Vehicle-mounted multi-laser radar automatic calibration method and system based on ground fitting

Abstract

The invention discloses a vehicle-mounted multi-laser radar automatic calibration method and system based on ground fitting, and relates to unmanned driving. The method comprises the steps that original point cloud data of vehicle-mounted multi-laser radars are collected, one laser radar is selected from the multi-laser radars to serve as a reference laser radar, and the rest laser radars serve as to-be-calibrated laser radars; dividing the original point cloud data of each laser radar into a ground point set and a non-ground point set; performing plane fitting by adopting an RANSAC (Random Sample Consensus) algorithm, and selecting a plane equation with the maximum number of inner points as a ground model through iterative calculation; respectively constructing a ground loss function Lg and a non-ground loss function Lng; according to the ground loss function Lg and the non-ground loss function Lng, a total optimization function Ltotal is constructed through weighted fusion; solving the total optimization function Ltotal by adopting a numerical optimization algorithm to obtain an optimal transformation matrix T; and performing multi-laser radar calibration according to the optimal transformation matrix T. According to the method, drift of the yaw angle and the horizontal displacement in the optimization process is avoided, and the multi-laser calibration stability is improved.

Description

Technical Field

[0001] This application relates to the field of autonomous driving, and in particular to an automatic calibration method and system for vehicle-mounted multi-LiDAR based on ground fitting. Background Art

[0002] Vehicle-mounted multi-LiDAR systems are widely used in autonomous driving, high-precision map building, and other fields. However, due to installation errors, manufacturing differences, and other reasons, there are deviations between the coordinate systems of different LiDARs, resulting in inaccurate spatial alignment of point cloud data collected by different LiDARs. This seriously affects the accuracy and reliability of subsequent perception, localization, and decision-making tasks. Therefore, an effective method is needed to calibrate vehicle-mounted multi-LiDAR systems so that the point cloud data collected by each LiDAR can be accurately aligned in the same coordinate system, thereby improving the performance of the entire vehicle-mounted multi-LiDAR system.

[0003] Currently, the calibration methods for vehicle-mounted multi-LiDAR systems are mainly divided into two categories: manual calibration and automatic calibration.

[0004] Manual calibration typically requires operators to use specific tools and equipment, such as high-precision measuring instruments, to measure the relative position and attitude between lidar units. This method has the following disadvantages: low efficiency, as the calibration process requires operators to perform a large amount of measurement and adjustment work, resulting in long calibration times that cannot meet the needs of large-scale production and rapid deployment; limited accuracy, as human factors make it difficult to guarantee calibration accuracy, and calibration results from different operators may vary significantly; and high cost, as the need for high-precision measuring equipment increases the cost of calibration.

[0005] Existing automatic calibration methods are mainly based on feature matching and optimization algorithms. Among them, automatic calibration methods based on ground features are commonly used due to the prevalence of ground in most scenarios. However, these methods suffer from a serious constraint imbalance problem: as a planar feature, the ground can only provide strong constraints on parameters perpendicular to the ground (roll, pitch, and altitude), while having almost no constraint capability on parameters parallel to the ground (yaw and horizontal displacements x and y). This constraint imbalance causes the yaw and horizontal displacement parameters to drift during the optimization algorithm's solution process, resulting in multiple local optima and non-unique calibration results, which seriously affects the stability and reliability of the calibration.

[0006] In addition, automatic calibration methods based on feature matching also suffer from poor environmental adaptability. In environments lacking obvious features, such as open roads or indoor parking lots, feature extraction and matching are difficult, leading to calibration failure.

[0007] Therefore, there is an urgent need for an automatic calibration method for vehicle-mounted multi-LiDAR that can solve the problem of constraint imbalance and achieve balanced constraints of six degrees of freedom parameters. Summary of the Invention

[0008] To address the poor stability of vehicle-mounted multi-LiDAR calibration, this application provides an automatic calibration method and system for vehicle-mounted multi-LiDAR based on ground fitting. The joint constraints of ground points and non-ground points avoid drift of yaw angle and horizontal displacement during the optimization process, thereby improving the stability of multi-LiDAR calibration.

