A method for self-adaptive extraction and precision self-evaluation of road cross section based on vehicle-mounted laser point cloud

Abstract

The present application relates to the technical field of road survey, in particular to a kind of road cross section self-adapting extraction and precision self-evaluation method based on vehicle-mounted laser point cloud, multiple survey point cloud data are dynamically filtered by space distribution law model, then the static point cloud data obtained after filtering is registered and spliced to carry out vectorization curve fitting, and the road geometric model generated can more accurately reflect the actual morphological characteristics of road.On this basis, by setting slice plane on the road center line and extracting cross section model, while the intersection point cloud is projected to the plane to form cross section slice point cloud, so that the extraction process of cross section has high adaptability, and can be accurately cut for any interested position on the road center line.Further, by calculating the distance statistics between road cross section model and cross section slice point cloud as precision evaluation index, the precision of extracted road cross section model is objectively, quantitatively and automatically evaluated.

Description

Technical Field

[0001] This invention relates to the field of road measurement technology, specifically to a method for adaptive extraction and accuracy self-evaluation of road cross-sections based on vehicle-mounted laser point clouds. Background Technology

[0002] During road construction and renovation, detailed road data is essential, with cross-sections being a key component. Cross-section measurement involves obtaining the natural undulations of the original ground relative to a center stake on both sides of the road, perpendicular to the route direction. This data is crucial for calculating construction quantities. Traditional manual measurement methods using total stations require measuring a cross-section at least every 20 meters to meet accuracy requirements, resulting in a massive workload.

[0003] Currently, some existing technologies incorporate laser point cloud analysis into road surveying. LiDAR scans the road, and point cloud modeling is used to obtain a mathematical model of the road, which in turn extracts the road's cross-section. However, in practical applications, when scanning roads using vehicle-mounted LiDAR, obstacles (such as vehicles and pedestrians) often obstruct the point cloud data, resulting in the inclusion of points that do not belong to the road, and also causing the loss of point clouds in some areas of the road. Because obstacles have complex and varied shapes, they are difficult to automatically identify using fixed thresholds or patterns. During the road model building process, these non-road noises need to be manually removed, which still requires a significant amount of manpower in large-scale road surveying work. Summary of the Invention

[0004] To address the technical problem of existing laser point cloud modeling techniques requiring manual handling of obstacle interference, this application provides an adaptive extraction and accuracy self-evaluation method for road cross-sections based on vehicle-mounted laser point clouds, comprising the following steps:

[0005] Multiple laser scans are performed on the same road area to obtain multiple sets of point cloud data;

[0006] A spatial distribution pattern model is used to dynamically filter multiple rounds of point cloud data, while retaining static point cloud data.

[0007] Multiple static point cloud data obtained after filtering are registered and stitched together to form overall point cloud data. Vectorized curve fitting is then performed on the overall point cloud data to generate a road geometric model.

[0008] The road centerline is obtained based on the road geometry model. A slicing plane is set at the target position on the road centerline. The road cross-section model is extracted. The point cloud data that intersects with the slicing plane in the overall point cloud data is projected onto the slicing plane to form a cross-section slicing point cloud.

[0009] Calculate the distance statistics between the road cross-section model and the cross-section slice point cloud, and use the distance statistics as the accuracy evaluation index of the road cross-section model.

