feature point mismatches are prone to occur, resulting in geometric distortion of the three-dimensional model framework and failure to meet high-precision detection requirements.

[0025] Furthermore, the convolutional network model includes: The input layer is used to input images taken by three drones at the same time to generate the first feature map, the second feature map, and the third feature map; The convolution layer performs convolution processing on the input first feature map, second feature map, and third feature map to extract the first latent feature, the second latent feature, and the third latent feature; The pooling layer is connected to the convolutional layer and is used to pool the first latent feature, the second latent feature, and the third latent feature; The probability calculation layer generates the probability information of each pixel belonging to a feature point from the first latent feature, the second latent feature, and the third latent feature; The first attention mechanism network inputs the first feature map and probability information to generate the first transformed feature; The second attention mechanism network inputs the second feature map and probability information to generate the second transformed feature; The third attention mechanism network inputs the third feature map and probability information to generate the third transformed feature; The mapping network maps the first transformation feature, the second transformation feature, and the third transformation feature to each other to generate feature points and mapping relationships between the feature points.

[0026] This solution has achieved a breakthrough in resolving the contradiction between computing power consumption and matching accuracy by constructing a feature point processing system of "SIFT precise labeling + convolutional network efficient reasoning". First, the videos are grouped based on flight speed to ensure that videos in the same group have similar motion characteristics; high-precision sample matching sets are generated through equidistant sampling combined with the SIFT algorithm to provide high-quality training data for the convolutional network, greatly improving the network's feature recognition accuracy in complex scenarios. This solution not only uses the SIFT algorithm to ensure the matching accuracy of basic feature points, but also realizes efficient processing of batch data through neural networks, reducing the distortion rate of the three-dimensional model framework. It provides a high-density, low-error feature point set for subsequent beam network adjustment, significantly improving the efficiency and reliability of three-dimensional road modeling and meeting the real-time monitoring needs of engineering sites.

[0027] There are two core technical challenges in the construction of a three-dimensional road model framework: First, traditional model rendering methods rely on a single data source or initial parameters, making it difficult to accurately process the spatial geometric relationship between multi-view images, resulting in large errors in the calculation of the ground coordinates of the connection points and distortion problems such as distortion and misalignment in the model framework; Second, if simple iterative calculations are used, continuous calculations when the accuracy requirements are not met will consume a lot of time and computing power resources, while terminating the calculations early cannot guarantee model accuracy. At the same time, too much intermediate calculation data will take up a lot of storage space, making it difficult to strike a balance between accuracy, efficiency and storage costs.

