A method and system for inversion of a pavement PBR material

Abstract

The present application provides a kind of inversion method and system of road surface PBR material, belong to computer vision and physical simulation field, this method includes: the IMU data, three-dimensional road surface data and two-dimensional road surface texture image space-time alignment are obtained, reconstruct vehicle trajectory;Parameter combination sample point is generated in the multidimensional parameter space of the initial simulation model constructed;Loss value between simulation image and real image is calculated for parameter combination sample point, and dataset is constructed;Parameter combination sample point is rendered using dataset machine learning agent model training, and the model is identified after training, and then sensitivity analysis is carried out to obtain the parameter subset with the strongest coupling effect;Simulation image is rendered using parameter subset, the difference between simulation image and real image is calculated, and parameter subset is updated synchronously, and road surface PBR material graph and system parameter are output when convergence occurs.Based on the method, a kind of inversion system of road surface PBR material is also proposed, and the present application realizes high-fidelity road surface PBR material graph and system parameter output.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision and physical simulation, and particularly relates to a road surface PBR material inversion method and system. BACKGROUND

[0002] PBR material is a material that can realize a rendering algorithm or process close to physical reality. It contains a set of properties of the surface of an object, such as albedo, metallic, roughness, and normal map, etc., which together determine the visual performance of the object under different lighting conditions.

[0003] Automatic visual detection of road surface diseases is a key technology to ensure the safety of transportation infrastructure. The mainstream method has developed from traditional image processing to semantic segmentation based on deep learning. However, the performance and generalization ability of these methods are severely limited by the imaging quality and the diversity of training data. In order to break through this bottleneck, academia and industry have begun to explore the use of physical simulation to generate data or enhance models. This requires us to be able to establish a physically accurate digital twin model for the real acquisition scene. The specific implementation is as follows: place a test object in a closed sphere or dome filled with programmable light sources in a device called "Lighting Stage"; perform a high-precision scan of the three-dimensional geometry of the test object; turn on the light sources in different positions one by one, and take hundreds of high-definition photos under different lighting angles from a fixed camera position; based on the known geometry, known camera pose, and accurate known light source information, the BRDF (PBR material) parameters of each point on the surface of the object are solved by an optimization algorithm. The technical problems of this method include: it is severely dependent on the expensive, static "lighting stage" equipment in the laboratory, and it is completely impossible to apply to outdoor, large-scale, dynamic moving scenes such as road surfaces; it requires taking a large number of (hundreds of) images under different lighting conditions, and the data acquisition and processing time is extremely long, and it is usually assumed that the camera and light source parameters are pre-accurately calibrated and fixed, which cannot handle the problem of uncertain system pose parameters caused by vibration, installation error, etc. in real acquisition. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide a road surface PBR material inversion method and system, which realizes high-fidelity road surface PBR material map and system parameter output.

[0005] In a first aspect, the application provides a road surface PBR material inversion method, comprising: acquiring IMU data, three-dimensional road surface data, and two-dimensional road surface texture images;

[0006] The IMU data, three-dimensional road surface data and two-dimensional road surface texture image are spatio-temporally aligned and fused to obtain the IMU data and three-dimensional road surface data for vehicle trajectory reconstruction;

[0007] An initial simulation model is constructed by parameterizing modeling of a physical system, parameter combination sample points are generated in a multi-dimensional parameter space of the initial simulation model by using a Latin hypercube sampling method, a loss value between a simulation image and a real image is calculated by simulation of each parameter combination sample point, a data set corresponding to the parameter combination sample points and the loss value is constructed, a machine learning agent model is trained by using the data set, a parameter subset with the strongest coupling effect is obtained by identification using the trained machine learning agent model and sensitivity analysis;

[0008] A simulation image is rendered by using the parameter subset, a difference from a realistic image is calculated, the parameter subset is synchronously updated and convergence is judged, and an optimized road PBR material map and system parameters are output.

[0009] In a second aspect, the present application provides a road PBR material inversion system, comprising a data acquisition module, a data processing module, a modeling analysis module and an optimization inversion module.

[0010] The data acquisition module is configured to acquire IMU data, three-dimensional road surface data and two-dimensional road surface texture images.

[0011] The data processing module is configured to spatio-temporally align the IMU data, three-dimensional road surface data and two-dimensional road surface texture image, and fuse the IMU data and three-dimensional road surface data to reconstruct a vehicle trajectory.

