A method for extracting road marking and constructing road triangulation structure based on road laser point cloud

Abstract

This invention discloses a method for constructing a road triangulation structure based on road markings extracted from road laser point clouds, belonging to the field of intelligent transportation and high-precision map construction. The method includes the following steps: First, robustly separating low points on the road surface from high points on non-road surfaces through a local slice reference system and tangential neighborhood soft support test; second, proposing a minimum-area rectangle centerline fitting strategy, utilizing the geometric prior of the marking strips to suppress lateral fitting drift, and constructing a high-precision road skeleton; third, reconstructing the trajectory of curb contour points using a constraint generation method based on the road skeleton, ensuring the parallelism and continuity of edges and lane lines; finally, constructing a semantically meaningful TIN triangulation network through unified diagonal rules. This invention can effectively address point cloud quality fluctuations and ensures strict alignment between the road skeleton and edge facilities through geometric constraints, and implements an automated modeling method based on topological logic, making it suitable for scenarios such as autonomous driving simulation, high-precision map production, and road holographic perception.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and high-precision map construction technology, and in particular to a method for constructing a road triangular network structure based on extracting road markings from road laser point clouds. Background Technology

[0002] With the rapid development of autonomous driving technology and intelligent transportation systems, increasingly higher demands are being placed on the digital representation of the road environment. High-fidelity, geometrically accurate, and semantically clear road surface models are the cornerstone of traffic digital twins. LiDAR, with its all-weather, high-precision 3D measurement capabilities, has become a core sensor for acquiring road geometric information. Constructing a Triangulated Irregular Network (TIN) using LiDAR point cloud data that reflects the microscopic undulations and semantic logic of the road surface is of great significance for vehicle dynamics simulation and lane-level path planning.

[0003] While 3D laser scanning technology can rapidly acquire massive amounts of road point cloud data, several key technical challenges remain in practical engineering for automatically constructing high-quality road triangulation networks based on point clouds. First, existing projection methods suffer from accuracy deficiencies. To reduce computational complexity, traditional methods often project or rasterize 3D point clouds into 2D images for processing. This dimensionality reduction inevitably introduces pixelation errors, leading to blurred or lost boundaries for road markings and curbs. Second, direct 3D extraction algorithms lack robustness. Existing direct fitting methods typically treat lane lines, curbs, and other elements as independent objects. However, limited by sensor scanning angles and differences in road surface reflectivity, point cloud data often exhibits sparseness, noise, or even localized gaps in distant or specific areas. In such weakly textured regions, traditional straight-line fitting algorithms (such as RANSAC) are highly susceptible to local noise interference, causing lateral drift or directional oscillations in the fitted trajectory.

[0004] Furthermore, existing methods lack geometric-semantic closed-loop constraints. Due to the lack of unified skeleton constraints, it is often difficult for separately extracted road markings and road edges (such as curbs) to maintain a strict parallel relationship, resulting in topological cracks, overlaps, or semantic confusion in the generated triangular mesh. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a method for constructing a road triangular network structure based on road laser point cloud extraction of road markings. This method utilizes a high-confidence road marking skeleton to constrain the generation of sparse high point trajectories, thereby solving common problems in road surface modeling such as geometric breaks, fitting drift, and semantic missingness.

[0006] Technical solution: A method for constructing a road triangular network structure based on road laser point cloud extraction of road markings, comprising the following steps:

[0007] S1. After preprocessing the original vehicle-mounted LiDAR point cloud, local slices are divided along the vehicle's driving direction to establish a unified local reference system. Based on the quantile elevation reference and tangential neighborhood soft support test of the local slices, the point cloud is separated into road surface point cloud and high point point cloud. S2, perform intensity threshold segmentation and clustering on the road point cloud, construct the minimum area bounding rectangle for each cluster, and fit the road marking centerline based on the geometric features of the rectangle to generate the road skeleton line segment; S3 uses the extracted road skeleton line segments as the main direction and longitudinal range constraints, and generates high point edge trajectories parallel to the lane lines through a lateral extreme value anchoring strategy. S4 sorts and aligns the generated road skeleton lines and high point edge trajectories laterally under a unified reference plane, uses a fixed diagonal rule to triangulate adjacent trajectory lines, and maps semantic labels to the generated mesh vertices and faces, finally outputting a topologically consistent highway pavement triangular network structure.

