A LiDAR-GPS / IMU self-calibration method for movable carriers

Abstract

The present invention discloses a LiDAR-GPS / IMU self-calibration method for a movable carrier. The method includes two parts: extrinsic parameter initialization and extrinsic parameter refinement. The extrinsic parameter initialization part calculates the initial value of the extrinsic parameter by decoupling the rotation and translation terms, calculates the reliable initial rotation value, calculates the initial translation value by rotation constraint, and removes the accumulated error of the LiDAR odometer through loop closure detection to complete map construction. The extrinsic parameter refinement part tightly couples the LiDAR odometer and the GPS / IMU odometer through the scan-to-map method, and obtains the refined extrinsic parameter value by constraining the extrinsic parameter value by the absolute position under the map. The present invention provides a more automated and lower-cost extrinsic parameter self-calibration method. By adopting more strategies, such as loop closure detection and tight coupling strategies, higher calibration accuracy, smaller error, and more complete strategies are obtained for the extrinsic parameter values ​​of multiple sensors of the movable carrier.

Description

Technical Field

[0001] The present invention relates to the technical field of sensor calibration, and in particular to a LiDAR-GPS / IMU self-calibration method for a movable carrier. Background Art

[0002] In the fields of autonomous driving and robotics, commonly used sensors include LiDAR, cameras, GPS / IMU, and wheel speedometers. GPS / IMU is a comprehensive system that combines an inertial measurement unit (IMU) with a global positioning system (GPS), providing centimeter-level absolute positioning accuracy. A single IMU, depending on its hardware performance, can only provide relatively accurate motion estimates for a short period of time, but positioning errors increase rapidly over time. LiDAR, the primary ranging sensor for autonomous vehicles, can estimate the vehicle's relative motion pose in real time through point cloud registration. However, LiDAR-based SLAM (Simultaneous Local Mapping) can degrade in areas with sparse geometric feature information. Cameras project a 3D scene onto a 2D image, which is then used to estimate the motion state of visual odometry. However, visual odometry suffers from scale uncertainty, and images are also susceptible to exposure and lighting conditions.

[0003] To leverage the strengths and weaknesses of individual sensors, many tasks require more than one sensor. Multi-sensor fusion has become a trend in autonomous driving. However, because each sensor has a different coordinate system, the data from each sensor must be unified into a common coordinate system for fusion. This requires calculating both intrinsic parameters for each sensor and extrinsic parameters for multiple sensors. Since intrinsic parameters for IMUs and LiDARs are typically provided by the manufacturer at the time of shipment, extrinsic calibration between LiDARs and GPS / IMUs is often necessary. Calibration of extrinsic parameters for LiDARs and GPS / IMUs currently faces numerous challenges. For example, vehicle motion in autonomous driving is unlike that of a robotic arm; it primarily involves three-degree-of-freedom planar motion with relatively few constraints on the other degrees of freedom, resulting in relatively large errors in the remaining three degrees of freedom. Due to significant differences in the measurement principles of multi-axis LiDARs and GPS / IMUs, many calibration methods require specific vehicle operating conditions and manual calibration scenes to ensure that the different sensors detect the same objects for calibrating their extrinsic parameters. This results in high costs and low automation.

[0004] The paper "Vision Meets Robotics: The KITTI Dataset" proposes hand-eye calibration using point-to-surface LiDAR odometry and GPS / IMU odometry. However, this approach lacks initial extrinsic parameter calculations, making the calibration results more dependent on the LiDAR odometry registration accuracy. The paper "3D LiDAR-GPS / IMU Calibration Based on Hand-Eye Calibration Model for Unmanned Vehicle" uses ground point optimization to solve the z translation, roll, and pitch angle transformations. The three-dimensional problem is then converted to a two-dimensional one, and the remaining three degrees of freedom are estimated through hand-eye calibration. This method can be viewed as using the three optimized degrees of freedom from ground points as initial values, followed by hand-eye calibration. However, the initialization presented is crude and rudimentary, resulting in relatively low accuracy. The Baidu Apollo project divides calibration into two steps: first, initial extrinsic parameters are calculated using trajectories, and then refined using feature-based methods. However, the specific implementation strategy is not comprehensive. The paper "Improved Vehicle LiDAR Calibration With Trajectory-Based Hand-Eye Method" uses large trajectories to solve the problem that planar motion has weak constraints on other degrees of freedom. However, for a calibration, the requirement for recording kilometer-level data is relatively demanding. Summary of the Invention

