Lidar external parameter calibration method and system based on mixed multi-type reference objects

Abstract

This application relates to the field of laser detection technology and discloses a method and system for extrinsic parameter calibration of lidar based on a mixture of multiple types of reference objects. The method includes: reading initial extrinsic parameters from a TF transform tree; separating a single-frame laser point cloud into a calibration board point set and a cylindrical point set; fitting the calibration board normal vector, the point set inside the calibration board, the coordinates of the cylinder center, and the point set inside the cylinder, respectively; applying the initial extrinsic parameters to the two types of point sets to construct residual relative deviations, and obtaining corrected initial values ​​after directional pre-correction; constructing a joint residual objective function using the corrected initial values ​​and obtaining the calibration extrinsic parameters through nonlinear optimization. This application improves the convergence reliability and calibration accuracy of single-frame lidar extrinsic parameter calibration under conditions where the initial extrinsic parameters have large deviations.

Description

Technical Field

[0001] This application relates to the field of laser detection technology, and in particular to a method and system for calibrating the extrinsic parameters of a lidar based on a mixture of multiple types of reference objects. Background Technology

[0002] LiDAR extrinsic parameter calibration is a fundamental step for AGVs to achieve accurate navigation and positioning. The goal of extrinsic parameter calibration is to determine the translation and yaw angle of the LiDAR coordinate system relative to the AGV base coordinate system. Existing target-based calibration methods typically place a single type of geometric reference object within the LiDAR's field of view. The extrinsic parameters are solved by extracting the geometric features of the reference object and constructing a nonlinear optimization problem. Common reference object types include planar calibration plates and cylinders. Calibration plates provide linear feature constraints, while cylinders provide circular feature constraints. The two types of reference objects have different directional constraint capabilities on the translation and yaw angle components, respectively.

[0003] However, when only a single type of reference object is used, its geometric constraint direction is singular, and its constraint ability in a certain parameter direction is weak, resulting in low calibration accuracy in that direction. To compensate for the insufficient constraint of a single type of reference object, some methods introduce the mixed use of multiple types of reference objects. However, existing mixed calibration methods of multiple types of reference objects directly superimpose the residuals of different types of reference objects into a unified objective function, without distinguishing and utilizing the differences in constraint ability of different types of reference objects in each parameter direction. As a result, the complementary constraint advantage of the mixed use of multiple types of reference objects is not fully utilized.

[0004] Existing multi-reference hybrid calibration methods all directly use the mechanical installation nominal values ​​stored in the TF transform tree as the initial extrinsic parameters for nonlinear optimization. However, the response mechanisms of the normal residual of the calibration plate and the radial residual of the cylinder to the initial extrinsic parameter error are fundamentally different. The normal residual of the calibration plate mainly reflects the error projection of the translation component in the normal direction, while the radial residual of the cylinder is simultaneously affected by the combined effects of translation error and yaw angle error. The relative deviation between the two types of residuals naturally carries the distribution information of the initial extrinsic parameter error in different parameter directions. Existing technologies do not make full use of this information and directly use the initial extrinsic parameters with systematic deviations as the starting point for optimization. This causes the nonlinear optimization to start from a position far from the true extrinsic parameters, which carries the risk of convergence to a local optimum. When the initial extrinsic parameter deviation is large, the reliability of the calibration results is difficult to guarantee. Summary of the Invention

[0005] This application provides a method and system for extrinsic parameter calibration of lidar based on the mixing of multiple types of reference objects. It solves the problem that existing multi-type reference object mixing calibration methods fail to use the relative deviation of the residuals of the two types of reference objects to identify the dominant direction of the initial extrinsic parameter error, resulting in low optimization starting point quality. It improves the convergence reliability and calibration accuracy of single-frame lidar extrinsic parameter calibration under the condition that there is a large deviation in the initial extrinsic parameters.

[0006] Firstly, this application provides a method for calibrating the extrinsic parameters of a lidar based on a mixture of multiple types of reference objects, the method comprising:

[0007] Step S1: Read the initial extrinsic parameters from the TF transform tree, and separate the single-frame laser point cloud of the lidar on the AGV into a calibration plate point set and a cylindrical point set according to the nominal position of the reference object;

[0008] Step S2: Obtain the normal vector of the calibration plate and the set of points inside the calibration plate by fitting the point set of the calibration plate; obtain the coordinates of the cylinder center and the set of points inside the cylinder by fitting the point set of the cylinder.

[0009] Step S3: Apply the initial extrinsic parameters to the point set inside the calibration plate and the point set inside the cylinder respectively to obtain the mean value of the normal residual and the mean value of the radial residual. Construct the relative deviation of the residual using the difference between the two. Use the weighted mean direction of the calibration plate normal vector as the translation correction direction and the relative deviation of the residual as the amplitude to perform orientation pre-correction on the translation component in the initial extrinsic parameters. Perform weighted least squares decomposition on the yaw angle contribution in the relative deviation of the residual based on the coordinates of the cylinder center to obtain the yaw angle correction amount. Superimpose the translation correction direction and the yaw angle correction amount onto the initial extrinsic parameters to obtain the corrected initial value.

[0010] Step S4: Starting from the modified initial value, construct a joint residual objective function containing the point set inside the calibration plate and the point set inside the cylinder, and obtain the calibration extrinsic parameters through nonlinear optimization.

[0011] Secondly, this application provides a lidar extrinsic parameter calibration system based on a mixture of multiple types of reference objects, the lidar extrinsic parameter calibration system based on a mixture of multiple types of reference objects includes:

[0012] The separation module is used to read the initial extrinsic parameters from the TF transform tree and separate the single-frame laser point cloud of the lidar on the AGV into a calibration plate point set and a cylindrical point set according to the nominal position of the reference object.

[0013] The fitting module is used to fit the calibration plate point set to obtain the calibration plate normal vector and the calibration plate in-point set, and to fit the cylinder point set to obtain the cylinder center coordinates and the cylinder in-point set.

[0014] The weighting module is used to apply the initial extrinsic parameters to the point set inside the calibration plate and the point set inside the cylinder, respectively, to obtain the mean normal residual and the mean radial residual, and to construct the relative residual deviation by the difference between the two. The translation component in the initial extrinsic parameters is pre-corrected using the weighted mean direction of the calibration plate normal vector as the translation correction direction and the relative residual deviation as the amplitude. The yaw angle contribution in the relative residual deviation is then decomposed using weighted least squares based on the cylinder center coordinates to obtain the yaw angle correction amount. The translation correction direction and the yaw angle correction amount are then superimposed on the initial extrinsic parameters to obtain the corrected initial value.

