Method and system for registering different-source lane lines to point cloud

Abstract

The invention discloses a method for registering a different-source lane line to a point cloud, and the method comprises the following steps: inputting point cloud data and different-source lane line data to be registered, the different-source lane line data comprising a plurality of lane line-shaped points; preprocessing the point cloud data, and classifying, identifying and extracting all ground point clouds through the point clouds; based on the average elevation of all the ground point clouds, elevation initialization is carried out on the plurality of lane line-shaped points, so that the plurality of lane lines are initially aligned to a point cloud plane; a class iteration nearest point ICP registration main process is executed, a variable-length search radius mechanism and a divided lane optimization strategy are integrated and adopted in the main process, and the variable-length search radius mechanism gradually reduces the search radius along with the increase of the number of iterations; according to the divided lane optimization strategy, independent calculation is carried out on each lane line in each iteration, and the optimal rotation matrix and the optimal translation vector of each lane line are recorded and applied; and updating the linear point positions of the plurality of lanes based on the optimal rotation matrix and the optimal translation vector, and outputting registered different-source lane line data.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and high-precision mapping, specifically to a method and system for registering heterogeneous lane lines to point clouds. Background Technology

[0002] In the field of high-precision map creation and lane marking for autonomous driving, accurately registering lane line data from different sources (heterogeneous sources) with collected laser point cloud data is a key technical step to improve mapping efficiency and accuracy. However, existing registration methods face many technical bottlenecks in practical applications.

[0003] Registration methods based on Normal Distribution Transform (NDT) rely on point cloud intensity information to construct a probability density function. This method is easily affected by the granularity of grid division and the accuracy of initial position estimation, which often leads to the registration process getting stuck in local optima and makes it difficult to achieve high registration accuracy.

[0004] While the traditional Iterative Closest Point (ICP) method is widely used, its registration accuracy heavily relies on accurate initial pose and is extremely sensitive to the search radius setting. More importantly, the traditional ICP method only supports overall rigid transformations and cannot effectively handle complex scenarios where scale inconsistencies or significant local deformations are common between heterogeneous data sources, leading to systematic biases in the registration results.

[0005] Another approach involves first identifying and extracting lane line features from the point cloud, and then aligning them with heterogeneous lane line data. While this method avoids the difficulties of direct registration, it is complex, introduces intermediate processing steps, and its final result largely depends on the accuracy of point cloud lane line identification and extraction. This not only increases the number of processing steps but also introduces new sources of error. Summary of the Invention

[0006] To overcome the above-mentioned technical defects, the present invention provides a method and system for registering heterogeneous lane lines to point clouds.

[0007] To solve the above problems, the present invention is implemented according to the following technical solution:

[0008] In a first aspect, the present invention provides a method for registering heterogeneous lane lines to a point cloud, characterized by comprising the following steps: inputting point cloud data and heterogeneous lane line data to be registered, wherein the heterogeneous lane line data includes multiple lane line points; preprocessing the point cloud data by identifying and extracting all ground point clouds through point cloud classification; initializing the elevation of the multiple lane line points based on the average elevation of all ground point clouds, so that the multiple lane lines are initially aligned to the point cloud plane; executing a class iterative nearest point (ICP) registration main process, wherein the main process integrates a variable-length search radius mechanism and a lane-specific optimization strategy, wherein the variable-length search radius mechanism gradually decreases the search radius as the number of iterations increases, and the lane-specific optimization strategy independently calculates, records, and applies the optimal rotation matrix and translation vector for each lane line in each iteration; updating the positions of the multiple lane line points based on the optimal rotation matrix and translation vector, and outputting the registered heterogeneous lane line data.

[0009] In conjunction with the first aspect, the present invention provides a first specific implementation of the first aspect. Specifically, the variable-length search radius mechanism includes: in the initial iteration stage, using a first preset radius as the search radius for coarse matching; in the intermediate iteration stage, dynamically reducing the search radius according to the ratio of the length of the translation vector in the current iteration to the length of the initial translation vector; and in the final iteration stage, using a second preset radius as the search radius for fine matching; wherein the second preset radius is smaller than the first preset radius.