[0010] The technical effects and advantages of the invention are as follows: By introducing redundant data in both time and space dimensions and combining it with a spatial distribution pattern model for dynamic filtering, dynamic interference factors in the point cloud, such as moving vehicles and pedestrians, can be effectively eliminated, thus significantly improving the purity of the static point cloud data and laying a high-quality data foundation for subsequent road geometry model construction. After registering and stitching together static point cloud data from multiple measurements, vectorized curve fitting is performed, and the resulting road geometry model more accurately reflects the actual morphological characteristics of the road. Based on this, by setting a slicing plane on the road centerline and extracting the cross-sectional model, while simultaneously projecting intersecting point clouds onto this plane to form a cross-sectional slice point cloud, the cross-sectional extraction process becomes highly adaptive, enabling precise cutting at any target location of interest on the road centerline. Furthermore, by calculating the distance statistics between the road cross-section model and the cross-section slice point cloud as an accuracy evaluation index, an objective, quantitative, and automated evaluation of the accuracy of the extracted road cross-section model is achieved without relying on additional manual comparison or external reference data. The entire process, from data acquisition, processing, model construction to accuracy evaluation, forms a complete closed loop, which not only improves the efficiency and accuracy of road cross-section extraction but also provides reliable technical support and quality assurance for the design, construction, and maintenance of road engineering. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the overall process of the method provided by the present invention.

[0012] Figure 2 This is a partial view of the overall point cloud data obtained by the present invention.

[0013] Figure 3 This is a partial view of the overall point cloud data after completion according to the present invention. Detailed Implementation

[0014] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. 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.

[0015] This invention proposes an adaptive extraction and accuracy self-evaluation method for road cross-sections based on vehicle-mounted laser point clouds. In the field of road surveying and design, cross-section data is the core foundational data for calculating earthwork volume, designing road alignment, and evaluating pavement smoothness. However, traditional measurement methods such as total stations and RTK (real-time dynamic differential positioning) often rely on manual point marking, which is not only extremely labor-intensive but also poses significant safety hazards on operational roads with high traffic volume. In recent years, vehicle-mounted mobile mapping systems (MMS) have been able to acquire high-density 3D point clouds of roads with extremely high efficiency by integrating 3D LiDAR, Global Navigation Satellite System (GNSS), and Inertial Navigation System (INS). However, during the scanning process, vehicle-mounted LiDAR inevitably scans dynamic obstacles (such as moving vehicles and pedestrians) and static obstacles (such as illegally parked vehicles and roadblocks), which generate a large amount of noise in the point cloud data. At the same time, occlusion by obstacles can lead to large-area gaps in the road surface point cloud. Directly extracting cross-sections from point clouds containing noise and missing data can lead to distorted cross-sectional contours and severely inaccurate calculation results. Furthermore, existing technologies lack a comprehensive self-evaluation mechanism for cross-section extraction accuracy, often relying on manual sampling, which makes it difficult to guarantee overall quality in large-scale data processing.

[0016] Based on this, embodiments of the present invention provide an adaptive extraction and accuracy self-evaluation method for road cross-sections based on vehicle-mounted laser point clouds, to solve the problem that existing laser point cloud modeling techniques require manual handling of obstacle interference. (Refer to...) Figure 1 The method includes the following steps:

[0017] S1. Perform multiple laser scans on the same road area to obtain multiple measurement point cloud data;

[0018] Due to the fixed viewing angle and random traffic flow, a single laser scan inevitably results in severe blind spots caused by dynamic vehicles and pedestrians, making the point cloud data acquired in a single scan physically incomplete. By conducting multiple scans at different time periods or from different viewing angles, the redundancy in both the temporal and spatial dimensions can be utilized to provide a rich data foundation for subsequent dynamic filtering and blind spot filling.

[0019] S2. Use a pre-trained spatial distribution model to dynamically filter multiple rounds of point cloud data, retaining static point cloud data;

[0020] S3. Register the multiple static point cloud data obtained after filtering, stitch them together to form overall point cloud data, perform vectorized curve fitting on the overall point cloud data, and generate a road geometric model.

[0021] S4. Obtain the road centerline based on the road geometry model, set a slicing plane at the target location of interest on the road centerline of the road geometry model, extract the road cross-section model, and project the point cloud data that intersects with the slicing plane in the overall point cloud data onto the slicing plane to form a cross-section slicing point cloud.

[0022] S5. Calculate the distance statistics between the road cross-section model and the cross-section slice point cloud, and use the distance statistics as the accuracy evaluation index of the road cross-section model.

