Lane line vectorization pre-labeling method and system based on multi-frame laser radar point cloud

Abstract

This invention discloses a method and system for lane line vectorization pre-annotation based on multi-frame LiDAR point clouds. It constructs a global dense point cloud map by performing spatial coarse filtering and spatiotemporal synchronous stitching on the original point cloud data; extracts the main road surface point cloud through elevation fine filtering and plane segmentation; suppresses and enhances non-uniform backgrounds based on morphological top-hat transform; extracts candidate lane line point clouds using adaptive threshold segmentation combining statistical lower bound and recursive maximum inter-class variance method; extracts structured candidate lane lines and geometric features based on weighted iterative RANSAC line fitting guided by both intensity and longitudinal density; and constructs a global topology graph model and a multi-dimensional scoring mechanism to screen and resolve conflicts among candidate lane lines, outputting the final lane line vector map. This invention effectively overcomes uneven intensity and local noise interference, achieving high-purity and high-robustness lane line vectorization reconstruction, and significantly reducing the cost of manual annotation.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving and high-precision map construction technology, and in particular to a method and system for lane line vectorization pre-annotation based on multi-frame LiDAR point clouds. Background Technology

[0002] With the rapid development of autonomous driving technology and high-precision maps, 3D environmental perception and vector mapping based on LiDAR point clouds have become core research directions. Among them, lane lines, as the most critical geometric and semantic elements in road structure, are of great significance for vehicle localization, path planning, and the safe operation of autonomous driving systems due to their high-precision and robust automated extraction.

[0003] In existing technologies, lane line extraction using lidar point clouds mainly relies on the significant difference in reflectivity between road markings and ordinary asphalt / cement pavement. Typically, the absolute reflectance intensity of the point cloud is used for threshold segmentation, combined with geometric rules and straight-line fitting algorithms to extract lane lines. However, in real-world wide-area mapping and complex road scenarios, existing technologies generally face the following common technical obstacles.

[0004] First, point cloud intensity features are easily affected by environmental and acquisition conditions. Due to the attenuation of radar scanning distance, the continuous change of laser incident angle, and the cumulative effect of stitching together multiple point clouds, the absolute reflection intensity of the global point cloud often exhibits significant non-uniformity. Traditional global threshold segmentation methods based on absolute intensity often fail in such complex scenarios, and the calculated threshold is often too low. This leads to a large number of bright road surface areas being incorrectly classified as foreground, causing not only broken lane lines but also introducing more road surface noise, which seriously interferes with subsequent geometric fitting.

[0005] Secondly, post-processing and fitting methods based on local geometric features have poor robustness. In actual roads, due to vehicle wear, occlusion, or a decrease in point cloud density with distance, the geometry of lane lines often becomes irregular and fragmented, no longer meeting the preset conventional geometric thresholds, easily leading to missed detections. Furthermore, curb edges and long strips of road repair marks are geometrically easily confused with lane lines.

[0006] Finally, there is a lack of effective global topology identification methods. After extracting highlighted candidate points, existing technologies typically use the standard random sampling consensus algorithm for line fitting. However, traditional algorithms sample all points uniformly, ignoring the longitudinal continuity of lane lines. This easily leads to the selection of discrete road markings or noise points at lane line edges, causing the fitted line to deviate and convergence to be slow. More importantly, relying solely on local geometric features cannot distinguish between real lane lines and road edges, stitching artifacts, and parallel interference lines, resulting in insufficient purity of the final output vector map, making it difficult to meet the stringent topology accuracy requirements of high-precision maps.

[0007] In summary, existing technologies generally suffer from weak anti-interference capabilities, low extraction purity, and lack of global topological constraints when dealing with complex scenarios. How to propose a comprehensive core processing mechanism that can simultaneously overcome multiple shortcomings such as uneven point cloud intensity, local noise interference, and confusion of complex road markings, and achieve high-purity and highly robust lane line vectorization reconstruction, is a pressing technical challenge in the field of high-precision map construction. Summary of the Invention

[0008] This application provides a lane line vectorization pre-annotation method and system based on multi-frame LiDAR point clouds, which solves the problems of sparse single-frame point clouds, uneven global point cloud intensity, and lane line extraction being easily interfered with by road signs in the prior art. It achieves high purity and high robustness of lane line vectorization reconstruction, and can be applied to technical fields such as vehicle positioning, path planning, and autonomous driving systems.

[0009] This application provides a lane line vectorization pre-annotation method based on multi-frame LiDAR point clouds, including the following steps: Step S1: Acquire multiple frames of raw point cloud data from LiDAR, perform spatial coarse filtering and spatiotemporal synchronous stitching to construct a global dense point cloud map; Step S2: Perform elevation fine filtering and plane segmentation on the global dense point cloud map to extract the main road surface point cloud; Step S3: Perform non-uniform background suppression and intensity enhancement on the main road surface point cloud based on morphological top-hat transformation; Step S4: Based on the enhanced reflection intensity, an adaptive threshold segmentation method combining statistical lower bound and recursive maximum inter-class variance method is used to extract lane line candidate point clouds. Step S5: Perform weighted iterative RANSAC straight line fitting based on intensity and longitudinal density dual guidance on the lane line candidate point cloud to extract structured candidate lane lines and geometric features. Step S6: Construct a global topology map model and a multi-dimensional scoring mechanism to filter candidate lane lines and resolve conflicts, and output the final lane line vector map.

[0010] Preferably, step S1 specifically includes: Step S11: Obtain raw point cloud data; Step S12: Perform radial distance filtering on the single-frame raw point cloud: Calculate the radial distance of each point on the horizontal plane according to a preset distance threshold range, and filter out points whose radial distance is outside the threshold range; perform coarse filtering on the point cloud after radial distance filtering based on the local elevation histogram, extract its elevation data to construct a one-dimensional histogram, use the elevation corresponding to the peak of the histogram as the local ground reference, set the elevation region of interest based on the reference, and filter out points whose elevation is outside the region; Step S13: Analyze the translation vector and rotation matrix corresponding to each frame of point cloud from the pose record, transform the local point cloud of each frame to the world coordinate system through homogeneous coordinate transformation, and stitch and fuse all the transformed point clouds to generate a global dense point cloud map.

[0011] Preferably, step S2 specifically comprises: Step S21: Extract elevation data of all points from the global dense point cloud map, construct an elevation histogram within a preset percentile interval, and use the elevation corresponding to the peak value of the histogram as the global ground reference; set an elevation region of interest based on this reference, filter out points whose elevation is outside the region, and obtain a finely filtered candidate point cloud for the main road surface. Step S22: For the finely filtered point cloud, the random sampling consensus algorithm is used to iteratively fit the road surface plane. The number of interior points of each candidate plane is counted with a preset distance tolerance. The plane with the most interior points is selected as the optimal road surface model, and the road surface point cloud is extracted based on the model. Step S23: Voxel downsample the road point cloud, estimate the local normal vector of each point and unify the direction, calculate the angle between the normal vector and the global reference normal, and filter out non-smooth points with an angle greater than a preset threshold; perform density clustering on the remaining point cloud and extract the connected component with the most points as the main road point cloud.

[0012] Preferably, step S3 specifically involves: projecting the main road surface point cloud to construct a two-dimensional bird's-eye view intensity map; setting the grid resolution, mapping the main road surface point cloud to a two-dimensional horizontal grid, taking the maximum reflection intensity of the point cloud within the grid as the pixel value, and generating the BEV intensity map; setting the physical size of the structural element and converting it to the pixel size, and performing a morphological opening operation on the BEV intensity map using the structural element; subtracting the result of the opening operation from the BEV intensity map to obtain an enhanced two-dimensional intensity map; inversely mapping the pixel values ​​of the enhanced two-dimensional intensity map to the corresponding three-dimensional point cloud, performing linear stretching, and updating the reflection intensity value of the main road surface point cloud.

[0013] Preferably, step S4 specifically involves: calculating the global mean and standard deviation of the point cloud intensity of the main road surface, and constructing a statistical lower limit threshold; performing initial global segmentation using the maximum inter-class variance method, and calculating the retention rate of candidate foregrounds; if the retention rate is greater than a preset tolerance threshold, triggering recursive local subdivision in the candidate foregrounds to find a local optimal threshold; and determining the final segmentation threshold by combining the statistical lower limit threshold, the initial global threshold, and the local optimal threshold, and extracting candidate point clouds for lane lines.