[0005] In view of the above-mentioned deficiencies in the prior art, the object of the present invention is to provide a LiDAR-GPS / IMU self-calibration method for a movable carrier, that is, through extrinsic parameter initialization and extrinsic parameter refinement, and the adoption of more strategies, to obtain a LiDAR-GPS / IMU self-calibration method for a movable carrier with higher calibration accuracy and more complete strategies.

[0006] The technical solutions of the present invention are as follows:

[0007] A LiDAR-GPS / IMU self-calibration method for a movable carrier, characterized in that the method includes extrinsic parameter initialization and extrinsic parameter refinement; wherein:

[0008] The external parameter initialization includes the following steps:

[0009] S1: Obtain feature-based LiDAR odometry and GPS / IMU odometry interpolated based on LiDAR timestamps to construct hand-eye calibration equations;

[0010] S2: Decoupling reliable initial values ​​of rotational extrinsic parameters and initial values ​​of translational extrinsic parameters according to the hand-eye calibration equation to obtain initial extrinsic parameter values;

[0011] S3: Tightly couple the LiDAR and IMU using the initial extrinsic parameter values, and then use factor graph optimization to eliminate the accumulated error of the LiDAR odometry to complete the construction of the global map;

[0012] The external parameter refinement comprises the following steps:

[0013] S4: filtering the global map constructed by the external parameter initialization process to obtain a local map;

[0014] S5: Obtaining the absolute position and posture of the LiDAR odometry by registering the real-time scanning point cloud with the local map;

[0015] S6: Constructing a cost function based on the absolute posture and the initial extrinsic parameter value, and performing nonlinear optimization to obtain a more accurate refined extrinsic parameter value.

[0016] Furthermore, the specific process of step S1 includes:

[0017] The predicted pose is obtained by fusing the six degrees of freedom of the constant velocity model with the three degrees of freedom of the ground point optimization.

[0018] Calculate the pitch angle of each point, perform coarse extraction based on the pitch angle being less than the threshold, and then perform fine extraction using the random sampling consistency RANSAC algorithm;

[0019] Feature points are extracted by the size of the curvature, and the residual is constructed based on the feature points for alignment to obtain the inter-frame LiDAR odometry. The inter-frame LiDAR odometry is used as the initial value for the alignment of the back-end frame to the local map, thereby refining the LiDAR odometry to obtain a feature-based LiDAR odometry.

[0020] Furthermore, the specific process of step S1 also includes:

[0021] At the same time, linear interpolation is performed on the adjacent high-frequency GPS / IMU based on the LiDAR timestamp to obtain the GPS / IMU odometer;

[0022] The hand-eye calibration equation AX = XB is constructed based on the relative pose constraints of the LiDAR odometer and the interpolated GPS / IMU odometer obtained above:

[0023]

[0024] Where, {L k} and {I k} are the LiDAR pose of the kth frame and the interpolated GPS / IMU pose at this time; For {I k} to {Ik+1}'s relative pose; It is from {L k} to {L k+1}'s relative pose.

[0025] Furthermore, the specific process of step S2 is as follows:

[0026] The hand-eye calibration equation established by formula (1) is decomposed into two parts: rotation term and translation term:

[0027]

[0028]

[0029] After the rotation term of formula (2) is written in the form of quaternion and converted into matrix multiplication using the left and right quaternion multiplication matrices, the following formula is obtained:

[0030]

[0031] in is the quaternion multiplication operator, and and are the matrix representations of left and right quaternion multiplications, respectively.

[0032] Furthermore, the specific process of step S2 also includes:

[0033] After accumulating measurement data at different times, the following overdetermined equation is obtained:

[0034]

[0035] where K is the number of rotation pairs in the overdetermined equation, is a robust weight used to better handle outliers;

[0036] The difference angle of the current rotation pair on the angle axis is calculated by formula (4) and used as the parameter of Huber loss, and its derivative is the weight:

[0037]

[0038] Among them, ρ() represents Huber loss, q ω is the real part of the quaternion q, () * Indicates taking the inverse of a quaternion.