[0009] One aspect of this application provides an automatic calibration method for vehicle-mounted multiple lidars based on ground fitting, comprising: S1, collecting raw point cloud data of multiple vehicle-mounted lidars, and selecting one lidar as a reference lidar, with the rest as lidars to be calibrated; S2, dividing the raw point cloud data of each lidar into ground point sets and non-ground point sets; S3, performing plane fitting using the RANSAC algorithm based on the ground point set of the reference lidar, and selecting the plane equation with the most interior points as the ground model through iterative calculation; S4, constructing ground loss functions L based on the ground point sets of all lidars to be calibrated and the non-ground point sets of all lidars to be calibrated, combined with the ground model. g Non-ground loss function L ng S5, based on the ground loss function L g Non-ground loss function L ng By weighted fusion, the overall optimization function L is constructed. total S6, The overall optimization function L is solved using a numerical optimization algorithm. total S7. Obtain the optimal transformation matrix T; S8. Perform multi-laser radar calibration based on the optimal transformation matrix T.

[0010] Furthermore, S2, the raw point cloud data of each lidar is divided into ground point set and non-ground point set, including: fitting the plane equation ax+by+cz+d=0 using the RANSAC algorithm based on the raw point cloud data of each lidar; calculating the distance from each point in the raw point cloud data to the fitted plane equation, and classifying points with a distance less than a preset threshold as ground point set, and otherwise classifying them as non-ground point set;

[0011] Among them: the ground point set is used to constrain parameters perpendicular to the ground, including roll angle, pitch angle, and altitude z; the non-ground point set is used to constrain parameters parallel to the ground, including yaw angle and horizontal displacement x and y.

[0012] The RANSAC (Random Sample Consensus) algorithm is used in this application as an iterative algorithm to robustly fit a ground plane from lidar point cloud data containing a large amount of noise and outliers. This algorithm constructs a plane hypothesis by randomly selecting three points, then counts the number of interior points that conform to this plane. After multiple iterations, the plane with the most interior points is selected as the optimal ground model, effectively avoiding interference from non-ground points (such as vehicles, pedestrians, buildings, etc.) on the ground fitting.

[0013] Roll angle is the angle at which a lidar rotates around its forward direction (x-axis), representing the degree of left-right tilt of the lidar. When there is a roll angle deviation in the lidar installation, the ground point cloud it collects will appear tilted left and right. This parameter can be accurately calibrated by constraining the ground point set.

[0014] Pitch is the angle by which a lidar rotates around its lateral axis (y-axis), representing the degree of forward or backward tilt of the lidar. When there is a pitch deviation in the installation of the lidar, the ground point cloud it collects will appear tilted forward or backward. This parameter can be accurately calibrated by constraining the ground point set.

[0015] Yaw is the angle by which a lidar rotates around its vertical axis (z-axis), representing its horizontal rotational deviation. When a lidar is installed with a yaw deviation, the vertical structures it acquires (such as walls and columns) will be horizontally misaligned. This parameter can be accurately calibrated by using matching constraints on these vertical structures through a set of non-ground points.

[0016] Height z is the vertical displacement of the lidar relative to a reference lidar. When there is a deviation in the installation height of the lidar, the ground point cloud it collects will shift vertically as a whole. This parameter can be accurately calibrated by constraining the distance from the ground point set to the reference ground.

[0017] In this application, the ground point set refers to the set of all points whose distance to the ground plane is less than a preset threshold after fitting the ground plane using the RANSAC algorithm. These points are mainly distributed on the road surface. Due to the planar characteristics of the ground, the ground point set can provide strong constraints on the three parameters (roll, pitch, z) perpendicular to the ground, but has weaker constraints on the parameters parallel to the ground.

[0018] In this application, the non-ground point set refers to the set of all points whose distance to the fitted ground plane is greater than a preset threshold. These points include vertical or three-dimensional structures such as building walls, utility poles, traffic signs, vehicles, and pedestrians. Due to their rich distribution characteristics in three-dimensional space, the non-ground point set can provide effective constraints for parameters (yaw, x, y) parallel to the ground.

[0019] Furthermore, in step S3, based on the ground point set of the reference lidar, the RANSAC algorithm is used for plane fitting. Through iterative calculation, the plane equation with the most interior points is selected as the ground model. This includes: randomly selecting three points p1(x1,y1,z1), p2(x2,y2,z2), and p3(x3,y3,z3) from the ground point set of the reference lidar, and calculating the plane equation ax + by + cz + d = 0 formed by the three points; calculating the distance d' from all points in the ground point set to the plane equation. If the distance d' is less than a threshold d... t If the corresponding point is marked as an interior point, then the plane equation with the largest number of interior points is selected as the ground model through iterative calculation.