[0010] In conjunction with the first aspect, the present invention provides a second specific implementation of the first aspect. Specifically, the lane optimization strategy includes: calculating the matching degree between each lane line and the point cloud, wherein the matching degree is determined based on the average intensity value of all adjacent ground point clouds of the lane line point; recording the optimal matching degree of each lane line and its corresponding rotation matrix and translation vector during the iteration process; and updating the position of the lane line point of each lane line using its recorded optimal rotation matrix and translation vector after registration is completed.

[0011] In conjunction with the first aspect, the present invention provides a third specific implementation of the first aspect. Specifically, the landmark registration steps are as follows: the landmark elements are registered using an alternating direction iterative method, including: fixing the first direction, searching along the second direction and determining the optimal offset to maximize the sum of the ground point cloud intensity values ​​covered by the landmark; fixing the second direction, searching along the first direction and determining the optimal offset to maximize the sum of the ground point cloud intensity values ​​covered by the landmark; iteratively executing the above bidirectional search steps until the convergence condition is met.

[0012] In conjunction with the first aspect, the present invention provides a fourth specific implementation of the first aspect. Specifically, the main process of execution class iterative nearest point ICP registration includes: for each lane alignment point, determining its neighboring point cloud set from the ground point cloud according to the search radius of the current iteration; calculating the corresponding weighted center point based on the intensity value of each point cloud in the neighboring point cloud set; constructing a matching pair between the lane alignment point and the corresponding weighted center point; and using singular value decomposition (SVD) to solve for the rotation matrix and translation vector that minimizes the distance between the matching pairs based on all matching pairs.

[0013] Secondly, the present invention also provides a system for registering heterogeneous lane lines to a point cloud, comprising: a data input module for inputting point cloud data and heterogeneous lane line data to be registered, wherein the heterogeneous lane line data includes multiple lane line points; a preprocessing module for preprocessing the point cloud data, identifying and extracting ground point cloud data through point cloud classification; an elevation initialization module for initializing the elevation of the multiple lane line points based on the average elevation of the ground point cloud, so that the multiple lane lines are initially aligned to the point cloud plane; a registration main process module for executing an iterative nearest point (ICP) registration main process, wherein the main process integrates a variable-length search radius mechanism and a lane-specific optimization strategy, wherein the variable-length search radius mechanism gradually decreases the search radius as the number of iterations increases, and the lane-specific optimization strategy independently calculates, records, and applies the optimal rotation matrix and translation vector for each lane line in each iteration; and an output module for updating the positions of the multiple lane line points based on the optimal rotation matrix and translation vector, and outputting the registered heterogeneous lane line data. In conjunction with the second aspect, the present invention provides a first specific implementation of the second aspect. Specifically, the variable-length search radius mechanism in the registration main process module includes: in the initial iteration stage, using a first preset radius as the search radius for coarse matching; in the intermediate iteration stage, dynamically reducing the search radius according to the ratio of the length of the translation vector in the current iteration to the length of the initial translation vector; and in the final iteration stage, using a second preset radius as the search radius for fine matching; wherein the second preset radius is smaller than the first preset radius.

[0014] In conjunction with the second aspect, the present invention provides a second specific implementation of the second aspect. Specifically, the lane optimization strategy executed by the lane optimization module includes: calculating the matching degree between each lane line and the point cloud, wherein the matching degree is determined based on the average intensity value of all adjacent ground point clouds of the lane line point; recording the optimal matching degree of each lane line and its corresponding rotation matrix and translation vector during the iteration process; and after registration is completed, each lane line independently adopts its recorded optimal rotation matrix and translation vector.

[0015] In conjunction with the second aspect, the present invention provides a third specific implementation of the second aspect. Specifically, the landmark registration module is used to register landmark elements using an alternating direction iterative method, including: fixing the first direction, searching along the second direction and determining the optimal offset to maximize the sum of the ground point cloud intensity values ​​covered by the landmark; fixing the second direction, searching along the first direction and determining the optimal offset to maximize the sum of the ground point cloud intensity values ​​covered by the landmark; iteratively executing the above bidirectional search steps until the convergence condition is met.