[0012] The modeling analysis module is configured to construct an initial simulation model by parameterizing modeling of a physical system, generate parameter combination sample points in a multi-dimensional parameter space of the initial simulation model by using a Latin hypercube sampling method, calculate a loss value between a simulation image and a real image by simulation of each parameter combination sample point, construct a data set corresponding to the parameter combination sample points and the loss value, train a machine learning agent model by using the data set, identify by using the trained machine learning agent model, and obtain a parameter subset with the strongest coupling effect by sensitivity analysis.

[0013] The optimization inversion module is configured to render a simulation image by using the parameter subset, calculate a difference from a realistic image, synchronously update the parameter subset and judge convergence, and output an optimized road PBR material map and system parameters.

[0014] The effects provided in the summary are only the effects of the embodiments, not all the effects of the application, and one of the above technical solutions has the following advantages or beneficial effects:

[0015] This invention proposes an inversion method and system for PBR (Propagation Backplane) material of road surface, belonging to the field of computer vision and physical simulation technology. The method includes the following steps: acquiring IMU data, 3D road surface data, and 2D road surface texture images; spatiotemporally aligning the IMU data, 3D road surface data, and 2D road surface texture images, and fusing the acquired IMU data and 3D road surface data to reconstruct vehicle trajectory; parametrically modeling the physical system to construct an initial simulation model; generating parameter combination sample points in the multidimensional parameter space of the initial simulation model using the Latin hypercube sampling method; performing simulation calculations on each parameter combination sample point to determine the loss value between the simulated image and the real image; constructing a dataset corresponding to the parameter combination sample points and the loss value; training a machine learning proxy model using the dataset, identifying the parameter subset with the strongest coupling effect using the trained machine learning proxy model, and then performing sensitivity analysis; rendering a simulated image using the parameter subset, calculating the difference between the simulated image and the real image, synchronously updating the parameter subset and determining convergence, and outputting the optimized road surface PBR material map and system parameters. Based on this method, an inversion system for pavement PBR material is also proposed. This invention solves the problems of low parameter inversion accuracy, strong coupling and low computational efficiency in traditional methods through a series of steps such as multimodal data acquisition, parametric modeling, sensitivity analysis and closed-loop optimization, and realizes high-fidelity pavement PBR material map and system parameter output. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0017] Figure 1 This is a flowchart of an inversion method for PBR road surface material proposed in Embodiment 1 of the present invention;

[0018] Figure 2 This refers to the road lighting pattern proposed in Embodiment 1 of the present invention;

[0019] Figure 3 This refers to the 3D light stripe visualization image proposed in Embodiment 1 of the present invention;

[0020] Figure 4 This is a schematic diagram of the line scan camera simulation and the actual ground conditions proposed in Embodiment 1 of the present invention;

[0021] Figure 5 This is a schematic diagram of an inversion system for PBR material of road surface proposed in Embodiment 2 of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. The drawing numbers in this application are only used to distinguish the various steps in the solution and are not used to limit the execution order of the various steps. The specific execution order is subject to the description in the specification.

[0023] Example 1

[0024] Embodiment 1 of this invention proposes an inversion method for PBR material of road surface, aiming to solve the technical problem of automated, high-precision, and robust inversion of the physical and optical properties (PBR material) of the road surface and other detection objects and the precise state parameters of the data acquisition system in a complex vehicle-mounted dynamic acquisition environment.

[0025] Figure 1 This is a flowchart of an inversion method for PBR road surface material proposed in Embodiment 1 of the present invention;

[0026] In step S100, the process begins.

[0027] In step S110, multimodal data is acquired simultaneously, including IMU data, three-dimensional road surface data, and two-dimensional road surface texture images.

[0028] Specifically, the vehicle's IMU data is acquired through an inertial measurement unit, and the IMU data includes the vehicle's acceleration and angular velocity.

[0029] Three-dimensional road surface data is acquired using lidar; the three-dimensional road surface data includes road surface depth data and reflectivity data. Figure 2 This refers to the road lighting pattern proposed in Embodiment 1 of the present invention;

[0030] Two-dimensional road surface texture images are acquired using line scanning equipment or area scan cameras.

[0031] In step S120, the IMU data, 3D road surface data, and 2D road surface texture image are spatiotemporally aligned and fused to obtain the IMU data and 3D road surface data for vehicle trajectory reconstruction.