[0008] Furthermore, the specific steps for separating the point cloud into road surface point cloud and high point point cloud include: The preprocessed point cloud is divided into a series of local slices along the vehicle's driving direction; In each slice First, calculate the principal direction normal vector of the slice. Next, calculate the normal vector of all points within the slice. The projected values ​​in the direction are used to form a set of projected elevations; the values ​​in this set are sorted in ascending order, and those values ​​located at the top of the total number are selected. The projected value at the percentile is used as the reference road surface height for this slice. , ; For any candidate point whose elevation is higher than the reference surface Define candidate points Tangential neighborhood on the local road surface plane The tangential neighborhood Including all tangential projection distances less than the radius Neighboring points The expression is ; Based on tangential neighborhood Calculate candidate points High point confidence The formula is as follows: , in, , Representing candidate points respectively With neighboring points Elevation value in a local reference frame Indicates the current local slice Road surface reference height For point Tangential distance between them The Gaussian kernel scaling parameter; This indicates an indicator function that returns 1 if the condition is true and 0 if the condition is false. Represents the natural constant; The normal vector representing the local plane is obtained by local PCA or plane fitting. like Then candidate points It is confirmed as a true high point; otherwise, it is classified as a road surface point. The final output consists of two sets of point clouds: road surface point cloud. High Point Cloud .

[0009] Furthermore, the midline of the calibrated line is fitted based on the geometric features of the minimum area circumscribed rectangle, specifically including: Intensity normalization and Otsu thresholding were performed on the road surface point cloud to extract high-reflectivity road marking point clouds; Euclidean clustering was then used to divide the road marking point cloud into independent road marking clusters. Each marking cluster is projected onto a local road surface plane, and the two-dimensional convex hull is calculated. Using the rotating caliper method to find the minimum area of ​​the circumscribed rectangle of the convex hull Get rectangle center Major axis direction ,length and width ; Connecting rectangles The midpoints of the two shorter sides are taken as the initial median. The parametric equation of the initial median is... Defined as: ,in, ; With rectangle width Assuming scale priors, construct a strip-shaped neighborhood. Weighted least squares optimization is performed on the initial centerline within the strip-shaped neighborhood to obtain the final road skeleton segment; where p represents a candidate point within the strip-shaped neighborhood. This represents the initial median line corresponding to the k-th rectangle. η represents the shortest distance from point p to l_k; η represents the bandwidth coefficient, which controls the extent to which the strip neighborhood expands relative to the width of the rectangle.

[0010] Furthermore, it also includes deduplication and directional coordination of the optimized initial median, specifically including: For each optimized line segment Calculate the overall score : , in, Indicates the score for the line segment. This represents the maximum number of support points among all candidate line segments. This represents the maximum length among all candidate line segments. Indicates the distance scale; Point to line segment The number of points whose vertical distance is less than a preset distance threshold; The length of the line segment; The mean square residual; , , These are the weighting coefficients; For any two line segments, if the angle between their directions is less than a preset angle threshold and the lateral distance is less than a preset distance threshold, they are considered parallel overlapping candidates and only retained. The higher one; The median of the main direction of all retained line segments is calculated, abnormal line segments with excessive directional deviation are removed, and the line segments are extended in a restricted manner along the main direction within the road surface projection range to form a continuous road surface skeleton.

[0011] Furthermore, generating high-point edge trajectories parallel to the lane lines specifically includes: Read the generated road skeleton line segments and directly inherit their principal direction vectors. and vertical extension range This serves as a strong constraint for generating the trajectory of high points; On the cross section of each local slice, search for the extreme points of the lateral distance of the high point cloud relative to the road skeleton line: for the left curb, take the point with the smallest lateral distance; for the right curb, take the point with the largest lateral distance; and determine these extreme points as lateral anchor points. Using the determined lateral anchor points as reference positions, strictly follow the main direction of the road skeleton line segments. A parallel straight line trajectory is generated and resampled at equal intervals to generate a high point edge line that is geometrically consistent with the lane line. The high point edge line is the edge trajectory of the curb and the central divider that is parallel to the lane line.