[0028] The model rendering module generates the model framework based on the following steps: Z1: Obtain the feature point mapping set and the relative position information corresponding to the feature point mapping set, and convert the relative position information into the exterior orientation element E. The exterior orientation element E includes the coordinates of the photography center S in the ground coordinate system (X S 、Y S , Z S ) and the rotation angle of the image coordinate system relative to the ground coordinate system , χ S Indicates the rotation angle of the image coordinate system around the X axis, Indicates the rotation angle of the image coordinate system around the Y axis, Indicates the rotation angle of the image coordinate system around the Z axis; the feature point in the feature point set is used as the connection point i (x ij ,y ij ), j represents the sequence number of the image, j=1, 2, 3; Z2: For each tie point i, based on the exterior orientation elements of image j containing tie point i and the pixel coordinates of tie point i in image j, perform forward intersection using the collinearity equation to solve the initial ground coordinate approximation TP of tie point i ; Z3: According to TP and the exterior orientation elements to construct the parameter vector Xk ; ; Among them, k represents the number of iterations, initially k=0, represents the exterior orientation element of the first aerial photography UAV at the kth iteration, represents the exterior orientation element of the third aerial photography UAV at the kth iteration, represents the approximate initial ground coordinate of connection point 1 at the kth iteration, represents the approximate initial ground coordinate of the connection point n at the kth iteration, n represents the total number of connection points, T represents the matrix transpose sign, 1E represents the exterior orientation element E of the first aerial photography UAV, and 3E represents the exterior orientation element E of the third aerial photography UAV; TP1 represents the approximate initial ground coordinate of the first connection point, and TPn represents the approximate initial ground coordinate of the nth connection point; Z4: For each connection point i (x ij ,y ij ), extract the current exterior orientation elements of image j and the current coordinates of the connection point i , calculate the predicted image point coordinates ; represents the exterior orientation element of the j-th image at the k-th iteration; ; Where f represents the focal length of the camera, 、 、 、 、 、 、 、 、 Represents the rotation matrix element of the jth image. The two digits in the subscript represent the row and column numbers of the rotation matrix respectively. x0 and y0 represent the pixel coordinates. 、 、 represents the approximate initial ground coordinate of the connection point u at the kth iteration, represents the exterior orientation element of the jth aerial photography UAV at the kth iteration, 、 Indicates the distortion correction value; The observation value is used as the constant term I c , calculate the observation residual l x and l y , arrange the residuals in order to generate a constant vector I im ; ; ; 、 represents the coordinates obtained by measuring the connection point i, 、 Represents the coordinate value of the theoretical connection point calculated by the collinear equation, represents the pixel coordinates of the connection point i after correction on image j, Represents the residual component of the observation value of the connection point i in the horizontal direction of image j, Represents the residual component of the observation value of the connection point i in the longitudinal direction of image j; Extract the control point g from the feature point set. For each ground control point g, according to the known observation value Generate control point ground coordinate residual vector I ct ; ; Indicates the current approximate coordinate value of the ground control point g, 、 represents the residual of the observation value of the ground control point g, 、 Represents the coordinates of the ground control point g measured, and the residuals are arranged in order to generate a constant vector to form the constant term vector I of the control point ct ; Connect the pixel coordinate residual vector I im and the control point ground coordinate residual vector I ct Vertical splicing; combined into a complete constant vector I; Z5: parameter vector X k Perform Taylor series expansion to obtain the linearized error equation of the observation value: ; Among them, v represents the residual vector, which contains the residuals of the coordinate observations of all tie points and control points. represents the correction vector, which contains the correction values of all exterior orientation elements E. Represents the design matrix; each element of the design matrix A is the predicted value of the collinear equation; Z6: Construct the normal equation: ; Where I represents the complete constant vector, A represents the design matrix, T represents the matrix transpose symbol, and P represents the weight matrix of the observation value; Z7: Solve the equation to generate the correction vector , the correction vector Input to parameter vector X k , update the new parameter vector Xk+1 ; Based on the new parameter vector X k+1 Generate the geographic location coordinates of each connection point and render the model frame according to the geographic location coordinates.

[0029] This solution effectively breaks through the technical bottleneck of traditional model rendering by constructing a model framework generation system based on iterative optimization. The solution converts relative position information into exterior orientation elements, combines collinear equations with forward intersection to obtain the initial coordinates of the connection points, and lays an accurate foundation for model construction; through iterative least squares adjustment, the geographic coordinates of the connection points are continuously optimized, and the errors between the exterior orientation elements and the ground coordinates are continuously corrected, so that the spatial positioning accuracy of the model framework is improved to the sub-centimeter level, significantly reducing the model distortion rate. At the same time, the solution supports flexible control of rendering time according to actual needs, and terminates the calculation in time after reaching the preset accuracy standard. Compared with traditional methods, it not only ensures the accuracy of model rendering, but also realizes the efficient use of computing resources, providing reliable technical support for the fast and accurate construction of road three-dimensional models.

[0030] A pavement health information extraction method based on large model analysis uses the aforementioned pavement health information extraction system based on large model analysis to extract pavement health information.

[0031] The beneficial effects of this application are: (1) In the process of collecting 3D road information, the inefficient mode of traditional single UAV reciprocating flight is abandoned, and an innovative strategy of synchronous collection by three UAVs is adopted. The core advantages are: (2) Synchronous multi-angle acquisition: During the flight, the three drones not only collect video data synchronously, but also continuously update and record the real-time relative position (including attitude) information between each other.

[0032] (3) Accurate spatial positioning: Based on accurate real-time relative position information, the relative coordinates and viewing angle of each frame in the video in three-dimensional space can be reliably determined.

[0033] (4) Efficient 3D fusion modeling: Utilizing multi-view synchronized videos and their precise spatial position relationships, 3D scenes can be fused and reconstructed more quickly and accurately, significantly improving the generation speed and modeling accuracy of road 3D models.

[0034] (5) Improved assessment efficiency: The more accurate and timely three-dimensional pavement model obtained directly improves the accuracy and efficiency of subsequent pavement health status assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A schematic diagram of the structure of the road health information extraction system based on large model analysis provided in Example 1 of the present application; Figure 2 Schematic diagram of an aerial photography drone.

[0036] Figure 3 This is a flowchart of extracting feature points from a three-dimensional image in Example 2.