[0023] The beneficial effect of the above method lies in its ability to effectively eliminate dynamic interference factors in point clouds, such as moving vehicles and pedestrians, by introducing redundant data in both time and space dimensions and combining them with a spatial distribution pattern model for dynamic filtering. This significantly improves the purity of static point cloud data and lays a high-quality data foundation for subsequent road geometry model construction. After registering and stitching together multiple measured static point cloud data sets, vectorized curve fitting is performed, resulting in a road geometry model that more accurately reflects the actual morphological characteristics of the road. Furthermore, by setting a slicing plane on the road centerline and extracting a cross-sectional model, while simultaneously projecting intersecting point clouds onto this plane to form a cross-sectional slice point cloud, the cross-sectional extraction process becomes highly adaptive, enabling precise cutting at any point of interest on the road centerline. Furthermore, by calculating the distance statistics between the road cross-section model and the cross-section slice point cloud as an accuracy evaluation index, an objective, quantitative, and automated evaluation of the accuracy of the extracted road cross-section model is achieved without relying on additional manual comparison or external reference data. The entire process, from data acquisition, processing, model construction to accuracy evaluation, forms a complete closed loop, which not only improves the efficiency and accuracy of road cross-section extraction but also provides reliable technical support and quality assurance for the design, construction, and maintenance of road engineering.

[0024] Specifically, in step S1, a single laser scan inevitably results in severe blind spots caused by dynamic vehicles and pedestrians due to the fixed viewing angle and randomly occurring traffic flow. The point cloud data acquired in a single scan is physically incomplete. By performing multiple scans at different time periods or from different viewing angles, the redundancy in the time and spatial dimensions can be utilized to provide a rich data foundation for subsequent dynamic filtering and blind spot filling.

[0025] Specifically, in step S1, multiple scans of the road are performed under the same lighting conditions as much as possible to ensure that the intensity information of the acquired laser point cloud remains consistent. For example, the acquisition vehicle equipped with lidar first completes the first scan by traveling at a constant speed along the forward lane of the road, acquiring the first round of point cloud data. Subsequently, after a preset interval (e.g., 15 to 30 minutes), the acquisition vehicle travels at a constant speed again along the reverse lane of the road to complete the second scan, acquiring the second round of point cloud data. The advantage of this implementation method is that it not only utilizes the time difference to expose most of the road surface that was originally obscured by dynamic vehicles during the second scan, but also provides opposite scanning perspectives by traveling in both directions, which can effectively eliminate unidirectional scanning shadows on the back side of three-dimensional structures such as curbs and medians. The second implementation method is multi-vehicle platooning cooperative scanning in the same direction. In scenarios such as highways or urban expressways with extremely high traffic volume and difficulty in reverse travel, this invention uses two or more acquisition vehicles equipped with the same license plate, forming a platoon with a fixed longitudinal spacing (e.g., 50 meters) and lateral stagger (e.g., traveling in the first lane and the third lane respectively) to travel in the same direction. The vehicle in front acquires the first set of point cloud data, and the vehicle behind acquires the second set. Because there is a physical distance between the vehicles, the relative position of the same vehicle changes when it is scanned by the vehicle in front and the vehicle behind. This allows for the rapid acquisition of multiple set of point cloud data with spatiotemporal differences within a short period of time, greatly improving the data acquisition efficiency.

[0026] After acquiring multiple rounds of point cloud data, the present invention performs step S2, which uses a spatial distribution law model to dynamically filter the multiple rounds of point cloud data and retain the static point cloud data.

[0027] Specifically, the spatial distribution pattern model is obtained through the following steps:

[0028] S21. Perform voxelization on each round of point cloud data using the same voxel partitioning method. Voxel feature values ​​containing point cloud data are recorded as the first value (e.g., binary 1), and voxel feature values ​​without point cloud data are recorded as the second value (e.g., binary 0). This voxelization process can transform complex point cloud data into a binary mask model, thereby clearly identifying inconsistent voxels (i.e., point clouds) across multiple rounds of point cloud data.