[0032] Specifically, IMU data, 3D road surface information, and 2D road surface texture images are uniformly interpolated onto a preset frequency time axis to achieve timestamp alignment; IMU data (100Hz), LiDAR point cloud (10Hz), and line scan image (20kHz) are uniformly interpolated onto a 100Hz time axis, and bilinear interpolation is used to preserve high-frequency details.

[0033] The vehicle trajectory is reconstructed by fusing IMU data and 3D road surface data using a factor graph optimization method. The initial trajectory is obtained using IMU integration; the LiDAR point cloud is aligned with the initial trajectory, and the pose is optimized using the NDT (Normal Distribution Transform) algorithm; the global trajectory is optimized using factor graphs to eliminate accumulated errors (e.g., trajectory drift <10cm after traveling 1km).

[0034] In step S130, the physical system is parametrically modeled to construct an initial simulation model; the Latin hypercube sampling method is used to generate parameter combination sample points in the multidimensional parameter space of the initial simulation model; the loss value between the simulated image and the real image is calculated for each parameter combination sample point; and a dataset corresponding to the parameter combination sample points and the loss value is constructed.

[0035] An initial simulation model was constructed by parametrically modeling the road surface, lighting system, and camera system. The road surface modeling parameters included height field, roughness, albedo, and metallicity; the lighting system modeling parameters included light source position, light source intensity, beam divergence angle, and polarization degree; and the camera system modeling parameters included intrinsic parameters, extrinsic parameters, exposure time, gain, and ISO value. Figure 3 This is a 3D light stripe visualization image proposed in Embodiment 1 of the present invention.

[0036] The process of generating parameter combination sample points in the multidimensional parameter space of the initial simulation model using the Latin hypercube sampling method includes: dividing the value range of each parameter into a preset number of intervals to ensure that each interval is sampled once; optimizing the distribution of sample points to maximize the minimum distance between samples and generating a preset number of parameter combinations.

[0037] In this application, the value range of each parameter is divided into N intervals (N=100) to ensure that each interval is sampled only once; the distribution of sample points is optimized by the Maximin criterion to maximize the minimum distance between samples and improve spatial coverage; finally, 100 sets of parameter combinations are generated, covering more than 90% of the parameter space variation.

[0038] Pixel loss: ;

[0039] Based on the Unity HDRP rendering pipeline, a custom PBR material shader is used, integrating the road surface height field and BRDF model.

[0040] The ray tracing sampling rate is set to 1024 samples / pixel to ensure the physical correctness of reflections and shadows;

[0041] The output image is 1920×1080 resolution, 32-bit floating-point format (preserving HDR lighting information).

[0042] In step S140, the dataset is used to train the machine learning agent model, the trained machine learning agent model is used for identification, and then sensitivity analysis is performed to obtain the parameter subset with the strongest coupling effect.

[0043] Machine learning agent models employ gradient boosting tree-based machine learning algorithms, such as XGBoost, to simulate behavior at low cost and high cost in physics.

[0044] By analyzing the surrogate model using SHAP values ​​and partial dependency graphs, the main and interaction effects of parameters on the loss value are quantified, and key parameters (such as the influence weight of roughness on reflected light intensity) are identified.

[0045] High-sensitivity parameters (such as light pitch angle, road surface roughness, and camera extrinsic parameters) are retained, while low-impact parameters (such as minor illumination spectral components) are eliminated to form a streamlined set of optimization variables.

[0046] In step S150, a simulated image is rendered using a subset of parameters, the difference between the simulated image and the real image is calculated, the subset of parameters is updated synchronously, convergence is determined, and the optimized road surface PBR material map and system parameters are output. Figure 4 This is a schematic diagram of the line scan camera simulation and the actual ground situation proposed in Embodiment 1 of the present invention.

[0047] By using a simulation-optimization closed loop, the PBR material and system parameters are updated synchronously to minimize the error between the simulated image and the real image.

[0048] Input a subset of parameters into the initial simulation model to render the simulation image;

[0049] A hybrid loss function consisting of pixel loss, structural loss, and perceptual loss is used to calculate the difference between simulated images and realistic images;

[0050] An automatic differentiation and optimization algorithm is used to calculate the gradient of the loss with respect to all key adjustable parameters and update them synchronously.

[0051] Determine if convergence has occurred. If convergence has occurred, output the optimized road surface PBR material map and system parameters.