[0037] Figure 4 Schematic diagram of the structure of the convolutional network model in Example 2. Reference numerals

[0038] 1. Drone body; 2. Laser transmitter; 3. Laser receiver; 4. High-speed camera. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific implementation methods. The same figure marks in the accompanying drawings represent the same components. It should be noted that the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the described embodiments of this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0040] Compared to the embodiments shown in the drawings, feasible embodiments within the scope of protection of the present application may have fewer components, other components not shown in the drawings, different components, differently arranged components, or differently connected components, etc. In addition, two or more components in the drawings may be implemented in a single component, or a single component shown in the drawings may be implemented as multiple separate components.

[0041] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning understood by persons of ordinary skill in the field to which this application belongs. The words "first", "second" and similar terms used in the specification and claims of this application do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "a" or "an" do not necessarily indicate a quantitative limitation. "Up", "down" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0042] refer to Figure 1 , Example 1: The first embodiment of the present application discloses a road health information extraction system based on large model analysis, including: a drone information collection device, a drone control device, a three-dimensional modeling device, and a model information collection and analysis device. Among them, the drone information collection device includes three aerial photography drones, which are deployed parallel to the road to be tested and fly synchronously along the road to be tested to obtain three-dimensional video information, and the three-dimensional video information includes: video data synchronously collected by the three drones; relative position information corresponding to each frame of the video data; the relative position information at least includes the real-time relative position between the three aerial photography drones; a drone control device is used to control the flight direction and posture of each aerial photography drone in real time to guide it to fly along a predetermined route; a three-dimensional modeling device receives three-dimensional video information in real time, and generates a three-dimensional road surface model based on the video data and its corresponding relative position information; a model information collection and analysis device collects three-dimensional road surface models generated in different time periods, and extracts and generates road health information representing changes in road surface conditions by comparing and analyzing the current three-dimensional road surface model with the historical three-dimensional road surface model. The road health information extraction system based on large-scale model analysis also includes an image preprocessing device for preprocessing the 3D video information. This preprocessing includes noise denoising and lens correction. Noise denoising and lens correction are existing technologies and will not be described in detail here.

[0043] refer to Figure 2 An aerial photography drone consists of a drone body, a high-speed camera, a wireless positioning module, a laser transmitter, a laser receiver, and an information processor. The drone body is equipped with a high-speed camera (mounted at a fixed downward angle). The laser transmitter is fixedly mounted below the drone body, and the laser receivers are fixedly mounted on either side of the drone body. The laser receiver is a fixed-area laser receiver that can identify the wavelength of the received light signal and thus locate the light spot on the laser receiver.

[0044] Each aerial photography drone uses millimeter-wave positioning technology via a wireless positioning module to generate initial relative position parameters, quickly establishing a relative position reference. Subsequently, the drone's laser transmitter sends a laser signal to the laser receiving planes on both sides of the adjacent drones. The laser receiver transmits the position information of the laser irradiating the receiving plane to the information processor. The information processor corrects the initial relative position parameters based on the laser landing point, accurately measuring the subtle differences in the relative positions between drones and achieving high-precision positioning at the centimeter or even millimeter level. This process establishes a drone relative position acquisition system of "wireless positioning initial measurement + laser refinement", effectively solving the problem of drone relative position error caused by increased flight altitude, significantly reducing feature point matching deviations, and improving the accuracy of road surface three-dimensional model construction and the efficiency and quality of road inspection.

[0045] Furthermore, the aerial photography drones on both sides obtain the relative positions of the aerial photography drone in the middle, and the aerial photography drone in the middle obtains its own current geographical position.

[0046] The drone information collection device also includes: a video information screening module, which is connected to the aerial photography drone signal and is used to obtain the synchronously collected video data of each aerial photography drone, and screen out image frames when the aerial photography drone's flight posture is stable to generate three-dimensional video information.

[0047] Specifically, the video information screening module connects to each drone via wireless or wired signals, acquiring real-time video data collected synchronously by high-speed cameras (for example, a continuous video stream recorded synchronously when drones fly in formation over a section of highway). The video information screening module first checks the laser receiver to determine whether the drone's flight posture is stable. For example, if the drone is shaking due to airflow, the laser receiver cannot receive the laser signal. At this time, the video information screening module automatically filters the video frames for the corresponding period. When the drone is in a stable flight state, it selects high-definition image frames in that state (for example, 10-15 stable frames per second), thereby generating three-dimensional video information.

