Multi-sensor-based agv i-beam wheel loading and unloading pose calibration method and system

Abstract

The application discloses a multi-sensor-based AGV I-beam wheel loading and unloading pose calibration method and system, the method comprising: collecting three-dimensional point cloud data of a hanging bracket cantilever through a 3D camera; scanning annular profile data of an I-beam inner hole through a 2D laser sensor; preprocessing the three-dimensional point cloud data and the annular profile data and converting them to the same vehicle body coordinate system through a coordinate system; executing a double-target pose collaborative matching algorithm based on the converted data to determine the relative pose deviation of the hanging bracket cantilever and the I-beam; including: performing feature extraction and feature constraint matching on the converted data, and the constraints including a circle center alignment constraint and an axis parallel constraint, calculating the relative pose deviation of the hanging bracket cantilever and the I-beam based on the matching features; and performing dynamic deviation real-time calibration according to the relative pose deviation, adjusting the vehicle body position or the gear cutting angle. The application can improve the accurate and stable AGV I-beam wheel automatic loading and unloading, and improve the efficiency and quality of the entire industrial production.

Description

Technical Field

[0001] This application relates to the technical field of posture calibration of automated logistics equipment, specifically to a method and system for calibrating the posture of AGV loading and unloading based on multiple sensors. Background Technology

[0002] In the field of industrial automation, AGVs (Automated Guided Vehicles) are being used more and more widely, especially in the automated loading and unloading of I-beams, where they can effectively improve production efficiency and reduce labor costs. With the continuous expansion of industrial production scale and the increasing demands for production precision, the accuracy and reliability of AGVs in the loading and unloading process of I-beams have become crucial.

[0003] In existing technologies, the conventional approach for automated loading and unloading of AGV I-beams is to rely solely on camera recognition of the cantilever's position. This method primarily involves capturing images of the cantilever and then using image processing techniques to determine its position and orientation. However, this approach only considers the cantilever's position and does not effectively measure the real-time orientation of the I-beams on the AGV's gears within the vehicle's coordinate system. In practice, knowing only the cantilever's position without understanding the I-beam's own orientation makes accurate loading and unloading difficult. This is because the real-time position and tilt angle of the I-beams all affect the accuracy of the loading and unloading process.

[0004] Clearly, existing technologies have many shortcomings. Firstly, the lack of vehicle body pose measurement for the I-beams on the AGV's gears prevents the acquisition of key parameters, creating an information gap where "only the bracket position is known, but the I-beam's own posture is unknown." Secondly, relying solely on a single camera to identify the bracket lacks the support of I-beam posture data, making it impossible to establish a relative posture relationship between the bracket cantilever and the I-beams. This makes it impossible to correct alignment deviations caused by AGV driving errors and I-beam placement misalignments during loading and unloading. Furthermore, insufficient positioning accuracy is also a serious problem. Due to the lack of I-beam posture measurement and dual-target collaborative calibration, alignment accuracy is low, easily leading to misalignment between the gears and cantilever holes, I-beam jamming, and collisions, directly affecting the stability of automated loading and unloading. Summary of the Invention

[0005] To achieve precise and stable automatic loading and unloading of AGV I-beam wheels and improve the efficiency and quality of the entire industrial production, this application provides a multi-sensor-based AGV I-beam wheel loading and unloading posture calibration method and system.

[0006] In a first aspect, this application provides a multi-sensor-based method for calibrating the loading and unloading posture of an AGV (Automated Guided Vehicle) with I-beam wheels, including: The 3D point cloud data of the cantilever arm of the hanger is acquired using a 3D camera; Scan the annular contour data of the inner hole of the I-beam wheel using a 2D laser sensor; The 3D point cloud data and the ring contour data are preprocessed and transformed to the same vehicle coordinate system using coordinate system one; Based on the transformed data, a dual-target pose collaborative matching algorithm is executed to determine the relative pose deviation between the cantilever arm and the I-beam wheel. The steps of executing the dual-target pose collaborative matching algorithm include: performing feature extraction and feature constraint matching on the transformed data, and the constraints include center alignment constraints and axis parallel constraints; calculating the relative pose deviation between the cantilever arm and the I-beam wheel based on the matching features; the relative pose deviation includes the translation compensation amount required by the vehicle body and the angle compensation amount required by the gear. The dynamic deviation is calibrated in real time based on the relative pose deviation, and the vehicle body position or gear angle is adjusted. The dynamic deviation real-time calibration steps include: fusing three-dimensional point cloud data and ring contour data in real time through Kalman filter tracking algorithm, dynamically updating the pose parameters of the cantilever and the I-beam wheel, and correcting the compensation amount in real time based on the calculated new relative pose deviation.