[0016] In conjunction with the second aspect, the present invention provides a fourth specific implementation of the second aspect. Specifically, the main registration process module executes an iterative nearest point (ICP) registration main process including: for each lane alignment point, determining its neighboring point cloud set from the ground point cloud based on the search radius of the current iteration; calculating the corresponding weighted center point based on the intensity value of each point cloud in the neighboring point cloud set; constructing a matching pair between the lane alignment point and the corresponding weighted center point; and using singular value decomposition (SVD) to solve for the rotation matrix and translation vector that minimizes the distance between the matching pairs based on the matching pairs.

[0017] Compared with the prior art, the beneficial effects of the present invention are:

[0018] By preprocessing and extracting ground point clouds, a reference benchmark is provided for registration, avoiding interference from non-ground objects such as vehicles and vegetation, thus improving registration accuracy from the source. Traditional methods treat all lane lines as a whole for optimization; when some lane lines have poor data quality or are occluded, it will drag down the overall registration effect. This strategy calculates transformation parameters independently for each lane line, allowing lane lines with good data quality to quickly converge to the correct position, while lane lines with poor data quality will not affect the registration of other lines. For complex road sections with curves, slopes, or non-parallel lanes, different lane lines have unique local transformation relationships. Lane-by-lane optimization can accurately capture and adapt to these local changes, thus achieving excellent registration results even in complex scenarios. Attached Figure Description

[0019] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, wherein:

[0020] Figure 1 This is a flowchart of a method for registering heterogeneous lane lines to point clouds according to the present invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] like Figure 1As shown, this invention provides a method and system for registering heterogeneous lane lines to point clouds.

[0023] Example 1

[0024] A method for registering heterogeneous lane lines to a point cloud includes the following steps: inputting point cloud data and heterogeneous lane line data to be registered, wherein the heterogeneous lane line data contains multiple lane line points; preprocessing the point cloud data by identifying and extracting all ground point clouds through point cloud classification; initializing the elevation of the multiple lane line points based on the average elevation of all ground point clouds, so that the multiple lane lines are initially aligned to the point cloud plane; executing a class-iterative nearest point (ICP) registration main process, in which a variable-length search radius mechanism and a lane-specific optimization strategy are integrated, wherein the variable-length search radius mechanism gradually decreases the search radius as the number of iterations increases, and the lane-specific optimization strategy independently calculates, records, and applies its optimal rotation matrix and translation vector for each lane line in each iteration; updating the positions of the multiple lane line points based on the optimal rotation matrix and translation vector, and outputting the registered heterogeneous lane line data.

[0025] In a preferred embodiment, input point cloud data and heterogeneous lane line data to be registered, wherein the heterogeneous lane line data includes multiple lane line points.

[0026] Specifically, the input includes point cloud data of the target scene and lane line data from other heterogeneous sensors (such as cameras, radar, or high-precision maps). This data is stored in .json format and contains multiple lane lines, each lane line consisting of a series of ordered points with X, Y, and Z coordinates in the world coordinate system. Preferably, the point cloud data comes from a 64-line LiDAR, with a point cloud density of approximately 200 points / m²; the data is stored in .pcd format and contains the three-dimensional coordinates and reflection intensity information of all objects in the scene.

[0027] In a preferred embodiment, the point cloud data is preprocessed, and all ground point clouds are identified and extracted through point cloud classification.

[0028] Specifically, a ground extraction algorithm is used to classify and identify the raw point cloud data, removing non-ground points such as vehicles, buildings, and vegetation. The advanced point cloud semantic segmentation network RandLA-Net is employed, which can efficiently process large-scale point clouds. The point cloud data of the entire scene is input into this network, and its output assigns a semantic label to each point, such as "ground," "vegetation," "building," "vehicle," or "road marking." Then, all points classified as "ground" are extracted to form a clean ground point cloud target set.

[0029] In a preferred embodiment, the elevation of the multiple lane line points is initialized based on the average elevation of all the ground point clouds, so that the multiple lane lines are initially aligned with the point cloud plane.