[0014] Preferably, step S5 specifically involves: calculating the longitudinal density distribution of the point cloud along the main direction, weighting and fusing it with the enhanced reflection intensity to construct a dual sampling weight; performing non-uniform random sampling based on this weight, prioritizing the extraction of features with high connectivity and high intensity; extracting multiple candidate line models through an iterative RANSAC algorithm, and calculating the physical length, geometric center, and other structured geometric features of each candidate line; and introducing a quadratic polynomial fitting mechanism with curvature constraints to adaptively fit the candidate lines to curves.

[0015] Preferably, step S6 specifically involves: calculating the road geometric centerline based on the candidate line set, strictly eliminating close-range parallel road edges and overlapping interference using the center deviation criterion; constructing a multi-dimensional scoring model that includes physical length score, edge protection score, and global parallel topology score; performing non-maximum suppression with conflict resolution radius based on the score to resolve spatial conflicts and output the final real lane lines.

[0016] Preferably, the determination of the final segmentation threshold by combining the comprehensive statistical lower limit threshold, the initial global threshold, and the local optimum threshold, and the extraction of lane line candidate point clouds specifically involves: if the retention rate is greater than a preset tolerance threshold, then the larger of the local optimum threshold and the statistical lower limit threshold is taken as the final segmentation threshold; if the retention rate is not greater than the preset tolerance threshold, then the larger of the initial global threshold and the statistical lower limit threshold is taken as the final segmentation threshold.

[0017] Preferably, the multidimensional scoring model specifically involves: calculating the road centerline based on the geometric center of the candidate lines; for parallel line pairs with a spacing less than the proximity interference threshold, retaining the one closer to the center based on the center deviation; constructing a multidimensional topology score, including a basic geometric length score positively correlated with the line segment length, an edge protection score with a fixed reward assigned to the outermost line segment, and a global topology score based on the relationship between the lateral spacing and the standard lane width; and finally, performing non-maximum suppression with a conflict resolution radius based on the total score to eliminate low-scoring conflict line segments.

[0018] This application also provides a lane line vectorization pre-annotation system based on multi-frame LiDAR point clouds, used to implement the above-mentioned lane line vectorization pre-annotation method, including: The point cloud acquisition and stitching module is used to acquire multiple frames of raw point cloud data from LiDAR, perform spatial coarse filtering and spatiotemporal synchronous stitching, and construct a global dense point cloud map. The road surface extraction module is used to perform elevation fine filtering and plane segmentation on the global dense point cloud map to extract the main road surface point cloud. The intensity enhancement module is used to perform non-uniform background suppression and intensity enhancement on the main road surface point cloud based on morphological top-hat transformation; The threshold segmentation module is used to extract lane line candidate point clouds based on the enhanced reflection intensity using an adaptive threshold segmentation method that combines statistical lower bound and recursive maximum inter-class variance method. The geometric fitting module is used to perform weighted iterative RANSAC straight line fitting based on both intensity and longitudinal density guidance on the lane line candidate point cloud, and extract structured candidate lane lines and geometric features. The topology filtering module is used to construct a global topology graph model and a multi-dimensional scoring mechanism, filter candidate lane lines and resolve conflicts, and output the final lane line vector map. Compared with the prior art, the advantages of this invention are: This invention proposes a comprehensive core processing mechanism that, through a collaborative processing logic and a complete technical solution, simultaneously addresses several common technical obstacles prevalent in existing technologies, such as uneven point cloud intensity, local noise interference, and confusion of complex road markings. This achieves high-purity and highly robust lane line vectorization reconstruction. Specifically, it is reflected in: Overcoming the shortcomings of uneven global point cloud intensity: The innovative introduction of BEV morphological top-hat transformation converts absolute reflection intensity into relative local contrast, effectively suppressing the intensity of large-area, slowly changing road background, highlighting the linear features of lane lines, and laying a robust foundation for subsequent segmentation.

[0019] To improve the adaptability and accuracy of threshold segmentation, a segmentation strategy combining statistical lower limit protection and recursive Otsu's method is proposed. When the retention rate of the initial segmentation is abnormal, local subdivision is adaptively triggered, effectively filtering out interference from large areas of bright road surface and significantly improving the purity of lane line extraction.

[0020] The robustness and accuracy of model fitting are improved: The traditional RANSAC uniform sampling strategy is improved by constructing dual sampling weights using enhanced intensity values ​​and point cloud vertical density distribution, so that the center points with high connectivity and high brightness have a higher probability of being selected. This achieves fitting driven by both radiation and geometric features, effectively suppressing discrete noise interference and improving fitting accuracy.

[0021] Finally, a highly robust anti-interference topology filtering was achieved: a multi-dimensional graph model was constructed that includes near-field interference removal, edge protection, and global topology scoring. The road edge was filtered out by the center deviation criterion, and the real lane line score was enhanced by the standard vehicle width multiple relationship.

[0022] Combining 300 consecutive frames, divided into 30 processing batches, and processing over 60 million point clouds of complex real-world road scene test data, the quantitative technological advancements of this invention are specifically verified through the following experimental data: Significant suppression of background noise and substantial improvement in purity: Traditional single-batch global Otsu segmentation suffers from a candidate point retention rate as high as 50%~56% when dealing with uneven point clouds due to the influence of bright road surface noise. The proposed mechanism, combining statistical lower bounds and recursive Otsu adaptive segmentation, reliably triggers local subdivision when the retention rate is abnormal, stabilizing the effective candidate point retention rate within the target range of 6.4%~7.0%. In a single batch processing of approximately 430,000 road surface points, the number of false candidate points was successfully reduced from over 230,000 to around 28,000, effectively eliminating the severe interference of large-area road surface noise on subsequent geometric fitting, resulting in a significant improvement in purity.

[0023] Accurate resolution of topological conflicts and low feature false detection rate: Traditional uniform RANSAC methods, under the influence of discrete noise, are prone to generating 14-15 line segments containing close-range curbs and scattered signs in a single batch. The multidimensional topology filtering model constructed in this invention can accurately remove 4-5 close-range parallel curb interferences per batch on average, and effectively resolve the remaining overlapping conflict line segments. In 30 consecutive test batches, this system ultimately extracted more than 160 effective lane line segments stably, with only 4 topology fitting deviations occurring globally. Based on the preprocessing principle guided by high reliability, the spacing between adjacent main lane lines output by the system is distributed within the real physical topology range of 3.3m to 4.6m, achieving high-precision reconstruction from high-noise discrete data to physical structure.

[0024] Robustness of Adaptive Compensation and Multi-Frame Spatiotemporal Feature Fusion in Long-Tail Scenarios: In long-distance real-world road tests, for complex long-tail scenarios such as sparse point cloud distribution or passageway entrances / exits, the algorithm of this invention prioritizes accuracy while rejecting low-confidence fitting, effectively avoiding the generation of false lane lines. In such edge scenarios, occasional missed detections within a single batch of the field of view can be effectively filled in by stable features from adjacent batches in the spatial geometric topology, thanks to the high-precision map multi-frame temporal processing mechanism of this system. This mechanism ensures both the accuracy of feature extraction at a single moment and the integrity of the global road network spatiotemporal topology, significantly reducing the cost of manual backend verification and error correction.

[0025] In this invention, top-hat enhancement, recursive Otsu's algorithm, dual-weighted priors, and multi-dimensional topology verification constitute a deeply coupled feedback chain. As confirmed by comparative experiments, conventional combinations lacking the mechanisms of this invention cause extracted lane lines to deviate significantly from road signs or curbs; while the full-link coupling of this invention not only significantly increases the number of effective candidate interior points and locks into the true physical skeleton, but also ensures that the overall road network topology conforms to the real physical world. This data enhancement based on physical mechanism restoration produces and verifies significant technical effects.