[0015] The optimization module is used to construct a joint residual objective function containing the point set inside the calibration plate and the point set inside the cylinder, starting from the modified initial value, and obtain the calibration extrinsic parameters through nonlinear optimization.

[0016] Thirdly, a lidar extrinsic parameter calibration device based on a mixture of multiple types of reference objects is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the lidar extrinsic parameter calibration device based on a mixture of multiple types of reference objects to execute the aforementioned lidar extrinsic parameter calibration method based on a mixture of multiple types of reference objects.

[0017] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the aforementioned lidar extrinsic parameter calibration method based on a mixture of multiple types of reference objects.

[0018] In the technical solution provided in this application, the initial extrinsic parameters are read from the TF transform tree, and the single-frame laser point cloud is separated into a calibration board point set and a cylinder point set according to the nominal position of the reference object. This allows the calibration process to automatically classify and extract the reference object point cloud under single-frame data conditions, without manual intervention or accumulation of multiple frames of data. On this basis, RANSAC linear fitting is performed on the calibration board point set to obtain the calibration board normal vector and the point set inside the calibration board. RANSAC circle fitting is performed on the cylinder point set under the nominal radius constraint of the cylinder to obtain the cylinder center coordinates and the point set inside the cylinder. The random sampling consistency mechanism of the RANSAC algorithm enables the fitting process to naturally suppress noise points and outliers mixed in with the point cloud. The introduction of the nominal radius constraint of the cylinder reduces the degree of freedom of circle fitting from three to two. Stable center coordinates can still be obtained when the visible arc segment of the cylinder is short. The fitting results of the two types of reference objects carry their respective in-point sets and observation quality information, providing reliable geometric feature inputs for subsequent residual calculation.

[0019] The most significant technical contribution of this application lies in the construction and directional pre-correction mechanism of the residual relative deviation: Initial extrinsic parameters are applied to the point sets within the calibration plate and the cylinder, respectively. Utilizing the inherent difference that the calibration plate normal residual primarily reflects translation error while the cylinder radial residual is simultaneously affected by both translation and yaw angle errors, the difference between the two is used to construct the residual relative deviation. This allows the distribution information of the initial extrinsic parameter error in different parameter directions to be extracted from existing fitting results and utilized. Furthermore, the weighted mean direction of the calibration plate normal vector is used as the translation correction direction to adjust the relative deviation. The shift component is pre-corrected orientedly, and the contribution of the yaw angle in the relative deviation of the residual is decomposed into a weighted least squares value based on the geometric sensitivity coefficient from the cylinder center coordinates to the base origin to obtain the yaw angle correction. The two types of corrections are superimposed on the initial extrinsic parameters to obtain the corrected initial value. Then, a joint residual objective function containing the inner point set of the calibration plate and the inner point set of the cylinder is constructed starting from the corrected initial value. The calibration extrinsic parameters are obtained through nonlinear optimization, so that the nonlinear optimization starts from a position closer to the true extrinsic parameters. Even when there is a large systematic deviation in the initial extrinsic parameters, it can still reliably converge to the correct result. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of an embodiment of the lidar extrinsic parameter calibration method based on the mixing of multiple types of reference objects in this application.

[0022] Figure 2 This is a schematic diagram comparing the calibration errors before and after orientation pre-correction in an embodiment of this application. Detailed Implementation

[0023] This application provides a method and system for extrinsic parameter calibration of lidar based on a mixture of multiple types of reference objects. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data used can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the lidar extrinsic parameter calibration method based on a mixture of multiple types of reference objects in this application includes:

[0025] Step S1: Read the initial extrinsic parameters from the TF transform tree, and separate the single-frame laser point cloud of the lidar on the AGV into a calibration plate point set and a cylindrical point set according to the nominal position of the reference object;

[0026] It is understood that the executing entity of this application can be a lidar extrinsic parameter calibration system based on a mixture of multiple types of reference objects, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiments use a server as an example for illustration.

[0027] Specifically, the initial extrinsic parameters refer to the set of translation and yaw components of the LiDAR coordinate system relative to the AGV base coordinate system. These are derived from the transformation record from the LiDAR coordinate system to the base coordinate system in the robot operating system's TF transformation tree, typically written from the nominal values ​​on the mechanical installation drawings. The translation component accuracy generally has a deviation of ±30mm, and the yaw component has a deviation of ±2°. The nominal reference position refers to the designed installation coordinates of each calibration plate and cylinder in the base coordinate system. The region of interest for the calibration plate is defined with this nominal position as the center, with a half-side length of 0.3m, and the region of interest for the cylinder is defined with a radius of 0.2m. The side length and radius are set to cover the entire drift range of the reference point cloud under the maximum deviation of the initial extrinsic parameters.

[0028] Step S2: Obtain the normal vector of the calibration plate and the set of points inside the calibration plate by fitting the point set of the calibration plate; obtain the coordinates of the cylinder center and the set of points inside the cylinder by fitting the point set of the cylinder.

[0029] Specifically, the calibration plate normal vector refers to the unit normal vector of the fitted reference line, whose direction is perpendicular to the calibration plate reflective surface. Physically, it represents the shortest distance direction from the laser point to the calibration plate surface. The cylinder center coordinates refer to the two-dimensional coordinates of the center obtained under the constraint of a fixed nominal cylinder radius, where the nominal cylinder radius is the known design radius of the cylinder cross-section. The normal distance threshold and the radial distance threshold are both set to 0.02m, which are based on twice the typical ranging noise of 2D LiDAR (±10mm). The RANSAC iteration count is set to 100, calculated based on the theoretical number of iterations required to obtain a pure subset of interior points with a 99% probability when the interior point rate is not less than 50%.

[0030] Step S3: Apply the initial extrinsic parameters to the point set inside the calibration plate and the point set inside the cylinder respectively to obtain the mean value of the normal residual and the mean value of the radial residual. Construct the relative deviation of the residual based on the difference between the two. Use the weighted mean value of the calibration plate normal vector as the translation correction direction and the relative deviation of the residual as the amplitude to perform orientation pre-correction on the translation component in the initial extrinsic parameters. Based on the coordinates of the cylinder center, perform weighted least squares decomposition on the yaw angle contribution in the relative deviation of the residual to obtain the yaw angle correction amount. Superimpose the translation correction direction and the yaw angle correction amount onto the initial extrinsic parameters to obtain the corrected initial value.