[0039] Furthermore, the step S3 is specifically as follows:

[0040] The initial values ​​of rotational and translational extrinsic parameters are initialized by the LiDAR-GPS / IMU, the LiDAR and IMU are tightly coupled, and the point clouds of two consecutive frames are aligned. The relative pose is then combined with the z translation, roll angle, and pitch angle obtained through ground point optimization to optimize the relative pose through loop detection, thereby completing the construction of the global map.

[0041] Furthermore, the local map is composed of key frames that are adjacent in time and key frames that are adjacent in position.

[0042] Furthermore, the step S4 further includes:

[0043] Determine the key frame; the key frame can be regarded as a new key frame as long as the relative displacement or relative rotation angle of the current frame relative to the previous frame is greater than a set threshold.

[0044] Furthermore, the step S5 specifically includes:

[0045] The LiDAR odometry is obtained by registering it with the local map. The LiDAR odometry and IMU are tightly coupled by initial extrinsic parameters. The LiDAR point cloud is dedistorted by a pre-integration method and used as the initial guess pose of the LiDAR odometry. The IMU deviation is corrected by optimizing the LiDAR odometry based on a factor graph.

[0046] Furthermore, the step S6 specifically includes:

[0047] The LiDAR odometry is constructed using the following formula Constraints on external parameters:

[0048]

[0049] in is the GPS / IMU pose corresponding to the current LiDAR timestamp;

[0050] Decoupled into new rotation part and translation part, namely:

[0051]

[0052] The cost function of absolute pose and extrinsic parameters is constructed by formula (12) to achieve the refinement and optimization of extrinsic parameters:

[0053]

[0054] The cost function is optimized by the Levenberg-Marquardt algorithm, and when the iteration converges, the final accurate extrinsic parameters can be obtained.

[0055] Beneficial effects: The present invention discloses a LiDAR-GPS / IMU self-calibration method for a movable carrier, which is used on a movable carrier, such as an unmanned vehicle and a robot. Compared with the prior art, the embodiment of the present invention provides an external parameter self-calibration method with a higher degree of automation and a lower cost. By adopting more strategies, such as loop detection and tight coupling strategy, the external parameters of multiple sensors of the movable carrier with higher calibration accuracy, smaller error and more complete strategy can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0057] Figure 1 A flowchart of the LiDAR-GPS / IMU self-calibration method for a movable carrier provided in an embodiment of the present invention.

[0058] Figure 2 A schematic diagram of the process of the LiDAR-GPS / IMU self-calibration method for a movable carrier provided in an embodiment of the present invention.

[0059] Figure 3 A diagram showing the relative pose relationships of hand-eye calibration in the extrinsic parameter initialization portion of the LiDAR-GPS / IMU self-calibration method for a movable carrier provided in an embodiment of the present invention.

[0060] Figure 4 A pose relationship diagram of the extrinsic parameter refinement part of the LiDAR-GPS / IMU self-calibration method for a movable carrier provided in an embodiment of the present invention.

[0061] Figure 5 A schematic diagram of a test unmanned vehicle used in one embodiment of the present invention. DETAILED DESCRIPTION

[0062] To make the objectives, technical solutions, and effects of the present invention more clear and distinct, the present invention is further described in detail below. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. The embodiments of the present invention are described below with reference to the accompanying drawings.

[0063] The LiDAR-GPS / IMU self-calibration method for a movable carrier provided in this embodiment is used on movable carriers, such as unmanned vehicles and robots, in some vehicles, robotic equipment, and scenarios that require autonomous driving functions. Currently, multi-sensor fusion has become a trend in autonomous driving. However, since the coordinate systems of each sensor are different, they need to be unified into a coordinate system to fuse the data of each sensor. Specifically, it is necessary to calculate the intrinsic parameters of a single sensor and the extrinsic parameters of multiple sensors. Since the intrinsic parameters of IMU and LiDAR are usually provided by the manufacturer at the time of shipment, the present invention mainly solves the extrinsic parameter calibration between LiDAR and GPS / IMU.