[0052] Optimization algorithms such as Adam are adopted to automatically differentiate the gradient of loss with respect to key parameters, iteratively update parameters (such as increasing roughness to diffuse the reflected light spot), and simultaneously optimize PBR material (such as albedo and metallicity) and system parameters (such as light position).

[0053] When the loss function drops below the threshold or the number of iterations reaches the upper limit, the optimization is terminated to ensure that the parameter inversion results converge to the vicinity of the global optimum.

[0054] Road surface PBR material map: including albedo map, roughness map, normal map, etc., accurately describes the optical and geometric properties of the road surface;

[0055] System parameters: Optimized lighting system parameters (such as light source angle and intensity) and camera internal and external parameters can be directly used for subsequent simulation or reconstruction tasks.

[0056] In step S150, the process ends.

[0057] The method for inverting PBR material of road surface proposed in Embodiment 1 of this invention, through multimodal data acquisition, integrates IMU, three-dimensional road surface and two-dimensional texture information, and combines factor graph optimization to achieve spatiotemporal alignment, providing an accurate data foundation for parameter inversion; by adopting physical system parameterized modeling and sensitivity analysis based on surrogate model, key parameters can be effectively identified, redundant calculations can be avoided, and inversion accuracy can be improved.

[0058] The inversion method for PBR material of road surface proposed in Embodiment 1 of this invention uses SHAP, partial dependency graph and other methods to analyze the main effect and interaction effect of parameters, screen out the key parameter set with strong coupling effect, and adopt group update strategy and projection operation in joint optimization process to effectively solve the strong coupling problem between parameters and improve inversion efficiency.

[0059] The inversion method for PBR road surface material proposed in Embodiment 1 of this invention introduces Latin hypercube sampling and surrogate models (such as XGBoost), which significantly reduces the number of physical simulations, reduces the time of a single simulation from 3s to 10ms, and shortens the total time of the optimization stage from 24 hours to 4 hours, thereby reducing the computational cost.

[0060] The method for inverting road surface PBR material proposed in Embodiment 1 of this invention outputs a high-fidelity road surface PBR material map and system parameters, which can be widely used in fields such as intelligent driving scenario modeling, road maintenance and inspection, and computer graphics. For example, it can improve the simulation realism of sensors in intelligent driving simulation and improve the crack recognition rate in road defect detection.

[0061] Example 2

[0062] Based on the inversion method for pavement PBR material proposed in Embodiment 1 of the present invention, Embodiment 2 of the present invention proposes an inversion system for pavement PBR material. Figure 5 This is a schematic diagram of an inversion system for PBR material of road surface proposed in Embodiment 2 of the present invention. The system includes: a data acquisition module, a data processing module, a modeling and analysis module, and an optimization inversion module.

[0063] The data acquisition module is used to acquire IMU data, 3D road surface data, and 2D road surface texture images;

[0064] The data processing module is used to spatiotemporally align IMU data, 3D road surface data, and 2D road surface texture images, and fuse the IMU data and 3D road surface data to reconstruct vehicle trajectories.

[0065] The modeling and analysis module is used to parametrically model the physical system and construct an initial simulation model. It uses the Latin hypercube sampling method to generate parameter combination sample points in the multidimensional parameter space of the initial simulation model. For each parameter combination sample point, it performs simulation calculations to determine the loss value between the simulated image and the real image. It then constructs a dataset corresponding to the parameter combination sample points and the loss value. Using this dataset, it trains a machine learning proxy model, identifies the parameters after training, and performs sensitivity analysis to obtain the subset of parameters with the strongest coupling effect.

[0066] The optimization inversion module is used to render a simulation image using the parameter subset, calculate the difference between the simulation image and the real image, update the parameter subset synchronously and determine convergence, and output the optimized road surface PBR material map and system parameters.

[0067] In the data acquisition module, vehicle IMU data is acquired through an inertial measurement unit, including the vehicle's acceleration and angular velocity; three-dimensional road surface data is acquired using a lidar, including road surface depth data and reflectivity data; and two-dimensional road surface texture images are acquired through a line scan device or an area scan camera.

[0068] In the data processing module, IMU data, 3D road surface information, and 2D road surface texture images are uniformly interpolated onto a preset frequency time axis to achieve timestamp alignment; the factor graph optimization method is used to fuse IMU data and 3D road surface data to reconstruct vehicle trajectories.