[0031] Specifically, the relative deviation of the residuals is the difference between the mean radial residuals and the mean normal residuals. Its physical meaning is as follows: the mean normal residuals are mainly driven by the projection of the initial translation error onto the normal vector direction of the calibration plate; the mean radial residuals are simultaneously affected by the combined effects of translation error and yaw angle error. The difference between the two is the additional contribution of the yaw angle component in the radial direction of the cylinder from the initial translation error. The translation correction direction is the weighted average direction of the normal vectors of each calibration plate, weighted by the inlier rate. The relative deviation of the residuals is decomposed around this direction as an axis. To the horizontal and vertical coordinate axes, correct the horizontal and vertical translation components of the initial extrinsic parameters respectively; the weighted least squares decomposition of the yaw angle contribution refers to: the geometric sensitivity coefficient of each cylinder is calculated from the Euclidean distance from the cylinder center coordinate to the origin of the base coordinate system, and the internal point rate of each cylinder is used as the weight. The residual relative deviation of each cylinder and its geometric sensitivity coefficient are used to form an overdetermined linear equation system. The weighted least squares solution is used to obtain the yaw angle correction, which is then superimposed on the yaw angle component after translation pre-correction to obtain the corrected initial value.

[0032] Step S4: Starting from the corrected initial value, construct a joint residual objective function containing the point set inside the calibration plate and the point set inside the cylinder, and obtain the calibration extrinsic parameters through nonlinear optimization.

[0033] Specifically, the joint residual objective function is composed of a weighted sum of calibration plate residuals and cylindrical residuals, with the weight coefficient of each residual term being the in-point rate of the corresponding reference object; the Huber threshold is set to 0.01m, based on the typical residual amplitude of residual outliers after RANSAC refinement. This threshold retains terms with absolute residual values ​​less than 0.01m in a quadratic penalty form, and terms with absolute residual values ​​not less than 0.01m are converted to a linear penalty form; the convergence threshold is set to 1× The iteration terminates when the norm of the change in the extrinsic parameters between two adjacent iterations is lower than this value. The self-checking mechanism performs nonlinear optimization once each, starting from the corrected initial value and the original initial extrinsic parameters. It compares the root mean square of the termination residuals of the two optimizations and takes the optimization result corresponding to the smaller root mean square of the termination residuals as the calibration extrinsic parameter output.

[0034] In one specific embodiment, step S1 includes:

[0035] The initial extrinsic parameters are obtained by reading the translation and yaw components of the lidar relative to the base coordinate system from the TF transform tree.

[0036] Based on the initial extrinsic parameters, the nominal positions of each calibration plate and each cylinder in the base coordinate system are inversely transformed to the lidar coordinate system to obtain the lidar coordinate system nominal positions of each reference object.

[0037] The region of interest is defined centered on the nominal position of the lidar system of each reference object. Points falling into the region of interest of each calibration plate in a single frame of lidar point cloud are extracted as calibration plate point sets, and points falling into the region of interest of each cylinder are extracted as cylinder point sets.

[0038] Specifically, the initial extrinsic parameters consist of three components: the translation of the lidar installation position relative to the center of the AGV base in the longitudinal direction and the translation in the lateral direction, and the yaw angle of the lidar installation orientation relative to the front of the base coordinate system. All three are read from the TF transformation tree of the robot operating system. The TF transformation tree is a tree-shaped database in the robot operating system specifically used to store the position and orientation transformation relationships between various coordinate systems. The transformation node from the lidar coordinate system to the base coordinate system stores the nominal installation parameters written at the factory according to the mechanical installation drawings. The execution logic of the nominal position inverse transformation is as follows: subtract the longitudinal and lateral translations from the initial extrinsic parameters from the known design installation coordinates of each calibration plate and each cylinder in the base coordinate system to obtain the relative coordinates with the lidar installation position as the origin. Then, rotate the relative coordinates by the corresponding angle in the opposite direction of the initial extrinsic parameter yaw angle to obtain the nominal position of each reference object in the lidar coordinate system. This position is the theoretical center coordinate of each reference object from the lidar viewpoint.

[0039] Regions of interest are defined centered on the nominal position of the lidar system for each reference object: For the calibration board, a square region of interest with a side length of 0.6m is defined with the nominal position of the lidar system as the center and extending forward, backward, left, and right by 0.3m. Points in the single-frame lidar point cloud whose horizontal and vertical coordinates fall within the boundary of this square are extracted as the calibration board point set; For the cylinder, a circular region of interest is defined with the nominal position of the lidar system as the center and a radius of 0.2m. Points whose straight-line distance to the nominal position of the lidar system does not exceed 0.2m are extracted as the cylinder point set. The size of the region of interest (ROI) is set as follows: the maximum deviation of the translation component in the initial extrinsic parameters is ±30 mm, and the maximum deviation of the yaw angle is ±2 degrees. When the reference object is 3 meters away from the lidar, the tangential position drift caused by the yaw angle deviation at the reference object is the sine value corresponding to 3 meters multiplied by 2 degrees, which is approximately 0.105 meters. After superimposing this drift amount with the translation deviation, the maximum comprehensive drift amount is approximately 0.135 meters. This value is doubled and rounded up to 0.3 meters as half the length of the ROI, ensuring that the reference object point cloud still falls completely within the ROI even when there is a maximum nominal deviation in the initial extrinsic parameters.

[0040] In one specific embodiment, step S2 includes:

[0041] RANSAC line fitting is performed on the calibration board point set. Points with normal distance less than the normal distance threshold are marked as candidate interior points of the calibration board. Least square refinement is performed based on the candidate interior points of the calibration board to obtain the calibration board normal vector and the set of interior points of the calibration board.

[0042] RANSAC circle fitting is performed on the point set of the cylinder under the constraint of the nominal radius of the cylinder. Points with radial distance deviation less than the radial distance threshold are marked as candidate interior points of the cylinder. Least squares refinement is performed based on the candidate interior points of the cylinder to obtain the coordinates of the cylinder center and the set of interior points of the cylinder.

[0043] The calibration plate in-point rate is obtained by the ratio of the number of points in the calibration plate to the number of points in the calibration plate point set. The cylinder in-point rate is obtained by the ratio of the number of points in the cylinder point set to the number of points in the cylinder point set. The calibration plate in-point rate and the cylinder in-point rate are used to weight the calibration plate in-point set and the cylinder in-point set, respectively, to obtain the weighted calibration plate in-point set and the weighted cylinder in-point set.