[0026] In summary, the global map construction and preprocessing method and system based on multi-frame LiDAR point clouds provided by this invention converts absolute intensity into relative contrast through morphological top-hat transformation, and effectively overcomes the interference of uneven point cloud intensity and local noise by combining recursive Otsu's algorithm and weighted RANSAC algorithm. Furthermore, through a global topology graph model and a multi-dimensional scoring mechanism, it achieves accurate removal of road edges and complex road surface markers. This invention can achieve high-purity and high-robustness lane line vectorization reconstruction in complex road scenarios. It is not a simple patchwork combination of existing commonly used image processing or point cloud algorithms, but rather the first to propose a physical reconstruction model driven by structured radiative properties and connectivity, providing reliable data support that follows the real spatial structure for the construction and pre-annotation of high-precision maps for autonomous driving, significantly reducing manual annotation costs and improving mapping efficiency. Attached Figure Description

[0027] Figure 1 This is an overall flowchart of the method provided in the embodiments of the present invention; Figure 2 This is a comparison of the effects of a single frame of original point cloud and a global point cloud map after multiple frames are stitched together in an embodiment of the present invention. The left image is the intensity rendering image of the single frame of original point cloud, and the right image is the intensity rendering image of the global point cloud after multiple frames are stitched together. Figure 3 This is a diagram showing the effect of global point cloud pre-filtering and road surface segmentation in an embodiment of the present invention; Figure 4 This is a point cloud effect diagram of the main road surface after normal constraint and density clustering in an embodiment of the present invention; Figure 5 The image shows a comparison of the intensity enhancement effect based on morphological top-hat transformation in an embodiment of the present invention. The left image is the original BEV intensity image, and the right image is the local contrast image after the top-hat transformation enhancement. Figure 6 The image shows a comparison between traditional single Otsu segmentation and recursive Otsu segmentation in this invention. The left image shows the result of single global Otsu segmentation extraction, and the right image shows the result of recursive Otsu adaptive segmentation extraction in this invention. Figure 7The image shows a comparison of the extraction results between the traditional conventional process and the full feedback chain mechanism of this invention. The left image shows the candidate lane lines extracted using absolute intensity threshold segmentation and traditional RANSAC uniform sampling, which have serious geometric offsets and incorrect topologies. The right image shows the extraction results after using the combined mechanism of BEV morphological top-hat transformation, recursive Otsu and dual-weighted RANSAC of this invention, which significantly increases the number of interior points and accurately restores the real physical topology. Figure 8 This is the final vectorized pre-annotation result of lane lines after the entire process of this invention; Figure 9 This is a comparison chart of point cloud retention rates in continuous batch processing using traditional single Otsu and recursive Otsu methods in an embodiment of the present invention. Figure 10 This is a stacked bar chart showing the distribution of lane line extraction quantity before and after multi-frame topology filtering in an embodiment of the present invention. Figure 11 This is a scatter plot comparing the physical consistency of the lateral spacing between the lane lines output in the final embodiment of the present invention and the lane lines output without topology constraints. Figure 12 This is a long-distance global lane line vectorized physical road network overview image constructed after aggregating 300 consecutive frame point clouds in an embodiment of the present invention. Detailed Implementation

[0028] This application provides a lane line vectorization pre-annotation method and system based on multi-frame LiDAR point clouds, which solves the problems of uneven point cloud intensity, local noise interference, and confusion of complex road signs in the vectorization mapping process due to the complexity of actual wide-area mapping and road scenes. It achieves high purity and high robustness of lane line vectorization reconstruction.

[0029] The technical solution in this application is to solve the above problems, and the overall approach is as follows: By introducing BEV morphological top-hat transformation, absolute reflection intensity is converted into relative local contrast, suppressing large-area, slowly changing road background intensity and highlighting the linear features of lane lines, laying a robust foundation for subsequent segmentation. A segmentation strategy combining statistical lower limit protection and recursive Otsu's method is proposed, which adaptively triggers local subdivision when the initial segmentation retention rate is abnormal, filtering out interference from large areas of bright road surface and improving the purity of lane line extraction. The traditional RANSAC uniform sampling strategy is improved by constructing dual sampling weights using enhanced intensity values ​​and point cloud longitudinal density distribution, giving high connectivity and high brightness center points a higher probability of being selected, achieving fitting driven by both radiation and geometric features, effectively suppressing discrete noise interference and improving fitting accuracy. A multi-dimensional graph model including near-field interference removal, edge protection, and global topology scoring is constructed, filtering out road edges through the center deviation criterion and enhancing the score of real lane lines using the standard vehicle width multiple relationship.

[0030] This application proposes a physical reconstruction model driven by both structured radiation properties and connectivity, which provides reliable data support for the construction and pre-annotation of high-precision maps for autonomous driving, following the real spatial structure, significantly reducing the cost of manual annotation and improving mapping efficiency.

[0031] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0032] Example 1

[0033] like Figure 1 As shown in the figure, this application provides a lane line vectorization pre-annotation method based on multi-frame LiDAR point clouds, including the following steps: Step S1: Acquire multiple frames of raw point cloud data from LiDAR, perform spatial coarse filtering and spatiotemporal synchronous stitching to construct a global dense point cloud map.

[0034] Step S11, acquire raw point cloud data. Acquire the raw point cloud data stream collected by the LiDAR, and define the first... The original point cloud set of the frame is , which includes _____ discrete sampling points. For any point in the set... Its eigenvectors are defined as seven-dimensional space vectors:

[0035] in, Represents Euclidean geometric coordinates based on the radar local coordinate system; Indicates the original reflection intensity. For sampling timestamps; For laser emission beam index (Ring); This is the hardware channel identifier (Block ID). Based on the PCD data structure of the radar used in this embodiment, the hardware channel identifier... The value space covers .

[0036] Step S12: Perform radial distance filtering on the single-frame original point cloud: Calculate the radial distance of each point on the horizontal plane according to the preset distance threshold range, and filter out points whose radial distance is outside the threshold range; perform coarse filtering on the point cloud after radial distance filtering based on the local elevation histogram, extract its elevation data to construct a one-dimensional histogram, use the elevation corresponding to the peak of the histogram as the local ground reference, set the elevation region of interest based on the reference, and filter out points whose elevation is outside the region.

[0037] Calculate the radial Euclidean distance of the point cloud in the horizontal plane (XY plane). Construct a distance mask Filtering distance less than the threshold and greater than the threshold Points:

[0038] Local Elevation Coarse Filtering Based on Statistical Histograms: Extracting the Elevation Set of Point Clouds that Satisfies Distance Filtering Conditions To adaptively determine the local ground reference elevation, for Construct a one-dimensional elevation histogram within the effective percentile interval. Let the elevation corresponding to the peak value of the histogram be the local ground level estimate. :

[0039] Based on this local ground benchmark, upper and lower boundaries of the region of interest (ROI) for elevation are defined. In this embodiment Construct a local elevation mask. :

[0040] This local coarse filtering mechanism utilizes histogram statistical properties to effectively remove non-road features such as trees, streetlights, and buildings at a high level within a single frame. While retaining road surfaces, vehicles, and low walls, it significantly reduces the amount of point cloud data in a single frame, thereby substantially reducing the computational overhead and memory usage for subsequent multi-frame stitching. The effective point cloud set after comprehensive spatial coarse filtering is denoted as... .

[0041] Step S13: Analyze the translation vector and rotation matrix corresponding to each frame of point cloud from the pose record, transform the local point cloud of each frame to the world coordinate system through homogeneous coordinate transformation, and stitch and fuse all the transformed point clouds to generate a global dense point cloud map.

[0042] Pose parsing and quaternion reconstruction: parsing the first quaternion from a synchronized pose record file Frame translation vector and rotation quaternions Convert quaternions to rotation matrix .

[0043] Homogeneous coordinate transformation: Constructing Global pose transformation matrix :

[0044] Convert local point cloud coordinates to homogeneous coordinates Perform global space mapping:

[0045] in This is the calibration extrinsic parameter matrix for the radar reaching the center of the vehicle body.

[0046] Global map fusion: Traverse all consecutive frames The converted point cloud and its corresponding intensity features The points are stitched together to generate a global dense point cloud map. .

[0047] In this embodiment, through the processing of steps S1 to S3 described above, the construction of a multi-frame global dense point cloud map from a single-frame local point cloud is achieved. For example... Figure 2 The comparison image showing the stitching of single frames and multiple frames clearly demonstrates that: (1) Coordinate system one: The multi-frame point cloud was successfully transformed to the global world coordinate system, eliminating the local coordinate system drift problem caused by the vehicle's motion in the single-frame point cloud, and forming a continuous and complete road geometry.

[0048] (2) Feature enhancement: When rendering by reflection intensity in point cloud visualization software, the lane line features are weak and discontinuous due to sparse sampling of single-frame point cloud; however, after multi-frame stitching and fusion, the density of road point cloud is greatly improved, and the high reflection intensity features of lane lines are significantly enhanced, making them visually clear and distinguishable.

[0049] Step S2 involves performing elevation fine filtering and plane segmentation on the global dense point cloud map to extract the main road surface point cloud. Based on the global elevation histogram, an adaptive estimation of the ground benchmark is performed, followed by elevation fine filtering. A random sampling consensus algorithm is used to fit the road surface plane, extracting the road surface point cloud. The road surface point cloud is then voxelized and downsampled, and the maximum connected component is extracted using a combination of normal smoothness constraints and density clustering algorithms to obtain a clean main road surface point cloud.

[0050] Step S21: Extract elevation data of all points from the global dense point cloud map, construct an elevation histogram within a preset percentile interval, and use the elevation corresponding to the peak value of the histogram as the global ground reference; set an elevation region of interest based on this reference, filter out points whose elevations are outside the region, and obtain a finely filtered candidate point cloud for the main road surface.