[0064] In order to realize a more automated and lower-cost external parameter self-calibration method, and to achieve a higher calibration accuracy, smaller error, and more perfect strategy for the external parameter effect of multiple sensors on a movable carrier, the present invention provides a LiDAR-GPS / IMU self-calibration method for a movable carrier, please refer to Figure 1-Figure 5 , the implementation process of the embodiment is as follows Figure 1 As shown, Figure 2 The overall idea of ​​this method is presented, which includes the following steps:

[0065] External parameter initialization and external parameter refinement, including:

[0066] The external parameter initialization includes the following steps:

[0067] S1: Obtain feature-based LiDAR odometry and GPS / IMU odometry interpolated based on LiDAR timestamps to construct hand-eye calibration equations;

[0068] S2: Decoupling reliable initial values ​​of rotational extrinsic parameters and initial values ​​of translational extrinsic parameters according to the hand-eye calibration equation to obtain initial extrinsic parameter values;

[0069] S3: Tightly couple the LiDAR and IMU using the initial extrinsic parameter values, and then use factor graph optimization to eliminate the accumulated error of the LiDAR odometry to complete the construction of the global map;

[0070] The external parameter refinement includes the following steps:

[0071] S4: filtering the global map constructed by the external parameter initialization process to obtain a local map;

[0072] S5: Obtaining the absolute position and posture of the LiDAR odometry by registering the real-time scanning point cloud with the local map;

[0073] S6: Constructing a cost function based on the absolute posture and the initial extrinsic parameter value, and performing nonlinear optimization to obtain a more accurate refined extrinsic parameter value.

[0074] In the embodiment of the present invention, the specific process of step S1 is:

[0075] The six degrees of freedom of the constant velocity model are fused with the three degrees of freedom optimized for ground points to obtain the predicted pose. Since LiDAR is usually installed in a horizontal manner, the pitch angle of each point can be calculated, and then coarse extraction is performed based on the pitch angle being less than the threshold, and then fine extraction is performed using the random sampling consistency RANSAC algorithm.

[0076] After obtaining the predicted pose, we extract feature points based on the size of the curvature, construct residuals based on the feature points for alignment, and obtain the inter-frame LiDAR odometry. The inter-frame LiDAR odometry is used as the initial value for the back-end frame-to-local map alignment, thereby refining the LiDAR odometry to obtain a feature-based LiDAR odometry.

[0077] At the same time, linear interpolation is performed on the adjacent high-frequency GPS / IMU based on the LiDAR timestamp to obtain the GPS / IMU odometer.

[0078] The hand-eye calibration equation AX=XB is constructed based on the relative pose constraints of the LiDAR odometer and the interpolated GPS / IMU odometer obtained above. The relative position relationship between the LiDAR odometer and the interpolated GPS / IMU odometer is as follows: Figure 3 As shown:

[0079]

[0080] Where, {L k} and {I k} are the LiDAR pose of the kth frame and the interpolated GPS / IMU pose at this time, For {I k} to {I k+1}'s relative pose. It is from {L k} to {L k+1}'s relative pose.

[0081] In an embodiment, the specific process of step S2 is:

[0082] The hand-eye calibration equation established by formula (1) can be decomposed into two parts: rotation term and translation term:

[0083]

[0084]

[0085] The translation term can be obtained by simply solving the rotation term. At this time, a reliable initial rotation value is required. After the rotation term of formula (2) is written in the form of quaternion and converted into matrix multiplication using the left and right quaternion multiplication matrices, the following formula is obtained:

[0086]

[0087] in is the quaternion multiplication operator, and and are the matrix representations of left and right quaternion multiplications, respectively.

[0088] After accumulating measurement data at different times, we obtain the following overdetermined equation:

[0089]

[0090] where K is the number of rotation pairs in the overdetermined equation, It is a robust weight used to better handle outliers. The difference angle of the current rotation pair on the angle axis is calculated by formula (4) and used as a parameter of Huber loss, and its derivative is the weight:

[0091]

[0092] Among them, ρ() represents Huber loss, q ω is the real part of the quaternion q, () * Indicates taking the inverse of a quaternion.