[0012] Furthermore, the specific steps to obtain the triangular network structure of the highway pavement include: Based on the lateral coordinates of each trajectory line on the cross section, adjacent trajectory line pairs are established. For each sorted trajectory line, set a uniform vertical step size. Resampling is performed along the trajectory line to generate a regularly ordered list of points. ,in For trajectory line index, For vertical point indexing; In two adjacent trajectory lines and The quadrilateral strip formed between the two sides is divided into two triangles by using a fixed diagonal. Based on the source attributes of the trajectory lines, semantic labels are automatically assigned to the generated triangular patches, and the final output is a TIN model containing geometric coordinates and semantic attributes.

[0013] Compared with the prior art, the significant advantages of this invention are as follows: 1. This invention combines robust statistical modeling, geometric morphological constraints, and semantic association to automatically extract road skeletons from vehicle-mounted LiDAR point clouds, generate high-point edge trajectories, and construct high-quality road triangulation networks. It can effectively cope with point cloud quality fluctuations and ensure strict alignment between road skeletons and edge facilities and topological logic self-consistency through geometric constraints. This automated modeling method is suitable for application scenarios such as autonomous driving simulation, high-precision map production, and road holographic perception. 2. When constructing road skeleton segments, this invention directly extracts the initial centerline using rectangular geometric features, avoiding the rotational offset caused by noise in traditional least squares fitting. 3. In the line system extraction stage, this invention adopts a robust strategy of multiple RANSAC segmentation within the cluster, global main direction consistency screening, and line segment deduplication and merging. It can stably recover a unique line system even under the conditions of broken lines, occlusion, and noise interference. This solves the problems of duplicate lines, pseudo lines, and short segment stacking that are easy to occur in existing methods, thereby avoiding subsequent network construction errors and improving the continuity and consistency of the line system. 4. This invention automatically extracts high points based on height difference and local support relationship and performs clustering segmentation. Furthermore, it extracts cluster edge samples from each high point cluster and extends them into edge trajectories, realizing stable modeling of edge facilities such as curbs and medians. This solves the problems of boundary loss, insufficient coverage and boundary drift caused by relying solely on road markings, and significantly enhances the lateral constraints and topological self-consistency of roads. Attached Figure Description

[0014] Figure 1 This is a schematic flowchart of the method of the present invention; Figure 2 Road point cloud images collected by vehicle-mounted LiDAR; Figure 3 The image shows the fitting result of the road horizontal point cloud markings. Figure 4 This is a graph showing the fitting results of the high-point point cloud. Figure 5 This is a combined image of the road surface and elevation point fitting results; Figure 6 This is the final generated road triangulation network structure diagram. Detailed Implementation

[0015] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0016] The principle of this invention is as follows: a "geometric-semantic closed loop" processing flow is established. First, a clean point cloud of the road surface and facilities is obtained by separating high and low point fields. A high-precision road surface skeleton is fitted using the strip-shaped geometric features of the road markings. Using this skeleton as the main directional constraint, parallel high point trajectories such as curb stones and median strips are generated, thereby avoiding fitting drift of sparse point clouds. Finally, a semantic TIN triangulation network is constructed through a unified diagonal rule.

[0017] like Figure 1 The diagram shows a flowchart of a method for constructing a road triangulation network structure based on road laser point cloud extraction of road markings, including the following steps: Step 1: Robust separation of high and low point fields; After preprocessing the original vehicle-mounted LiDAR point cloud, local slices are divided along the vehicle's driving direction to establish a unified local reference system. Based on the quantile elevation reference of the local slices and tangential neighborhood soft support testing, the point cloud is robustly separated into a road surface point cloud set (low points) and a non-road surface high point set (high points). Specific implementation steps include: Step 11, establish a unified local reference frame; transfer the preprocessed point cloud... Divide into a series of local slices along the direction of vehicle travel. To standardize the sampling density, voxel filtering is first performed, and the voxel size is set. .