[0051] Global elevation fine filtering based on statistical histograms for acquiring global dense point cloud maps Subsequently, since cumulative pose error is inevitably introduced during the multi-frame stitching process, and the single-frame coarse filtering retains some low walls and vehicles, global elevation fine filtering is required to further extract clean road surface candidate areas.

[0052] Global elevation statistical analysis: extraction The set of Z-axis coordinates of all points To avoid interference from extreme outliers, select... The 1% to 99% percentile range is used as the effective statistical range.

[0053] Adaptive global ground benchmark estimation: Constructing an elevation histogram within the effective statistical range In this embodiment, the area is divided into 100 intervals. By finding the global peak value of the histogram, the global ground reference elevation is adaptively estimated. :

[0054] In actual testing, the estimated ground elevation of a certain road section It is -3.44m.

[0055] Fine-grained Region of Interest (ROI) extraction: Set a narrower upper and lower floating threshold around the center. (0.5m is used in this embodiment), construct a global elevation mask. :

[0056] Applying this mask yields a finely filtered set of point clouds. In this embodiment, the original global point cloud contains 3,985,274 points, which are then processed through an elevation range. After fine filtering, 2,247,246 points were retained, with a retention rate of 56.39%. Step S4 and the single-frame coarse filtering in step S2 form a progressive relationship of first coarse filtering and then fine filtering. While effectively dealing with splicing errors, it further eliminates non-road interference and significantly reduces the computational complexity of subsequent algorithms.

[0057] Step S22: For the finely filtered point cloud, the random sampling consensus algorithm is used to iteratively fit the road surface plane. The number of interior points of each candidate plane is counted with a preset distance tolerance. The plane with the most interior points is selected as the optimal road surface model, and the road surface point cloud is extracted based on the model.

[0058] Point cloud after pre-filtering for accurate road surface segmentation based on random sampling consistency It still contains some low obstacles, and further precise road surface geometry needs to be extracted.

[0059] Planar model definition: The ideal road surface equation in three-dimensional space is defined as follows:

[0060] in Let be a plane normal vector, and satisfy . .

[0061] RANSAC iterative fitting: in Three non-collinear points are randomly selected from the data. Calculate the parameters of the candidate plane they form. .calculate Orthogonal distance from all points to the candidate plane :

[0062] Set distance tolerance threshold In this embodiment, 0.30m is used, and the statistics satisfy... The number of inliers.

[0063] Optimal model selection: After In this iteration, in this embodiment The planar model with the most interior points is selected as the final road surface model. In actual testing, the optimal plane equation obtained through fitting is: .

[0064] Road surface point cloud extraction: Based on the optimal planar model, extract all interior points to form the final road surface point cloud set. In this embodiment, the number of input points was 2,247,246, and the number of road surface points extracted after RANSAC segmentation was 2,128,488, with a retention rate of 94.72%.

[0065] In this embodiment, through the processing of steps S4 to S5 described above, the road surface point cloud is accurately extracted from the global dense point cloud. For example... Figure 3 As shown, it can be clearly observed that: (1) Non-road surface interference removal: The elevation pre-filtering in step S4 effectively removes tall structures such as buildings and tree canopies on both sides of the road. The RANSAC segmentation in step five further removes interference objects such as low vegetation that are not coplanar with the road surface.

[0066] (2) Road surface feature purification: After two-step spatial geometric filtering, the final retained features are... It exhibits highly flat geometric characteristics and fully preserves the high reflectivity of lane lines on the road surface, eliminating spatial geometric interference for subsequent intensity-based lane line extraction.

[0067] Step S23: Voxel downsample the road point cloud, estimate the local normal vector of each point and unify the direction, calculate the angle between the normal vector and the global reference normal, and filter out non-smooth points with an angle greater than a preset threshold; perform density clustering on the remaining point cloud and extract the connected component with the most points as the main road point cloud.

[0068] Road surface region growth based on normal constraints and density clustering, followed by RANSAC segmentation of the road surface point cloud. It may still contain interference objects that are coplanar with the road surface but not part of the road, such as curb slopes, vehicle underside reflection points on the road surface, and isolated noise clusters. To extract a continuous and smooth main road surface, this embodiment employs an efficient region growing algorithm based on normal smoothness constraints and DBSCAN density clustering.

[0069] Voxelization downsampling and intensity preservation: To significantly improve the computational efficiency of normal estimation and clustering, voxelization downsampling and intensity preservation are performed. Perform voxel size Voxelization downsampling, in this embodiment During downsampling, the reflection intensity of the point cloud is... Normalized and mapped to color channels, the intensity is smoothly transitioned and preserved using Open3D's built-in voxelization color averaging mechanism. The downsampled point cloud is denoted as... In this embodiment, the number of points was reduced from 2,128,488 to 543,457, with a compression rate of 25.5%, which significantly accelerated subsequent processing.

[0070] Local normal estimation and smoothness filtering: In Above, with To find the search radius, the local normal vector of each point is estimated using a KD-Tree. The uniform normal direction is upward, that is... Then, the median of all normal vectors is calculated as the global reference road surface normal. Calculate the angle between the normal vector at each point and the reference normal. :

[0071] Set smoothness angle threshold In this embodiment, we take Filter out The abrupt change in normals, corresponding to the edge of the curb or the sidewall of a vehicle tire, yields a smooth point cloud. .

[0072] DBSCAN density clustering extracts the main road surface: Due to the removal of curb edges, the main road surface is spatially broken from the sidewalk or adjacent clutter. Execute the DBSCAN clustering algorithm and set the neighborhood search radius. That is, 0.80m, the minimum number of core points. The algorithm divides the point cloud into several connected clusters. The number of points contained in each cluster is counted, and the largest connected component with the most points is extracted as the final main road point cloud. :

[0073] In this embodiment, DBSCAN clustering identified 31 clusters. After extracting the largest cluster, the final number of main road points retained was 454,659. Figure 4 As shown, from a visual perspective, this step successfully eliminated debris on the side of the road, car remnants on the road surface, and curb structures, leaving only a clean, continuous physical road surface.

[0074] Step S3: Perform non-uniform background suppression and intensity enhancement based on morphological white top-hat transformation on the main road point cloud. Project the main road point cloud to construct a two-dimensional bird's-eye view intensity map; perform morphological white top-hat transformation on the BEV intensity map to extract the highlight features of local abrupt changes and convert the absolute intensity into relative local contrast; inversely map the enhanced two-dimensional intensity map back to the three-dimensional point cloud to update the reflection intensity value of the main road point cloud.

[0075] Non-uniform background suppression and intensity enhancement based on morphological top-hat transformation for obtaining clean main road point clouds Subsequently, due to differences in radar scanning distance, incident angle, and multi-frame stitching in the stitched ground map, the absolute reflection intensity distribution of the road surface and lane lines is uneven. Traditional global threshold segmentation fails when the reflection intensity of the far-end lane lines is lower than that of the near-end ordinary road surface. To address this, this embodiment introduces morphological top-hat transformation from computer vision into the bird's-eye view projection of the 3D point cloud to extract locally abrupt tubular / linear highlight features.

[0076] BEV intensity map construction: Main road surface point cloud Project onto a 2D horizontal grid. Set the grid resolution. In this embodiment, 0.1m is used. The boundaries of the point cloud in the X and Y directions are calculated, and a size of [size missing] is constructed. A two-dimensional grid. Each point... Mapped to the corresponding grid cell middle:

[0077] For multiple points falling into the same grid cell, the maximum reflection intensity is taken as the pixel value of that grid cell, thereby generating a two-dimensional BEV intensity map. .

[0078] Morphological top-hat transformation: To filter out large areas of slowly changing road surface background intensity while preserving the highlight features of lane lines with abrupt changes, the following transformations are performed: Perform a morphological white-hat transformation. Set the physical dimensions of the structuring element. In this embodiment, 1.5m is used, slightly larger than the standard lane width. This is then converted to pixel dimensions. The top-hat transformation is defined as the original image minus its morphological opening operation (erosion followed by dilation).

[0079] in This indicates the opening operation. For size The structural element. This operation is equivalent to a spatial high-pass filter, converting absolute intensity into relative local contrast.

[0080] Intensity inverse mapping and fusion: transforming the enhanced 2D intensity map Inverse mapping back to 3D point cloud In this process, the reflection intensity value of each point is updated. To maintain a consistent numerical range, the enhanced intensity is linearly stretched so that its maximum value aligns with the original maximum value. This step significantly improves the relative contrast of the lane line region, laying a robust foundation for subsequent adaptive thresholding segmentation. Figure 5 As shown, the top cap transformation effectively suppresses the uneven road surface background and highlights the linear features of the lane lines.