[0048] The video information screening module includes: a video information acquiring unit, a video information screening unit and a video information cutting unit.

[0049] The video information screening module first synchronously receives the video data collected by the high-speed cameras of three aerial drones through the video information acquisition unit, and arranges them in chronological order based on the time tags of the video data to form a regular video frame matrix. For example, when three drones are shooting the same section of road, the unit will align the video frames collected by each drone at the same time. Next, the video information screening unit monitors the working status of the laser receivers of each aerial drone in real time, obtains the laser signal reception time period, and determines the period when the laser receivers of the three drones all receive the laser signal as the ideal time period, which means that the three drones fly smoothly during this period. Finally, the video information cropping unit accurately crops the corresponding part from the video frame matrix based on the ideal time period to generate three-dimensional video information for three-dimensional modeling.

[0050] The three-dimensional modeling device includes: a pixel matching module, a three-dimensional model construction module and a model rendering module. The pixel matching module first analyzes the three-dimensional video information, uses the SIFT algorithm to extract feature points in the video frame, and then performs coordinate conversion on the feature points to map and generate a feature point mapping set containing the spatial position information of each feature point. The three-dimensional model construction module matches and optimizes the points in the feature point mapping set based on the bundle network adjustment algorithm, and constructs a preliminary three-dimensional model framework by calculating the intersection relationship of light under different perspectives to ensure the accuracy of the model's geometric structure. The model rendering module extracts rendering materials from three-dimensional images containing texture, color and other information, such as the rough texture of the road surface, the reflective properties of road markings, etc., and assigns these materials to the model framework. Through lighting calculation, texture mapping and other technologies, the model is rendered and processed, and finally a realistic and accurate three-dimensional road model is generated, providing intuitive and effective data presentation for road detection and analysis.

[0051] The drone control device is mainly used to control the flight direction and flight route of the drone. In practice, if the interval between adjacent ideal time periods is too long, the drone control device controls the drone to fly around the area between the adjacent ideal time periods.

[0052] The model information collection and analysis device primarily compares the currently generated 3D model with historical 3D models, noting areas where the 3D models differ. The more areas of difference there are and the larger the volume of the difference, the worse the road health. Conversely, the healthier the road. Thus, road health information is the differential information between two adjacent 3D models.

[0053] refer to Figure 3 , Example 2: Example 2 provides a method for extracting feature points based on Example 1, specifically: The pixel matching module uses the following steps to extract feature points from a 3D image and map them: Step 1: Extract the flight speed of the drone from the 3D video information and divide the 3D video information into several video groups based on the flight speed. Each video group includes video streams captured by three drones in the same time period. In practice, the drone's flight speed data is first obtained from the 3D video information. For example, the drone's GPS module or other speed measurement equipment can be used to record speed changes during flight. Based on flight speed, the 3D video information is divided into several video groups. For example, if the drone flies at a relatively stable speed during a long highway aerial video, the video can be divided into multiple small segments based on speed and time intervals. Each video group contains video streams captured by three drones during the same time period. This division is intended to centrally process videos with similar shooting conditions and scenes, facilitating subsequent feature point extraction and matching.

[0054] Step 2: Extract several sample images at equal intervals from the video stream, extract feature points from the sample images based on the SIFT algorithm, and match the feature points one by one based on cosine distance to generate a sample matching set.

[0055] From the video stream of each video group, several sample images are extracted at equal intervals, based on a fixed time or frame interval. For example, a sample image is extracted every 10 frames. These sample images can, to a certain extent, represent the overall visual features of the video group. The sample images are then processed using the SIFT (Scale-Invariant Feature Transform) algorithm, which detects stable feature points in different scale spaces and generates unique descriptors for each feature point. Feature points in different sample images are then matched one-to-one based on the cosine distance, which measures the similarity between feature point descriptors. Feature points with high similarity are paired to generate a sample matching set. In this way, the feature correspondence between different images is preliminarily determined.

[0056] For example, three aerial photography drones took pictures A, B, and C respectively. The feature points in pictures A, B, and C are extracted respectively, and similarity matching is performed on these feature points. If the similarity exceeds the threshold, it means that these feature points are the same feature points, and then the corresponding mapping relationship is established.

[0057] Step 3: Use the sample matching set to train the convolutional network model and adjust the weight parameters in the convolutional network model.