[0018] Step 12, Road surface reference height estimation; in each slice First, calculate the local slice. All slice normal vectors Next, calculate the normal vector of all points within the slice. The projected values ​​in the direction form a set of projected elevations; the set of projected elevations The expression is as follows: (1) in, A three-dimensional coordinate vector of a point. Point The elevation.

[0019] Sort the values ​​in the projected elevation set in ascending order and select the values ​​that are at the top of the total. The projected value at the percentile is used as the reference road surface height for this slice. By utilizing low quantile statistical characteristics, underground noise can be effectively filtered and aligned with the road surface bottom, overcoming overall elevation fluctuations caused by vehicle bumps. In this embodiment, the projected elevation set... Sort the elements in the array from smallest to largest to obtain an ordered sequence. Where N is the total number of point clouds in the local slice, then the reference road surface height of the slice is... Defined as: (2) in, , This indicates rounding down. This range can effectively filter out noise from road potholes and fits the bottom of the road surface.

[0020] Step 13, preliminary screening of candidate points; defining relative height thresholds. (Usually set to the level of curb height), if a certain point elevation If it is positive, mark it as a high point candidate; otherwise, classify it as a road surface candidate.

[0021] Step 14, Tangential Neighborhood Soft Support Test; To remove false high points in the air (such as leaves), construct the tangential projection matrix. Calculation points Tangential neighborhood on the tangential plane Based on this tangential neighborhood Calculate the confidence level of the high point of soft support. The formula is as follows: (3) in, , Representing candidate points respectively With neighboring points Elevation value in a local reference frame Indicates the current local slice Road surface reference height For point Tangential distance between them For Gaussian kernel scaling parameters, This is the soft support threshold; This indicates an indicator function that returns 1 if the condition is true and 0 if the condition is false. Represents the natural constant. Represents the identity matrix. Indicates matrix transpose; The normal vector representing a local plane is generally obtained through local PCA or plane fitting. Indicates confirmation Candidate points The distance threshold of neighboring points.

[0022] like Then candidate points Points are identified as true high points; otherwise, they are classified as road surface points. The final output consists of two sets of point clouds: road surface point cloud. High Point Cloud .

[0023] In this embodiment, the original vehicle-mounted LiDAR point cloud is denoised and registered (e.g. Figure 2 As shown), local slices are divided along the vehicle's direction of travel. Within each slice, the lower quantile of the elevation is calculated (taking...). This serves as a reference height for the road surface. A relative height threshold is set. First, high point candidates are initially selected. Then, the tangential neighborhood soft support algorithm is used to calculate the weighted number of low-level neighbors around the candidate point. If the support is lower than the set threshold, it is judged as a suspended noise point such as a leaf and removed. Finally, the point cloud is separated into road surface point cloud (low point cloud) and non-road surface high point cloud (high point cloud).

[0024] Step 2: Extract the midline of the minimum area rectangle and fit it to the skeleton; The road surface point cloud is segmented and clustered using intensity thresholds. For each cluster, a minimum area bounding rectangle is constructed, and the road marking centerline is fitted based on the geometric features of the rectangle. Through density-adaptive robust fitting and coordination with the principal direction, road skeleton segments resistant to lateral drift are generated. Specific steps include: Step 21, Marking Clustering and Preprocessing; For road surface point cloud... Intensity normalization and Otsu thresholding are performed to extract high-reflectivity datum point clouds. Euclidean clustering was used to divide the gradation point cloud into independent gradation clusters. .

[0025] Step 22: Construct the minimum area rectangle; for each marking cluster, project it onto the local road surface plane and calculate the 2D convex hull. Use the rotating calipers method to search for the minimum area bounding rectangle of the convex hull. Get the rectangle Geometric parameters: center of rectangle Major axis direction ,length and width .