[0081] Step S4: Based on the enhanced reflection intensity, an adaptive threshold segmentation method combining the statistical lower bound and the recursive Otsu's method is used to extract candidate point clouds for lane lines. The global mean and standard deviation of the main road surface point cloud intensity are calculated to construct a statistical lower bound threshold. Initial global segmentation is performed using the Otsu's method to calculate the retention rate of candidate foregrounds. If the retention rate is greater than a preset tolerance threshold, recursive local subdivision is triggered in the candidate foreground to find a locally optimal threshold. The final segmentation threshold is determined by combining the statistical lower bound threshold, the initial global threshold, and the locally optimal threshold, and the candidate point clouds for lane lines are extracted.

[0082] Adaptive intensity threshold segmentation based on statistical lower bound and recursive maximum inter-class variance method for obtaining the main road point cloud after intensity enhancement Subsequently, the high-reflectivity lane lines (foreground) and low-reflectivity asphalt / cement pavement (background) need to be separated using the reflection intensity characteristics of the point cloud. Due to significant differences in intensity distribution across different road sections and material conditions, traditional global fixed thresholds or single-run Otsu's algorithm often fail to balance robustness and accuracy. Therefore, this embodiment proposes a robust adaptive threshold segmentation method combining statistical lower bound protection and recursive Otsu's algorithm.

[0083] Global Strength Statistical Analysis and Lower Limit Protection: Extraction The set of reflection intensities of all points Calculate the global intensity mean of this set. and standard deviation :

[0084] To prevent subsequent adaptive algorithms from outputting excessively low thresholds in special cases and introducing significant background noise, a hard lower threshold based on statistical distribution is constructed. :

[0085] in For the confidence coefficient, in this embodiment, we take... In actual testing, the average intensity of the input road surface point cloud was used. Standard deviation The calculated statistical lower limit threshold .

[0086] Initial global segmentation based on the Otsu's method: For Constructing an intensity histogram Quantization is performed to levels 0-255. Let the current candidate threshold be... The point cloud is divided into background classes. ( ) and foreground ( ). Calculate the pixel proportions of the two classes. and within-class mean Define the between-class variance. for:

[0087] By iterating through all possible intensity levels, we find the initial Otsu threshold that maximizes the inter-class variance. :

[0088] In actual testing, the threshold initially calculated by Otsu... .

[0089] Foreground Retention Rate Assessment and Recursive Local Segmentation: Utilizing an Initial Threshold The point cloud is initially segmented to obtain a set of candidate foreground point clouds. Calculate the retention rate of candidate prospects. Since lane markings typically occupy a very small area of ​​the physical road surface, if If the value is too high, it indicates that the initial Otsu segmentation failed to effectively separate the high-strength road surface from the lane markings. Set a retention rate tolerance threshold. In this embodiment, we take .like This triggers a recursive mechanism: extracting the candidate foreground set. Local intensity subset ,exist The Otsu's method is executed again to find a local optimum threshold. :

[0090] In actual testing, the initial retention rate reached 54.88%, far exceeding the target value. The system determined that road surface noise was suspected, and therefore performed a second Otsu subdivision in the foreground to calculate the local threshold. .

[0091] Multi-strategy fusion for final threshold decision-making and application: To ensure the absolute robustness of the final threshold, a comprehensive statistical lower limit is used. Initial global threshold and secondary local threshold The final segmentation threshold is determined using a max pooling strategy. :

[0092] In actual testing, the threshold used for the system's final decision was... . use The main road surface point cloud is subjected to a final binarization masking operation, and combined with near-field blind spot filtering, the final lane line point cloud set is extracted. In this embodiment, 454,659 road surface points were input. After the robust adaptive threshold segmentation described above, 27,327 high-intensity points were extracted, and the final retention rate was reduced to 6.01%.

[0093] like Figure 6 As shown, to verify the superiority of the recursive Otsu algorithm in this embodiment, the effects of single Otsu segmentation and recursive Otsu segmentation are compared: (1) Single Otsu segmentation effect: Due to the uneven reflection intensity of the road surface, the global threshold calculated by the single Otsu algorithm is too low. This resulted in a large number of bright road surface areas being incorrectly classified as foreground, and the extracted point cloud containing a large number of non-lane line noise points with a retention rate as high as 54.88%, which seriously interfered with subsequent straight line fitting.

[0094] (2) Recursive Otsu segmentation effect: After detecting the initial retention rate anomaly, the recursive Otsu algorithm in this embodiment adaptively triggers secondary subdivision in the foreground, raising the threshold to 63.00. It can be clearly observed from the figure that large areas of road surface noise are significantly filtered out, and only the real lane lines with high reflection intensity are retained. The point cloud presents a clear and slender linear topology structure.

[0095] Step S5 involves performing weighted iterative RANSAC line fitting based on both intensity and longitudinal density guidance on the candidate lane point cloud to extract structured candidate lane lines and geometric features. Specifically, the longitudinal density distribution of the point cloud along the main direction is calculated and weighted with the enhanced reflection intensity to construct a dual sampling weight. Non-uniform random sampling is performed based on this weight, prioritizing the extraction of features with high connectivity and high intensity. Multiple candidate line models are extracted using the iterative RANSAC algorithm, and the physical length, geometric center, and other structured geometric features of each candidate line are calculated. A quadratic polynomial fitting mechanism with curvature constraints is introduced to adaptively fit the candidate lines to curves.

[0096] Structured lane line fitting and geometric feature extraction based on weighted iterative random sampling consistency guided by both intensity and longitudinal density to obtain high-purity lane line candidate point clouds. Subsequently, to extract a mathematically expressed structured lane line model from the discrete point cloud, this embodiment employs a weighted iterative two-dimensional RANSAC algorithm that combines longitudinal connectivity and reflection intensity, while simultaneously extracting the geometric features of each line segment. Traditional RANSAC, with its equal probability during fitting, is susceptible to noise at lane line edges or isolated bright road markings, leading to fitting bias and missed detections. This embodiment optimizes the traditional strategy by evaluating the longitudinal clustering density of the point cloud through principal direction projection and combining it with point cloud intensity to construct a dual sampling weight. This ensures that regions with good longitudinal connectivity and high intensity have a higher probability of being selected, effectively suppressing interference from isolated bright spots.

[0097] Solving and normalizing the longitudinal density distribution: Since lane lines usually extend along a specific longitudinal direction, we first calculate the current remaining point set. In the two-dimensional covariance matrix on the horizontal plane, the eigenvector corresponding to its largest eigenvalue is extracted as the first principal direction vector of the road point cloud. For all points at Projecting onto the surface yields a one-dimensional sequence of projected coordinates. The grid width for the vertical density statistics is then set. In this embodiment, a value of 0.20m is used. A one-dimensional histogram is constructed within the projected coordinate interval, and the frequency of point clouds in each grid is counted as the local density. Each point is mapped to the density value of its corresponding grid, and normalized using the global maximum density to obtain the vertical connectivity weight of each point. :

[0098] in To prevent small constants with a denominator of zero (such as...) ).

[0099] Double-weighted random sampling of fusion intensity: extracting the enhancement intensity value of the current point cloud. The intensity weights are then obtained by performing min-max normalization. Design the vertical connectivity weight ratio parameter. (In this embodiment, connectivity is emphasized) ), calculate the final combined weights :

[0100] All weights that produce negative values ​​are truncated to 0, and then the sum is calculated. The final sampling probability distribution is obtained by normalizing the weights. In each iteration, two points are randomly selected with weighted probability to generate a line seed, thus driving the algorithm from both the physical connectivity and radiation intensity levels.

[0101] Iterative RANSAC multi-segment extraction: Calculate the direction vector formed by two points. The unit direction vector is obtained by normalizing it. Therefore, a two-dimensional linear equation can be constructed. The model parameters are defined as follows: , , Calculate all points in the remaining point cloud. Orthogonal distance to the linear model .like Less than the set distance threshold In this embodiment, a value of 0.20m is used, and this point is determined as an inlier. Within the set number of iterations, in this embodiment... The line model with the most interior points is retained as the current optimal fitted line. If the number of interior points of the optimal model exceeds the set minimum number of points threshold... In this embodiment, a value of 30 is used. The model is then considered a valid candidate line, and all interior points belonging to that model are removed from the current point cloud set. This process is repeated for the remaining point cloud until the number of remaining points is insufficient. Or reach the maximum allowed number of extraction lines In this embodiment, 15 is chosen to achieve the stripping of multiple lane lines one by one. In actual testing, the system successfully extracted 15 candidate lines iteratively. For example, candidate line 0 contains 5703 points and has a physical length of 79.69m; candidate line 1 contains 5016 points and has a physical length of 68.62m, etc.