[0044] Specifically, the execution logic of RANSAC line fitting is as follows: Two points are randomly selected from the calibration plate point set each time to form a hypothetical straight line. The perpendicular distance from all remaining points in the calibration plate point set to this straight line is calculated, i.e., the normal distance. Points with a normal distance less than the normal distance threshold are marked as candidate interior points of the calibration plate. The normal distance threshold is set to 0.02 meters, which is twice the typical ranging noise of 2D LiDAR (±10 mm) to cover legal interior points within the normal measurement error range. The above random sampling and candidate interior point statistics process is repeated 100 times. All candidate interior points corresponding to the hypothetical straight line with the largest number of candidate interior points are taken as the optimal interior point set for this iteration. Then, least squares refinement is performed on all points in this optimal interior point set, i.e., the straight line parameters are refitted with the goal of minimizing the sum of the squares of the normal distances from all candidate interior points to the straight line, resulting in the refined calibration plate normal vector and the calibration plate interior point set. The calibration plate normal vector is a unit direction vector perpendicular to the calibration plate reflective surface. The execution logic of RANSAC circle fitting is as follows: Three points are randomly selected from the set of cylindrical points each time. Under the constraint of a fixed nominal cylinder radius, a unique center position is determined. The nominal cylinder radius refers to the known design radius of each cylindrical section. This value is used as a fixed parameter during the fitting process and does not participate in the optimization. The radial distance deviation from all other points in the cylindrical point set to the assumed edge of the circle is calculated, which is the absolute value of the difference between the straight-line distance from each point to the center and the nominal cylinder radius. Points with radial distance deviations less than the radial distance threshold are marked as candidate inner points of the cylinder. The radial distance threshold is also set to 0.02 meters, for the same reason as the normal distance threshold. After repeating the above process 100 times, the center position with the most candidate inner points is selected. Least squares refinement is performed on all candidate inner points of the cylinder under the constraint of a fixed nominal cylinder radius. The center coordinates are re-solved with the objective of minimizing the sum of the squares of the differences between the distances from all candidate inner points to the center and the nominal cylinder radius, resulting in the refined cylinder center coordinates and the set of inner points.

[0045] The calibration plate in-point ratio is the ratio of the number of points in the calibration plate in-point set to the total number of points in the calibration plate in-point set. This ratio reflects the proportion of points in the calibration plate in-point set that conform to the linear model; a higher value indicates better quality of the calibration plate in-point set. The cylinder in-point ratio is the same as the calibration plate in-point ratio, calculated by dividing the number of points in the cylinder in-point set by the total number of points in the cylinder in-point set. Weighting refers to using the calibration plate in-point ratio as the quality weight coefficient for each calibration plate in-point set and the cylinder in-point ratio as the quality weight coefficient for each cylinder in-point set. These weight coefficients are bound and stored with their corresponding in-point sets, allowing each reference object's in-point set to carry weight information reflecting its observation quality. This results in weighted calibration plate in-point sets and weighted cylinder in-point sets. These weight coefficients are used as the basis for adjusting the contribution ratio of each reference object in the subsequent weighted calculation of the mean normal residual and the mean radial residual, and in the construction of the joint residual objective function.

[0046] Figure 2This is a schematic diagram comparing the calibration errors before and after orientation pre-correction in an embodiment of this application. Figure 2 In Figure (a), the horizontal axis represents the initial error of the translation component in the initial extrinsic parameters, and the vertical axis represents the residual error of the translation component in the calibration extrinsic parameters output by the calibration process. In Figure (b), the horizontal axis represents the initial error of the yaw angle component in the initial extrinsic parameters, and the vertical axis represents the residual error of the yaw angle component in the calibration extrinsic parameters. The dashed broken lines in both figures represent the calibration residual error after nonlinear iterative optimization with the initial extrinsic parameters as the optimization starting point. The solid broken lines represent the calibration residual error after nonlinear iterative optimization with the initial value obtained after orientation pre-correction and then optimization starting from the initial value after correction. The numerical difference between the two broken lines reflects the improvement of the calibration extrinsic parameter accuracy by the orientation pre-correction step.

[0047] In one specific embodiment, step S3 applies the initial extrinsic parameters to the point set inside the calibration plate and the point set inside the cylinder, respectively, to obtain the mean normal residual and the mean radial residual, including:

[0048] The translation and yaw angle components in the initial extrinsic parameters are used to construct a rotation and translation transformation matrix. Based on the rotation and translation transformation matrix, the coordinates of each point in the calibration plate point set are transformed to obtain the base coordinate system coordinates of each point in the calibration plate.

[0049] Based on the reference line normal vector corresponding to each calibration plate, calculate the signed normal distance from the base coordinate system coordinates of each point in the calibration plate to the reference line. The average value of the signed normal distance of all points in the calibration plate is obtained by weighting the points in the calibration plate according to the rate of the points in the calibration plate.

[0050] Based on the rotation and translation transformation matrix, the coordinates of each point in the cylinder's inner point set are transformed. The difference between the Euclidean distance from the base coordinate system coordinates of each inner point to the corresponding reference circle center and the nominal radius of the cylinder is calculated. The distance deviations of all inner points are weighted according to the inner point ratio of the cylinder and the mean value of the radial residual is obtained.

[0051] Specifically, the rotation-translation transformation matrix is ​​composed of the yaw angle component and the translation component from the initial extrinsic parameters: the yaw angle component determines the rotation part, specifically a two-dimensional rotation matrix with the yaw angle as the rotation angle. This matrix rotates the direction vector in the lidar coordinate system to the corresponding direction in the base coordinate system; the translation component determines the translation part, specifically by adding the forward / backward and left / right translations to the rotated coordinates to obtain the coordinates in the base coordinate system. The lidar coordinate system coordinates of each point in the calibration plate point set are sequentially input into this rotation-translation transformation matrix. The rotation transformation is performed first, and then the translation components are superimposed to obtain the coordinates of each point in the calibration plate in the base coordinate system, denoted as the base coordinate system coordinates of each point in the calibration plate. The reference line normal vector refers to the unit direction vector perpendicular to the reflective surface of the calibration plate in the base coordinate system, which is directly given by the calibration plate normal vector obtained from the calibration plate fitting process. The signed normal distance refers to the perpendicular distance from the base coordinate system coordinates of each point within the calibration plate to the reference line, retaining the positive and negative signs. A positive sign indicates that the point is located on the side pointed to by the normal vector of the reference line, and a negative sign indicates that it is located on the opposite side. The purpose of retaining the sign is to ensure that positive and negative deviations can cancel each other out when the weighted average is calculated, thereby reflecting the systematic offset direction of the point cloud relative to the reference line rather than the offset magnitude. The signed normal distances of all points within all calibration plates are weighted and averaged using the in-point rate of each calibration plate as the weight. That is, the sum of the signed normal distances of all points within each calibration plate is multiplied by the in-point rate of that calibration plate, and then divided by the sum of the in-point rates of all calibration plates to obtain the average normal residual.

