Highway pavement roadbed detection method and system based on laser three-dimensional scanning

Abstract

The invention belongs to the technical field of pavement detection, and discloses a highway pavement roadbed detection method and system based on laser three-dimensional scanning. By establishing an analytic optical or radar scattering physical model, a distance measurement deviation initial value conforming to a multi-path reflection physical mechanism is provided for each laser measurement point, and then refined correction is performed on the initial value through a residual error regression network fusing local point cloud context information, so that advantage complementation of physical prior and data driving is realized, and the accuracy of distance measurement is improved. And finally, in combination with adaptive filtering based on a confidence map, ranging errors caused by a multi-path effect can be effectively distinguished and corrected in the presence of complex road surface conditions such as slippery road surface, greasy dirt or thin water film, and meanwhile, geometric details such as real road surface pits, textures and the like are completely reserved. Therefore, the reconstruction error and the disease false alarm rate of the pavement digital elevation model are obviously reduced, and the precision and the reliability of highway pavement roadbed detection are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of road surface detection, in particular to a highway pavement and subgrade detection method and system based on laser three-dimensional scanning. BACKGROUND

[0002] The rapid and accurate detection and maintenance of highway pavement and subgrade are crucial. Laser three-dimensional scanning technology, especially vehicle-mounted mobile measurement systems, has become a core means of modern highway digitization and disease investigation due to its high precision, high efficiency, and rich data acquisition characteristics. This technology can quickly acquire massive three-dimensional point cloud data by emitting laser pulses to the target surface and receiving the echoes, thereby reconstructing a high-precision digital elevation model of the road surface to identify common diseases such as potholes, ruts, cracks, etc.

[0003] However, traditional laser scanning data processing methods perform well on ideal dry pavements, but when the pavement is wet, oily, or has a thin water film (thickness below millimeters), the laser beam will interact with these media in a complex way. Specifically, part of the laser energy is specularly reflected at the surface of the medium, while the other part of the energy penetrates the medium and is reflected again at the real pavement solid surface. This multi-path reflection effect causes the laser echo signal to be mixed or broadened, introducing additional propagation delay, which ultimately manifests as local height measurement bias in the point cloud data, i.e., abnormal points with false high or false low.

[0004] The existing technology mainly relies on the following two methods to deal with the above problems: (1) Geometric feature-based filtering and denoising algorithm: This method identifies and removes abnormal points by analyzing the geometric distribution characteristics of the point cloud (such as elevation discontinuity, normal vector change, local curvature, etc.). However, since the reflection error caused by the water film may be similar in spatial distribution to the real small potholes and rough texture height, simple geometric filtering cannot effectively distinguish between the two. In the denoising process, it is often unavoidable to smooth out the real road surface details, leading to missed detection of real diseases or retention of some error points, resulting in false positives.

[0005] (2) Simple threshold method based on echo intensity information: This method uses the characteristic that wet surfaces often cause changes in echo intensity to filter out suspicious points by setting an intensity threshold. However, this method is too empirical, and the threshold setting relies on manual work and has poor adaptability, making it difficult to accurately model the complex physical process of multi-path reflection, and the treatment effect is limited for thin water film or oil stain areas with insignificant intensity changes.

[0006] In summary, existing technologies face a core technical bottleneck when dealing with wet and contaminated road surfaces, as they struggle to distinguish between real road surface details and multipath reflection errors. This results in a loss of accuracy in the reconstructed digital elevation model, which in turn affects the accuracy and reliability of subsequent automatic identification of road surface defects, leading to persistently high false alarm and false negative rates.

[0007] Therefore, there is an urgent need for a new detection method that can understand and correct the above-mentioned physical measurement errors in order to achieve high-precision road surface and subgrade detection under various complex road conditions. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for detecting highway pavement and subgrade based on laser three-dimensional scanning, so as to solve the problems mentioned in the background art.

[0009] This invention provides a method for detecting road surface subgrade based on laser three-dimensional scanning, comprising the following steps: Step S10, acquiring the original laser scanning data of the target road surface subgrade, including at least the spatial coordinates, original measurement distance, intensity, echo width, and scanning angle of each laser measurement point; Step S20, based on the scanning angle and intensity of each laser measurement point, calculating the corresponding initial value of the predicted distance deviation for each measurement point using a pre-established analytical optical model or radar scattering model; Step S30, organizing multiple measurement points into a local point cloud block, and storing the spatial coordinates, intensity, and echo width of all laser measurement points within the local point cloud block. The scanning angle and the corresponding initial value of the predicted distance deviation are input into the residual regression network, which predicts a residual correction value for each laser measurement point. In step S40, the original measurement distance of the corresponding laser measurement point is corrected according to the sum of the initial value of the predicted distance deviation and the residual correction value to obtain the corrected measurement distance. In step S50, a three-dimensional point cloud is generated based on the corrected measurement distance of all laser measurement points. Geometric consistency analysis and uncertainty estimation are performed on the three-dimensional point cloud to generate a confidence map. The confidence map is then filtered to reconstruct the digital elevation model of the road surface for roadbed quality assessment.