[0102] Structured geometric feature calculation: For each successfully extracted candidate line, the system not only saves its point cloud data, but also simultaneously calculates and structurally stores its key geometric features: Physical length calculation: Calculate the bounding box of the point set within the line segment on a two-dimensional plane, and characterize the physical length of the line segment by calculating the maximum span. .

[0103] Geometric center calculation: Calculate the average coordinates of all interior points of the line segment in a two-dimensional plane to obtain the geometric center point. .

[0104] Adaptive Curve Fitting with Curvature Constraints: Considering the presence of curves in real roads, this embodiment further introduces a quadratic polynomial fitting mechanism with curvature constraints on top of straight line extraction. The spans of the point set along the X and Y axes are compared to determine the dominant direction. If the Y-axis span is greater than the X-axis span, then Y is used as the independent variable for fitting, i.e. Conversely, X is used as the independent variable, i.e. By performing a quadratic polynomial fit on the point set, the equation is obtained. Extracting the coefficients of the quadratic term. As a value representing curvature. If Greater than the set maximum curvature threshold In this embodiment, the value is 0.005, indicating that the fitting result may be affected by noise interference or overfitting. In this case, the system actively degenerates the model into a linear fit; otherwise, the quadratic curve model is retained. At the same time, the Z-axis coordinate is fitted independently to ensure that the final smooth lane line model can closely fit the three-dimensional undulation slope of the road surface.

[0105] Spatial topology calculation: To perform global topology verification in subsequent steps, the system defines the following spatial relationship calculation logic: Parallelism verification: For any two candidate lines A and B, extract the normal vector of their linear model. and Calculate the cosine of the angle between the two normal vectors. If the included angle If the angle is less than the set parallel angle threshold, the two lines are considered to be approximately parallel.

[0106] Lateral spacing calculation: using the geometric center point of line B Straight line model to line A The formula for the distance from a point to a line is used to calculate the horizontal and vertical distance between two lines. .

[0107] Step S6 involves constructing a global topology map model and a multi-dimensional scoring mechanism to filter candidate lane lines and resolve conflicts, outputting the final lane line vector map. Specifically, the road geometric centerline is calculated based on the candidate line set, and close-range parallel road edges and overlapping interference are strictly eliminated using the center deviation criterion; a multi-dimensional scoring model is constructed, including physical length score, edge protection score, and global parallel topology score; based on the scoring, non-maximum suppression with conflict resolution radius is performed to resolve spatial conflicts and output the final real lane lines.

[0108] Lane selection and conflict resolution based on a global topology graph model and multi-dimensional scoring mechanism (GlobalTopology Filter): The candidate lane set extracted in step nine. In addition to real lane lines, there are often false line segments caused by road markings (such as directional arrows and printed text), curb edges, and splicing artifacts. To accurately remove these interferences, this embodiment constructs a screening method based on a global topology graph model and a multi-dimensional scoring mechanism.

[0109] Spatial ordering and road centerline estimation: Extracting all candidate lines x-coordinate of the geometric center point Assuming the road mainly extends along the Y-axis, the candidate line set is divided into... By spatially sorting from smallest to largest, we obtain an ordered set. Calculate the lateral average of the geometric centers of all candidate lines as an estimate of the lateral coordinate of the global centerline of the road. :

[0110] Strict removal of near-distance interference based on center deviation (removal of near-distance parallel curbs and signs): After RANSAC or the thresholding stage, interference segments that are close to the actual lane lines may be generated (such as parallel segments caused by the curb slope remaining due to the lower limit of height filtering). Traverse adjacent and parallel candidate line pairs in the ordered set. Calculate its horizontal and vertical spacing. Set the near-field interference threshold. In this embodiment, the setting is within 1.0m or even as low as 0.65m, and will be adjusted according to data quality; currently, it is set to 1.0m. If the two line segments are closely matched, they are considered a "close" conflict pair, meaning one is a real lane line and the other is likely a curb or shaded edge. The deviations of the two lines from their estimated positions relative to the road centerline are compared: the line with the smaller deviation, closer to the road center, is retained, while the line with the larger deviation is discarded. This cleaning process yields a deduplicated set of candidate lines. .

[0111] Multidimensional topological scoring graph model construction, core anti-interference mechanism: for set Each candidate line in Initialize topology confidence score The score consists of three parts: basic geometry score, edge protection score, and global topology score. Basic geometric length classification: the physical length of a line segment Calculate the basic weights and scores:

[0112] in To enhance the length weighting factor (0.06 in this embodiment), long and continuous line segments are significantly encouraged to survive.

[0113] Edge protection: in ordered sets In the middle, the leftmost (index 0) and the rightmost (index 1) The line segment is highly likely to be the actual physical boundary of the road, and is assigned a fixed edge protection score. In this embodiment, 10.0 is used:

[0114] In actual testing, the leftmost line 0 and the rightmost line 10 successfully obtained edge protection points.

[0115] Global topology analysis: Real lane lines strictly adhere to multiples of the standard lane width, while road markings are randomly distributed and lack this topological prior. For any two parallel candidate lines... and ( ), calculate its horizontal spacing Set standard lane width. In this embodiment, 3.6m and width tolerance are used. In this embodiment, 0.8m is used. If the spacing satisfies:

[0116] This means and This constitutes a valid lane topology pair, for Add topology points:

[0117] in To add points for topology fundamentals, this example uses 5.0. The number of lanes crossed is used to determine the score. This mechanism allows real lane lines to receive higher scores due to mutual corroboration, while isolated road markings receive lower scores because they cannot form standard spacing with other lane segments. In actual testing, after multi-dimensional scoring, real lane lines, line 0, scored 28.28 and line 6 scored 21.42, significantly higher than interfering road marking lines, such as line 9, which scored only 14.62.

[0118] Non-maximum suppression based on conflict resolution radius: After establishing the aforementioned score, a local suppression with conflict distance is performed to remove misfit lines that are closely spaced and interfere with each other. The mutual exclusion resolution radius for parallel lines is set. In this embodiment, the distance is set to 2.5m, which is less than the width of a single lane, thus avoiding interference across lanes. Within the set, searches are conducted on elements with a distance less than [a certain value]. For parallel line pairs, when a conflict occurs, the topological confidence score is always removed. The lower line:

[0119] Repeat this conflict checking and removal process until the distance between any two lines in the set is greater than or equal to the specified distance. Parallel mutually exclusive lines exist.

[0120] Final Lane Line Output and Smoothing Reconstruction: After the above topology filtering and conflict resolution, the remaining line segments in the set are the final true lane lines. The corresponding interior point set is extracted and substituted into the quadratic polynomial model with curvature constraints from step nine for smooth reconstruction, outputting the final high-precision, structured, and interference-free global lane line vector map. (Example:) Figure 8 As shown, in actual testing on a typical urban elevated road section, the system successfully retained 6 actual lane lines from 12 candidate lines. The spacing between adjacent lane lines were 3.60m, 3.88m, 3.86m, 4.50m, and 4.39m, respectively. The spacing between the main lanes fully conforms to the physical topology of real roads (3.3m-4.6m), effectively eliminating interference from curbs and road markings.

[0121] like Figure 7 As shown, to verify the overall superiority and inventiveness of the algorithm in this embodiment, the extraction effects of the traditional conventional process and the coupling mechanism of this invention are compared in detail: (1) Results of the conventional process: Absolute intensity threshold segmentation and conventional RANSAC uniform sampling were used, but no rigorous topology filtering was performed. Due to radar scanning distance attenuation, threshold segmentation only extracted 15,031 candidate points, and a large number of dark lane lines were mistakenly filtered out. Due to the extremely sparse point cloud and noise interference, uniform RANSAC suffered severe geometric shift, mistakenly fitting discrete road markings and road edges as lane lines. The final output had 12,653 points within the lines, of which 15 lines were retained, with an average of only about 843 points per line. This proves that the simple concatenation of conventional algorithms is prone to topology logic errors in complex scenarios.

[0122] (2) Effect of the full feedback chain mechanism of this invention: By transforming absolute intensity into local contrast through BEV morphological top-hat transformation, the number of potential effective candidate points in the threshold segmentation stage increases to 27,327. Furthermore, by constructing dual sampling weights using the enhanced intensity values ​​and the longitudinal density distribution of the point cloud, the RANSAC algorithm is guided to avoid being induced by isolated discrete features, thereby tightly locking the real lane line skeleton. Then, by using the global topology graph model and multi-dimensional scoring mechanism for screening and conflict resolution, the number of effective lane line points extracted is significantly improved. The final output has 16,583 lane line points, of which 6 lines are retained, with an average of approximately 2,763 points per line. The lane line spacing conforms to the real physical world, achieving a significant technical improvement from a messy point cloud to a rigorous physical feature distribution.