[0052] The reference center refers to the theoretical center position of each cylindrical section in the base coordinate system, directly given by the cylinder center coordinates obtained from the cylinder fitting process. The coordinates of each point in the cylinder's inner point set are input into the same rotation and translation transformation matrix as the inner points on the calibration plate to obtain the coordinates of each inner point in the base coordinate system. The straight-line distance from the base coordinate system coordinates of each inner point to the corresponding reference center is calculated, and then the nominal radius of the cylinder is subtracted to obtain the radial distance deviation of each inner point. A positive deviation indicates that the point is located outside the reference circle, and a negative deviation indicates that it is located inside the reference circle. The radial distance deviations of all inner points of all cylinders are weighted and averaged using the inner point rate of each cylinder as a weight. That is, the sum of the radial distance deviations of all inner points of each cylinder is multiplied by the inner point rate of that cylinder, and then divided by the sum of the inner point rates of all cylinders to obtain the mean radial residual. The mean normal residual mainly reflects the systematic error of the translation component in the direction of the calibration plate normal vector in the initial external parameters. The mean radial residual is simultaneously affected by the combined effects of translation component error and yaw angle error. The difference in the numerical meaning of the two constitutes the calculation basis for orientation pre-correction.

[0053] In one specific embodiment, step S3 involves constructing the residual relative deviation based on the difference between the two, using the weighted mean direction of the calibration plate normal vector as the translation correction direction, and using the residual relative deviation as the amplitude to perform directional pre-correction on the translation components in the initial extrinsic parameters, including:

[0054] The relative deviation of the residuals is obtained by subtracting the mean of the radial residuals from the mean of the normal residuals.

[0055] Using the in-point rate of each calibration plate as a weight, the normal vectors of all calibration plates are weighted and averaged to obtain the translation correction direction;

[0056] The negative of the product of the residual relative deviation and the horizontal component of the translation correction direction is added to the horizontal translation component of the initial extrinsic parameter, and the negative of the product of the residual relative deviation and the vertical component of the translation correction direction is added to the vertical translation component of the initial extrinsic parameter to obtain the initial translation pre-correction value.

[0057] Specifically, the relative deviation of the residuals is obtained by subtracting the mean of the normal residuals from the mean of the radial residuals. Its physical meaning is as follows: the mean of the normal residuals is mainly driven by the error of the translation component of the initial extrinsic parameters in the direction of the normal vector of the calibration plate, while the mean of the radial residuals is simultaneously affected by the combined effects of translation error and yaw angle error. The relative deviation of the residuals obtained by subtracting the two removes the contribution of translation error to the radial residuals and mainly retains the projection of the yaw angle error in the radial direction of the cylinder. This amount is used as input in the weighted least squares decomposition of the subsequent yaw angle correction. The translation correction direction is obtained by weighting the calibration board normal vectors of all calibration boards by their respective in-point rates. Specifically, the calculation is as follows: multiply the horizontal component of the calibration board normal vector of each calibration board by the in-point rate of that calibration board, sum the results, and then divide by the sum of the in-point rates of all calibration boards to obtain the horizontal component of the translation correction direction; perform the same operation on the vertical component to obtain the vertical component of the translation correction direction. The physical meaning of the translation correction direction is the comprehensive normal direction of all calibration board normal vectors weighted by observation quality. This direction is the dominant direction of the systematic shift of the point cloud relative to the reference line.

[0058] The geometric basis for orientation pre-correction is as follows: the mean of the normal residual is numerically equal to the projection component of the initial extrinsic parameter translation error vector in the translation correction direction. Therefore, multiplying the mean of the normal residual by the horizontal component of the translation correction direction and then inverting it yields the correction amount to be superimposed on the horizontal translation component of the initial extrinsic parameter. Multiplying the mean of the normal residual by the vertical component of the translation correction direction and then inverting it yields the correction amount to be superimposed on the vertical translation component of the initial extrinsic parameter. Superimposing these two correction amounts onto the horizontal and vertical translation components of the initial extrinsic parameter respectively yields the initial translation pre-correction value. This initial translation pre-correction value refers to the intermediate extrinsic parameter value where only the translation component has undergone orientation correction, and the yaw angle component has not yet been corrected. This value is used as the base value in subsequent yaw angle correction superposition steps. The basis for inversion is: a positive mean of the normal residual indicates that the point cloud as a whole is located on the side pointed to by the normal vector of the reference line, indicating that the initial extrinsic parameter translation component has a positive deviation in that direction and needs to be corrected in the opposite direction; therefore, it is inverted before superposition.

[0059] In one specific embodiment, step S3 involves performing a weighted least squares decomposition on the contribution of the yaw angle to the relative deviation of the residual based on the coordinates of the cylinder center to obtain the yaw angle correction. The translation correction direction and the yaw angle correction are then superimposed on the initial extrinsic parameters to obtain the corrected initial value, including:

[0060] Based on the Euclidean distance from the center coordinates of each cylinder to the origin of the base coordinate system, the geometric sensitivity coefficient of each cylinder to the yaw angle error is calculated. Using the in-cylinder ratio of each cylinder as the weight, the linear equation system composed of the residual relative deviation and the geometric sensitivity coefficient is solved by weighted least squares to obtain the yaw angle correction.

[0061] The yaw angle correction is superimposed on the yaw angle component of the translation pre-correction initial value to obtain the corrected initial value.