[0029] S22. Compare the feature values ​​of each voxel in any two rounds of point cloud data. When the feature values ​​of the corresponding voxels in the two rounds of point cloud data are different (i.e., the voxel is hit by a laser in one round, while the voxel is empty in the other round), it indicates that the object at that spatial location is dynamically changing, and both voxels are marked as pending. When the feature values ​​of the two voxels are the same (i.e., the voxel is occupied by the point cloud in both rounds), it indicates that the object at that spatial location remains stationary during multiple scans, and both voxels are marked as static voxels.

[0030] S23. Extract the feature vector of each static voxel, train a neural network model, and when the error between the feature vector of the voxel reconstructed by the neural network model and the feature vector of the static voxel is less than the accuracy threshold, output the neural network model as a spatial distribution law model.

[0031] Specifically, the neural network model can employ a 3D sparse autoencoder. This autoencoder comprises an encoder network and a decoder network. Static voxels and their neighborhoods (e.g., ...) are extracted. The local geometric features (such as point density, normal vector variance, and local curvature) within the neighborhood are used to form the feature vector input to the encoder. The encoder compresses the feature vector into a low-dimensional latent space through multiple layers of 3D convolution, extracting the essential geometric distribution patterns of the static road structure; the decoder then maps the latent space vector back to the original feature space through 3D deconvolution. During training, the mean squared error (MSE) is used as the loss function, i.e., the optimization objective is to minimize the following formula:

[0032]

[0033] The network weights are obtained, where, As input features, For reconstruction features.

[0034] The model training is complete when the reconstruction error on the validation set stabilizes and falls below the preset accuracy threshold. Since the training data comes purely from voxels that are identified as static backgrounds, the model implicitly learns the spatial distribution patterns of normal road surfaces (such as smoothness and continuity).

[0035] On the other hand, neural network models can also employ Generative Adversarial Networks (GANs). The local spatial distribution of static voxels is input as real data to the discriminator, while the generator generates a fake voxel distribution. Through alternating adversarial training, the discriminator can eventually determine with extreme precision whether the input voxel blocks conform to the spatial distribution characteristics of a "real static road." After training, the discriminator network can be extracted and used as a model of spatial distribution patterns.

[0036] After obtaining the spatial distribution pattern model, this invention performs dynamic filtering through the following steps:

[0037] S24. The voxels marked as undetermined are screened using a pre-trained spatial distribution model. For the autoencoder in the first implementation, the features of the undetermined voxels are input into the model. If the reconstruction error is less than a threshold, it indicates that the geometric features of the undetermined voxel conform to the pattern of a static road (it may be because it was blocked in a certain scan, but it is still a road). In this case, it is re-marked as a static voxel. If the reconstruction error is huge, it indicates that it belongs to a chaotic dynamic obstacle. For the discriminator in the second implementation, the undetermined voxels are input. If the output confidence is greater than a threshold, it is re-marked as a static voxel.

[0038] S25. For each measured point cloud data, strictly retain the original point cloud corresponding to the static voxel, and directly delete the point cloud in the voxel marked as pending, so as to obtain clean static point cloud data that removes all non-road noise.

[0039] After dynamically filtering the original multi-review point cloud data, step S3 is executed to register the multiple static point cloud data obtained after filtering, stitch them together to form overall point cloud data, and then perform vectorized curve fitting on the overall point cloud data to generate a road geometric model.

[0040] Specifically, in step S3, registration is performed based on the scanning path trajectory data acquired synchronously during laser scanning, combined with the geometric features of the point cloud. The specific process is as follows:

[0041] S31. During laser scanning, the scanning path trajectory data is acquired simultaneously. This trajectory data is calculated in real time by the vehicle-mounted POS system (integrating RTK and high-precision IMU) and includes the vehicle's three-dimensional coordinates, heading angle, pitch angle and roll angle at each time stamp.