[0058] The generated sample matching set is used as training data and fed into a convolutional network model. Convolutional networks are powerful deep learning models that automatically learn characteristic patterns in images. During training, the model continuously adjusts its weight parameters based on the feature points in the sample matching set and their matching relationships. For example, through the backpropagation algorithm, the error between the predicted and actual matching results is propagated backwards to update the model weights, enabling the model to better capture the distribution and matching patterns of feature points in the image. After multiple iterations of training, the convolutional network model is gradually optimized, improving the accuracy of image feature point extraction and matching.

[0059] refer to Figure 4 ,Specifically, the convolutional network model includes: input layer, convolution layer, pooling layer, ,probability calculation layer, first attention mechanism network, second ,attention mechanism network and mapping network.

[0060] The input layer is used to input images taken by three aerial photography drones at the same time to generate the first feature map, the second feature map, and the third feature map. The input layer is the entrance to the convolutional network model, and its main function is to receive images taken by three aerial photography drones at the same time.

[0061] The convolution layer performs convolution processing on the input first feature map, second feature map and third feature map to extract the first latent feature, second latent feature and third latent feature.

[0062] The convolution layer, as the core component of the model, uses multiple convolution kernels of varying sizes and numbers (e.g., 3×3 and 5×5 kernels) to convolve the first, second, and third input feature maps. Taking a 3×3 convolution kernel as an example, the kernel slides across the feature map, performing element-wise multiplication and accumulation operations to extract local features from the image. During the convolution process, different strides (e.g., 1 or 2) can be set to control the interval at which the kernel slides, allowing for the extraction of features at different scales.

[0063] The pooling layer, connected to the convolutional layer, is used to pool the first, second, and third latent features. The pooling layer is connected to the convolutional layer and its main function is to average the first, second, and third latent features, reducing the amount of data and computational complexity while retaining key features.

[0064] The probability calculation layer generates the probability information that each pixel belongs to a feature point from the first, second, and third latent features. Based on the pooled first, second, and third latent features, the probability calculation layer uses a fully connected layer and an activation function (such as the Softmax function) to calculate the probability information that each pixel belongs to a feature point.

[0065] The first attention mechanism network inputs the first feature map and probability information to generate the first transformed feature; The second attention mechanism network inputs the second feature map and probability information to generate the second transformed feature; The third attention mechanism network inputs the third feature map and probability information to generate the third transformed feature; The first, second, and third attention networks have the same structure. The attention network first fuses the input feature map and probability information through a linear transformation and an activation function to generate an attention weight matrix. This weight matrix reflects the importance of different locations in the image. Regions with higher weights contain more critical features. The attention weight matrix is then weighted with the first feature map to produce the transformed features.

[0066] The mapping network maps the first, second, and third transformed features to generate feature points and mapping relationships between them. The mapping network takes the first, second, and third transformed features as input and performs nonlinear transformations, such as through a multilayer perceptron. During this process, the network learns the spatial relationships and similarities between features from different perspectives, thereby mapping these transformed features to generate feature points and mapping relationships between them.

[0067] Step 4: Input the remaining images in the video stream into the trained convolutional network model to extract feature points and the mapping relationship between feature points; The remaining images in the video stream, excluding the sample image, are fed into a trained and optimized convolutional network model. The trained model can then use the learned feature patterns to automatically extract feature points in the image and determine the mapping relationships between these feature points.

[0068] Step 5: The mapping relationship between the feature points extracted in step 2 and step 4 is used as a feature point mapping set.

[0069] The feature points and their mapping relationships of the sample images obtained through the SIFT algorithm and cosine distance matching in step 2 are combined with the feature points and mapping relationships of the remaining images extracted by the convolutional network model in step 4. These feature points and mapping relationships are aggregated to form a complete feature point mapping set. This set contains the feature points of each image in the 3D video information and their corresponding relationships, providing an important data foundation for subsequent 3D model construction, enabling the model to accurately restore the 3D structure and characteristics of the road surface.

[0070] Example 3: Example 3 provides a model framework generation solution based on Example 1; The model rendering module generates the model framework based on the following steps: Z1: Obtain the feature point mapping set and the relative position information corresponding to the feature point mapping set, and convert the relative position information into the exterior orientation element E. The exterior orientation element E includes the coordinates of the photography center S in the ground coordinate system (X S 、Y S , Z S ) and the rotation angle of the image coordinate system relative to the ground coordinate system , χ S Indicates the rotation angle of the image coordinate system around the X axis, Indicates the rotation angle of the image coordinate system around the Y axis, Indicates the rotation angle of the image coordinate system around the Z axis; the feature point in the feature point set is used as the connection point i (x ij ,y ij ), j represents the sequence number of the image, j=1, 2, 3.