[0020] Furthermore, the distance d' is calculated using the following formula: Where a, b, c, and d represent the coefficients of the plane equation.

[0021] Furthermore, the ground loss function L g The expression is: Where, N g Let n be the number of ground points in each lidar to be calibrated, and T be the ground normal vector of the reference lidar. i p is the transformation matrix from the lidar to the reference lidar to the lidar to be calibrated. k These are the coordinates of the ground point in the current lidar to be calibrated;

[0022] In this application, the ground loss function L g This measure assesses the degree of fit between the transformed ground points of the lidar to be calibrated and the ground model of the reference lidar. The function is constructed by calculating the sum of the squared distances from each transformed ground point to the reference ground plane; the expression is the sum of the distances from the point to the plane. Due to the planar nature of the ground, L... g It has strong constraints on three parameters perpendicular to the ground (roll, pitch, and altitude z). When these parameters deviate, the distance from the ground point to the reference plane will increase significantly, thereby driving the optimization algorithm to adjust these parameters.

[0023] Furthermore, the non-ground loss function L ng The expression is: Where, N ng T represents the number of non-ground points in each lidar to be calibrated. i Let q be the transformation matrix from the lidar to the reference lidar. j Let q′ be the non-ground coordinates of the lidar to be calibrated. j For q j After transformation matrix T i The coordinates of the nearest point in the reference lidar after conversion.

[0024] In this application, the non-ground loss function L ng This measure assesses the degree of matching between the transformed non-ground points (buildings, vehicles, pillars, etc.) of the lidar to be calibrated and the corresponding structure of the reference lidar. The function is constructed by calculating the sum of squared Euclidean distances from each transformed non-ground point to its nearest neighbor in the reference point cloud. Due to the rich distribution characteristics of non-ground points in three-dimensional space, L... ng It has the ability to effectively constrain three parameters parallel to the ground (yaw angle, horizontal displacement x and y). When these parameters deviate, the vertical structure will be misaligned, resulting in an increase in the nearest neighbor distance.

[0025] Furthermore, S5, according to the ground loss function L g Non-ground loss function L ng By weighted fusion, a total optimization function Ltotal is constructed, including: obtaining the number N of ground points in all ground point sets in step S2. g And the number N of all non-ground points in the set of non-ground points ng Based on the number of ground points N g Number of non-ground points N ng Calculate the adaptive weighting coefficient α; based on the adaptive weighting coefficient α and the ground loss function L... g Non-ground loss function L ng By weighted fusion, the overall optimization function L is constructed. total :L total =αL g +(1-α)L ng ;

[0026] This application employs vertical parameter constraints provided by ground loss and horizontal parameter constraints provided by non-ground loss. An adaptive weighting coefficient α ensures the balance of these two types of constraints, guaranteeing that all six degrees of freedom parameters receive sufficient and appropriate constraints during the optimization process. The overall optimization function L is minimized. total The optimal transformation matrix can then be obtained, enabling precise calibration of the lidar. This separate processing and adaptive fusion design solves the problem of constraint imbalance.

[0027] Furthermore, the adaptive weighting coefficient α is calculated using the following formula: Among them, the adaptive weighting coefficient α is used to balance the parameters perpendicular to the ground and the parameters parallel to the ground to avoid the yaw angle and horizontal displacement x and y from drifting during the optimization process, and to ensure the stability of the calibration results.

[0028] Furthermore, in S6, a numerical optimization algorithm is used to solve the overall optimization function L. totalThe optimal transformation matrix T is obtained, including: initializing the transformation matrix T of the lidar to be calibrated, which contains six degrees of freedom parameters: roll, pitch, yaw, x, y, and z; in each iteration, based on the current transformation matrix T, transforming the ground point set and non-ground point set of the lidar to be calibrated to the reference coordinate system; and based on the ground loss function L... g Calculate the distance from the transformed ground points to the reference lidar ground model; based on the non-ground loss function L... ng Calculate the nearest neighbor distance from the transformed non-ground points to the reference lidar point cloud; based on the overall optimization function L... total =αL g +(1-α)L ng Calculate the loss value for the current iteration; use the Levenberg-Marquardt algorithm to optimize the overall function L. total Calculate the partial derivatives of the six parameters of the transformation matrix T, construct the Jacobian matrix, and update the six degrees of freedom parameters of the transformation matrix T based on the Jacobian matrix; iterate until the total optimization function L is obtained. total The change in the transformation matrix is ​​less than the preset convergence threshold, or the maximum number of iterations is reached; the converged transformation matrix T is output as the optimal transformation matrix.