[0010] This invention also provides a highway pavement and subgrade detection system based on laser three-dimensional scanning. The system includes an acquisition subunit, an estimation and prediction subunit, a residual prediction subunit, a distance correction subunit, and a reconstruction subunit. The acquisition subunit is used to: acquire the original laser scanning data of the target highway pavement and subgrade, including at least the spatial coordinates, original measurement distance, intensity, echo width, and scanning angle of each laser measurement point. The estimation and prediction subunit is used to: calculate the initial value of the predicted distance deviation for each measurement point based on the scanning angle and intensity of each laser measurement point, using a pre-established analytical optical model or radar scattering model. The residual prediction subunit is used to: organize multiple measurement points into a local point cloud block, and within this local point cloud block... The spatial coordinates, intensity, echo width, and scanning angle of all laser measurement points, along with their respective initial predicted ranging deviations, are input into a residual regression network. This network predicts a residual correction value for each laser measurement point. The distance correction subunit is used to correct the original measurement distance of the corresponding laser measurement point based on the sum of the initial predicted ranging deviation and the residual correction value, thus obtaining a corrected measurement distance. The reconstruction subunit is used to generate a three-dimensional point cloud based on the corrected measurement distances of all laser measurement points, perform geometric consistency analysis and uncertainty estimation on the point cloud, generate a confidence map, and filter the confidence map to reconstruct a digital elevation model of the road surface for roadbed quality assessment.

[0011] The present invention also provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the electronic device performs the method as described in any of the preceding claims.

[0012] The present invention also provides a non-volatile storage medium on which a computer program stored is executed by a processor to implement the method as described in the preceding one.

[0013] This invention establishes an analytical optical or radar scattering physical model to provide an initial value for ranging error that conforms to the physical mechanism of multipath reflection for each laser measurement point. Then, a residual regression network that integrates local point cloud context information is used to refine this initial value, achieving a complementary advantage of physical prior and data-driven approaches. Finally, combined with adaptive filtering based on confidence maps, it can effectively distinguish and correct ranging errors caused by multipath effects under complex road surface conditions such as wet, slippery, oily, or thin water films. At the same time, it fully preserves the true geometric details of the road surface, such as potholes and textures, thereby significantly reducing the reconstruction error and false alarm rate of the road surface digital elevation model and improving the accuracy and reliability of highway pavement and subgrade detection. Attached Figure Description

[0014] Figure 1This is a flowchart illustrating a method for detecting highway pavement and subgrade based on laser three-dimensional scanning, as disclosed in an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of the residual regression network disclosed in an embodiment of the present invention.

[0016] Figure 3 This is a schematic diagram of the internal structure of the analytical optical model or radar scattering model disclosed in the embodiments of the present invention.

[0017] Figure 4 This is a schematic diagram of a highway pavement and subgrade detection system based on laser three-dimensional scanning, as disclosed in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides a method 100 for detecting road surface subgrade based on laser three-dimensional scanning, comprising the following steps: Step S10, acquiring the original laser scanning data of the target road surface subgrade, including at least the spatial coordinates, original measurement distance, intensity, echo width, and scanning angle of each laser measurement point; in this step, the original laser scanning data of the target road surface subgrade is acquired through a vehicle-mounted laser three-dimensional scanning system. Specifically, the vehicle-mounted laser three-dimensional scanning system emits laser pulses to the road surface and receives echo signals, recording key parameters such as the spatial coordinates, original measurement distance, intensity, echo width, and scanning angle of each laser measurement point.

[0020] Among them, spatial coordinates are used to locate the three-dimensional position of the laser measurement point, the original measurement distance is the initial observation value from the sensor to the target surface, the intensity reflects the reflection characteristics of the target surface, the echo width characterizes the time broadening of the laser pulse reflection on the target surface, and the scanning angle describes the angle between the laser beam and the road surface normal.

[0021] Step S20: Based on the scanning angle and intensity of each laser measurement point, calculate the initial value of the predicted ranging deviation for each measurement point using a pre-established analytical optical model or radar scattering model. This step addresses the ranging error caused by laser multipath reflection under complex road surface conditions such as wet, slippery, oily, or thin water film. It uses a pre-established analytical optical model or radar scattering model to calculate a preliminary estimate of the ranging deviation for each laser measurement point from a physical perspective, i.e., to predict the initial value of the ranging deviation.

[0022] Specifically, the calculations are based on an analytical optics model: This model, based on classical geometric optics and thin-film interference theory, treats the road surface thin film (such as a thin water film) as an ideal optical layering medium. The core of the model is to describe the multiple reflections and transmissions of the laser beam at the upper and lower interfaces of the thin film. These behaviors generate multiple beams with a fixed optical path difference, which is converted into a phase difference or time delay of the echo signal at the sensor receiving end, thus causing ranging errors.

[0023] The calculation process involves the model taking the scanning angle and echo intensity of the laser measurement point as input. The scanning angle is used to calculate the actual optical path length and incident angle, while the echo intensity is related to the Fresnel reflection formula and used to estimate optical parameters such as the refractive index of the thin film. Finally, the model solves the interference equation of the multilayer medium to accurately calculate the expected ranging deviation caused by the theoretical optical path difference, and uses this as the initial value for predicting the ranging deviation.

[0024] Calculation based on radar scattering model: This model is based on electromagnetic wave scattering theory and focuses on the statistical description of the interaction between laser (as an electromagnetic wave) and rough road surface. It regards the road surface as a scatterer with specific dielectric properties and roughness, and the laser echo is the result of incoherent scattering of electromagnetic waves with the surface.

[0025] The calculation process involves using the scanning angle and echo intensity as inputs. The scanning angle directly affects the scattering cross-section, while the echo intensity is closely related to the surface dielectric constant (affected by humidity and oil) and roughness. By establishing empirical or semi-empirical relationships between these observed parameters and the scattering characteristics under typical road surface conditions (such as wetness and oil), the model can predict the average ranging deviation caused by changes in dielectric properties and scattering paths, and use this as the initial value for predicting the ranging deviation.

[0026] It should be noted that both the analytical optics model based on wave optics and the radar scattering model based on electromagnetic scattering utilize physical priors to make an initial, interpretable estimate of the systematic error caused by multipath effects.