[0093] In the embodiment of the present invention, the overdetermined equation (5) is solved by SVD, and its closed-form solution is the right unit singular vector corresponding to the minimum singular value. At the same time, in order to ensure sufficient rotation constraints, it is necessary to ensure that the second smallest singular value is greater than the set threshold. The rapid increase of , through the priority queue to eliminate the minimum rotation constraint, thus obtaining a reliable initial rotation value. At this time, the relative poses of different time periods are accumulated to solve the translation term:

[0094]

[0095] However, in general, the movement of a vehicle is usually planar motion with three degrees of freedom: x, y, and yaw. Therefore, the z-axis is usually not observable. At the same time, since the acceleration of the IMU is coupled with gravity, it is related to rotation. Therefore, it is not reliable to calculate the initial translation value based on the initial rotation value obtained by the IMU measurement. When the calculated z-axis translation value has a large deviation, we can make a plane assumption and rewrite the equation as follows:

[0096]

[0097] Equation (8) is the planar motion constraint generated by the k+1th relative pose, where γ is the yaw angle, t x and t y are the translations of the x and y axes, R k+1 yes The 2x2 block matrix in the upper left corner, and yes and The first two elements of the column vector.

[0098] The equation can be rewritten as AX = b:

[0099]

[0100] in, [t1k+1] i Represents the i-th element of the column vector [t1k+1].

[0101] Similar to equation (5), by superimposing the measurements at different times according to equation (9), we can obtain the final matrix equation AX = b, which can be solved by the least squares method:

[0102]

[0103] The obtained yaw angle is fused with the initial rotation value, and the z-axis translation is set to the hand-measured translation value and further optimized in the refinement part.

[0104] In a further embodiment of the present invention, step S3 is specifically as follows:

[0105] The rotation and translation extrinsics obtained from the LiDAR-GPS / IMU are initialized, the LiDAR and IMU are tightly coupled, and the point clouds of two consecutive frames are aligned. The obtained relative pose is then combined with the z translation, roll angle, and pitch angle obtained through ground point optimization. This loop detection optimizes the relative pose and completes the construction of the global map.

[0106] In an embodiment, the step S4 is specifically as follows:

[0107] The local map is obtained by filtering the global map obtained in the initialization process. That is, the local map is composed of key frames that are adjacent in time and position. For temporal proximity, only the key frames adjacent to the current timestamp need to be extracted. For position proximity, the position of each key frame needs to be set to the xyz of a point. A point cloud is constructed through all key frames and added to the KD tree. The nearest neighbor search is performed through the KD tree to obtain key frames that are adjacent in position.

[0108] Furthermore, the S4 further includes:

[0109] Determine the key frame; for the determination of the key frame, as long as the relative displacement or relative rotation angle of the current frame relative to the previous frame is greater than a set threshold, it can be regarded as a new key frame.

[0110] In a further embodiment, step S5 is specifically as follows:

[0111] The LiDAR odometry is obtained by registering it with the local map obtained above. The LiDAR odometry and IMU are tightly coupled by the initial extrinsic parameters. The LiDAR point cloud is dedistorted using the pre-integration method and used as the initial guess pose of the LiDAR odometry. The IMU deviation is corrected by optimizing the LiDAR odometry based on the factor graph.

[0112] In a further embodiment, step S6 is specifically as follows:

[0113] The LiDAR odometer obtained above The constraints on external parameters can be constructed using the following formula. The relative position relationship between the LiDAR odometer and the GPS / IMU odometer is as follows: Figure 4 As shown:

[0114]

[0115] in is the GPS / IMU pose corresponding to the current LiDAR timestamp.

[0116] According to the method of decoupling the rotation part and translation part of the hand-eye calibration equation in step S2, formula (11) can also be decoupled into a new rotation part and translation part, namely:

[0117]

[0118] The construction of the overdetermined equations for rotation and translation is similar to the initialization method, but the assumption of planar motion is not required at this time. The translation parameters of the extrinsic parameters can be directly optimized through the absolute pose. The cost function of the absolute pose and the extrinsic parameters can be constructed by formula (12) to achieve the refined optimization of the extrinsic parameters:

[0119]

[0120] The cost function is optimized by the Levenberg-Marquardt algorithm, and when the iteration converges, the final accurate extrinsic parameters can be obtained.