[0029] Another aspect of this application provides an automatic calibration system for multiple vehicle-mounted lidars based on ground fitting, comprising: a data acquisition module for acquiring raw point cloud data of multiple vehicle-mounted lidars and selecting one lidar as a reference lidar, with the remainder as lidars to be calibrated; a point cloud division module for dividing the raw point cloud data of each lidar into ground point sets and non-ground point sets; the point cloud division module, based on the raw point cloud data of each lidar, fits the plane equation ax + by + cz + d = 0 using the RANSAC algorithm, calculates the distance from each point in the raw point cloud data to the fitted plane equation, and divides points with distances less than a preset threshold into ground point sets, and vice versa; wherein, the ground point set is used to constrain parameters perpendicular to the ground, including roll angle, pitch angle, and altitude z; the non-ground point set is used to constrain parameters parallel to the ground, including yaw angle and horizontal displacements x and y;

[0030] The ground fitting module performs plane fitting using the RANSAC algorithm based on the ground point set of the reference lidar. Through iterative calculation, it selects the plane equation with the most interior points as the ground model. The loss function construction module constructs ground loss functions L based on the ground point sets of all lidars to be calibrated, as well as the non-ground point sets of all lidars to be calibrated, and the ground model. g Non-ground loss function L ng;

[0031] The optimization function fusion module is used to optimize the ground loss function L. g Non-ground loss function L ng By weighted fusion, the overall optimization function L is constructed. total The optimization function fusion module obtains the number N of ground points in the set of all ground points. g And the number N of all non-ground points in the set of non-ground points ng According to the formula Calculate the adaptive weighting coefficient α, and according to formula L total =αL g +(1-α)L ng Construct the overall optimization function;

[0032] The optimization solution module is used to solve the overall optimization function L using numerical optimization algorithms. total The optimal transformation matrix T is obtained; the numerical optimization algorithm adopts the Levenberg-Marquardt algorithm; the calibration module is used to calibrate multiple lidars based on the optimal transformation matrix T.

[0033] Compared with the existing technology, the advantages of this application are:

[0034] Existing vehicle-mounted multi-LiDAR calibration methods that rely solely on ground constraints suffer from an imbalance in constraints. Specifically, the ground, as a planar feature, only provides strong constraints on parameters perpendicular to the ground (roll angle, pitch angle, and altitude z), while offering almost no constraints on parameters parallel to the ground (yaw angle and horizontal displacement x, y). This leads to problems such as yaw angle and horizontal displacement parameters drifting, multiple local optima, and non-unique calibration results during the optimization algorithm's solution process. This application provides an automatic calibration method for vehicle-mounted multi-LiDAR based on ground fitting. By separating the original point cloud data into ground point sets and non-ground point sets, a ground loss function L is constructed for each. g Non-ground loss function L ng The overall optimization function L is formed by weighted fusion using adaptive weight coefficient α. total The ground point set is used to constrain parameters perpendicular to the ground, while the non-ground point set (including vertical structures such as walls and columns) is used to constrain parameters parallel to the ground. Through this separate processing and adaptive weight fusion mechanism, it can achieve balanced constraints on the six degrees of freedom parameters, effectively avoid the drift of yaw angle and horizontal displacement during the optimization process, and ensure the uniqueness and stability of the calibration results. Finally, a high-precision transformation matrix between lidars is obtained through numerical optimization algorithm, completing the automatic calibration of multiple lidars. Attached Figure Description

[0035] This application will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein:

[0036] Figure 1 This is an exemplary flowchart illustrating an automatic calibration method for vehicle-mounted multi-LiDAR based on ground fitting, according to some embodiments of this application. Detailed Implementation