[0027] Step S30: Organize multiple measurement points into a local point cloud block, and input the spatial coordinates, intensity, echo width, and scanning angle of all laser measurement points in the local point cloud block, as well as their respective initial values ​​of the predicted ranging deviation, into the residual regression network. The residual regression network then predicts a residual correction value for each laser measurement point. In this step, the local point cloud block and its corresponding initial value of the predicted ranging error are input into the residual regression network, which then predicts the residual correction value for each measurement point.

[0028] Specifically, the residual regression network takes the spatial coordinates, intensity, echo width, scanning angle, and initial predicted ranging bias of all measurement points within a local point cloud block as input features. It extracts spatial geometric correlation features and reflection characteristic correlation features through graph convolutional layers or point cloud feature extraction layers, and fuses these features with physical priors. Finally, it outputs the residual correction value for each point through a fully connected layer. Understandably, this residual regression network is trained using a loss function that simultaneously constrains geometric continuity and reflection consistency, ensuring that it maintains realistic road surface details while correcting multipath errors.

[0029] Please refer to Figure 2 The residual regression network includes an input layer, a graph construction layer, a hierarchical feature extraction backbone network, and a feature fusion and output layer, as follows: 1. Input layer.

[0030] Input: A local point cloud patch consisting of N points. Normalize the input features so that their mean is 0 and their variance is 1.

[0031] 2. Graph construction layer.

[0032] Construct a k-nearest neighbor graph for the input point cloud patch. Each point is treated as a node in the graph and connected by edges to its k nearest neighbors in Euclidean space (e.g., k=16 or 32). This graph structure defines the topology for information propagation in the network.

[0033] 3. Hierarchical feature extraction backbone network.

[0034] It consists of multiple cascaded graph attention convolutional modules, used to extract multi-level features.

[0035] Each graph attention convolutional module includes: edge feature encoding: for center point i and its neighboring points j, calculate edge feature e. ij It includes relative coordinates (pj-pi), Euclidean distance, etc.

[0036] Attention mechanism: Calculate the attention coefficient α between the center point and each neighbor point. ij This coefficient indicates how much neighboring point j contributes to the feature update of center point i. This allows the residual regression network to adaptively focus on more important neighbors (e.g., points belonging to the same smooth surface) while suppressing interference from outliers or irrelevant points.

[0037] Feature aggregation: Using the calculated attention coefficient as weights, the features of all neighboring points are summed in a weighted manner, and the features of the center point are updated.

[0038] The hierarchical structure of this backbone network is as follows: First layer: With a small receptive field, it learns fine-grained local features, such as subtle smoothness between points and changes in normal vectors. Middle layers: As the layers deepen, through repeated aggregation, the receptive field of each point gradually expands, enabling it to capture a wider range of geometric structures (such as the overall shape of a tiny pit). Last layer: Outputs point features F rich in contextual information. final It contains both the point's own properties and encodes the geometric and reflection characteristics of its surrounding local area.

[0039] 4. Feature fusion and output layer.

[0040] Skip connections: To preserve the fine details of the original points, the network boosts the initial input features (or features from a certain intermediate layer) through a shared multilayer perceptron, and then combines them with the final extracted features F. final Then, the parts are assembled.

[0041] Final feature F fused =Concat(MLP(F input ),F final ).

[0042] Residual prediction: The fused features F fused Through another shared multilayer perceptron, the MLP acts as a regressor, ultimately outputting a scalar value, which is the residual correction value δ for that point.

[0043] Step S40: Based on the sum of the initial value of the predicted ranging deviation and the residual correction value, the original measurement distance of the corresponding laser measurement point is corrected to obtain the corrected measurement distance. In this step, the initial value of the predicted ranging deviation of each measurement point is added to the residual correction value predicted by the network to obtain the total correction amount of the measurement point. Then, the total correction amount is subtracted from the original measurement distance to obtain the corrected measurement distance.

[0044] Understandably, this step integrates physical priors with data-driven corrections, which can significantly reduce systematic errors caused by multipath effects.

[0045] Step S50: Generate a three-dimensional point cloud based on the corrected measurement distance of all laser measurement points, perform geometric consistency analysis and uncertainty estimation on the three-dimensional point cloud, generate a confidence map, and filter the confidence map to reconstruct a digital elevation model of the road surface for use in road surface and subgrade quality assessment.

[0046] In this step, geometric consistency indices such as the consistency of the normal vector distribution and the smoothness of elevation changes of each measurement point in the local neighborhood are calculated. Uncertainty scores are also calculated by combining the magnitude of the residual correction value and the spatial dispersion of the neighborhood point cloud, and a comprehensive confidence map is generated.

[0047] Furthermore, adaptive filtering is applied to the point cloud based on the aforementioned comprehensive confidence map to simultaneously suppress noise and preserve realistic road surface details. Finally, a digital elevation model of the road surface is reconstructed based on the filtered point cloud for roadbed quality assessment.

[0048] This invention establishes an analytical optical or radar scattering physical model to provide an initial value for ranging error that conforms to the physical mechanism of multipath reflection for each laser measurement point. Then, a residual regression network that integrates local point cloud context information is used to refine this initial value, achieving a complementary advantage of physical prior and data-driven approaches. Finally, combined with adaptive filtering based on confidence maps, it can effectively distinguish and correct ranging errors caused by multipath effects under complex road surface conditions such as wet, slippery, oily, or thin water films. At the same time, it fully preserves the true geometric details of the road surface, such as potholes and textures, thereby significantly reducing the reconstruction error and false alarm rate of the road surface digital elevation model and improving the accuracy and reliability of highway pavement and subgrade detection.