[0121] Example 1

[0122] See also Figure 5 , the embodiment of the present invention is based on Figure 5The sensors were assembled as shown. The Ouster 128-axis LiDAR was deployed at the center of the top truss at the front of the vehicle. The GPS and LiDAR were equidistantly spaced at either end of the truss. The FDI integrated navigation system was installed on a panel above the center of the front wheel. Data was collected in outdoor scenarios. The Ouster-128 LiDAR output point clouds at a frequency of 10Hz, and the FDI integrated navigation system output GPS / IMU measurements at a frequency of 100Hz. Since there are no ground truth values ​​for LiDAR-GPS / IMU in real-world scenarios and no open-source LiDAR-GPS / IMU calibration algorithms, we used the ground truth values ​​provided by the CAD assembly drawings for verification and conducted multiple experiments to verify the repeatability and accuracy of the calibration results. The experimental results are shown in the following table:

[0123]

[0124] Experimental inspection and calibration results table

[0125] As can be seen, our proposed calibration system can still achieve high accuracy in the real world. In the rotation part, the pitch angle errors of the three scenes are all within 0.2 degrees, the errors of the original angles are all within 0.45 degrees, and the errors of the yaw angles are all within 0.6 degrees. In the translation part, all errors are within 0.05 meters.

[0126] In summary, the present invention discloses a LiDAR-GPS / IMU self-calibration method for a movable carrier, which is used on a movable carrier, such as an unmanned vehicle and a robot. The method includes extrinsic parameter initialization and extrinsic parameter refinement. The extrinsic parameter initialization portion decouples the extrinsic parameter from the rotation and translation portion, calculates a reliable initial rotation value, and then calculates the initial translation value through rotation constraints, and completes map construction by removing the accumulated error of the LiDAR odometer through loop closure detection; the extrinsic parameter refinement portion tightly couples the LiDAR odometer and the IMU in a scan-to-map manner, and refines the extrinsic parameter by constraining the external parameter with the absolute pose under the local map. The present invention provides a more automated and lower-cost external parameter self-calibration method. By adopting more strategies, such as loop closure detection and tight coupling strategies, the external parameters of multiple sensors of the movable carrier with higher calibration accuracy, smaller errors, and more complete strategies are obtained.

[0127] It should be understood that the application of the present invention is not limited to the above examples. For those skilled in the art, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A LiDAR-GPS / IMU self-calibration method for a movable carrier, characterized in that: The method includes external parameter initialization and external parameter refinement; wherein: The external parameter initialization includes the following steps: S1: Obtain feature-based LiDAR odometry and GPS / IMU odometry interpolated based on LiDAR timestamps to construct hand-eye calibration equations; S2: Decoupling reliable initial values ​​of rotational extrinsic parameters and initial values ​​of translational extrinsic parameters according to the hand-eye calibration equation to obtain initial extrinsic parameter values; S3: Tightly couple the LiDAR and IMU using the initial extrinsic parameter values, and then use factor graph optimization to eliminate the accumulated error of the LiDAR odometry to complete the construction of the global map; The external parameter refinement includes the following steps: S4: filtering the global map constructed by the external parameter initialization process to obtain a local map; S5: Obtaining the absolute position and posture of the LiDAR odometry by registering the real-time scanning point cloud with the local map; S6: Constructing a cost function based on the absolute posture and the initial extrinsic parameter value, and performing nonlinear optimization to obtain a more accurate refined extrinsic parameter value.