[0026] Step 23, Initial centerline extraction; Define rectangle The line connecting the midpoints of the two shorter sides is the initial median, and its parametric equation is: (4) in, .

[0027] This invention avoids the rotational shift caused by noise in traditional least-squares fitting by directly utilizing the geometric features of a rectangle.

[0028] Step 24, strip neighborhood weighted optimization; to further improve accuracy, the rectangle width is used as the weighted optimization method. As a priori scale, construct a strip-shaped neighborhood. Within this strip-shaped neighborhood, a weighted least squares (WLS) refinement is performed on the initial median parameters to obtain a robust center. and direction Where p represents a candidate point within the strip-shaped neighborhood, This represents the initial median line corresponding to the k-th rectangle. Indicates point p to The shortest distance; η represents the bandwidth coefficient, which controls the extent to which the strip neighborhood expands relative to the width of the rectangle.

[0029] Step 25, skeleton deduplication and coordination; perform MSAC estimation (M-estimator Sample Consensus) on all initial medians optimized in Step 24, and calculate the segment score: (5) in, Indicates the score for the line segment. This represents the maximum number of support points among all candidate line segments. This represents the maximum length among all candidate line segments; The distance scale is typically set to a threshold that allows for a specific fitting error. The number of interior points, i.e., the distance from a point to a line segment. The number of points whose vertical distance is less than a preset distance threshold; The length of the line segment; The mean square residual; , , These are the weighting coefficients.

[0030] If the angle between two line segments is less than a preset angle threshold and the lateral distance is less than a preset distance threshold, they are considered to overlap, and only the one with the higher score is retained. Finally, the median of the main direction of all retained line segments is calculated, abnormal line segments with excessive directional deviations are removed, and the segments are extended in a restricted manner along the main direction to form a continuous road surface skeleton.

[0031] In this embodiment, the road surface point cloud obtained in step 1 is segmented by intensity threshold and subjected to Euclidean clustering. For each marking cluster, a two-dimensional convex hull is constructed, and the minimum area bounding rectangle is calculated using the rotating caliper method. The midpoint of the short side of the rectangle is connected as the initial centerline, and the rectangle width is used as a scale constraint to perform weighted optimization within the strip neighborhood to prevent lateral drift. Finally, all fitted line segments are deduplicated and aligned with the main direction to generate continuous road surface skeleton line segments, such as... Figure 3 As shown.

[0032] Step 3: Generate high-point trajectories based on skeleton constraints; Using the road skeleton line segments extracted in step 2 as the main direction and longitudinal range constraints, the curb stones and median strip edge trajectories parallel to the lane lines are generated through a lateral extreme value anchoring strategy.

[0033] Step 31, Constraint Inheritance; Read the road skeleton line segments generated in Step 25 and directly inherit their principal direction vectors. and vertical extension range As a strong constraint for generating high-point trajectories, independent orientation fitting is no longer performed on sparse high-point point clouds.

[0034] Step 32, Lateral Extreme Anchoring; On the cross-section of the local slice, search for the extreme points of the lateral distance of the high point cloud (such as the curb point cloud) relative to the road surface skeleton line. For the left curb, take the point with the smallest lateral distance; for the right curb, take the point with the largest lateral distance; determine these extreme points as lateral anchor points.

[0035] Step 33, Constraint Trajectory Generation: Using the determined lateral anchor point as the reference position, strictly follow the inherited main direction. Generate parallel straight line trajectories. This is equivalent to transforming the free fitting problem into a constrained generation problem, effectively preventing trajectory deviations or directional oscillations in areas where the point cloud is missing.

[0036] In this embodiment, the road skeleton line segment generated in step 2 is read, and its principal direction vector is directly inherited as a strong constraint. On the transverse section, the extreme points of the transverse distance of the set of high points relative to the skeleton line are searched as anchor points. Based on the anchor points, parallel trajectories are generated strictly according to the inherited principal direction, and equal-interval resampling is performed, such as... Figure 4 As shown.