[0049] As an example, based on the scanning angle and intensity of each laser measurement point, a pre-established analytical optical model or radar scattering model is used to calculate the initial value of the predicted ranging deviation for each measurement point, including: Step S21: Based on the scanning angle, intensity and echo width of the laser measurement point, the initial equivalent optical thickness is obtained by inversion through the first sub-module of the analytical optical model or radar scattering model. The initial equivalent optical thickness is then corrected by using the intensity ratio of the first echo and the last echo originating from the same laser pulse at the laser measurement point, thus obtaining the final equivalent optical thickness. In this step, based on three key parameters—scanning angle, intensity, and echo width—the initial thickness is inverted through the first submodule. Specifically, the scanning angle is used to determine the actual propagation path and incident angle of the laser in the thin film medium, which are the fundamental geometric parameters for calculating the optical path. For example, when the scanning angle is 30 degrees, the actual propagation path length of the laser in the thin film will be the thin film thickness divided by the refractive index and then divided by cos30°.

[0050] The intensity parameter is then correlated with the refractive properties of the thin film through Fresnel's law of reflection. Taking a water film as an example, its refractive index is approximately 1.33. According to Fresnel's formula, the reflectivity is approximately 2.1% when the incident angle is 30 degrees. The difference between this theoretical value and the measured intensity can be used to infer the film thickness.

[0051] The echo width directly characterizes the propagation time delay of a laser pulse in a layered medium. For example, for a water film with a thickness of 1 mm, the round-trip propagation of the laser within it will produce a time delay of approximately 9 picoseconds, resulting in broadening of the echo signal.

[0052] Next, the intensity ratio of the first and last echoes of the same laser pulse is introduced to accurately correct the initial inversion results. Here, the same laser pulse refers to a complete laser energy packet emitted in the same direction within a single emission cycle by the lidar sensor. Laser energy attenuates as it penetrates the thin film. The thicker the film, the longer the path the laser energy travels, resulting in more absorption and scattering. Consequently, the intensity of the last echo (from the actual road surface) is weaker than the first echo (from the upper surface of the film). Therefore, the intensity ratio changes monotonically with increasing film thickness. For example, on a 0.5mm thick water film road surface, the intensity of the first echo is typically 3-5 times that of the last echo; this intensity ratio has a definite quantitative relationship with the film thickness.

[0053] During the correction process, a pre-defined calibration function weights and fuses the film thickness corresponding to the intensity ratio with the initial thickness estimate. For example, if the initial inversion thickness is 0.6 mm, while the intensity ratio indicates a thickness of 0.5 mm, the calibration function automatically adjusts the weights based on the difference between the two, outputting the corrected thickness value. It is understood that this calibration function is based on extensive experimental data, covering different film thicknesses from 0.1 mm to 2 mm, ensuring accurate thickness estimates under various practical operating conditions.

[0054] It should be noted that you should refer to [link / reference]. Figure 3 The first submodule is the core component of the physical model (i.e., the analytical optics model or the radar scattering model) responsible for accurately estimating the equivalent optical thickness of the road surface film. Its design goal is to integrate multi-source, heterogeneous observation data and solve the complex problem of thickness inversion by combining data-driven approaches with physical laws. The first submodule can be implemented as a trained machine learning model (such as a gradient boosting decision tree, neural network, etc.) or a physical empirical model, which learns the mapping relationship between the scanning angle, intensity, and echo width and the initial equivalent optical thickness from a large amount of training data with known thicknesses.

[0055] Step S22: Input the equivalent optical thickness and scanning angle into the second sub-module of the analytical optical model or radar scattering model, calculate the additional optical path difference caused by the multipath effect based on the multilayer medium interference principle, and convert the additional optical path difference into the initial value of the predicted ranging deviation.

[0056] In this step, the final equivalent optical thickness and scanning angle obtained above are input into the second submodule of the physical model to accurately calculate the initial value of the ranging deviation. The specific calculation process includes: the second submodule establishes an accurate optical model of laser reflection at the upper and lower interfaces of the thin film based on the principle of multilayer dielectric interference.

[0057] For example, assuming the measured equivalent optical thickness is 1 mm, the scanning angle is 45 degrees, and the film refractive index is 1.33 (water film); according to the law of refraction, the refraction angle θ' = arcsin(sin45° / 1.33) ≈ 32°; the theoretical optical path difference ΔR = (1.33 × 1 mm) / cos32° ≈ 1.57 mm. Using the principle of thin-film interference, the optical path difference is converted into a ranging deviation, resulting in an actual ranging deviation δ = ΔR / 2 ≈ 0.785 mm. This indicates that in the presence of a 1 mm water film, the laser rangefinder will produce a measurement deviation of approximately 0.785 mm.

[0058] It should be noted that further reading is required. Figure 3 The second submodule is the deterministic calculation component in the physical model responsible for converting physical thickness into ranging deviation. Essentially, it is a physical formula solver, the core of which is a mathematical model embedded with the principle of multilayer dielectric interference (e.g., based on Fresnel equations and interference optical path difference formulas). This mathematical model accurately describes the phase interference effect caused by the different propagation paths of the two reflected beams after the laser is reflected at the upper and lower interfaces of the thin film.

[0059] This embodiment constructs a two-level physical modeling architecture of equivalent optical thickness inversion-correction and multi-medium interferometry calculation, and innovatively introduces the multi-echo intensity ratio as direct physical evidence for thickness correction, achieving accurate estimation of the optical thickness of the road surface thin film. Then, based on the principle of multi-medium interferometry, the thickness parameter is accurately converted into an initial value for ranging deviation. This significantly improves the prediction accuracy of the physical model for systematic ranging deviations caused by multipath reflections under complex road surface conditions such as wet, slippery, oily, and thin water films, providing high-quality physical priors for the subsequent residual network.