[0007] By adopting the above scheme, data on the cantilever arm and the inner hole of the I-beam are collected by a 3D camera and a 2D laser sensor, respectively. After preprocessing and coordinate unification, a dual-target pose collaborative matching algorithm is executed. The circle center alignment constraint and axis parallel constraint are configured to perform feature matching and calculate the relative pose deviation, which solves the problem of relative pose mismatch between the cantilever arm and the I-beam, improves positioning accuracy, and then the deviation is calibrated in real time by a Kalman filter tracking algorithm to avoid static measurement errors and improve the accuracy of pose matching between the cantilever arm and the I-beam during the loading and unloading of the AGV I-beam.

[0008] Preferably, the preprocessing steps for the 3D point cloud data and the ring contour data include: adding hardware-level timestamps to the 3D point cloud data and the ring contour data; using IMU sensors installed on the vehicle body to provide vehicle operating status data, constructing motion trajectories within a sliding window, and interpolating and numerically compensating the corresponding 3D point cloud data or ring contour data for each frame.

[0009] By adopting the above scheme, hardware-level timestamps are added to 3D point cloud data and ring contour data. Motion trajectories are constructed using vehicle running status data provided by IMU sensors. Interpolation and numerical compensation are performed on each frame of data, which can reduce errors caused by inconsistent data acquisition time and vehicle movement, and improve the accuracy and reliability of the data.

[0010] Preferably, the preprocessing steps for the annular contour data further include: performing multi-threshold filtering and contour extraction on the collected annular contour data; performing distance-based spatial clustering tests based on the spatial distribution characteristics of the reflection mechanism to screen out and remove outlier noise points; using a point cloud hole information repair algorithm based on attention fusion network to detect and remove abnormal reflection stripe data in the annular contour data, and obtaining the center position of the laser line and repairing the center line in the stripe center extraction stage.

[0011] By adopting the above scheme, the collected annular contour data is effectively processed to remove outlier noise points and abnormal reflection stripes, repair the center line, and improve the quality and accuracy of the annular contour data, thereby providing more reliable data support for subsequent pose calibration.

[0012] Preferably, the coordinate system is one that converts the 3D camera coordinate system and the 2D laser sensor coordinate system into the vehicle coordinate system through a pre-established calibration matrix.

[0013] By adopting the above scheme, the data collected by the 3D camera and 2D laser sensor are converted into the same vehicle coordinate system, which facilitates the subsequent calculation and calibration of the relative pose deviation between the cantilever and the I-beam wheel, and provides a foundation for achieving accurate I-beam wheel loading and unloading pose calibration.

[0014] Preferably, the coordinate system employs a targetless calibration method combining offline calibration and online dynamic tracking to convert the 3D camera coordinate system and the 2D laser sensor coordinate system into a vehicle coordinate system; the offline calibration step includes: The design of the vehicle body running excitation trajectory, and the simultaneous acquisition of three-dimensional point cloud data collected by 3D camera, two-dimensional contour data collected by 2D laser sensor, and inertial data collected by IMU sensor installed on the vehicle body during the execution of the motion excitation trajectory; the excitation running trajectory includes: translational motion, rotational motion, and compound motion; Define calibration parameters and use them as optimization variables for graph optimization; the calibration parameters include: camera-IMU extrinsic parameters, laser-IMU extrinsic parameters, and time offset parameters; The design involves an optimization model for the graph. In this model, the nodes represent calibration parameters, and the edges represent the constraints between data collected by various sensors. The goal is to minimize the weighted sum of squares of all constraint residuals to find the optimal calibration parameters. These constraints include: IMU pre-integration constraints, camera-IMU extrinsic parameter constraints, and laser-IMU extrinsic parameter constraints. IMU pre-integration constraints refer to calculating the pose changes of the AGV at adjacent time points based on the angular velocity and acceleration of the IMU, forming constraints with the pose sequences of the nodes in the graph. Camera-IMU extrinsic parameter constraints refer to using feature points of the cantilever arm collected by the 3D camera, combined with camera-IMU extrinsic parameters, to determine the optimal calibration parameters. The feature points are transformed from the camera coordinate system to the IMU coordinate system, and then to the world coordinate system, forming a constraint with the world coordinates of the cantilever feature points in the graph node; the laser-IMU extrinsic constraint refers to fitting the I-shaped wheel circle based on the contour data acquired by 2D laser, and combining the laser-IMU extrinsic parameters to transform the center of the circle from the laser coordinate system to the IMU coordinate system, and then to the world coordinate system, forming a constraint with the world coordinates of the center of the I-shaped wheel circle in the graph node; the constraint residual refers to: the deviation between the IMU pre-integration result and the pose sequence of the graph node, the deviation from the world coordinates of the cantilever feature points in the graph node, and the deviation from the world coordinates of the center of the inner hole of the I-shaped wheel in the graph node; The optimal calibration parameters are obtained by using a graph optimization model. The optimized extrinsic parameters are saved to the AGV control system as initial extrinsic parameters. The 3D camera coordinate system and 2D laser sensor coordinate system are converted into the vehicle coordinate system.