[0037] The methods and systems provided in the embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0038] like Figure 1 As shown, raw point cloud data of multiple vehicle-mounted lidars are collected, and one lidar is selected as the reference lidar, while the others are used as lidars to be calibrated. The raw point cloud data of each lidar is divided into ground point sets and non-ground point sets. Based on the ground point set of the reference lidar, the RANSAC algorithm is used for plane fitting. Through iterative calculation, the plane equation with the most interior points is selected as the ground model. Based on the ground point sets of all lidars to be calibrated and the non-ground point sets of all lidars to be calibrated, combined with the ground model, ground loss functions L are constructed respectively. g Non-ground loss function L ng According to the ground loss function L g Non-ground loss function L ng By weighted fusion, the overall optimization function L is constructed. total The overall optimization function L is solved using a numerical optimization algorithm. total The optimal transformation matrix T is obtained; multi-laser radar calibration is performed based on the optimal transformation matrix T.

[0039] The vehicle-mounted multi-LiDAR system is activated, allowing the vehicle to drive in the scene. Each LiDAR continuously collects point cloud data of the surrounding environment and stores the collected point cloud data synchronously according to timestamps to ensure that the data collected by different LiDARs is consistent in time.

[0040] Select any lidar on the vehicle as the reference lidar, and construct a reference coordinate system with the upward direction of the reference lidar as the x-axis, the right direction as the y-axis, and the forward direction as the z-axis.

[0041] Specifically, by selecting a single lidar as a benchmark, a common reference frame is provided for all lidars, enabling point cloud data from different lidars to be compared and fused within the same coordinate system. Furthermore, compared to simultaneously optimizing the relative relationships between all lidars, using a single lidar as a benchmark simplifies the many-to-many calibration problem into a many-to-one calibration problem, reducing optimization complexity and improving algorithm stability.

[0042] For each lidar point cloud data point, the RANSAC (Random Sample Consensus) algorithm is used to fit the ground, solving the ground equation of that point cloud ground in the reference coordinate system. RANSAC effectively handles noise and outliers in the point cloud data. Through random sampling and iterative verification, the ground plane can be accurately extracted even in the presence of a large number of non-ground points.

[0043] Specifically, select any three points p1(x1,y1,z1), p2(x2,y2,z2), and p3(x3,y3,z3) from the point cloud, and calculate the equation of the plane formed by these three points.

[0044] ax + by + cz + d = 0; where a = (y2 - y1)(z3 - z1) - (z2 - z1)(y3 - y1);

[0045] b=(z2-z1)(x3-x1)-(x2-x1)(z3-z1); c=(x2-x1)(y3-x1)-(y2-y1)(x3-x1);

[0046] d = -(ax1 + by1 + cz1);

[0047] Calculate the distance d' from all points in the point cloud to the plane. If the distance is less than a threshold d... t If , then it is denoted as an interior point.

[0048] The loop is executed N times, and the plane equation with the largest number of interior points is taken as the ground equation of that point cloud.

[0049] Specifically, this is achieved by calculating the distance from the point to the fitted plane and setting a threshold d. t It can accurately divide the original point cloud into ground point sets and non-ground point sets. The ground point set contains points on the ground, which are mainly distributed on a plane and can provide strong constraints for parameters perpendicular to the ground (roll, pitch, z); the non-ground point set contains vertical structures such as walls and columns, which have rich distribution characteristics in three-dimensional space and can provide effective constraints for parameters parallel to the ground (yaw, x, y).

[0050] The optimization function is composed of a weighted sum of the ground point loss and non-ground point loss of each LiDAR point cloud to the reference LiDAR point cloud.

[0051] The ground point loss is the distance from each ground point in the lidar point cloud to the ground coordinate system of the reference lidar. First, the lidar point cloud is transformed to the reference coordinate system using a transformation matrix T. For the ground in-point of the current lidar point cloud in step 2, the ground point loss is the distance from that point to the ground coordinate system of the reference lidar. Where, N g Let n be the number of ground points in each lidar to be calibrated, and T be the ground normal vector of the reference lidar. i p is the transformation matrix from the lidar to the reference lidar to the lidar to be calibrated. k These are the coordinates of the ground point in the lidar system currently being calibrated.