[0060] As an example, the spatial coordinates, intensity, echo width, and scanning angle of all laser measurement points within the local point cloud block, along with their corresponding initial predicted ranging deviations, are input into a residual regression network. This residual regression network then predicts a residual correction value for each laser measurement point. The process includes: Step S31, constructing a multi-dimensional feature vector set from the spatial coordinates, intensity, echo width, and scanning angle of each laser measurement point within the local point cloud block, along with their corresponding initial predicted ranging deviations. In this step, the multi-source heterogeneous data of all laser measurement points within the local point cloud block are standardized and vectorized to construct a structured input suitable for neural network processing. The specific implementation process includes: reorganizing the features of each laser measurement point within the local point cloud block, normalizing the five dimensions of features: spatial coordinates (X, Y, Z), intensity, echo width, scanning angle, and initial predicted ranging deviation. These five feature parameters of each measurement point are then combined into a feature vector P. i =[X i ,Y i ,Zi,I iW i ,θ i biasphys i ], where i represents the index of the point in the local block. The feature vectors of all measured points together constitute the multidimensional feature vector set P={P1,P2,...,P} of the local point cloud block. n}, where n is the number of measurement points in the local point cloud block.

[0061] Step S32: The residual regression network, based on the multidimensional feature vector set, extracts the spatial geometric correlation features and reflection characteristic correlation features between all laser measurement points in the local point cloud block through at least one graph convolutional layer or point cloud feature extraction layer. This step extracts complex correlation features within the point cloud block using a deep learning network. The specific implementation process includes: the residual regression network employs a graph convolutional neural network (GCN) or PointNet++, a point cloud-specific architecture, to construct a k-nearest neighbor graph on the local point cloud block. For each center point, the graph convolutional layer aggregates the feature information of its neighboring points and performs feature transformation through a learnable weight matrix.

[0062] In spatial geometric correlation feature F geom In terms of extraction, the residual regression network learns geometric constraints such as the relative positional relationships between points, normal vector consistency, and local curvature. For example, it calculates the relative coordinate difference ΔP between pairs of points. ij =P i -P j By combining distance weights, geometric features characterizing the continuity and smoothness of the local surface are learned.

[0063] In the reflection property related feature F refl In terms of extraction, the residual regression network analyzes reflection characteristics such as intensity correlation, echo width distribution pattern, and scanning angle variation trend among neighboring points. For example, for wet areas, the residual regression network learns the correlation pattern between intensity attenuation and echo width increase; for real craters, the residual regression network identifies the consistency characteristics of intensity abrupt changes and geometric abrupt changes.

[0064] Step S33: Integrate the spatial geometric correlation features, the reflection characteristic correlation features, and the initial value of the predicted ranging deviation for each laser measurement point, and output the residual correction value corresponding to each laser measurement point through regression using at least one fully connected layer.

[0065] In this step, the spatial geometric correlation feature F geom Reflection characteristics related feature F refl The vector F is concatenated with the initial predicted ranging bias value (biasphys) of each point to form the enhanced feature vector. fused =[F geom ,F refl[biasphys]. The fused enhanced feature vector undergoes nonlinear transformation and dimensionality compression through a fully connected layer. The fully connected layer introduces nonlinearity through an activation function, gradually mapping the high-dimensional features to the residual correction value space. Finally, the residual regression network outputs the residual correction value δ for each point through a regression layer. i .

[0066] It should be noted that during training, the residual regression network optimizes its parameters by minimizing a loss function that considers both geometric continuity constraints and reflection consistency constraints. The geometric continuity constraint ensures that the corrected point cloud maintains a smooth surface manifold structure, while the reflection consistency constraint guarantees that the corrected result conforms to the laws of physical reflection. This dual constraint mechanism enables the residual regression network to accurately distinguish between real road surface details and reflection errors, performing targeted corrections only on outliers.

[0067] This embodiment constructs a multi-dimensional feature vector set that integrates physical priors and multi-source observation features. It then uses graph convolutional networks or PointNet++ to extract spatial geometric and reflection characteristics of local point cloud blocks. Finally, through a dual-constraint fusion regression mechanism, the residual regression network can fully learn local context information, thereby achieving accurate prediction of physical model residuals. It specifically corrects abnormal ranging errors caused by multipath effects while preserving the geometric details of the real road surface, which is beneficial to improving the overall accuracy and reliability of laser ranging under complex road surface conditions.

[0068] As an example, the original measurement distance of the corresponding laser measurement point is corrected based on the sum of the initial predicted ranging deviation and the residual correction value to obtain the corrected measurement distance. This includes step S41, adding the initial predicted ranging deviation of each laser measurement point to the residual correction value predicted for that point to obtain the total correction amount for that laser measurement point. In this step, for each laser measurement point, its corresponding initial predicted ranging deviation bias is algebraically added to the residual correction value δ predicted by the residual regression network to obtain the total correction amount Δtotal for that measurement point. i .

[0069] That is, Δtotal i =biasphys i +δ i .

[0070] Understandably, bias represents a preliminary estimate of systematic bias based on optical principles, while δ represents a fine-tuning of the physical model residuals based on data-driven methods. The sum of the two constitutes a complete estimate of the ranging bias.

[0071] Step S42: Subtract the total correction amount from the original measurement distance of the laser measurement point to obtain the corrected measurement distance of the laser measurement point.

[0072] In this step, the original measured distance includes positive or negative deviations caused by multipath effects. By subtracting the estimated total deviation, a measurement value closer to the actual road surface distance can be obtained. Specifically, the original measured distance Rraw from the laser measurement point... i Subtract the aforementioned total correction amount Δtotal i Obtain the corrected measurement distance Rcorrected at that point. i That is, R corrected i =Rraw i -Δtotal i .