[0071] The imaging center S is simplified to the coordinates of the drone itself, or in other words, the coordinates of the drone itself are replaced by the imaging center. Although the drone itself is flying steadily, the high-speed camera is tilted downward to achieve tilt measurement. The image number corresponds to the serial number of the drone itself.

[0072] Z2: For each tie point i, based on the exterior orientation elements of image j containing tie point i and the pixel coordinates of tie point i in image j, perform forward intersection using the collinearity equation to solve the initial ground coordinate approximation TP of tie point i ; Z3: According to TP and the exterior orientation elements to construct the parameter vector X k ; ; Among them, k represents the number of iterations, initially k=0, represents the exterior orientation element of the first aerial photography UAV at the kth iteration, represents the exterior orientation element of the third aerial photography UAV at the kth iteration, represents the approximate initial ground coordinate of connection point 1 at the kth iteration, represents the approximate initial ground coordinate of the connection point n at the kth iteration, n represents the total number of connection points, T represents the matrix transpose sign, 1E represents the exterior orientation element E of the first aerial photography UAV, and 3E represents the exterior orientation element E of the third aerial photography UAV; TP1 represents the approximate initial ground coordinate of the first connection point, and TPn represents the approximate initial ground coordinate of the nth connection point; Z4: For each connection point i (x ij ,y ij), extract the current exterior orientation elements of image j and the current coordinates of the connection point i , calculate the predicted image point coordinates ; represents the exterior orientation element of the j-th image at the k-th iteration; ; Where f represents the focal length of the camera, 、 、 、 、 、 、 、 、 Represents the rotation matrix element of the jth image. The two digits in the subscript represent the row and column numbers of the rotation matrix respectively. x0 and y0 represent the pixel coordinates. 、 、 represents the approximate initial ground coordinate of the connection point u at the kth iteration, represents the exterior orientation element of the jth aerial photography UAV at the kth iteration, 、 Indicates the distortion correction value; The observation value is used as the constant term I c , calculate the observation residual l x and l y , arrange the residuals in order to generate the connection point pixel coordinate residual vector I im ; ; ; 、 represents the coordinates obtained by measuring the connection point i, 、 Represents the coordinate value of the theoretical connection point calculated by the collinear equation, represents the pixel coordinates of the connection point i after correction on image j, Represents the residual component of the observation value of the connection point i in the horizontal direction of image j, Represents the residual component of the observation value of the connection point i in the longitudinal direction of image j; Extract the control point g from the feature point set. For each ground control point g, according to the known observation value Generate control point ground coordinate residual vector I ct ; ; Indicates the current approximate coordinate value of the ground control point g, 、 represents the residual of the observation value of the ground control point g, 、 Represents the coordinates of the ground control point g measured, and the residuals are arranged in order to generate a constant vector to form the constant term vector I of the control point ct ; Connect the pixel coordinate residual vector I im and the control point ground coordinate residual vector I ct vertical splicing; Z5: parameter vector X k Perform Taylor series expansion to obtain the linearized error equation of the observation value: ; Among them, v represents the residual vector, which contains the residuals of the coordinate observations of all tie points and control points. represents the correction vector, which contains the correction values of all exterior orientation elements E. Represents the design matrix; each element of the design matrix A is the predicted value of the collinear equation; Z6: Construct the normal equation: ; Where I represents the complete constant vector, A represents the design matrix, T represents the matrix transpose symbol, and P represents the weight matrix of the observation value; Z7: Solve the equation to generate the correction vector , the correction vector Input to parameter vector X k , update the new parameter vector X k+1 ; Based on the new parameter vector X k+1 Generate the geographic coordinates of each connection point and render the model framework based on the geographic coordinates The control points in this solution are points with known specific locations in advance, such as a ruler with known location and size placed on the road surface.

[0073] Example 4: A road health information extraction method based on large model analysis, using the aforementioned road health information extraction system based on large model analysis to extract road health information.