[0123] like Figure 8 The diagram illustrates the final output of the entire processing workflow of this invention. Against the background of the main road point cloud, the generated lane lines are accurately superimposed as vectorized line segments, exhibiting smoothness and correct topological relationships. This result can not only be directly used as ground values ​​for training autonomous driving perception algorithms but also as pre-annotation input for high-precision map vectorization, significantly reducing the cost of manual verification and drawing.

[0124] To further verify the technical effectiveness of this invention, a batch test was conducted using the system of this invention on 300 frames of raw point cloud data continuously collected from a road section in a city. The test data was divided into 30 batches, with 10 frames fused in each batch, resulting in an initial fused point cloud volume of approximately 2.1 million points per batch. The quantitative data statistics are as follows: Verification Experiment 1: Quantization Comparison of Recursive Otsu Adaptive Segmentation In all 30 batches of testing, the number of main road points processed each time was counted, averaging about 430,000 points.

[0125] Conventional comparison group: The traditional single-run Otsu algorithm was used, and the calculated global threshold was consistently low, with an intensity level of around 26-27. This resulted in an average foreground candidate retention rate as high as 53.2%, with a large number of ordinary asphalt pavements being mistaken for lane lines. Subsequent fitting often collapsed due to excessive noise.

[0126] The experimental group of this invention employs the recursive Otsu and statistical lower limit protection mechanism of this invention. The system successfully detected initial retention rate anomalies in each batch and adaptively triggered recursive local subdivision, intelligently raising the final segmentation threshold to 62-64. The final lane line candidate point cloud retention rate remained stable between 6.4% and 7.0%.

[0127] Conclusion: This invention reduces the noise level in candidate interior points by nearly 8 times, demonstrating the algorithm's superiority in removing large-area non-uniform background interference. Figure 9 The threshold segmentation noise reduction effect comparison chart shown has 30 consecutive test batches on the horizontal axis and point cloud retention rate on the vertical axis. The chart intuitively shows the traditional single Otsu error retention rate, represented by a dashed line with a cross, which is stable at a high bandwidth of 50%~56%. In contrast, the retention rate of the present invention, represented by a solid line with a solid dot, is stable at a low bandwidth of about 6.5%, showing a significant advantage in high purity.

[0128] Verification Experiment 2: Stability of Global Topology Graph Model Output and Quantization of True Distribution in Geometric Fitting and Screening Stages: Dynamic output quantity statistics: In 30 consecutive test batches, the sequence of the number of valid lane lines retained by this invention is as follows: [6,6,5,5,5,5,5,6,5,6,6,6,6,6,6,6,6,6,6,5,6,6,6,4,4,4,4,4,4,4,4].

[0129] Data shows that the system stably outputs 5-6 real lane lines on regular road sections, while in the final stage when the road network narrows, the algorithm can accurately adapt to 4 lane lines. For example... Figure 10The comparison chart of lane line counts before and after multi-frame topology filtering shows that the total height of the bars represents the total number of messy candidate lines extracted by the initial RANSAC, approximately 14-15 lines. Among them, the cross-grid texture (xx) segment represents the false curb lines that were accurately identified and eliminated by the algorithm based on center deviation, approximately 4-5 lines. The diagonal texture ( / / / / ) segment represents the interference lines that were eliminated due to global topology conflicts and low scores, approximately 4-5 lines. The unfilled blank segment represents the number of lane lines that were finally confirmed as real lane lines and retained, which is stable at 5-6 lines. This visually demonstrates the ability of the topology graph model to distinguish between real and false lane lines in complex environments.

[0130] Basic physics confirms that the algorithm's output distances between adjacent main lanes, such as 3.76m, 3.86m, and 4.20m, all fall within the reasonable range of real-world road design. Figure 11 As shown in the scatter plot, the lane spacing physical consistency is represented by a scatter plot. The horizontal axis represents the processing batch, and the vertical axis represents the lateral spacing between adjacent lane lines. The two dashed lines mark the actual physical spacing tolerance range of 3.3m to 4.6m. Traditional outputs without topological constraints, represented by crosses 'x', exhibit a disordered and chaotic distribution, largely scattered within various erroneous value ranges such as 0.1m, 0.5m, or 2.9m. In contrast, the outputs of this invention, represented by circles 'o', are highly cohesive and largely fall within the actual road safety distribution zone. It should be noted that... Figure 11 The scattered circles located at the 0m baseline correspond to a few topology fitting errors. In consecutive test batches, only four lane topology offsets occurred globally. Since a single lane offset can cause the calculation of the distance between two adjacent lanes on its left and right sides to fail, to objectively and systematically reflect the algorithm's true false detection status, this embodiment uniformly sets the erroneous distances caused by single-line topology deviations to zero and visually marks them on the 0m baseline in the chart. Despite these four extreme errors, and their accurate recording, the overall data distribution of the system still exhibits robust high cohesion, fully demonstrating the global reliability of the multi-dimensional topology filtering mechanism described in this patent when constructing a high-precision road network skeleton.

[0131] Verification Experiment 3: Robustness Analysis under Complex Road Morphology and Extreme Observation Conditions, Verification in Real-World Long-Tail Scenarios. In long-distance real-world road tests, point cloud quality degradation or abrupt changes in road topology are inevitable, such as at ramp entrances and exits. Combining the output sequences of the aforementioned 30 batches, the mechanism of this invention demonstrates a high-purity priority strategy with engineering value in such complex edge scenarios.

[0132] High-confidence feature extraction and multi-frame spatiotemporal correlation mechanism in sparse point cloud scenarios: In the output sequence, due to local radar physical obstruction or significant attenuation of point cloud reflectivity, the algorithm experienced a small number of feature omissions in individual batches such as Batch 3-7 and Batch 19. The weighted iterative RANSAC and multi-dimensional topology filtering mechanism of this invention adopts an accuracy-oriented preprocessing principle, actively rejecting low-confidence fitting requests, thereby effectively avoiding the generation of false features. More importantly, based on the feature correlation and restoration characteristics of multi-frame temporal information, this system enables such localized and occasional feature omissions to be recovered in the global map aggregation process by stable redundant features provided by adjacent batches in the physical topology, ensuring the integrity and coherence of the final global vectorized map.

[0133] Adaptive handling of topological abrupt changes at ramp entrances and exits: The algorithm demonstrates good adaptive robustness when handling topological abrupt changes in areas such as highway ramp entrances and exits. For Batch 22, although merging / diverging causes non-standard numerical fluctuations in lane spacing on the projection plane, the system successfully locked onto the true target, without erroneous line additions or subtractions due to geometrical abrupt changes, proving the algorithm's accuracy in recognizing non-parallel structures. For the extreme topological abrupt change scenario in Batch 23, although one fitting bias and three missed line segments occurred within a single batch, the spatiotemporal redundancy provided by Batch 22 and subsequent stable features allowed this area to be effectively completed and restored in the global aggregated map. Experimental data demonstrates that the conflict resolution mechanism of this invention can effectively filter temporary topological conflicts caused by geometrical abrupt changes, achieving robust compatibility with long-tailed complex scenarios while ensuring the continuity of the main road framework. Figure 12 The global mapping aggregation effect diagram shown in the figure seamlessly stitches together the outputs of the aforementioned 1 to 30 batches, intuitively displaying the complete physical road network under a long-distance environment for 300 consecutive frames.

[0134] After rigorous manual verification based on PCD point cloud maps, fitting / topology errors were found in four locations out of approximately 160 lane segments extracted from all 30 batches. The overall accuracy of the algorithm is high. Experimental results fully demonstrate that the processing mechanism described in this invention has extremely high reliability and engineering application value for lane line vectorization reconstruction.

[0135] Example 2

[0136] This embodiment provides a lane line vectorization pre-annotation system based on multi-frame LiDAR point clouds, used to implement the lane line vectorization pre-annotation method described above, including: The point cloud acquisition and stitching module is used to acquire multiple frames of raw point cloud data from LiDAR, perform spatial coarse filtering and spatiotemporal synchronous stitching, and construct a global dense point cloud map. The road surface extraction module is used to perform elevation fine filtering and plane segmentation on the global dense point cloud map to extract the main road surface point cloud. The intensity enhancement module is used to perform non-uniform background suppression and intensity enhancement on the main road surface point cloud based on morphological top-hat transformation; The threshold segmentation module is used to extract lane line candidate point clouds based on the enhanced reflection intensity using an adaptive threshold segmentation method that combines statistical lower bound and recursive maximum inter-class variance method. The geometric fitting module is used to perform weighted iterative RANSAC straight line fitting based on both intensity and longitudinal density guidance on the lane line candidate point cloud, and extract structured candidate lane lines and geometric features. The topology filtering module is used to construct a global topology graph model and a multi-dimensional scoring mechanism, filter candidate lane lines and resolve conflicts, and output the final lane line vector map.