[0052] Specifically, the ground point loss function is constructed based on the geometric principle of the distance from a point to a plane. When the ground points of the lidar to be calibrated are transformed by the matrix T... i After being converted to the reference coordinate system, it should ideally coincide with the ground plane of the reference lidar.

[0053] The distance from a point to a plane reflects the accuracy of the parameters perpendicular to the ground in the transformation matrix: when there is an error in the roll angle, the transformed ground will tilt around the x-axis, causing the ground point to deviate from the reference plane; when there is an error in the pitch angle, the transformed ground will tilt around the y-axis, producing a similar deviation; when there is an error in the height z, the transformed ground will shift vertically as a whole; therefore, minimizing the ground point loss function Lg can effectively constrain the three parameters perpendicular to the ground (roll, pitch, z).

[0054] The non-ground point loss is the Euclidean distance from each non-ground point in each lidar point cloud to the nearest point in the reference lidar point cloud. Where, N ng T represents the number of non-ground points in each lidar to be calibrated. i Let q be the transformation matrix from the lidar to the reference lidar. j Let q′ be the non-ground coordinates of the lidar to be calibrated. j For q j After transformation matrix T i The coordinates of the nearest point in the reference lidar after conversion.

[0055] In particular, the non-ground point loss function is based on nearest neighbor matching. Non-ground points mainly include vertical structures such as walls, columns, and vehicles, which have rich geometric features in three-dimensional space.

[0056] From the perspective of constraint mechanisms: the projection differences of vertical structures (such as walls and columns) under different viewpoints are extremely sensitive to the yaw angle. When there is an error in yaw, the transformed vertical structure will be rotated and misaligned. The spatial distribution of these structures effectively constrains the horizontal displacement (x, y). When there is an error in x and y, the transformed point cloud and the reference point cloud will have a significant mismatch in the horizontal direction. By minimizing the nearest neighbor distance from non-ground points to the reference point cloud, the parameters (yaw, x, y) parallel to the ground can be effectively constrained, making up for the insufficiency of ground constraints.

[0057] Based on the number N of each lidar ground point g Number of non-ground points N ng Establish adaptive weight coefficient α, Among them, the number of ground points N g When there are many points, increasing α gives the ground loss function a larger weight, making full use of ground constraints; when the number of non-ground points N... ng When the value is large, (1-α) increases, giving greater weight to the non-ground loss function and strengthening the constraints on the parallel parameters; this adaptive mechanism ensures that all six degrees of freedom parameters receive sufficient and balanced constraints. This weight design avoids optimization bias caused by excessively strong constraints of one type, enabling the optimization process to converge evenly in the six-dimensional parameter space.

[0058] The overall optimization function is the weighted sum of the ground loss and non-ground loss of each lidar, L total =αL g +(1-α)L ng Among them, the ground loss function L g It provides strong constraints on (roll, pitch, z), and the non-ground loss function L ng Effective constraints are provided for (yaw, x, y), and the adaptive weight α ensures the balance between the two types of constraints. This construction method makes the overall optimization function have good convexity in the six-dimensional parameter space, avoiding ill-conditioned optimization problems caused by a single constraint. When the optimization algorithm solves for min(Ltotal), all six parameters can obtain sufficient gradient information, ensuring the stable convergence of the optimization process and the uniqueness of the solution.

[0059] The transformation matrix T of the lidar to be calibrated is initialized by setting initial values ​​for the six degrees of freedom parameters. The rotation parameters (roll, pitch, and yaw) are initialized to zero or based on small angle values ​​measured with coarse means, while the translation parameters (x, y, z) are set to initial estimated values ​​according to the physical installation location of the lidar.

[0060] In each iteration, the point cloud data of the lidar to be calibrated is processed as follows: the coordinates p of each point in the ground point set are... k Perform coordinate transformation using the current transformation matrix T: p k '=T i ×p k ; Set the coordinates q of each point in the non-ground point set j Perform coordinate transformation using the current transformation matrix T: q j '=T i ×q j The transformed point cloud data is unified into the reference lidar coordinate system.

[0061] Distance calculation is performed on the transformed ground point set: the normal vector n and plane parameters of the reference lidar ground model are extracted; the distance of each transformed ground point p is calculated. k Distance to the reference ground: d k =|n×(T) i ×p k The ground loss L is obtained by summing the squared distances of all ground points. g .