[0030] Specifically, the average elevation value H of all extracted ground point clouds is calculated. avg The input lane alignment points are set to Z coordinates, so that the overall lane alignment is located at H. avg Elevation plane. Preferably, the elevation initialization uses the formula: Z′=Z+(H avg -H ine ); where H ine This represents the average elevation of the set of lane alignment points. This initialization process allows the lane lines to be initially aligned to the ground plane, pulling lane alignment points that were originally floating in the air or sunk underground to a plane with the average height of the point cloud ground, achieving coarse alignment in the elevation direction and laying a solid foundation for subsequent fine registration. Preferably, the least squares method is used to fit the plane equation ax + by + cz + d = 0, and the lane line points are projected onto this plane.

[0031] In a preferred embodiment, an iterative nearest point (ICP) registration main process is executed, in which a variable-length search radius mechanism and a lane-specific optimization strategy are integrated. The variable-length search radius mechanism gradually reduces the search radius as the number of iterations increases, and the lane-specific optimization strategy independently calculates, records, and applies the optimal rotation matrix and translation vector for each lane line in each iteration.

[0032] Specifically, an ICP-like registration algorithm is used to iteratively optimize the spatial correspondence between lane lines and ground point clouds. In this main process, a "variable-length search radius mechanism" and a "lane-specific optimization strategy" are introduced to iteratively update the rotation matrix R and translation vector T of each lane line until convergence. Preferably, the convergence condition is: the magnitude change of the translation vector is less than 1 cm between two consecutive iterations, or the maximum number of iterations reaches 50.

[0033] In a preferred embodiment, the positions of the multiple lane line points are updated based on the optimal rotation matrix and translation vector, and the registered heterogeneous lane line data is output.

[0034] Specifically, once the convergence condition is met, the final registered lane line data is output. This result can be used for subsequent tasks such as map fusion and localization correction.

[0035] In a preferred embodiment, the variable-length search radius mechanism includes: in the initial iteration stage, using a first preset radius as the search radius for coarse matching; in the intermediate iteration stage, dynamically reducing the search radius according to the ratio of the length of the translation vector in the current iteration to the length of the initial translation vector; and in the final iteration stage, using a second preset radius as the search radius for fine matching; wherein the second preset radius is smaller than the first preset radius.

[0036] Specifically, the execution process of the variable-length search radius mechanism includes three stages:

[0037] In the initial iteration phase (coarse matching), a first preset search radius R1 = 1.5m is set. At the start of registration, the deviation between the lane line points and the actual lane line positions in the point cloud may be large. Using a larger search radius ensures that potential corresponding ground points are found for the vast majority of lane lines, effectively preventing the algorithm from failing to converge in the initial stage due to the inability to find corresponding points, thus completing large-scale zooming and coarse localization.

[0038] Intermediate Iteration Stage (Adaptive Scaling): The search radius is dynamically reduced based on the real-time convergence status of the registration. The translation convergence ratio α is defined and calculated using the formula: α = norm(T) current ) / norm(T initial ); where T initial It is the average translation vector calculated in the first iteration, T current This is the average translation vector calculated in the current iteration. α reflects the convergence of the overall translation from the beginning to the present. The search radius R in the current iteration... current R is dynamically determined by the following formula: current =R coarse *α; The value of α starts from 1. As registration progresses, the overall displacement gradually decreases, and α decreases accordingly. This ensures that the shrinking of the search radius is synchronized with the registration progress. When the positional deviation is large, the radius shrinks slowly, ensuring convergence stability; when the position is close to accurate, the radius shrinks rapidly, accelerating the transition to fine matching. This mechanism effectively avoids the problem of registration progress becoming disconnected from the number of iterations that may occur with a fixed step size.

[0039] Final Iteration Stage (Refined Matching): When the search radius R current Dynamically reduce to a preset second preset radius (e.g., R) fine When the radius is 0.1m, the mechanism enters the final stage. In this stage, the search radius is locked at R. fine No longer changing. fineThis is a small value (usually smaller than the lane line width) set based on the point cloud density and lane line width, which is sufficient to capture the precise position of the lane line edge while effectively eliminating nearby interference points (such as other lane lines or road surface cracks). Several iterations are performed under this fixed small radius to achieve "fine-tuning" of the lane line position, ensuring that the final registration result achieves the highest accuracy.