[0062] Specifically, the geometric sensitivity coefficient reflects the degree of response of the radial residual of each cylinder to the yaw angle error. Its physical basis is as follows: when there is a yaw angle error in the initial extrinsic parameters, the point cloud of each cylinder undergoes a rotational offset around the origin in the base coordinate system. The offset is proportional to the straight-line distance from the cylinder's center to the origin of the base coordinate system. The farther the cylinder is, the greater its radial response to the yaw angle error. Therefore, the Euclidean distance from the cylinder's center coordinates to the origin of the base coordinate system is directly used as the geometric sensitivity coefficient of that cylinder to the yaw angle error. The logic for establishing the linear equation system is as follows: for each cylinder, its relative residual deviation is equal to the cylinder's geometric sensitivity coefficient multiplied by the yaw angle correction to be determined. When the number of cylinders is greater than one, each cylinder contributes an equation, forming an overdetermined linear equation system with the yaw angle correction as the only unknown. The steps of weighted least squares solution are as follows: using the in-point ratio of each cylinder as the weight, multiply the relative deviation of the residual of each cylinder by the in-point ratio of that cylinder and sum them up, then divide by the sum of the squared geometric sensitivity coefficient of each cylinder multiplied by the in-point ratio of that cylinder, and the resulting quotient is the yaw angle correction. This solution method makes the cylinders with high observation quality (i.e., high in-point ratio) contribute a larger proportion to the yaw angle correction, while the cylinders with low observation quality contribute a smaller proportion.

[0063] The physical meaning of the yaw angle correction is: the systematic deviation estimate of the yaw angle component in the initial extrinsic parameters relative to the true yaw angle. Its positive or negative sign indicates whether the initial yaw angle is too large or too small, and the magnitude indicates the angle amount that needs correction. The yaw angle correction is directly superimposed on the yaw angle component of the translation pre-correction initial value; that is, the yaw angle component in the translation pre-correction initial value is added to the yaw angle correction to obtain the corrected initial value. The corrected initial value refers to the complete extrinsic parameter initial value where both the translation and yaw angle components have undergone directional pre-correction. All three components are closer to the true extrinsic parameters than the initial extrinsic parameters. Using the corrected initial value as the starting point for subsequent nonlinear optimization allows the optimization iteration to start from a position closer to the true value, increasing the reliability of convergence to the global optimum compared to schemes that directly use the initial extrinsic parameters as the starting point.

[0064] In one specific embodiment, step S4 includes:

[0065] Starting with the corrected initial value as the starting point of the optimization variables, the signed normal distance from each point in the weighted calibration plate inner point set to the corresponding reference line after rotation and translation transformation is used to construct the calibration plate residual term. The difference between the Euclidean distance from each point in the weighted cylinder inner point set to the corresponding reference circle center and the nominal radius of the cylinder is used to construct the cylinder residual term. The inner point rate of the calibration plate and the inner point rate of the cylinder are used as the weight coefficients of the calibration plate residual term and the cylinder residual term, respectively. The weighted sum of the calibration plate residual term and the cylinder residual term is used to obtain the joint residual objective function.

[0066] Apply the Huber loss function to each residual term in the joint residual objective function, retain the quadratic form of the residual terms whose absolute residual value is less than the Huber threshold, and convert the residual terms whose absolute residual value is not less than the Huber threshold into the linear form to obtain the robust joint residual objective function;

[0067] Starting with the corrected initial value, the robust joint residual objective function is solved nonlinearly. The iteration is terminated when the parameter change between two adjacent iterations is less than the convergence threshold, and the calibration extrinsic parameters are obtained.

[0068] The root mean square of the optimization termination residual starting from the modified initial value is compared with the root mean square of the optimization termination residual starting from the initial extrinsic parameters. When the root mean square of the termination residual corresponding to the modified initial value is less than the root mean square of the termination residual corresponding to the initial extrinsic parameters, the calibrated extrinsic parameters are used as the output result; otherwise, the result obtained by re-optimization starting from the initial extrinsic parameters is used as the output result.

[0069] Specifically, the calibration plate residual term is constructed as follows: a two-dimensional rotation matrix is ​​constructed using the yaw angle component in the corrected initial value, and a translation vector is constructed using the translation component in the corrected initial value. The coordinates of the lidar coordinate system of each point in the weighted calibration plate point set are first rotated and then superimposed with the translation vector to obtain the coordinates of that point in the base coordinate system. The signed normal distance from that coordinate to the corresponding calibration plate reference line is calculated, and this signed normal distance is used as the calibration plate residual term for that point. The cylindrical residual term is constructed as follows: the coordinates of the lidar coordinate system of each point in the weighted cylindrical point set are rotated and translated by the same method. The coordinates in the base coordinate system are obtained by transformation. The difference between the Euclidean distance from the coordinate to the center of the corresponding cylindrical reference circle and the nominal radius of the cylinder is calculated. This difference is used as the cylindrical residual term at that point. The joint residual objective function is composed of the weighted sum of all calibration plate residual terms and all cylindrical residual terms. All residual terms of each calibration plate are multiplied by the in-point rate of that calibration plate as the weight coefficient, and all residual terms of each cylinder are multiplied by the in-point rate of that cylinder as the weight coefficient. The sum of all weighted residual terms yields the joint residual objective function. This function uses three components in the corrected initial value, namely the lateral translation, longitudinal translation, and yaw angle, as optimization variables.

[0070] The Huber loss function processes each residual term in the joint residual objective function as follows: when the absolute value of a residual term is less than the Huber threshold, the squared form of the residual term is retained as a penalty; when the absolute value of a residual term is not less than the Huber threshold, the residual term is converted into a linear form with the Huber threshold as the coefficient as the penalty. The specific calculation of the linear form is the Huber threshold multiplied by the absolute value of the residual term and then half of the square of the Huber threshold. The Huber threshold is set to 0.01 meters, which is based on the typical residual amplitude of the residual outliers after RANSAC refinement. Residuals smaller than this value are considered normal measurement errors and penalized in a quadratic form. Residuals smaller than this value are considered outliers and their penalty weights are reduced linearly. After the above piecewise processing, a robust joint residual objective function is obtained. The nonlinear iterative solution starts with the corrected initial value. In each iteration, the values ​​of the three optimization variables are updated according to the gradient direction of the robust joint residual objective function. The iteration terminates when the norm of the change in the three optimization variables in two adjacent iterations is less than the convergence threshold. The convergence threshold is set to the tenth decimal place, i.e., the order of ten to the power of negative ten. This value is set according to the parameter spatial resolution corresponding to the millimeter-level translation accuracy and 0.1-degree-level angle accuracy required for the calibration of the lidar extrinsic parameters. After the iteration terminates, the values ​​of the current three optimization variables are output as the calibration extrinsic parameters.