[0062] Perform nearest neighbor search and distance calculation on the transformed non-ground point set: construct a spatial index structure (such as a KD tree) for the baseline lidar point cloud; for each transformed non-ground point p j Search for the nearest neighbor q in the baseline point cloud. j ';Calculate the Euclidean distance: d j =||T i ×p j -q j '||;Sum the squared distances of all non-ground points to obtain the non-ground loss L. ng .

[0063] Count the current number of ground points N g Number of non-ground points N ng Calculate the adaptive weights: Execution weighted fusion: L total =αL g +(1-α)L ng Record the total loss value of the current iteration.

[0064] The Levenberg-Marquardt algorithm is used for parameter optimization: the partial derivatives of the total loss function with respect to the six parameters are calculated using numerical differentiation; a 6×N Jacobian matrix J is constructed, where N is the total number of points involved in the calculation; the parameter update is calculated as: ΔT = -(J T J+λI) -1 J T ×e, where e is the error vector and λ is the damping factor; update the six parameters of the transformation matrix: T'=T+ΔT.

[0065] Calculate the change in the total loss function between adjacent iterations: When ΔL total When the number of iterations is less than the preset convergence threshold ε, or when the number of iterations reaches the maximum value, the iteration is terminated; the converged transformation matrix T is output as the optimal transformation matrix.

[0066] Multiple lidar calibrations are performed based on the optimal transformation matrix T. For each lidar to be calibrated, its corresponding optimal transformation matrix T is stored. iA mapping relationship is established from the local coordinate system of each lidar to the reference coordinate system. For the real-time acquired point cloud data of each lidar, the corresponding transformation matrix T is used... i The data is transformed to a reference coordinate system, and all transformed point cloud data are fused to form a unified global point cloud. Redundant points in overlapping areas are removed to ensure the consistency of the fused point cloud. The average distance from each LiDAR ground point to the reference ground is calculated after calibration to verify the vertical parameter calibration accuracy. The registration error of non-ground points after calibration is calculated to verify the horizontal parameter calibration accuracy. A calibration quality assessment report is output to ensure that the calibration results meet the system requirements.

[0067] The foregoing illustrative description of the present application and its embodiments is not restrictive and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. The accompanying drawings are only one embodiment of the present application, and the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present application, such designs should fall within the scope of protection of this application. Furthermore, the word "comprising" does not exclude other elements or steps, and the word "a" preceding an element does not exclude the inclusion of "a plurality" of that element. Terms such as "first," "second," etc., are used to indicate names and do not indicate any specific order.

Claims

1. An automatic calibration method for vehicle-mounted multi-LiDAR based on ground fitting, characterized in that, include: S1: Collect raw point cloud data of multiple vehicle-mounted lidars, and select one lidar from the multiple lidars as the reference lidar, and the rest as lidars to be calibrated. S2 divides the raw point cloud data of each lidar into ground point sets and non-ground point sets; S3. Based on the ground point set of the reference lidar, the RANSAC algorithm is used for plane fitting. Through iterative calculation, the plane equation with the most interior points is selected as the ground model. S4. Based on the ground point set of all lidars to be calibrated, and the non-ground point set of all lidars to be calibrated, and in conjunction with the ground model, construct the ground loss function L respectively. g Non-ground loss function L ng ; S5, based on the ground loss function L g Non-ground loss function L ng By weighted fusion, the overall optimization function L is constructed. total ; S6, Solve the overall optimization function L using a numerical optimization algorithm. total The optimal transformation matrix T is obtained; S7. Perform multi-laser radar calibration based on the optimal transformation matrix T.

2. The automatic calibration method for vehicle-mounted multi-lidar radar based on ground fitting according to claim 1, characterized in that: S2 divides the raw point cloud data of each lidar into ground point sets and non-ground point sets, including: Based on the raw point cloud data of each lidar, the plane equation ax+by+cz+d=0 is fitted using the RANSAC algorithm; Calculate the distance from each point in the original point cloud data to the fitted plane equation, and classify points whose distance is less than a preset threshold as ground point set, and otherwise classify them as non-ground point set; in: The ground point set is used to constrain parameters perpendicular to the ground, including roll, pitch, and altitude (z). The non-ground point set is used to constrain parameters parallel to the ground. These parameters include the yaw angle and the horizontal displacements x and y.