[0040] Compared to simple linear decreasing, this three-stage dynamic adjustment mechanism has the advantage that the change in the search radius is no longer a function of time (number of iterations), but a state function (convergence ratio α). This allows the algorithm to adapt to scenarios with different initial deviations. For data with large initial deviations, it will spend more iterations in the coarse matching stage to achieve stable convergence; for data with small initial deviations, it will enter the fine matching stage more quickly, thereby improving the overall efficiency of the algorithm.

[0041] In a preferred embodiment, the lane division optimization strategy includes: calculating the matching degree between each lane line and the point cloud, the matching degree being determined based on the average intensity value of all adjacent ground point clouds of the lane line point; recording the best matching degree of each lane line and its corresponding rotation matrix and translation vector during the iteration process; and updating the position of the lane line point for each lane line using its recorded best rotation matrix and translation vector after registration is completed.

[0042] Specifically, before the iteration begins, the system will define each lane line L i Initialize an optimal state record to store the best match (Score_best) found throughout the entire iteration process. i And the corresponding optimal rotation matrix R_best i and the best translation vector T_best i Initially, Score_best i It can be set to a very small value.

[0043] In each iteration, for each lane line L i Perform the following steps: Same as the basic strategy, using the current search radius R current , for L i Establish a set of corresponding point pairs C i And solve for the transformation parameters (R) of this iteration. ik T ik Subsequently, this transformation is applied to obtain the lane line L. i At the temporary new position after this iteration, calculate the matching score of this lane line at the current temporary new position. current Iterate through L at its temporary new position. iAll shape points. For each shape point, search all ground point clouds within a certain radius (e.g., 0.15 meters, slightly larger than the lane line width) of its projection onto the ground. Calculate the average reflection intensity value of these neighboring ground point clouds as evidence of local matching at that shape point. Since the reflection intensity of real lane lines (especially painted markings) to laser light is much higher than that of ordinary asphalt pavement, this average intensity value can effectively characterize whether the shape point is aligned with the highlighted features (i.e., lane lines) in the point cloud. Finally, average the local matching degree of all shape points to obtain the lane line L. i Overall matching score in the current iteration current Set the current matching score. current Score_best (match with historical best) i Compare them.

[0044] If Score current core_best i If the result is positive, it indicates that the current pose is optimal. At this point, the system will update L. i The "best state record", namely: core_best i =Score current R_best i =R ik T_best i =T ik Otherwise, retain the original best state record unchanged.

[0045] When the main registration process terminates (reaching the maximum number of iterations or convergence), we do not directly use the pose from the last iteration as the final result. Instead, we perform the optimal pose backtracking application steps: for each lane line L... i The system reads the set of transformation parameters (R_best) that has proven to be the best match throughout the entire iteration history from its "best state record". i ,T_best i This optimal set of rotation matrices and translation vectors is then applied one last time to the original shape point position of the lane line after elevation initialization (or equivalently, its final pose is calculated through cumulative transformation), thereby obtaining the final and optimal registered position of the lane line.

[0046] In a preferred embodiment, the method further includes a landmark registration step: registering the landmark elements using an alternating direction iterative method, including: fixing a first direction, searching along a second direction and determining the optimal offset to maximize the sum of the ground point cloud intensity values ​​covered by the landmark; fixing the second direction, searching along the first direction and determining the optimal offset to maximize the sum of the ground point cloud intensity values ​​covered by the landmark; iteratively executing the above bidirectional search steps until the convergence condition is met.

[0047] Specifically, alternating direction iterative optimization refers to alternately fixing one coordinate axis direction while performing a one-dimensional search in another direction to find the optimal position that maximizes the sum of point cloud intensities below the landmark coverage area. Assuming that the landmark mainly undergoes translational transformations in the plane, the optimization problem is defined as finding the optimal offset (Δx, Δy).