[0073] As an example, geometric consistency analysis and uncertainty estimation are performed on the 3D point cloud to generate a confidence map. Based on this confidence map, filtering is applied to reconstruct the road surface digital elevation model. This includes step S51: For the corrected 3D point cloud, the geometric consistency index and uncertainty score of each laser measurement point within its local neighborhood are calculated. The geometric consistency index is determined by analyzing the consistency of the normal vector distribution or the smoothness of elevation changes in the neighborhood point cloud. The uncertainty score is determined by the magnitude of the residual correction value output by the residual regression network and the spatial dispersion of the neighborhood point cloud. This step assesses the quality of the corrected 3D point cloud, quantifying the reliability of each laser measurement point using dual indices. The specific implementation process is as follows: In terms of geometric consistency analysis, the geometric consistency index of each laser measurement point within its k nearest neighbors is calculated. The geometric consistency index is determined through two complementary methods: first, analyzing the consistency of the normal vector distribution of the neighborhood point cloud, and evaluating the smoothness of the local surface by calculating the standard deviation of the angle between the normal vectors. For example, on a continuous road surface, the normal vector changes gently and the standard deviation is small; second, calculating the smoothness of the elevation change within the neighborhood, and evaluating it by fitting the local surface and calculating the root mean square of the elevation residual. Real road surfaces exhibit lower elevation residuals, while noisy areas exhibit larger elevation fluctuations.

[0074] For uncertainty scoring, a comprehensive evaluation method based on the magnitude of residual correction values ​​and spatial dispersion is adopted. The magnitude of the residual correction value directly reflects the degree of deviation between the laser measurement point and the physical model; a larger correction magnitude implies higher uncertainty. Simultaneously, the spatial dispersion of the point cloud within the neighborhood of the laser measurement point is calculated, and the flatness of the local surface is calculated through eigenvalue analysis. Regions with high dispersion indicate uneven point cloud distribution, and their reliability is correspondingly lower.

[0075] Step S52: Integrate the geometric consistency index and uncertainty score of each laser measurement point to generate a comprehensive confidence map; wherein, each confidence level in the confidence map represents the reliability of the corrected measurement distance of the corresponding laser measurement point. In this step, the geometric consistency index and uncertainty score are first normalized, and then a weighted fusion algorithm is used to synthesize the two indices into a comprehensive confidence value for each point. The weighting coefficients are determined experimentally based on the actual application scenario. The confidence value ranges from 0 to 1, with a higher value indicating more reliable corrected measurement distance for that point.

[0076] The generation of the confidence map takes spatial continuity into account. Gaussian filtering is used to smooth the initial confidence scores, eliminating isolated outlier values ​​and ensuring a continuous confidence distribution across adjacent areas. The final confidence map is stored in matrix form, with each element corresponding to the confidence value of a laser measurement point.

[0077] Step S53: Adaptive filtering is performed on the 3D point cloud based on the confidence map. Specifically: for regions with confidence scores below a first preset threshold, filtering with noise suppression as the main focus is performed; for regions with confidence scores above a second preset threshold, filtering with structure preservation as the main focus is performed. This step involves intelligent adaptive filtering of the 3D point cloud based on the confidence map. The specific implementation process is as follows: For regions in the confidence map with a confidence level below a first preset threshold (e.g., 0.3), they are identified as low-quality point cloud regions, and a filtering strategy primarily focused on noise suppression is adopted. Specifically, a strong bilateral filter or statistical filter can be used, and by increasing the filter kernel size in the spatial and color domains, noise points can be effectively removed while preserving edge features.

[0078] For regions in the confidence map with a confidence level higher than the second preset threshold (e.g., 0.7), they are identified as high-quality point cloud regions, and a filtering strategy that prioritizes structure preservation is adopted. Specifically, a weaker guided filtering or anisotropic diffusion filtering can be used to remove minor noise while preserving point cloud details by reducing the filtering intensity.

[0079] In addition, for the region in the confidence graph where the confidence level is between two thresholds, adaptive filtering parameters can be used to calculate the filtering strength by linear interpolation based on the confidence level value, so as to achieve a smooth transition. No specific limitations are imposed.

[0080] Step S54: The road surface digital elevation model is obtained based on the filtered 3D point cloud reconstruction.

[0081] In this step, the filtered 3D point cloud is meshed, and an initial triangular mesh model is generated using Delaunay triangulation. Then, a confidence-weighted surface fitting algorithm is used to optimize the triangular mesh, more strictly preserving the elevation information of the original point cloud in high-confidence regions, while allowing appropriate smoothing in low-confidence regions. Finally, the optimized triangular mesh model is converted into a regular grid digital elevation model, and the elevation values ​​of the grid nodes are filled using bilinear interpolation.

[0082] The reconstructed digital elevation model can be directly used for roadbed quality assessment tasks such as pavement smoothness analysis and rut depth detection.

[0083] Please see Figure 4 This invention also provides a highway pavement and subgrade detection system 200 based on laser three-dimensional scanning. The system includes an acquisition subunit 2001, an estimation and prediction subunit 2002, a residual prediction subunit 2003, a distance correction subunit 2004, and a reconstruction subunit 2005. The acquisition subunit 2001 is used to acquire the original laser scanning data of the target highway pavement and subgrade, including at least the spatial coordinates, original measurement distance, intensity, echo width, and scanning angle of each laser measurement point. The estimation and prediction subunit 2002 is used to calculate the initial value of the predicted distance deviation for each measurement point based on the scanning angle and intensity of each laser measurement point, using a pre-established analytical optical model or radar scattering model. The residual prediction subunit 2003 is used to organize multiple measurement points... The system is configured to: generate a local point cloud block; and input the spatial coordinates, intensity, echo width, and scanning angle of all laser measurement points within the local point cloud block, along with their respective initial predicted ranging deviations, into a residual regression network. This residual regression network then predicts a residual correction value for each laser measurement point. The distance correction subunit 2004 is used to: correct the original measurement distance of the corresponding laser measurement point based on the sum of the initial predicted ranging deviation and the residual correction value, thereby obtaining a corrected measurement distance. The reconstruction subunit 2005 is used to: generate a three-dimensional point cloud based on the corrected measurement distances of all laser measurement points; perform geometric consistency analysis and uncertainty estimation on the three-dimensional point cloud; generate a confidence map; and perform filtering based on the confidence map to reconstruct a digital elevation model of the road surface for roadbed quality assessment.