[0037] Step 4: Construct geometric-semantic alignment and TIN triangulation; The generated road skeleton line segments and high point edge trajectories are horizontally sorted and aligned under a unified reference plane. Adjacent trajectory lines are triangulated and connected using a fixed diagonal rule. Semantic labels are mapped to the generated mesh vertices and faces, and finally, a topologically consistent highway pavement triangular network structure (i.e., TIN model) is output.

[0038] Step 41, Lateral Sorting and Alignment: Project the road marking skeleton generated in Step 2 and the high point edge trajectory generated in Step 3 onto a unified reference plane. Sort each trajectory line according to its lateral coordinates on the cross section (e.g., from left to right: left curb, left lane line, central divider, right lane line, right curb), and establish adjacent trajectory line pairs.

[0039] Step 42, longitudinal resampling at equal intervals. For each sorted trajectory line, a uniform longitudinal step size is set. (For example, 0.5 meters). Resampling is performed along the trajectory line to generate a regularly ordered list of points. ,in For trajectory line index, This is a vertical point index.

[0040] Step 43, unify diagonal triangulation; on two adjacent trajectory lines and The quadrilateral strips formed between these elements are divided using a fixed diagonal rule. For example, the points are always connected. and Divide the quadrilateral into two triangles and This rule ensures the consistency of the mesh topology and avoids cracks.

[0041] Step 44, Semantic Mapping: Based on the source attributes of the trajectory lines, the generated triangular patches are automatically assigned semantic labels. For example, the triangular region connecting the "lane centerline" and the "curb line" is labeled as "lane surface" or "shoulder". The final output is a TIN model containing geometric coordinates and semantic attributes.

[0042] In this embodiment, the road surface skeleton from step 2 and the high-point trajectory from step 3 are projected onto a unified reference plane, such as... Figure 5 As shown, the lines are sorted from left to right according to their horizontal coordinates. Semantic labels are automatically assigned based on the origin of the trajectory lines (markings or curbs). Adjacent trajectory lines are triangulated using a uniform diagonal rule, resulting in a triangular network structure as shown. Figure 6 As shown, the TIN model is crack-free and possesses complete semantic attributes.

Claims

1. A method for extracting road marking and constructing a road triangulated network structure based on a road laser point cloud, characterized in that, Includes the following steps: S1. After preprocessing the original vehicle-mounted LiDAR point cloud, local slices are divided along the vehicle's driving direction to establish a unified local reference system. Based on the quantile elevation reference and tangential neighborhood soft support test of the local slices, the point cloud is separated into road surface point cloud and high point point cloud. S2, perform intensity threshold segmentation and clustering on the road point cloud, construct the minimum area bounding rectangle for each cluster, and fit the road marking centerline based on the geometric features of the rectangle to generate the road skeleton line segment; S3 uses the extracted road skeleton line segments as the main direction and longitudinal range constraints, and generates high point edge trajectories parallel to the lane lines through a lateral extreme value anchoring strategy. S4 sorts and aligns the generated road skeleton lines and high point edge trajectories laterally under a unified reference plane, uses a fixed diagonal rule to triangulate adjacent trajectory lines, and maps semantic labels to the generated mesh vertices and faces, finally outputting a topologically consistent highway pavement triangular network structure.

2. The method of claim 1, wherein the method further comprises: The specific steps for separating a point cloud into a road surface point cloud and a high point point cloud include: The preprocessed point cloud is divided into a series of local slices along the vehicle's driving direction; In each slice First, calculate the principal direction normal vector of the slice. Next, calculate the normal vector of all points within the slice. The projected values ​​in the direction are used to form a set of projected elevations; the values ​​in this set are sorted in ascending order, and those values ​​located at the top of the total number are selected. The projected value at the percentile is used as the reference road surface height for this slice. , ; For any candidate point whose elevation is higher than the reference surface Define candidate points Tangential neighborhood on the local road surface plane The tangential neighborhood Including all tangential projection distances less than the radius Neighboring points The expression is ; Based on tangential neighborhood Computing candidate points High point confidence The formula is as follows: in, , Representing candidate points respectively With neighboring points Elevation value in a local reference frame Indicates the current local slice Road surface reference height Indicates the relative height threshold; For point Tangential distance between them The Gaussian kernel scaling parameter; This indicates an indicator function that returns 1 if the condition is true and 0 if the condition is false. Represents the natural constant; The normal vector representing the local plane is obtained by local PCA or plane fitting. If , the candidate point is confirmed as a real high point; otherwise, it is classified as a road point. Two sets of point clouds are output finally: road surface point cloud and high point point cloud .