[0048] The specific iteration steps are as follows:

[0049] a. Fix the Y-direction, optimize the X-direction: Keep the current Y-coordinate of the landmark unchanged. Move in the X-direction within a search window (e.g., ±0.5 meters) at a preset step size (e.g., 0.02 meters). For each candidate X offset Δx candidate This involves calculating the sum of the reflection intensities of all ground point clouds covering the lane line data (typically polygonal features) to be registered after it has been translated to a new location. A Δx is then selected that maximizes this total intensity. candidate As the optimal offset Δx in the X direction for this round optimal And apply this offset to update the X coordinate of the landmark.

[0050] b. Fix the X direction, optimize the Y direction: Maintain the current new X coordinates of the landmark (i.e., Δx has already been applied). optimal The position remains unchanged; it moves in the Y direction with the same preset step size and search window. Similarly, for each candidate Y offset Δy candidate The sum of point cloud intensities over the marked area is calculated. The value Δy is chosen to maximize the sum of intensities. candidate As the optimal offset Δy in the Y direction for this round optimal And apply this offset to update the Y coordinate of the landmark.

[0051] Repeating steps a and b constitutes one complete iteration. The iteration terminates when (Δx) equals 0.5% in two consecutive iterations. optimal ,Δy optimal The changes in all directions are less than a very small threshold (e.g., 0.001 meters) or reach the maximum number of iterations. After the alternating direction iterations converge, the final registered landmark vector data is output.

[0052] In a preferred embodiment, the main process of performing iterative nearest point (ICP) registration includes: for each lane alignment point, determining its neighboring point cloud set from the ground point cloud based on the search radius of the current iteration; calculating the corresponding weighted center point based on the intensity value of each point cloud in the neighboring point cloud set; constructing a matching pair between the lane alignment point and the corresponding weighted center point; and using singular value decomposition (SVD) to solve for the rotation matrix and translation vector that minimizes the distance between the matching pairs based on all matching pairs.

[0053] Specifically, in each iteration, for each lane line L i Perform the following steps:

[0054] Construct intensity-weighted corresponding points for each shape point: for L i Each shape point P on j Based on the search radius R of the current iteration current From the ground point cloud, determine all points within its neighborhood to form a set N of neighboring point clouds. j Calculate the weighted center point C of this set. j Instead of simply selecting the geometrically closest point, the weighted formula is as follows: C j =(∑I k *P k ) / ∑I k , where P k It is set N j The position of the k-th point cloud in I k This is its reflection intensity value. The intensity information from the point cloud is used to improve the quality of corresponding points. The intensity value of ordinary asphalt pavement is relatively low, while the intensity value of the painted sections of the lane markings is significantly higher. Therefore, the weighted center point C... j This will pull the point cloud towards a higher intensity point cloud, closer to the physical center of the actual lane lines. This is much more stable and accurate than simply relying on geometric distance to find the nearest point (which might just be the edge of the lane line or a noise point). Through this method, the shape point P... j A high-quality matching pair (P) was constructed. j C j ).

[0055] Calculate the transformation matrix: Collect the lane lines L for this lane. i The matching pairs of all points on the line form the matching pair set {(P)} of the lane line. j C j Based on this set of matching pairs, the optimal rotation matrix R is solved using the Singular Value Decomposition (SVD) algorithm. ik Translation vector T ik This minimizes the overall distance error between all matched pairs. (R) ik ,T ik ) Applied to lane lines L i Get its temporary new position. Iterate through L at the temporary new position. i For all shape points, calculate the average reflection intensity of the ground point cloud in their neighborhood, and then calculate the average value of all shape points as the matching score for this iteration. current The step is the same as before, but this time, due to the use of weighted centroids, the basis for calculating the matching degree is more reliable. The Score... currentScore_best (match with historical best) i Compare the results; if the result is better, update the stored (R_best) value. i ,T_best i and Score_best i .

[0056] Example 2

[0057] In a preferred embodiment, a system for registering heterogeneous lane lines to a point cloud includes: a data input module for inputting point cloud data and heterogeneous lane line data to be registered, wherein the heterogeneous lane line data includes multiple lane line points; a preprocessing module for preprocessing the point cloud data, identifying and extracting ground point cloud data through point cloud classification; an elevation initialization module for initializing the elevation of the multiple lane line points based on the average elevation of the ground point cloud, so that the multiple lane lines are initially aligned to the point cloud plane; a registration main process module for executing an iterative nearest point (ICP) registration main process, which integrates a variable-length search radius mechanism and a lane-specific optimization strategy. The variable-length search radius mechanism gradually decreases the search radius as the number of iterations increases, and the lane-specific optimization strategy independently calculates, records, and applies the optimal rotation matrix and translation vector for each lane line in each iteration; and an output module for updating the positions of the multiple lane line points based on the optimal rotation matrix and translation vector, and outputting the registered heterogeneous lane line data.