[0042] S32. Based on the scan path trajectory data, perform coordinate transformation on multiple measured point cloud data and project them onto the same global engineering coordinate system (such as CGCS2000);

[0043] Let the local coordinates of the measuring instrument be... The rotation matrix provided by the trajectory Translation vector The initial global coordinates are .

[0044] Due to the cumulative error in the POS system, there are still slight misalignments between the multiple measured point clouds after the initial projection.

[0045] S33. Use the static voxels in each measured point cloud data as registration primitives to construct a static voxel feature set. Use the ICP algorithm (Iterative Closest Point) to perform iterative coarse registration, and then use the original points in the voxels to perform fine registration.

[0046] The ICP algorithm can be a standard point-to-point ICP algorithm used to extract the centroid set from the static voxel set of the source point cloud. Extracting the centroid set from the target point cloud Using KD-trees in sets Search The nearest point The matching point pairs are constructed. The exact rotation matrix and translation vector that minimize the error function are obtained through singular value decomposition. This process is iterated until the change in the exact rotation matrix and translation vector is less than a set minimum value. This method is suitable for urban road environments with rich structural features (such as lampposts, guardrails, and traffic signs).

[0047] The ICP algorithm can also employ a point-to-surface ICP algorithm. Given that highways or rural roads often lack significant elevation features and are primarily composed of flat surfaces, point-to-point matching is prone to slippage errors. Therefore, principal component analysis (PCA) is first used to calculate the values ​​of each point in the target point cloud. Local normal vector Then optimize the error function, that is, minimize the vertical projection distance from the source point to the tangent plane of the target point.

[0048] The process of fine registration using original points within voxels is as follows: During the process of uniformly downsampling the original point cloud to obtain a voxel grid, the original point cloud within each voxel is retained for subsequent fine registration. The transformation matrix obtained from the initial registration is applied to the source point cloud, and a KD-tree is used to search for corresponding matching point pairs in the target point cloud. The optimal transformation matrix is ​​then solved, and the optimization is iterated until the termination condition is met, completing the fine registration and obtaining the final aligned vehicle-mounted laser point cloud. This method has higher registration convergence speed and alignment accuracy in road scenes with simple structures.

[0049] After completing the registration of the point cloud data using the above method, the overall point cloud data is obtained, and the generation of the road geometry model begins. The specific steps include:

[0050] S34. Identify and classify key feature point sets from the overall point cloud data, including road boundary points, lane line points, and main road surface points;

[0051] Using the reflection intensity information of LiDAR, high-intensity points are clustered and extracted as lane line points; using the elevation change and normal vector change features of the point cloud, the curb edge is extracted as the road boundary point; the remaining flat areas are identified as the main points of the road surface.

[0052] S35. Vector curve fitting is performed on the road boundary points and the lane line points respectively. After removing error points by using the least squares method combined with the Random Sampling Consensus Algorithm (RANSAC), a smooth road outline and lane lines are fitted using cubic spline curves.

[0053] S36. Reconstruct the surface of the main points of the road surface, and use the non-uniform rational B-spline (NURBS) technology to approximate the surface of the road surface points to generate a road surface model with high-order continuity.

[0054] S37. The fitted road contour line, the lane line and the reconstructed road surface model are topologically integrated to obtain a road geometric model containing complete geometry and topological relationships.

[0055] Specifically, the cross-section of a road is a cross-section perpendicular to the road centerline. Therefore, the cross-sectional model of the road is obtained based on the road centerline. In step S4, the steps for obtaining the road centerline include:

[0056] S41. Perform parameterized discrete sampling based on the generated road contour lines (usually including the left and right contour lines), and obtain dense sampling points according to a set step size (e.g., every 0.5 meters).

[0057] S42. For each sampling point on any road contour line (e.g., the left contour line), calculate the intersection point of the normal of the sampling point and another road contour line (the right contour line), calculate the center point of the line connecting the intersection point and the sampling point, and form a discrete sequence of center points.