[0069] In the modeling and analysis module, the process of constructing an initial simulation model for the physical system through parametric modeling includes: constructing an initial simulation model for the road surface, lighting system, and camera system through parametric modeling; among which, the road surface modeling parameters include height field, roughness, albedo, and metallicity; the lighting system modeling parameters include light source position, light source intensity, beam divergence angle, and polarization degree; and the camera system modeling parameters include intrinsic parameters, extrinsic parameters, exposure time, gain, and ISO value.

[0070] The process of generating parameter combination sample points in the multidimensional parameter space of the initial simulation model using the Latin hypercube sampling method includes: dividing the value range of each parameter into a preset number of intervals to ensure that each interval is sampled once; optimizing the distribution of sample points to maximize the minimum distance between samples and generating a preset number of parameter combinations.

[0071] The machine learning agent model employs a gradient boosting tree-based machine learning algorithm.

[0072] The sensitivity analysis process includes: obtaining the subset of parameters with the strongest coupling effect through SHAP value analysis and partial dependency graph analysis.

[0073] In the optimization and inversion module, the process of rendering a simulated image using the parameter subset, calculating the difference between the simulated image and the realistic image, synchronously updating the parameter subset, determining convergence, and outputting the optimized road surface PBR material map and system parameters includes: inputting the parameter subset into the initial simulation model to render the simulated image; using a hybrid loss function consisting of pixel loss, structural loss, and perceptual loss to calculate the difference between the simulated image and the realistic image; using an automatic differentiation and optimization algorithm to calculate the gradient of the loss with respect to all key adjustable parameters and updating them synchronously; determining whether convergence has occurred, and if convergence has occurred, outputting the optimized road surface PBR material map and system parameters.

[0074] The PBR material map includes an albedo map, roughness map, normal map, and metallicity map describing the optical and geometric properties of the road surface; the system parameters include optimized lighting system parameters and optimized camera internal and external parameters.

[0075] The inversion system for PBR road surface material proposed in Embodiment 2 of this invention integrates IMU, 3D road surface and 2D texture information through multimodal data acquisition, and achieves spatiotemporal alignment by combining factor graph optimization, providing an accurate data foundation for parameter inversion; by adopting physical system parameterized modeling and sensitivity analysis based on surrogate model, it can effectively identify key parameters, avoid redundant calculations and improve inversion accuracy.

[0076] The inversion system for PBR road surface material proposed in Embodiment 2 of this invention uses methods such as SHAP and partial dependency graphs to analyze the main effects and interaction effects of parameters, screen out the key parameter set with strong coupling effect, and adopt a group update strategy and projection operation in the joint optimization process to effectively solve the problem of strong coupling between parameters and improve the inversion efficiency.

[0077] The inversion system for PBR road surface material proposed in Embodiment 2 of this invention introduces Latin hypercube sampling and surrogate models (such as XGBoost), which significantly reduces the number of physical simulations, reduces the time of a single simulation from 3s to 10ms, and shortens the total time of the optimization stage from 24 hours to 4 hours, thereby reducing the computational cost.

[0078] The road surface PBR material inversion system proposed in Embodiment 2 of this invention outputs a high-fidelity road surface PBR material map and system parameters, which can be widely used in fields such as intelligent driving scenario modeling, road maintenance and inspection, and computer graphics. For example, it can improve the simulation realism of sensors in intelligent driving simulation and improve the crack recognition rate in road defect detection.

[0079] In the road surface PBR material inversion system provided in Embodiment 2 of this application, the aforementioned modules can also implement the road surface PBR material inversion method steps provided in the above method embodiments. Alternatively, the apparatus provided in this application embodiment may further include other modules besides the aforementioned modules to implement the method steps provided in the above method embodiments. Furthermore, the apparatus provided in this application embodiment can achieve the technical effects achievable by the above method embodiments.