[0058] In a preferred embodiment, the variable-length search radius mechanism in the registration main process module includes: in the initial iteration stage, using a first preset radius as the search radius for coarse matching; in the intermediate iteration stage, dynamically reducing the search radius according to the ratio of the length of the translation vector in the current iteration to the length of the initial translation vector; and in the final iteration stage, using a second preset radius as the search radius for fine matching; wherein the second preset radius is smaller than the first preset radius.

[0059] In a preferred embodiment, the lane optimization strategy executed by the lane optimization module includes: calculating the matching degree between each lane line and the point cloud, the matching degree being determined based on the average intensity value of all adjacent ground point clouds of the lane line point; recording the best matching degree of each lane line and its corresponding rotation matrix and translation vector during the iteration process; and after registration is completed, each lane line independently adopts its recorded best rotation matrix and translation vector.

[0060] In a preferred embodiment, the system further includes a landmark registration module for registering landmark elements using an alternating direction iterative method, comprising: fixing a first direction, searching along a second direction and determining the optimal offset to maximize the sum of the ground point cloud intensity values ​​covered by the landmark; fixing the second direction, searching along the first direction and determining the optimal offset to maximize the sum of the ground point cloud intensity values ​​covered by the landmark; iteratively executing the above bidirectional search steps until the convergence condition is met.

[0061] In a preferred embodiment, the registration main process module executes the iterative nearest point (ICP) registration main process, which includes: for each lane alignment point, determining its neighboring point cloud set from the ground point cloud based on the search radius of the current iteration; calculating the corresponding weighted center point based on the intensity value of each point cloud in the neighboring point cloud set; constructing a matching pair between the lane alignment point and the corresponding weighted center point; and using singular value decomposition (SVD) to solve for the rotation matrix and translation vector that minimizes the distance between the matching pairs based on the matching pairs.

[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for registering heterogeneous lane lines to a point cloud, characterized in that, Includes the following steps: Input point cloud data and heterogeneous lane line data to be registered, wherein the heterogeneous lane line data contains multiple lane line points; The point cloud data is preprocessed, and all ground point clouds are identified and extracted through point cloud classification. Based on the average elevation of all the ground point clouds, the elevation of the multiple lane line points is initialized so that the multiple lane lines are initially aligned to the point cloud plane. The main process of execution is the iterative nearest point ICP registration. The main process integrates a variable-length search radius mechanism and a lane optimization strategy. The variable-length search radius mechanism gradually reduces the search radius as the number of iterations increases. The lane optimization strategy independently calculates, records and applies the optimal rotation matrix and translation vector for each lane line in each iteration. The positions of the multiple lane line points are updated based on the optimal rotation matrix and translation vector, and the registered heterogeneous lane line data is output.

2. The method for registering heterogeneous lane lines to a point cloud according to claim 1, characterized in that, The variable-length search radius mechanism includes: In the initial iteration phase, a first preset radius is used as the search radius for coarse matching; During the intermediate iteration phase, the search radius is dynamically reduced based on the ratio of the current translation vector length to the initial translation vector length. In the final iteration stage, a second preset radius is used as the search radius for refined matching; Wherein, the second preset radius is smaller than the first preset radius.

3. The method for registering heterogeneous lane lines to a point cloud according to claim 1, characterized in that, The lane separation optimization strategy includes: For each lane line, calculate its matching degree with the point cloud, the matching degree being determined based on the average intensity value of all adjacent ground point clouds of that lane line point; During the iteration process, the optimal matching degree of each lane line and its corresponding rotation matrix and translation vector are recorded; After registration is complete, each lane line updates the position of its points using its best recorded rotation matrix and translation vector.