[0058] S43. Perform B-spline curve fitting on the center point sequence, assuming the center point sequence is... By defining a set of node vectors, basis functions are constructed using the Cox-de-Boor recurrence relation. Fit the parameterized road centerline:

[0059]

[0060] In the formula, To control the vertices, this mathematical process ensures the centerline of the road. Continuity eliminates centerline inflection points or fluctuations caused by local point cloud noise.

[0061] Specifically, in step S4, the step of setting the slice plane at the location of interest on the road centerline (e.g., at station K1+100) includes:

[0062] S44. Calculate the tangent vector of the road centerline at the slice point, and establish the slice plane using the slice point and the tangent vector.

[0063] Calculate the road centerline at this slice point. tangent vector By establishing a slicing plane that is strictly perpendicular to the tangent vector through this point, its spatial plane equation can be expressed as:

[0064]

[0065] Based on the obtained slice plane, the cross-sectional slice point cloud and the road cross-section model can be obtained:

[0066] S45. Using the slice plane as the center, extract local point clouds from the overall point cloud data that are less than the slice projection threshold (e.g., 0.05 meters) from the slice plane. For example, the KD tree spatial indexing method can be used. For each extracted point, use the orthographic projection formula to calculate its two-dimensional coordinates on the slice plane. The set of these two-dimensional points constitutes the cross-sectional slice point cloud representing the real scan data.

[0067] S46. Using the tangent plane, Boolean operations are performed directly on the road geometry model in three-dimensional space to find the intersection and cut. The intersection line generated by the cut is the road cross-section model in a purely mathematical sense.

[0068] The above steps complete the extraction of the cross-sectional model (estimated value) and the slice point cloud (measured value). Then, step S5 can be executed to evaluate the accuracy of the road cross-sectional model.

[0069] Specifically, the distance between each point in the statistical slice point cloud and the road cross-section model is calculated, and the distance statistics are used as an accuracy evaluation index for the road cross-section model.

[0070] Specifically, the mean Euclidean distance, maximum distance, or standard deviation of the distance can be used as accuracy evaluation indicators. The mean distance reflects the systematic deviation level of the cross-sectional model from the actual point cloud as a whole. The maximum distance is significant in engineering applications because it assesses whether there are local abnormal protrusions or depressions (such as potholes or unfiltered small stones) causing model fitting distortion. It represents the worst-case fitting situation for that cross-section and is an important red-line indicator for engineering safety assessment. The standard deviation evaluation indicator reflects the dispersion of the error distribution. If the standard deviation is extremely small, it indicates that the cross-sectional model fits very uniformly at all locations; if the standard deviation is large, it indicates that the model fits very well in some areas but has a large offset in others, which usually means that the original point cloud distribution at that cross-section is extremely uneven or has complex topological distortions.

[0071] Furthermore, during application, this invention discovered that in certain special scenarios, such as a large bus parked on the roadside, no matter how many scans are performed, a large physical blind spot always exists in that area of ​​the road surface. For example... Figure 2As shown, after the dynamic filtering in step S2 and the registration in step S3, the resulting overall point cloud data has large areas of missing data. If these missing areas are not filled in, the algorithm will forcibly stretch the surface during subsequent meshing or fitting, resulting in severe deformation of the geometric model. To solve this key technical problem, the method provided in this application also includes the following steps:

[0072] S6. Perform voxelization on the overall point cloud data. When a voxel does not contain point cloud data, the voxel is marked as a blank voxel.

[0073] S7. Extract the neighboring elements around each blank voxel using the spatial convolution operator (e.g., ... The feature vectors of non-blank voxels (including local surface normal vector distribution, rate of curvature change, elevation gradient, etc.) within the window are analyzed using the spatial distribution law model trained earlier. When the feature vectors of adjacent voxels conform to the spatial distribution law of the road structure (for example, they all show the same slope and extremely low curvature planar extension), the blank voxel is identified as the missing area to be filled. If the surrounding environmental features are messy, they are not filled to prevent the introduction of false features.