3. The method of claim 1, wherein the method further comprises: Fitting the calibrated midline based on the geometric features of the minimum area circumscribed rectangle specifically includes: Intensity normalization and Otsu threshold segmentation are performed on the road point cloud to extract the high-reflectivity target line point cloud; the target line point cloud is divided into independent target line clusters by using Euclidean clustering Each target line cluster is projected to a local road plane, and a two-dimensional convex hull is calculated. Search the minimum area circumscribed rectangle of the convex hull by using the rotary caliper method , obtain the center , long axis direction , length and width of the rectangle ; Connecting rectangle Midpoints of the two short sides as initial midline, parametric equation of the initial midline is defined as: wherein, ; with rectangular width a strip-shaped neighborhood is constructed with the rectangular width as a scale prior a weighted least square optimization is performed on the initial centerline in the strip-shaped neighborhood to obtain a final road skeleton segment; wherein p represents a candidate point in the strip-shaped neighborhood, l_k represents an initial centerline corresponding to the kth rectangle, d_p_k represents a shortest distance from the point p to l_k; η represents a strip width coefficient, which controls an expansion degree of the strip-shaped neighborhood relative to the rectangular width.

4. The method of claim 3, wherein the method further comprises: This also includes deduplication and directional coordination of the optimized initial median, specifically including: for each line segment of the optimized set of line segments calculating a line segment score : wherein, represents a line segment score, represents the maximum number of support points among all candidate line segments, represents the maximum length among all candidate line segments, represents a distance scale; is the number of points whose perpendicular distance to the line segment is less than a preset distance threshold; is the line segment length; is the mean square residual; , , is a weight coefficient; For any two line segments, if the included angle of their directions is less than a preset angle threshold and the transverse distance is less than a preset distance threshold, they are considered as parallel overlapping candidates, only one with a higher score is retained higher one; The median of the main direction of all retained line segments is calculated, abnormal line segments with excessive directional deviation are removed, and the line segments are extended in a restricted manner along the main direction to form a continuous road surface skeleton.

5. The method for constructing a road triangular network structure based on road laser point cloud extraction of road markings according to claim 1, characterized in that, Generate a high-point edge trajectory parallel to the lane line, specifically including: Read the generated road skeleton line segments and directly inherit their principal direction vectors. and vertical extension range This serves as a strong constraint for generating the trajectory of high points; On the cross section of each local slice, search for the extreme points of the lateral distance of the high point cloud relative to the road skeleton line: for the left curb, take the point with the smallest lateral distance; for the right curb, take the point with the largest lateral distance; and determine these extreme points as lateral anchor points. Using the determined lateral anchor points as reference positions, and following the main direction of the road skeleton line segments... A parallel straight line trajectory is generated and resampled at equal intervals to generate a high point edge line that is geometrically consistent with the lane line. The high point edge line is the edge trajectory of the curb and the central divider that is parallel to the lane line.

6. The method for constructing a road triangular network structure based on road laser point cloud extraction of road markings according to claim 1, characterized in that, The specific steps to obtain the triangular network structure of the highway pavement include: Based on the lateral coordinates of each trajectory line on the cross section, adjacent trajectory line pairs are established. For each sorted trajectory line, set a uniform vertical step size. Resampling is performed along the trajectory line to generate a regularly ordered list of points. ,in For trajectory line index, For vertical point indexing; In two adjacent trajectory lines and The quadrilateral strip formed between the two sides is divided into two triangles by using a fixed diagonal. Based on the source attributes of the trajectory lines, semantic labels are automatically assigned to the generated triangular patches, and the final output is a TIN model containing geometric coordinates and semantic attributes.

Images