[0084] As an example, the estimation and prediction subunit 2002 is specifically used to: based on the scanning angle, intensity, and echo width of the laser measurement point, invert the initial equivalent optical thickness through the first submodule of the analytical optical model or radar scattering model, and correct the initial equivalent optical thickness using the intensity ratio of the first echo and the last echo originating from the same laser pulse at the laser measurement point, to obtain the final equivalent optical thickness; input the equivalent optical thickness and scanning angle to the second submodule of the analytical optical model or radar scattering model, calculate the additional optical path difference caused by the multipath effect based on the multilayer medium interference principle, and convert the additional optical path difference into the initial value of the predicted ranging deviation.

[0085] As an example, the residual prediction subunit 2003 is specifically used to: construct a multi-dimensional feature vector set from the spatial coordinates, intensity, echo width, and scanning angle of each laser measurement point within the local point cloud block, as well as the corresponding initial value of the predicted ranging deviation; the residual regression network extracts the spatial geometric correlation features and reflection characteristic correlation features between all laser measurement points in the local point cloud block based on the multi-dimensional feature vector set through at least one graph convolutional layer or point cloud feature extraction layer; and fuses the spatial geometric correlation features, the reflection characteristic correlation features, and the initial value of the predicted ranging deviation for each laser measurement point, and outputs the residual correction value corresponding to each laser measurement point through at least one fully connected layer.

[0086] As an example, the distance correction subunit 2004 is specifically used to: add the initial value of the predicted ranging deviation of each laser measurement point to the residual correction value predicted for the point to obtain the total correction amount of the laser measurement point; and subtract the total correction amount from the original measurement distance of the laser measurement point to obtain the corrected measurement distance of the laser measurement point.

[0087] As an example, the reconstruction subunit 2005 is specifically used for: calculating the geometric consistency index and uncertainty score of each laser measurement point in its local neighborhood for the corrected 3D point cloud; wherein, the geometric consistency index is determined by analyzing the consistency of the normal vector distribution or the smoothness of the elevation change of the neighborhood point cloud, and the uncertainty score is determined by the magnitude of the residual correction value output by the residual regression network and the spatial dispersion of the neighborhood point cloud; fusing the geometric consistency index and the uncertainty score of each laser measurement point to generate a comprehensive confidence map; wherein, each confidence level in the confidence map represents the reliability of the corrected measurement distance of the corresponding laser measurement point; performing adaptive filtering on the 3D point cloud based on the confidence map, specifically: for areas with confidence levels below a first preset threshold, performing filtering mainly for noise suppression; for areas with confidence levels above a second preset threshold, performing filtering mainly for structure preservation; and reconstructing the road surface digital elevation model based on the filtered point cloud.

[0088] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the electronic device performs the method as described in any of the preceding claims.

[0089] This invention also provides a non-volatile storage medium on which a computer program stored is executed by a processor to implement the method described in any of the preceding claims.

[0090] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for detecting highway pavement and subgrade based on laser three-dimensional scanning, characterized in that, The method includes the following steps: Step S10, acquiring the original laser scanning data of the target highway roadbed, including at least the spatial coordinates, original measurement distance, intensity, echo width, and scanning angle of each laser measurement point; Step S20, based on the scanning angle and intensity of each laser measurement point, calculating the corresponding initial value of the predicted ranging deviation for each measurement point using a pre-established analytical optical model or radar scattering model; Step S30, organizing multiple measurement points into a local point cloud block, and storing the spatial coordinates, intensity, echo width, and scanning angle of all laser measurement points within the local point cloud block, along with their corresponding initial values ​​of the predicted ranging deviation. The initial value of the distance measurement deviation is input into the residual regression network, which predicts a residual correction value for each laser measurement point. In step S40, the original measurement distance of the corresponding laser measurement point is corrected according to the sum of the predicted initial value of the distance measurement deviation and the residual correction value to obtain the corrected measurement distance. In step S50, a three-dimensional point cloud is generated based on the corrected measurement distance of all laser measurement points. Geometric consistency analysis and uncertainty estimation are performed on the three-dimensional point cloud to generate a confidence map. The confidence map is then filtered to reconstruct the digital elevation model of the road surface for roadbed quality assessment.

2. The method for detecting highway pavement and subgrade based on laser three-dimensional scanning according to claim 1, characterized in that: Based on the scanning angle and intensity of each laser measurement point, the initial value of the predicted ranging deviation is calculated for each measurement point using a pre-established analytical optical model or radar scattering model. This includes: Step S21, based on the scanning angle, intensity, and echo width of the laser measurement point, the initial equivalent optical thickness is obtained by inversion using the first sub-module of the analytical optical model or radar scattering model, and the initial equivalent optical thickness is corrected by using the intensity ratio of the first echo and the last echo originating from the same laser pulse at the laser measurement point, to obtain the final equivalent optical thickness; Step S22, the equivalent optical thickness and scanning angle are input to the second sub-module of the analytical optical model or radar scattering model, and the additional optical path difference caused by the multipath effect is calculated based on the multilayer medium interference principle, and the additional optical path difference is converted into the initial value of the predicted ranging deviation.