[0080] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0081] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0082] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An inversion method for PBR road surface materials, characterized in that, Includes the following steps: Acquire IMU data, 3D road surface data, and 2D road surface texture images; IMU data, 3D road surface data, and 2D road surface texture images are spatiotemporally aligned and fused to obtain IMU data and 3D road surface data for vehicle trajectory reconstruction; The process of spatiotemporally aligning IMU data, 3D road surface information, and 2D road surface texture images, and fusing them to obtain IMU data and 3D road surface information for trajectory reconstruction includes: IMU data, 3D road surface information, and 2D road surface texture images are uniformly interpolated onto a preset frequency time axis to achieve timestamp alignment; the IMU data and 3D road surface data are fused using a factor graph optimization method to reconstruct the vehicle trajectory. A parametric modeling process is performed on the physical system to construct an initial simulation model. A Latin hypercube sampling method is used to generate parameter combination sample points in the multidimensional parameter space of the initial simulation model. For each parameter combination sample point, the loss value between the simulated image and the real image is calculated using simulation. A dataset corresponding to the parameter combination sample points and the loss value is constructed. A machine learning proxy model is trained using this dataset. The trained machine learning proxy model is then used for identification, and sensitivity analysis is performed to obtain the parameter subset with the strongest coupling effect. The simulation image is rendered using the parameter subset, the difference between it and the real image is calculated, the parameter subset is updated synchronously and convergence is determined, and the optimized road surface PBR material map and system parameters are output; the PBR material map includes an albedo map, roughness map, normal map and metallicity map describing the optical and geometric properties of the road surface; The system parameters include optimized lighting system parameters and optimized camera internal and external parameters.

2. The method as described in claim 1, characterized in that, The process of acquiring IMU data, 3D road surface information, and 2D road surface texture images includes: Vehicle IMU data is acquired through an inertial measurement unit, and the IMU data includes the vehicle's acceleration and angular velocity. Three-dimensional road surface data is acquired using lidar; the three-dimensional road surface data includes road surface depth data and reflectivity data. Two-dimensional road surface texture images are acquired using line scanning equipment or area scan cameras.

3. The method as described in claim 1, characterized in that, The process of constructing an initial simulation model by parametric modeling the physical system includes: constructing an initial simulation model by parametric modeling the road surface, lighting system, and camera system; among which, the road surface modeling parameters include height field, roughness, albedo, and metallicity; the lighting system modeling parameters include light source position, light source intensity, beam divergence angle, and polarization degree; and the camera system modeling parameters include intrinsic parameters, extrinsic parameters, exposure time, gain, and ISO value.

4. The method as described in claim 3, characterized in that, The process of generating parameter combination sample points in the multidimensional parameter space of the initial simulation model using the Latin hypercube sampling method includes: The range of values ​​for each parameter is divided into a preset number of intervals to ensure that each interval is sampled once. The sample point distribution is optimized to maximize the minimum distance between samples and generate a preset number of parameter combinations.

5. The method as described in claim 1, characterized in that, The machine learning agent model employs a gradient boosting tree-based machine learning algorithm.

6. The method as described in claim 5, characterized in that, The sensitivity analysis process includes: obtaining the parameter subset with the strongest coupling effect through SHAP value analysis and partial dependency graph analysis.

7. The method as described in claim 1, characterized in that, The process of rendering a simulated image using the parameter subset, calculating the difference between the simulated image and the realistic image, synchronously updating the parameter subset and determining convergence, and outputting the optimized road surface PBR material map and system parameters includes: The parameter subset is input into the initial simulation model to render the simulation image; The difference between the simulated image and the realistic image is calculated using a hybrid loss function consisting of pixel loss, structural loss and perceptual loss. An automatic differentiation and optimization algorithm is used to calculate the gradient of the loss with respect to all key adjustable parameters and update them synchronously. Determine if convergence has occurred. If convergence has occurred, output the optimized road surface PBR material map and system parameters.

8. An inversion system for pavement PBR material, used to execute the inversion method for pavement PBR material according to any one of claims 1 to 7, characterized in that, include: Data acquisition module, data processing module, modeling and analysis module, and optimization and inversion module; The data acquisition module is used to acquire IMU data, three-dimensional road surface data, and two-dimensional road surface texture images; The data processing module is used to spatiotemporally align IMU data, 3D road surface data, and 2D road surface texture images, and fuse the IMU data and 3D road surface data to reconstruct vehicle trajectories. The modeling and analysis module is used to perform parametric modeling of the physical system and construct an initial simulation model; it uses the Latin hypercube sampling method to generate parameter combination sample points in the multidimensional parameter space of the initial simulation model; it performs simulation calculations on each parameter combination sample point to obtain the loss value between the simulated image and the real image; it constructs a dataset corresponding to the parameter combination sample points and the loss value; it uses the dataset to train a machine learning agent model, uses the trained machine learning agent model for identification, and then performs sensitivity analysis to obtain the parameter subset with the strongest coupling effect; The optimization inversion module is used to render a simulation image using the parameter subset, calculate the difference between the simulation image and the real image, update the parameter subset synchronously and determine convergence, and output the optimized road surface PBR material map and system parameters.

Images