[0074] S8. For each missing region, the feature vectors of the voxels adjacent to the missing region (as contextual conditions) are input into the spatial distribution model. The encoder of the model compresses these vectors into the latent space, and then the point cloud generation module in the decoder (such as a fully connected network layer containing a multilayer perceptron or a folded network) directly maps the latent space features into three-dimensional spatial coordinates. Based on the average intensity of adjacent voxels, the corresponding reflection intensity is predicted, thereby generating complete point cloud data with coordinates and intensity within the missing region. This deep learning-based generation, compared to traditional linear interpolation, can better preserve the camber, longitudinal slope, and texture features of the road surface.

[0075] S9. Calculate the geometric continuity index at the boundary between the completed point cloud data and the point cloud of adjacent voxels. If the geometric continuity index does not reach the preset threshold, use the geometric continuity index as the optimization target to iteratively optimize the spatial distribution law model until the geometric continuity index of the generated completed point cloud data reaches the preset threshold.

[0076] The geometric continuity index is composed of the deviation of the normal vector angle, the elevation step difference, and the continuity of the second derivative. If, after calculation, the geometric continuity index at the boundary does not reach the preset smoothing threshold (e.g., the normal vector angle difference is greater than 3 degrees, or the elevation jump is greater than 2 centimeters), it indicates that the generated point cloud has edge discontinuities. In this case, the system automatically uses this geometric continuity index as a new optimization objective (the penalty term of the Loss Function) and performs fine-tuning iterative optimization of the latent variables or local weights of the spatial distribution law model through the backpropagation algorithm. After each iteration, the completed point cloud data is regenerated and edge detection is performed again until the generated completed point cloud data smoothly transitions at the boundary, and the geometric continuity index fully reaches the preset threshold. Figure 3 The overall point cloud data after registration shows that large areas of missing point cloud data have been effectively filled in. This iterative and optimized data completion not only effectively eliminates holes and discontinuities in the original point cloud data caused by occlusion, scanning blind spots, or noise interference, but also ensures that the completed point cloud data maintains a high degree of geometric consistency with the original point cloud in terms of overall shape and local details. This data completion strategy provides a high-quality, high-completeness point cloud data source for subsequent adaptive extraction of road cross-sections, enabling the extraction process to more accurately reflect the actual three-dimensional shape of the road and reduce extraction errors caused by missing or anomalies.

[0077] Through the detailed implementation steps described above, this invention completely streamlines the entire process from raw data acquisition from vehicle-mounted LiDAR, multi-source data filtering and stitching, continuous reconstruction of road surface geometry models, adaptive centerline extraction, to cross-section slice generation and accuracy statistical evaluation. This method eliminates the need for manually preset complex filtering thresholds and manual noise removal. It achieves a high degree of adaptability to dynamic environments by relying on a spatial distribution model and provides rigorous quantitative evaluation indicators.

[0078] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for adaptive extraction and accuracy self-evaluation of road cross-sections based on vehicle-mounted laser point clouds, characterized in that, Includes the following steps: Multiple laser scans are performed on the same road area to obtain multiple sets of point cloud data; A spatial distribution pattern model is used to dynamically filter multiple rounds of point cloud data, while retaining static point cloud data. Multiple static point cloud data obtained after filtering are registered and stitched together to form overall point cloud data. Vectorized curve fitting is then performed on the overall point cloud data to generate a road geometric model. The road centerline is obtained based on the road geometry model. A slicing plane is set at the target position on the road centerline. The road cross-section model is extracted. The point cloud data that intersects with the slicing plane in the overall point cloud data is projected onto the slicing plane to form a cross-section slicing point cloud. Calculate the distance statistics between the road cross-section model and the cross-section slice point cloud, and use the distance statistics as the accuracy evaluation index of the road cross-section model.