[0071] The execution method of the self-verification comparison logic is as follows: After completing the nonlinear iterative solution starting from the corrected initial value, the calibration extrinsic parameters are substituted into the robust joint residual objective function to calculate the root mean square value of all residual terms, obtaining the root mean square value of the optimization termination residual corresponding to the corrected initial value; simultaneously, starting from the initial extrinsic parameters, the same robust joint residual objective function is re-executed with the nonlinear iterative solution, and the extrinsic parameters after convergence are substituted into the robust joint residual objective function to calculate the root mean square value of all residual terms, obtaining the root mean square value of the optimization termination residual corresponding to the initial extrinsic parameters; The two root mean square (RMS) residuals are compared numerically. If the RMS residual corresponding to the corrected initial value is strictly less than the RMS residual corresponding to the initial extrinsic parameter, the orientation pre-correction is deemed effective, and the calibration extrinsic parameter is used as the final output. If the RMS residual corresponding to the corrected initial value is not less than the RMS residual corresponding to the initial extrinsic parameter, the orientation pre-correction direction is deemed to be biased, and the result obtained by re-optimization starting from the initial extrinsic parameter is used as the output. This self-checking mechanism ensures that the orientation pre-correction step does not degrade the calibration result under any circumstances.

[0072] The above describes the lidar extrinsic parameter calibration method based on the mixing of multiple types of reference objects in the embodiments of this application. The following describes the lidar extrinsic parameter calibration system based on the mixing of multiple types of reference objects in the embodiments of this application. One embodiment of the lidar extrinsic parameter calibration system based on the mixing of multiple types of reference objects in the embodiments of this application includes:

[0073] The separation module is used to read the initial extrinsic parameters from the TF transform tree and separate the single-frame laser point cloud of the lidar on the AGV into a calibration plate point set and a cylindrical point set according to the nominal position of the reference object.

[0074] The fitting module is used to fit the calibration plate point set to obtain the calibration plate normal vector and the calibration plate in-point set, and to fit the cylinder point set to obtain the cylinder center coordinates and the cylinder in-point set.

[0075] The weighting module is used to apply the initial extrinsic parameters to the point set inside the calibration plate and the point set inside the cylinder, respectively, to obtain the mean normal residual and the mean radial residual, and to construct the relative residual deviation by the difference between the two. The translation component in the initial extrinsic parameters is pre-corrected using the weighted mean direction of the calibration plate normal vector as the translation correction direction and the relative residual deviation as the amplitude. The yaw angle contribution in the relative residual deviation is then decomposed using weighted least squares based on the cylinder center coordinates to obtain the yaw angle correction amount. The translation correction direction and the yaw angle correction amount are then superimposed on the initial extrinsic parameters to obtain the corrected initial value.

[0076] The optimization module is used to construct a joint residual objective function containing the point set inside the calibration plate and the point set inside the cylinder, starting from the modified initial value, and obtain the calibration extrinsic parameters through nonlinear optimization.

[0077] This invention also provides a LiDAR extrinsic parameter calibration device based on a hybrid multi-type reference object system. This device can be a server and includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor, designed as a computer, provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores the data corresponding to this embodiment. The network interface communicates with external terminals via a network connection. The computer program, when executed by the processor, implements the above-described method.

[0078] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the lidar extrinsic parameter calibration method based on a mixture of multiple types of reference objects.

[0079] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0080] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a LiDAR extrinsic parameter calibration device (which may be a personal computer, server, or network device, etc.) based on a mixture of multiple types of reference objects to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0081] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for calibrating the extrinsic parameters of a lidar system based on a mixture of multiple types of reference objects, characterized in that, The method includes: Step S1: Read the initial extrinsic parameters from the TF transform tree, and separate the single-frame laser point cloud of the lidar on the AGV into a calibration plate point set and a cylindrical point set according to the nominal position of the reference object; Step S2: Obtain the normal vector of the calibration plate and the set of points inside the calibration plate by fitting the point set of the calibration plate; obtain the coordinates of the cylinder center and the set of points inside the cylinder by fitting the point set of the cylinder. Step S3: Apply the initial extrinsic parameters to the point set inside the calibration plate and the point set inside the cylinder respectively to obtain the mean value of the normal residual and the mean value of the radial residual. Construct the relative deviation of the residual using the difference between the two. Use the weighted mean direction of the calibration plate normal vector as the translation correction direction and the relative deviation of the residual as the amplitude to perform orientation pre-correction on the translation component in the initial extrinsic parameters. Perform weighted least squares decomposition on the yaw angle contribution in the relative deviation of the residual based on the coordinates of the cylinder center to obtain the yaw angle correction amount. Superimpose the translation correction direction and the yaw angle correction amount onto the initial extrinsic parameters to obtain the corrected initial value. Step S4: Starting from the modified initial value, construct a joint residual objective function containing the point set inside the calibration plate and the point set inside the cylinder, and obtain the calibration extrinsic parameters through nonlinear optimization.

2. The lidar extrinsic parameter calibration method based on a mixture of multiple types of reference objects according to claim 1, characterized in that, Step S1 includes: The initial extrinsic parameters are obtained by reading the translation and yaw components of the lidar relative to the base coordinate system from the TF transform tree. Based on the initial extrinsic parameters, the nominal positions of each calibration plate and each cylinder in the base coordinate system are inversely transformed to the lidar coordinate system to obtain the lidar system nominal positions of each reference object. The region of interest is defined centered on the nominal position of the lidar system of each reference object. Points falling into the region of interest of each calibration plate in a single frame of lidar point cloud are extracted as calibration plate point sets, and points falling into the region of interest of each cylinder are extracted as cylinder point sets.

3. The lidar extrinsic parameter calibration method based on a mixture of multiple types of reference objects according to claim 1, characterized in that, Step S2 includes: RANSAC line fitting is performed on the calibration board point set, and points whose normal distance is less than the normal distance threshold are marked as candidate interior points of the calibration board. Least squares refinement is performed based on the candidate interior points of the calibration board to obtain the calibration board normal vector and the calibration board interior point set. The cylindrical point set is fitted with RANSAC circle under the nominal radius constraint of the cylinder. Points with radial distance deviation less than the radial distance threshold are marked as candidate interior points of the cylinder. Least squares refinement is performed based on the candidate interior points of the cylinder to obtain the cylinder center coordinates and the cylinder interior point set. The calibration plate in-point rate is obtained by the ratio of the number of points in the calibration plate in-point set to the number of points in the calibration plate in-point set. The cylinder in-point rate is obtained by the ratio of the number of points in the cylinder in-point set to the number of points in the cylinder in-point set. The calibration plate in-point rate and the cylinder in-point rate are used to weight the calibration plate in-point set and the cylinder in-point set, respectively, to obtain the weighted calibration plate in-point set and the weighted cylinder in-point set.