[0074] The above are merely preferred embodiments of the present application and are not intended to limit the present application. Those skilled in the art will readily appreciate that various modifications and variations are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A road health information extraction system based on large model analysis, characterized in that: include: The drone information collection device includes three drones for aerial photography, which are deployed parallel to the road to be measured and fly synchronously along the road to be measured to obtain three-dimensional video information. The three-dimensional video information includes: video data synchronously collected by the three drones; relative position information corresponding to each frame of the video data; the relative position information includes at least the real-time relative position between the three drones; UAV control device: used to control the flight direction and attitude of each aerial photography UAV in real time, guiding it to fly along the predetermined route; 3D modeling device: receives 3D video information in real time and generates a 3D road surface model based on the video data and its corresponding relative position information; Model information collection and analysis device: collects pavement 3D models generated in different time periods, compares and analyzes the current pavement 3D model with the historical pavement 3D model, and extracts and generates pavement health information that characterizes changes in pavement status.

2. The road health information extraction system based on large model analysis according to claim 1 is characterized in that: Aerial photography drones include: The drone body is used to provide an aerial flight platform; A high-speed camera is mounted on the drone body at a fixed angle, tilted downward; The wireless positioning module generates initial relative position parameters with other aerial photography drones through millimeter wave positioning; Laser transmitter, fixedly installed on the drone; The laser receiver forms two laser receiving planes on both sides of the drone body; An information processor is connected to the wireless positioning module and the laser receiver signal to generate relative position information; Among them, the laser transmitter of the aerial photography drone sends a laser signal to the laser receiver of the adjacent aerial photography drone, and the information processor corrects the relative position information based on the position where the laser irradiates the laser receiving plane.

3. The road health information extraction system based on large model analysis according to claim 1 is characterized in that: The aerial photography drones on both sides obtain the relative positions of the aerial photography drone in the middle, and the aerial photography drone in the middle obtains its own current geographical location.

4. The road health information extraction system based on large model analysis according to claim 2 is characterized in that: The drone information collection device also includes: The video information screening module is connected to the aerial photography drone signal and is used to obtain the video data collected synchronously by each aerial photography drone, and to screen out the image frames when the aerial photography drone's flight posture is stable to generate three-dimensional video information.

5. The road health information extraction system based on large model analysis according to claim 4 is characterized in that: The video information screening module includes: A video information acquisition unit is used to acquire the video data collected synchronously by each aerial photography drone, and synchronously arrange the video data acquired by the three aerial photography drones based on the time tags to generate a video frame matrix; The video information screening unit obtains the time period when the laser receiver of each aerial photography UAV receives the laser signal, and defines the time period when all laser receivers receive the laser signal as the ideal time period; The video information cropping unit takes the portion corresponding to the ideal time period in the video frame matrix as the three-dimensional video information.

6. The road health information extraction system based on large model analysis according to claim 1 is characterized in that: The 3D modeling device includes: A pixel matching module is used to extract feature points from 3D video information and map the feature points to generate a feature point mapping set; 3D model building module, which matches feature points based on bundle network adjustment to build the model framework; The model rendering module extracts rendering materials from the three-dimensional image to render the model frame and generate a three-dimensional road model.

7. The road health information extraction system based on large model analysis according to claim 6 is characterized in that: The pixel matching module uses the following steps to extract feature points from a 3D image and map them: Step 1: Extract the flight speed of the drone from the 3D video information and divide the 3D video information into several video groups based on the flight speed. Each video group includes video streams captured by three drones in the same time period. Step 2: Extract several sample images with equal spacing from the video stream, extract feature points from the sample images based on the SIFT algorithm, and perform one-to-one matching of the feature points based on cosine distance to generate a sample matching set; Step 3: Use the sample matching set to train the convolutional network model and adjust the weight parameters in the convolutional network model; Step 4: Input the remaining images in the video stream into the trained convolutional network model to extract feature points and the mapping relationship between feature points; Step 5: The mapping relationship between the feature points extracted in step 2 and step 4 is used as a feature point mapping set.

8. The road health information extraction system based on large model analysis according to claim 7 is characterized in that: Convolutional network models include: The input layer is used to input images taken by three drones at the same time to generate the first feature map, the second feature map, and the third feature map; The convolution layer performs convolution processing on the input first feature map, second feature map, and third feature map to extract the first latent feature, the second latent feature, and the third latent feature; The pooling layer is connected to the convolutional layer and is used to pool the first latent feature, the second latent feature, and the third latent feature; The probability calculation layer generates the probability information of each pixel belonging to a feature point from the first latent feature, the second latent feature, and the third latent feature; The first attention mechanism network inputs the first feature map and probability information to generate the first transformed feature; The second attention mechanism network inputs the second feature map and probability information to generate the second transformed feature; The third attention mechanism network inputs the third feature map and probability information to generate the third transformed feature; The mapping network maps the first transformation feature, the second transformation feature, and the third transformation feature to each other to generate feature points and mapping relationships between the feature points.