3. The automatic calibration method for vehicle-mounted multi-lidar radar based on ground fitting according to claim 2, characterized in that: S3, based on the ground point set of the reference lidar, uses the RANSAC algorithm for plane fitting. Through iterative calculation, the plane equation with the most interior points is selected as the ground model, including: From the ground point set of the reference lidar, three points p1(x1,y1,z1) and p2(x2,y2,z2) are randomly selected. p3(x3,y3,z3), calculate the equation of the plane formed by the three points: ax + by + cz + d = 0; Calculate the distance d' from all points in the ground point set to the plane equation. If the distance d' is less than the threshold d... t If so, then the corresponding point will be marked as an interior point; Iterative calculations are performed, and the plane equation with the largest number of interior points is selected as the ground model.

4. The automatic calibration method for vehicle-mounted multi-LiDAR based on ground fitting according to claim 3, characterized in that: Distance d' is calculated using the following formula: Where a, b, c, and d represent the coefficients of the plane equation.

5. The automatic calibration method for vehicle-mounted multi-lidar radar based on ground fitting according to claim 2, characterized in that: Ground loss function L g The expression is: Where, N g Let n be the number of ground points in each lidar to be calibrated, and T be the ground normal vector of the reference lidar. i p is the transformation matrix from the lidar to the reference lidar to the lidar to be calibrated. k These are the coordinates of the ground point in the lidar system currently being calibrated.

6. The automatic calibration method for vehicle-mounted multi-LiDAR based on ground fitting according to claim 5, characterized in that: Non-ground loss function L ng The expression is: Where, N ng T represents the number of non-ground points in each lidar to be calibrated. i Let q be the transformation matrix from the lidar to the reference lidar. j Let q' be the non-ground point coordinates of the lidar to be calibrated. j For q j After transformation matrix T i The coordinates of the nearest point in the reference lidar after conversion.

7. The automatic calibration method for vehicle-mounted multi-LiDAR based on ground fitting according to claim 6, characterized in that: S5, based on the ground loss function L g Non-ground loss function L ng By weighted fusion, the total optimization function Ltotal is constructed, including: Obtain the number N of ground points in the set of all ground points in step S2. g And the number N of all non-ground points in the set of non-ground points ng ; Based on the number of ground points N g Number of non-ground points N ng Calculate the adaptive weighting coefficient α; Based on the adaptive weighting coefficient α and the ground loss function L g Non-ground loss function L ng By weighted fusion, the overall optimization function L is constructed. total : L total =αL g +(1-α)L ng 。 8. The automatic calibration method for vehicle-mounted multi-LiDAR based on ground fitting according to claim 7, characterized in that: The adaptive weighting coefficient α is calculated using the following formula: The adaptive weighting coefficient α is used to balance the parameters perpendicular to the ground and the parameters parallel to the ground to avoid drift of the yaw angle and horizontal displacement x and y during the optimization process, and to ensure the stability of the calibration results.

9. The automatic calibration method for vehicle-mounted multi-lidar radar based on ground fitting according to claim 2, characterized in that: S6, Solve the overall optimization function L using a numerical optimization algorithm. total The optimal transformation matrix T is obtained, including: The Levenberg-Marquardt algorithm was used for numerical optimization.

10. A vehicle-mounted multi-LiDAR automatic calibration system based on ground fitting, characterized in that, include: The data acquisition module collects raw point cloud data from multiple vehicle-mounted lidars and selects one lidar as the reference lidar, while the others are lidars to be calibrated. The point cloud segmentation module divides the raw point cloud data of each lidar into ground point sets and non-ground point sets; The ground fitting module uses the RANSAC algorithm to perform plane fitting based on the ground point set of the reference lidar. Through iterative calculation, it selects the plane equation with the most interior points as the ground model. The loss function module constructs ground loss functions L based on the ground point set of all lidars to be calibrated and the non-ground point set of all lidars to be calibrated, combined with the ground model. g Non-ground loss function L ng ; The optimization solution module, based on the ground loss function L g Non-ground loss function L ng Calculate the weighted fusion coefficients and the overall optimization function L. total The overall optimization function L is solved using a numerical optimization algorithm. total The optimal transformation matrix T is obtained; The calibration module performs calibration of multiple lidar systems based on the optimal transformation matrix T.

Images