4. The method for registering heterogeneous lane lines to a point cloud according to claim 1, characterized in that, It also includes the landmark registration step: The landmark elements are registered using an alternating direction iterative method, including: With the first direction fixed, search along the second direction and determine the optimal offset to maximize the sum of the ground point cloud intensity values ​​covered by the landmark; Fix the second direction, search along the first direction and determine the optimal offset to maximize the sum of the ground point cloud intensity values ​​covered by the landmark; The bidirectional search steps described above are executed iteratively until the convergence condition is met.

5. The method for registering heterogeneous lane lines to a point cloud according to claim 1, characterized in that, The main process for execution-type iterative nearest point ICP registration includes: For each lane alignment point, its neighboring point cloud set is determined from the ground point cloud based on the search radius of the current iteration; Based on the intensity value of each point cloud in the adjacent point cloud set, the corresponding weighted center point is calculated; Construct matching pairs between the lane alignment points and their corresponding weighted center points; Based on all matching pairs, use Singular Value Decomposition (SVD) to find the rotation matrix and translation vector that minimizes the distance between the matching pairs.

6. A system for registering heterogeneous lane lines to a point cloud, characterized in that, include: The data input module is used to input point cloud data and heterogeneous lane line data to be registered, wherein the heterogeneous lane line data includes multiple lane line points; The preprocessing module is used to preprocess the point cloud data, identify and extract ground point clouds through point cloud classification; The elevation initialization module is used to initialize the elevation of the multiple lane line points based on the average elevation of the ground point cloud, so that the multiple lane lines are initially aligned to the point cloud plane. The main registration process module is used to execute the iterative nearest point (ICP) registration main process. The main process integrates a variable-length search radius mechanism and a lane optimization strategy. The variable-length search radius mechanism gradually decreases the search radius as the number of iterations increases. The lane optimization strategy independently calculates, records, and applies the optimal rotation matrix and translation vector for each lane line in each iteration. The output module is used to update the positions of the multiple lane line points based on the optimal rotation matrix and translation vector, and output the registered heterogeneous lane line data.

7. The system for registering heterogeneous lane lines to point clouds according to claim 1, characterized in that, The variable-length search radius mechanism in the registration main process module includes: In the initial iteration phase, a first preset radius is used as the search radius for coarse matching; During the intermediate iteration phase, the search radius is dynamically reduced based on the ratio of the current translation vector length to the initial translation vector length. In the final iteration stage, a second preset radius is used as the search radius for refined matching; Wherein, the second preset radius is smaller than the first preset radius.

8. The system for registering heterogeneous lane lines to point clouds according to claim 1, characterized in that, The lane optimization strategy executed by the lane optimization module includes: For each lane line, calculate its matching degree with the point cloud, the matching degree being determined based on the average intensity value of all adjacent ground point clouds of that lane line point; During the iteration process, the optimal matching degree of each lane line and its corresponding rotation matrix and translation vector are recorded; After registration, each lane line independently uses its recorded optimal rotation matrix and translation vector.

9. The system for registering heterogeneous lane lines to point clouds according to claim 1, characterized in that, It also includes a landmark registration module: This method is used for registration of landmark features using an alternating direction iterative method, including: With the first direction fixed, search along the second direction and determine the optimal offset to maximize the sum of the ground point cloud intensity values ​​covered by the landmark; Fix the second direction, search along the first direction and determine the optimal offset to maximize the sum of the ground point cloud intensity values ​​covered by the landmark; The bidirectional search steps described above are executed iteratively until the convergence condition is met.

10. The system for registering heterogeneous lane lines to point clouds according to claim 1, characterized in that, The main registration process module executes an iterative, nearest-point (ICP) registration main process, including: For each lane alignment point, its neighboring point cloud set is determined from the ground point cloud based on the search radius of the current iteration; Based on the intensity value of each point cloud in the adjacent point cloud set, the corresponding weighted center point is calculated; Construct matching pairs between the lane alignment points and their corresponding weighted center points; Based on the matching pair, the rotation matrix and translation vector that minimize the distance between the matching pairs are solved using singular value decomposition (SVD).

Images