2. The LiDAR-GPS / IMU self-calibration method for a movable carrier according to claim 1, characterized in that: The specific process of step S1 includes: The predicted pose is obtained by fusing the six degrees of freedom of the constant velocity model with the three degrees of freedom of the ground point optimization. Calculate the pitch angle of each point, perform coarse extraction based on the pitch angle being less than the threshold, and then perform fine extraction using the random sampling consistency RANSAC algorithm; Feature points are extracted by the size of the curvature, and the residual is constructed based on the feature points for alignment to obtain the inter-frame LiDAR odometry. The inter-frame LiDAR odometry is used as the initial value for the alignment of the back-end frame to the local map, thereby refining the LiDAR odometry to obtain a feature-based LiDAR odometry.

3. The LiDAR-GPS / IMU self-calibration method for a movable carrier according to claim 1, characterized in that: The specific process of step S1 also includes: At the same time, linear interpolation is performed on the adjacent high-frequency GPS / IMU based on the LiDAR timestamp to obtain the GPS / IMU odometer; The hand-eye calibration equation AX = XB is constructed based on the relative pose constraints of the LiDAR odometer and the interpolated GPS / IMU odometer obtained above: Where, {L k } and {I k } are the LiDAR pose of the kth frame and the interpolated GPS / IMU pose at this time; For {I k } to {I k+1 }'s relative pose; It is from {L k } to {L k+1 }'s relative pose.

4. The LiDAR-GPS / IMU self-calibration method for a movable carrier according to claim 3, characterized in that: The specific process of step S2 includes: The hand-eye calibration equation established by formula (1) is decomposed into two parts: rotation term and translation term: After the rotation term of formula (2) is written in the form of quaternion and converted into matrix multiplication using the left and right quaternion multiplication matrices, the following formula is obtained: in is the quaternion multiplication operator, and and are the matrix representations of left and right quaternion multiplications, respectively.

5. The LiDAR-GPS / IMU self-calibration method for a movable carrier according to claim 3, characterized in that: The specific process of step S2 also includes: After accumulating measurement data at different times, the following overdetermined equation is obtained: where K is the number of rotation pairs in the overdetermined equation, is a robust weight used to better handle outliers; The difference angle of the current rotation pair on the angle axis is calculated by formula (4) and used as the parameter of Huber loss, and its derivative is the weight: Among them, ρ() represents Huber loss, q ω is the real part of the quaternion q, and ()* represents the inverse of the quaternion.

6. The LiDAR-GPS / IMU self-calibration method for a movable carrier according to claim 3, characterized in that: The step S3 specifically includes: The initial values ​​of rotational and translational extrinsic parameters are initialized by the LiDAR-GPS / IMU, the LiDAR and IMU are tightly coupled, and the point clouds of two consecutive frames are aligned. The relative pose is then combined with the z translation, roll angle, and pitch angle obtained through ground point optimization to optimize the relative pose through loop detection, thereby completing the construction of the global map.

7. The LiDAR-GPS / IMU self-calibration method for a movable carrier according to claim 1, wherein: The local map is composed of key frames that are adjacent in time and key frames that are adjacent in position.

8. The LiDAR-GPS / IMU self-calibration method for a movable carrier according to claim 7, characterized in that: Said S4 further comprises: Determine the key frame; the key frame can be regarded as a new key frame as long as the relative displacement or relative rotation angle of the current frame relative to the previous frame is greater than a set threshold.

9. The LiDAR-GPS / IMU self-calibration method for a movable carrier according to claim 1, characterized in that: The step S5 specifically includes: The LiDAR odometry is obtained by registering it with the local map. The LiDAR odometry and IMU are tightly coupled by initial extrinsic parameters. The LiDAR point cloud is dedistorted by a pre-integration method and used as the initial guess pose of the LiDAR odometry. The IMU deviation is corrected by optimizing the LiDAR odometry based on a factor graph.

10. The LiDAR-GPS / IMU self-calibration method for a movable carrier according to claim 1, characterized in that: The step S6 specifically includes: The LiDAR odometry is constructed using the following formula Constraints on external parameters: in is the GPS / IMU pose corresponding to the current LiDAR timestamp; Decoupled into new rotation part and translation part, namely: The cost function of absolute pose and extrinsic parameters is constructed by formula (12) to achieve the refinement and optimization of extrinsic parameters: The cost function is optimized by the Levenberg-Marquardt algorithm, and when the iteration converges, the final accurate extrinsic parameters can be obtained.

Images