3. The method for detecting highway pavement and subgrade based on laser three-dimensional scanning according to claim 1, characterized in that: The spatial coordinates, intensity, echo width, and scanning angle of all laser measurement points within the local point cloud block, along with their respective initial predicted ranging deviations, are input into a residual regression network. This residual regression network then predicts a residual correction value for each laser measurement point. The process includes: Step S31, constructing a multidimensional feature vector set from the spatial coordinates, intensity, echo width, and scanning angle of each laser measurement point within the local point cloud block, along with their corresponding initial predicted ranging deviations; Step S32, the residual regression network extracts spatial geometric correlation features and reflection characteristic correlation features between all laser measurement points in the local point cloud block based on the multidimensional feature vector set through at least one graph convolutional layer or point cloud feature extraction layer; Step S33, fusing the spatial geometric correlation features, the reflection characteristic correlation features, and the initial predicted ranging deviation for each laser measurement point, and outputting the residual correction value for each laser measurement point through at least one fully connected layer.

4. The method for detecting highway pavement and subgrade based on laser three-dimensional scanning according to claim 1, characterized in that: The original measurement distance of the corresponding laser measurement point is corrected based on the sum of the initial value of the predicted ranging deviation and the residual correction value to obtain the corrected measurement distance. This includes: step S41, adding the initial value of the predicted ranging deviation of each laser measurement point to the residual correction value predicted for the point to obtain the total correction amount for the laser measurement point; step S42, subtracting the total correction amount from the original measurement distance of the laser measurement point to obtain the corrected measurement distance for the laser measurement point.

5. The method for detecting highway pavement and subgrade based on laser three-dimensional scanning according to claim 1, characterized in that: Geometric consistency analysis and uncertainty estimation are performed on the 3D point cloud to generate a confidence map. Based on this confidence map, filtering is applied to reconstruct the road surface digital elevation model. This includes step S51: For the corrected 3D point cloud, the geometric consistency index and uncertainty score of each laser measurement point within its local neighborhood are calculated. The geometric consistency index is determined by analyzing the consistency of the normal vector distribution or the smoothness of elevation changes in the neighborhood point cloud. The uncertainty score is determined by the magnitude of the residual correction value output by the residual regression network and the spatial dispersion of the neighborhood point cloud. Step S5 2. Integrate the geometric consistency index and uncertainty score of each laser measurement point to generate a comprehensive confidence map; wherein, each confidence level in the confidence map represents the reliability of the corrected measurement distance of the corresponding laser measurement point; Step S53, perform adaptive filtering on the three-dimensional point cloud based on the confidence map, specifically: for areas with confidence levels below a first preset threshold, perform filtering with noise suppression as the main focus; for areas with confidence levels above a second preset threshold, perform filtering with structure preservation as the main focus; Step S54, reconstruct the road surface digital elevation model based on the filtered point cloud.

6. A highway pavement and subgrade inspection system based on laser three-dimensional scanning, characterized in that: The system includes an acquisition subunit, an estimation and prediction subunit, a residual prediction subunit, a distance correction subunit, and a reconstruction subunit. The acquisition subunit is used to: acquire the original laser scanning data of the target highway pavement subgrade, including at least the spatial coordinates, original measurement distance, intensity, echo width, and scanning angle of each laser measurement point. The estimation and prediction subunit is used to: calculate the initial value of the predicted distance deviation for each measurement point based on the scanning angle and intensity of each laser measurement point, using a pre-established analytical optical model or radar scattering model. The residual prediction subunit is used to: organize multiple measurement points into a local point cloud block, and reconstruct the spatial coordinates, intensity, and echo width of all laser measurement points within this local point cloud block. The degree, echo width, and scanning angle, along with their respective initial predicted ranging deviations, are input into a residual regression network, which predicts a residual correction value for each laser measurement point. The distance correction subunit is used to correct the original measurement distance of the corresponding laser measurement point based on the sum of the initial predicted ranging deviation and the residual correction value, thereby obtaining a corrected measurement distance. The reconstruction subunit is used to generate a three-dimensional point cloud based on the corrected measurement distances of all laser measurement points, perform geometric consistency analysis and uncertainty estimation on the three-dimensional point cloud, generate a confidence map, and filter the confidence map to reconstruct a digital elevation model of the road surface for roadbed quality assessment.

7. A highway pavement and subgrade inspection system based on laser three-dimensional scanning according to claim 6, characterized in that: The estimation and prediction subunit is specifically used for: based on the scanning angle, intensity, and echo width of the laser measurement point, inverting the initial equivalent optical thickness through the first submodule of the analytical optical model or radar scattering model, and correcting the initial equivalent optical thickness using the intensity ratio of the first echo and the last echo originating from the same laser pulse at the laser measurement point, to obtain the final equivalent optical thickness; inputting the equivalent optical thickness and scanning angle into the second submodule of the analytical optical model or radar scattering model, calculating the additional optical path difference caused by the multipath effect based on the multilayer medium interference principle, and converting the additional optical path difference into the initial value of the predicted ranging deviation.

8. A highway pavement and subgrade inspection system based on laser three-dimensional scanning according to claim 6, characterized in that: The residual prediction subunit is specifically used to: construct a multi-dimensional feature vector set from the spatial coordinates, intensity, echo width, and scanning angle of each laser measurement point in the local point cloud block and the corresponding initial value of the predicted ranging deviation; the residual regression network extracts the spatial geometric correlation features and reflection characteristic correlation features between all laser measurement points in the local point cloud block based on the multi-dimensional feature vector set through at least one graph convolutional layer or point cloud feature extraction layer. By integrating the spatial geometric correlation features, the reflection characteristic correlation features, and the initial value of the predicted ranging deviation for each laser measurement point, the residual correction value corresponding to each laser measurement point is output through regression using at least one fully connected layer.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it causes the electronic device to implement the method as described in any one of claims 1-5.

10. A non-volatile storage medium, characterized in that: The computer program stored thereon is executed by a processor to implement the method as described in any one of claims 1-5.

Images