[0137] The embodiments described herein are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made to the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention. Although preferred embodiments of the present invention have been described, those skilled in the art, once they understand the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the present invention. Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A lane line vectorization pre-annotation method based on multi-frame LiDAR point clouds, characterized in that, Includes the following steps: Step S1: Acquire multiple frames of raw point cloud data from LiDAR, perform spatial coarse filtering and spatiotemporal synchronous stitching to construct a global dense point cloud map; Step S2: Perform elevation fine filtering and plane segmentation on the global dense point cloud map to extract the main road surface point cloud; Step S3: Perform non-uniform background suppression and intensity enhancement on the main road surface point cloud based on morphological top-hat transformation; Step S4: Based on the enhanced reflection intensity, an adaptive threshold segmentation method combining statistical lower bound and recursive maximum inter-class variance method is used to extract lane line candidate point clouds. Step S5: Perform weighted iterative RANSAC straight line fitting based on intensity and longitudinal density dual guidance on the lane line candidate point cloud to extract structured candidate lane lines and geometric features. Step S6: Construct a global topology map model and a multi-dimensional scoring mechanism to filter candidate lane lines and resolve conflicts, and output the final lane line vector map.

2. The lane line vectorization pre-annotation method based on multi-frame LiDAR point clouds as described in claim 1, characterized in that, Step S1 specifically involves: Step S11: Obtain raw point cloud data; Step S12: Perform radial distance filtering on the single-frame raw point cloud: Calculate the radial distance of each point on the horizontal plane according to a preset distance threshold range, and filter out points whose radial distance is outside the threshold range; perform coarse filtering on the point cloud after radial distance filtering based on the local elevation histogram, extract its elevation data to construct a one-dimensional histogram, use the elevation corresponding to the peak of the histogram as the local ground reference, set the elevation region of interest based on the reference, and filter out points whose elevation is outside the region; Step S13: Analyze the translation vector and rotation matrix corresponding to each frame of point cloud from the pose record, transform the local point cloud of each frame to the world coordinate system through homogeneous coordinate transformation, and stitch and fuse all the transformed point clouds to generate a global dense point cloud map.

3. The lane line vectorization pre-annotation method based on multi-frame LiDAR point clouds as described in claim 1, characterized in that, Step S2 specifically involves: Step S21: Extract elevation data of all points from the global dense point cloud map, construct an elevation histogram within a preset percentile interval, and use the elevation corresponding to the peak value of the histogram as the global ground reference; set an elevation region of interest based on this reference, filter out points whose elevation is outside the region, and obtain a finely filtered candidate point cloud for the main road surface. Step S22: For the finely filtered point cloud, the random sampling consensus algorithm is used to iteratively fit the road surface plane. The number of interior points of each candidate plane is counted with a preset distance tolerance. The plane with the most interior points is selected as the optimal road surface model, and the road surface point cloud is extracted based on the model. Step S23: Voxel downsample the road point cloud, estimate the local normal vector of each point and unify the direction, calculate the angle between the normal vector and the global reference normal, and filter out non-smooth points with an angle greater than a preset threshold; perform density clustering on the remaining point cloud and extract the connected component with the most points as the main road point cloud.

4. The lane line vectorization pre-annotation method based on multi-frame LiDAR point clouds as described in claim 1, characterized in that, Step S3 specifically involves: projecting the main road surface point cloud to construct a two-dimensional bird's-eye view intensity map; setting the grid resolution, mapping the main road surface point cloud to a two-dimensional horizontal grid, taking the maximum reflection intensity of the point cloud within the grid as the pixel value, and generating a BEV intensity map; setting the physical size of the structural element and converting it to pixel size, and performing a morphological opening operation on the BEV intensity map using the structural element; subtracting the result of the opening operation from the BEV intensity map to obtain an enhanced two-dimensional intensity map; inversely mapping the pixel values ​​of the enhanced two-dimensional intensity map to the corresponding three-dimensional point cloud, performing linear stretching, and updating the reflection intensity value of the main road surface point cloud.

5. The lane line vectorization pre-annotation method based on multi-frame LiDAR point clouds as described in claim 1, characterized in that, Step S4 specifically involves: calculating the global mean and standard deviation of the point cloud intensity of the main road surface, and constructing a statistical lower limit threshold; performing initial global segmentation using the maximum inter-class variance method, and calculating the retention rate of candidate foregrounds; if the retention rate is greater than a preset tolerance threshold, triggering recursive local subdivision in the candidate foregrounds to find the local optimal threshold; and determining the final segmentation threshold by combining the statistical lower limit threshold, the initial global threshold, and the local optimal threshold, and extracting candidate point clouds for lane lines.

6. The lane line vectorization pre-annotation method based on multi-frame LiDAR point clouds as described in claim 1, characterized in that, Step S5 specifically involves: calculating the longitudinal density distribution of the point cloud along the main direction, weighting and fusing it with the enhanced reflection intensity to construct a dual sampling weight; performing non-uniform random sampling based on this weight, prioritizing the extraction of features with high connectivity and high intensity; extracting multiple candidate line models through the iterative RANSAC algorithm, and calculating the physical length, geometric center, and other structured geometric features of each candidate line; and introducing a quadratic polynomial fitting mechanism with curvature constraints to adaptively fit the candidate lines to curves.

7. The lane line vectorization pre-annotation method based on multi-frame LiDAR point clouds as described in claim 1, characterized in that, Step S6 specifically involves: calculating the road geometric centerline based on the candidate line set, strictly eliminating close-range parallel road edges and overlapping interference using the center deviation criterion; constructing a multi-dimensional scoring model that includes physical length score, edge protection score, and global parallel topology score; performing non-maximum suppression with conflict resolution radius based on the score to resolve spatial conflicts and output the final real lane lines.

8. The lane line vectorization pre-annotation method based on multi-frame LiDAR point clouds as described in claim 5, characterized in that, The method of determining the final segmentation threshold by combining the comprehensive statistical lower limit threshold, the initial global threshold, and the local optimum threshold, and extracting lane line candidate point clouds specifically involves: if the retention rate is greater than a preset tolerance threshold, then the larger of the local optimum threshold and the statistical lower limit threshold is taken as the final segmentation threshold; if the retention rate is not greater than the preset tolerance threshold, then the larger of the initial global threshold and the statistical lower limit threshold is taken as the final segmentation threshold.

9. The lane line vectorization pre-annotation method based on multi-frame LiDAR point clouds as described in claim 7, characterized in that, The multidimensional scoring model is as follows: the road centerline is calculated based on the geometric center of the candidate lines; for parallel line pairs with a spacing less than the proximity interference threshold, the one closer to the center is retained based on the center deviation; a multidimensional topology score is constructed, including a basic geometric length score positively correlated with the line segment length, an edge protection score with a fixed reward for the outermost line segment, and a global topology score based on the relationship between the lateral spacing and the standard lane width; finally, non-maximum suppression with a conflict resolution radius is performed based on the total score to eliminate low-scoring conflict line segments.

10. A lane line vectorization pre-annotation system based on multi-frame LiDAR point clouds, used to implement the lane line vectorization pre-annotation method based on multi-frame LiDAR point clouds as described in any one of claims 1-9, characterized in that, include: The point cloud acquisition and stitching module is used to acquire multiple frames of raw point cloud data from LiDAR, perform spatial coarse filtering and spatiotemporal synchronous stitching, and construct a global dense point cloud map. The road surface extraction module is used to perform elevation fine filtering and plane segmentation on the global dense point cloud map to extract the main road surface point cloud. The intensity enhancement module is used to perform non-uniform background suppression and intensity enhancement on the main road surface point cloud based on morphological top-hat transformation; The threshold segmentation module is used to extract lane line candidate point clouds based on the enhanced reflection intensity using an adaptive threshold segmentation method that combines statistical lower bound and recursive maximum inter-class variance method. The geometric fitting module is used to perform weighted iterative RANSAC straight line fitting based on both intensity and longitudinal density guidance on the lane line candidate point cloud, and extract structured candidate lane lines and geometric features. The topology filtering module is used to construct a global topology graph model and a multi-dimensional scoring mechanism, filter candidate lane lines and resolve conflicts, and output the final lane line vector map.

Images