9. The road health information extraction system based on large model analysis according to claim 6 is characterized in that: The model rendering module generates the model framework based on the following steps: Z1: Obtain the feature point mapping set and the relative position information corresponding to the feature point mapping set, and convert the relative position information into the exterior orientation element E. The exterior orientation element E includes the coordinates of the photography center S in the ground coordinate system (X S 、Y S , Z S ) and the rotation angle of the image coordinate system relative to the ground coordinate system , χ S Indicates the rotation angle of the image coordinate system around the X axis, Indicates the rotation angle of the image coordinate system around the Y axis, Indicates the rotation angle of the image coordinate system around the Z axis; the feature point in the feature point set is used as the connection point i (x ij ,y ij ), j represents the sequence number of the image, j=1, 2, 3; Z2: For each tie point i, based on the exterior orientation elements of image j containing tie point i and the pixel coordinates of tie point i in image j, perform forward intersection using the collinearity equation to solve the initial ground coordinate approximation TP of tie point i ; Z3: According to TP and the exterior orientation elements to construct the parameter vector X k ; ； Among them, k represents the number of iterations, initially k=0, represents the exterior orientation element of the first aerial photography UAV at the kth iteration, represents the exterior orientation element of the third aerial photography UAV at the kth iteration, represents the approximate initial ground coordinate of connection point 1 at the kth iteration, represents the approximate initial ground coordinate of the connection point n at the kth iteration, n represents the total number of connection points, T represents the matrix transpose sign, 1E represents the exterior orientation element E of the first aerial photography UAV, and 3E represents the exterior orientation element E of the third aerial photography UAV; TP1 represents the approximate initial ground coordinate of the first connection point, and TPn represents the approximate initial ground coordinate of the nth connection point; Z4: For each connection point i (x ij ,y ij ), extract the current exterior orientation elements of image j and the current coordinates of the connection point i , calculate the predicted image point coordinates ; represents the exterior orientation element of the j-th image at the k-th iteration; ； Where f represents the focal length of the camera, 、 、 、 、 、 、 、 、 Represents the rotation matrix element of the jth image. The two digits in the subscript represent the row and column numbers of the rotation matrix respectively. x0 and y0 represent the pixel coordinates. 、 、 represents the approximate initial ground coordinate of the connection point u at the kth iteration, represents the exterior orientation element of the jth aerial photography UAV at the kth iteration, 、 Indicates the distortion correction value; The observation value is used as the constant term I c , calculate the observation residual l x and l y , arrange the residuals in order to generate the connection point pixel coordinate residual vector I im ; ； ； 、 represents the coordinates obtained by measuring the connection point i, 、 Represents the coordinate value of the theoretical connection point calculated by the collinear equation, represents the pixel coordinates of the connection point i after correction on image j, Represents the residual component of the observation value of the connection point i in the horizontal direction of image j, Represents the residual component of the observation value of the connection point i in the longitudinal direction of image j; Extract the control point g from the feature point set. For each ground control point g, according to the known observation value Generate control point ground coordinate residual vector I ct ; ； Indicates the current approximate coordinate value of the ground control point g, 、 represents the residual of the observation value of the ground control point g, 、 Represents the coordinates of the ground control point g measured, and the residuals are arranged in order to generate a constant vector to form the constant term vector I of the control point ct ; Connect the pixel coordinate residual vector I im and the control point ground coordinate residual vector I ct Vertical splicing; combined into a complete constant vector I; Z5: parameter vector X k Perform Taylor series expansion to obtain the linearized error equation of the observation value: ； Among them, v represents the residual vector, which contains the residuals of the coordinate observations of all tie points and control points. represents the correction vector, which contains the correction values of all exterior orientation elements E. Represents the design matrix; each element of the design matrix A is the predicted value of the collinear equation; Z6: Construct the normal equation: ； Where I represents the complete constant vector, A represents the design matrix, T represents the matrix transpose symbol, and P represents the weight matrix of the observation value; Z7: Solve the equation to generate the correction vector , the correction vector Input to parameter vector X k , update the new parameter vector X k+1 ; Based on the new parameter vector X k+1 Generate the geographic location coordinates of each connection point and render the model frame according to the geographic location coordinates.

10. A road health information extraction method based on large model analysis, characterized in that: The pavement health information is extracted using a pavement health information extraction system based on large model analysis according to any one of claims 1 to 9.

Images