4. The lidar extrinsic parameter calibration method based on a mixture of multiple types of reference objects according to claim 3, characterized in that, In step S3, the initial extrinsic parameters are applied to the point set inside the calibration plate and the point set inside the cylinder, respectively, to obtain the mean normal residual and the mean radial residual, including: The translation and yaw angle components in the initial extrinsic parameters are used to form a rotation and translation transformation matrix. Based on the rotation and translation transformation matrix, the coordinates of each point in the calibration plate point set are transformed to obtain the base coordinate system coordinates of each point in the calibration plate. Based on the reference line normal vector corresponding to each calibration plate, calculate the signed normal distance from the base coordinate system coordinates of each point in the calibration plate to the reference line. The average value of the signed normal distance of all points in the calibration plate is obtained by weighting the average value of the points in the calibration plate according to the rate of the points in the calibration plate. Based on the rotation and translation transformation matrix, coordinate transformation is performed on each point in the cylinder's inner point set. The difference between the Euclidean distance from the base coordinate system coordinates of each inner point to the corresponding reference circle center and the nominal radius of the cylinder is calculated. The distance deviations of all inner points are weighted and averaged according to the cylinder inner point rate to obtain the mean radial residual.

5. The lidar extrinsic parameter calibration method based on a mixture of multiple types of reference objects according to claim 4, characterized in that, In step S3, the residual relative deviation is constructed based on the difference between the two. The translation correction direction is taken as the weighted mean of the calibration plate normal vector, and the amplitude is used to perform directional pre-correction on the translation components in the initial extrinsic parameters, including: The relative deviation of the residuals is obtained by subtracting the mean radial residual from the mean normal residual. Using the in-point rate of each calibration plate as a weight, the normal vectors of all calibration plates are weighted and averaged to obtain the translation correction direction; The opposite of the product of the residual relative deviation and the horizontal component of the translation correction direction is added to the horizontal translation component of the initial extrinsic parameter, and the opposite of the product of the residual relative deviation and the vertical component of the translation correction direction is added to the vertical translation component of the initial extrinsic parameter to obtain the initial translation pre-correction value.

6. The lidar extrinsic parameter calibration method based on a mixture of multiple types of reference objects according to claim 5, characterized in that, In step S3, the yaw angle contribution in the residual relative deviation is decomposed using weighted least squares based on the cylinder center coordinates to obtain the yaw angle correction. The translation correction direction and the yaw angle correction are then superimposed on the initial extrinsic parameters to obtain the corrected initial value, including: Based on the Euclidean distance from the center coordinates of each cylinder to the origin of the base coordinate system, the geometric sensitivity coefficient of each cylinder to the yaw angle error is calculated. Using the in-cylinder rate of each cylinder as the weight, the linear equation system formed by the residual relative deviation and the geometric sensitivity coefficient is solved by weighted least squares to obtain the yaw angle correction. The yaw angle correction is superimposed on the yaw angle component of the translation pre-correction initial value to obtain the corrected initial value.

7. The lidar extrinsic parameter calibration method based on a mixture of multiple types of reference objects according to claim 6, characterized in that, Step S4 includes: Using the modified initial value as the starting point of the optimization variables, the signed normal distance from each point in the weighted calibration plate inner point set to the corresponding reference line after rotation and translation transformation is used to construct the calibration plate residual term. The difference between the Euclidean distance from each point in the weighted cylinder inner point set to the corresponding reference circle center after rotation and translation transformation and the nominal radius of the cylinder is used to construct the cylinder residual term. The inner point rate of the calibration plate and the inner point rate of the cylinder are used as the weight coefficients of the calibration plate residual term and the cylinder residual term, respectively. The weighted sum of the calibration plate residual term and the cylinder residual term is used to obtain the joint residual objective function. Apply the Huber loss function to each residual term in the joint residual objective function, retain the quadratic form of the residual terms whose absolute residual value is less than the Huber threshold, and convert the residual terms whose absolute residual value is not less than the Huber threshold into linear form to obtain the robust joint residual objective function. Starting from the modified initial value, the robust joint residual objective function is solved nonlinearly. The iteration is terminated when the parameter change between two adjacent iterations is less than the convergence threshold, and the calibration extrinsic parameters are obtained. The root mean square of the optimization termination residual starting from the modified initial value is compared with the root mean square of the optimization termination residual starting from the initial extrinsic parameter. If the root mean square of the termination residual corresponding to the modified initial value is less than the root mean square of the termination residual corresponding to the initial extrinsic parameter, the calibrated extrinsic parameter is used as the output result; otherwise, the result obtained by re-optimization starting from the initial extrinsic parameter is used as the output result.

8. A lidar extrinsic parameter calibration system based on a mixture of multiple types of reference objects, characterized in that, For implementing the lidar extrinsic parameter calibration method based on the mixing of multiple types of reference objects as described in any one of claims 1-7, the lidar extrinsic parameter calibration system based on the mixing of multiple types of reference objects comprises: The separation module is used to read the initial extrinsic parameters from the TF transform tree and separate the single-frame laser point cloud of the lidar on the AGV into a calibration plate point set and a cylindrical point set according to the nominal position of the reference object. The fitting module is used to fit the calibration plate point set to obtain the calibration plate normal vector and the calibration plate in-point set, and to fit the cylinder point set to obtain the cylinder center coordinates and the cylinder in-point set. The weighting module is used to apply the initial extrinsic parameters to the point set inside the calibration plate and the point set inside the cylinder, respectively, to obtain the mean normal residual and the mean radial residual, and to construct the relative residual deviation by the difference between the two. The translation component in the initial extrinsic parameters is pre-corrected using the weighted mean direction of the calibration plate normal vector as the translation correction direction and the relative residual deviation as the amplitude. The yaw angle contribution in the relative residual deviation is then decomposed using weighted least squares based on the cylinder center coordinates to obtain the yaw angle correction amount. The translation correction direction and the yaw angle correction amount are then superimposed on the initial extrinsic parameters to obtain the corrected initial value. The optimization module is used to construct a joint residual objective function containing the point set inside the calibration plate and the point set inside the cylinder, starting from the modified initial value, and obtain the calibration extrinsic parameters through nonlinear optimization.

9. A lidar extrinsic parameter calibration device based on a mixture of multiple types of reference objects, characterized in that, It includes a memory and a processor, the memory storing a computer program that can run on the processor, and the processor executing the computer program to implement the lidar extrinsic parameter calibration method based on the hybrid use of multiple types of reference objects as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it causes the processor to execute the lidar extrinsic parameter calibration method based on the mixing of multiple types of reference objects as described in any one of claims 1 to 7.

Images