2. The method according to claim 1, characterized in that, The spatial distribution pattern model is obtained through the following steps: The same voxel partitioning method was used to voxelize each round of point cloud data. The voxel feature value containing point cloud data was recorded as the first value, and the voxel feature value without point cloud data was recorded as the second value. Compare the feature values ​​of each voxel in any two rounds of point cloud data. When the feature values ​​of the corresponding voxels in the two rounds of point cloud data are different, both voxels are marked as pending. When the feature values ​​of the two voxels are the same, both voxels are marked as static voxels. Extract the feature vector of each static voxel, train a neural network model, and output the neural network model as a spatial distribution law model when the error between the feature vector of the reconstructed voxel and the feature vector of the static voxel is less than the accuracy threshold.

3. The method according to claim 2, characterized in that, The steps for dynamic filtering of the spatial distribution law model include: The trained spatial distribution law model is used to screen the voxels that are marked as undetermined, and the voxels that satisfy the spatial distribution law are re-marked as static voxels. For each measured point cloud data, retain the static voxels and delete the rest to obtain the static point cloud data.

4. The method according to claim 2, characterized in that, The registration steps for the multiple static point cloud data include: Simultaneously acquire scanning path trajectory data during laser scanning; Based on the scan path trajectory data, multiple measured point cloud data are transformed and projected onto the same coordinate system; Static voxels from each measured point cloud data set are used as registration primitives to construct a static voxel feature set, and the ICP algorithm is used for iterative registration.

5. The method according to claim 1, characterized in that, The road geometry model is obtained through the following steps: Key feature point sets are identified and classified from the overall point cloud data. The key feature point sets include road boundary points, lane line points, and main road surface points. Vectorized curve fitting is performed on the road boundary points and the lane line points respectively to generate the corresponding road outline and lane lines; The main points of the road surface are reconstructed to generate a road surface model; The fitted road contour line, the lane line, and the reconstructed road surface model are topologically integrated to obtain a road geometric model containing complete geometry and topological relationships.

6. The method according to claim 5, characterized in that, The road centerline is obtained through the following steps: Parametric discrete sampling is performed based on the generated road contour lines, which typically include a left contour line and a right contour line. For each sampling point on the left contour line, calculate the intersection of the normal of the sampling point and the right contour line, and calculate the center point of the line connecting the intersection point and the sampling point to form a discrete sequence of center points; The road centerline is obtained by fitting a B-spline curve to the center point sequence.

7. The method according to claim 1, characterized in that, The steps to obtain the point cloud of a cross-section slice include: Calculate the tangent vector of the road centerline at the slice point, and establish a slice plane perpendicular to the tangent vector through the slice point; Extract point clouds from the overall point cloud data that are less than the slice projection threshold from the slice plane, and project these point clouds vertically onto the slice plane to obtain the cross-sectional slice point cloud.

8. The method according to claim 1, characterized in that, The steps for obtaining the road cross-section model include: Calculate the tangent vector of the road centerline at the slice point, establish the slice plane using the slice point and tangent vector, and use the slice plane to cut the road geometric model to form the road cross-section model.

9. The method according to claim 1, characterized in that, The distance statistic is any one of the following: average distance, maximum distance, or standard deviation.

10. The method according to claim 1, characterized in that, The method further includes: The overall point cloud data is voxelized. When a voxel does not contain point cloud data, the voxel is marked as a blank voxel. Extract the feature vectors of the non-blank voxels adjacent to each blank voxel, and analyze the feature vectors of the adjacent voxels using a spatial distribution law model. When the feature vectors of the adjacent voxels conform to the spatial distribution law of the road structure, the blank voxel is identified as a missing area to be filled. For each missing region, the feature vectors of the voxels adjacent to the missing region are input into the spatial distribution pattern model to generate complete point cloud data; Calculate the geometric continuity index at the boundary between the completed point cloud data and the point clouds of adjacent voxels. If the geometric continuity index does not reach a preset threshold, use the geometric continuity index as the optimization target to iteratively optimize the spatial distribution law model until the geometric continuity index of the generated completed point cloud data reaches the preset threshold.

Images