[0015] By adopting the above scheme, using the targetless calibration method of offline calibration and online dynamic tracking, combined with multi-sensor data obtained from the vehicle's running excitation trajectory, and using a graph optimization model to solve for the optimal calibration parameters, the accuracy of converting the 3D camera coordinate system and the 2D laser sensor coordinate system into the vehicle coordinate system is improved, thereby enhancing the positioning accuracy of the relative pose of the cantilever and the I-beam wheel.

[0016] Preferably, the online dynamic tracking step includes: Using initial extrinsic parameters as initialization input, the extrinsic parameters are updated in real time through an EKF extended Kalman filter; the real-time update step includes: Define state variables: use the offsets of extrinsic parameters as state variables in the EKF; Definition of observations: The offset of external parameters inferred from the world coordinate deviation of the cantilever feature points observed by the 3D camera and the world coordinate deviation of the inner hole center of the I-beam wheel observed by the 2D laser is used as the observation of EKF. Execution of prediction and observation update: During the online operation of the AGV, based on the dynamic characteristics of the slight deviation of the external parameters, the state variables and covariance are predicted; based on the three-dimensional point cloud data collected by the 3D camera and the two-dimensional contour data collected by the 2D laser sensor, the observation residuals are calculated, the Kalman gain is calculated, and the optimal estimates of the state variables and the optimal estimates of the state covariance are updated. Based on the updated optimal estimate of the state variables, the extrinsic parameters at the current moment are corrected, and the corrected extrinsic parameters are output to perform subsequent conversion of the 3D camera coordinate system and the 2D laser sensor coordinate system into the vehicle coordinate system.

[0017] By adopting the above scheme, the external parameters of the 3D camera and 2D laser sensor are updated and corrected in real time using the EKF extended Kalman filter to adapt to the slight deviation of the external parameters during the online operation of the AGV, thereby further improving the accuracy and real-time performance of the coordinate system and thus enhancing the accuracy and reliability of the pose calibration of the cantilever and I-beam wheel.

[0018] Preferably, the steps of executing the binocular pose cooperative matching algorithm further include: Based on the iterative nearest neighbor algorithm, a joint cost function is constructed that includes data point cloud registration terms, circle center alignment constraints, and axis parallelism constraints. The transformation containing relative pose deviation is solved, and the process is iterated until convergence, and the final relative pose deviation is output.

[0019] By adopting the above scheme, a joint cost function is constructed based on the iterative nearest neighbor algorithm, which more accurately solves the transformation containing relative pose deviation. After iterative convergence, the final relative pose deviation is output, further improving the accuracy of the calculation of the relative pose deviation between the cantilever and the I-beam wheel.

[0020] Secondly, this application provides a multi-sensor-based AGV (Automated Guided Vehicle) loading and unloading posture calibration system, comprising: The pose data acquisition module is used to acquire three-dimensional point cloud data of the cantilever arm of the hanger through a 3D camera; and to scan the annular contour data of the inner hole of the I-beam wheel through a 2D laser sensor. The pose processing module is used to preprocess 3D point cloud data and ring contour data and transform them to the same vehicle coordinate system through coordinate system one; The relative pose calculation module is used to execute a dual-target pose collaborative matching algorithm based on the converted data to determine the relative pose deviation between the cantilever arm and the I-beam wheel. The steps of executing the dual-target pose collaborative matching algorithm include: performing feature extraction and feature constraint matching on the converted data, and the constraints include center alignment constraints and axis parallel constraints; calculating the relative pose deviation between the cantilever arm and the I-beam wheel based on the matching features; the relative pose deviation includes the translation compensation amount required by the vehicle body and the angle compensation amount required by the gear. The pose calibration and adjustment module is used to perform dynamic deviation calibration in real time based on relative pose deviation, and adjust the vehicle body position or gear angle. The dynamic deviation real-time calibration steps include: fusing three-dimensional point cloud data and ring contour data in real time through Kalman filter tracking algorithm, dynamically updating the pose parameters of the cantilever and I-beam wheel, and correcting the compensation amount in real time based on the calculated new relative pose deviation.

[0021] By adopting the above scheme, accurate measurement of the three-dimensional pose of the cantilever and the pose of the I-beam wheel in the vehicle body coordinate system is achieved, which solves the problem of relative pose mismatch between the cantilever and the I-beam wheel. The pose parameters are dynamically updated and the compensation amount is corrected in real time, avoiding the lag of static measurement, improving positioning accuracy, and enhancing the stability and reliability of automatic loading and unloading of the I-beam wheel.

[0022] Thirdly, this application provides a computer-readable storage medium including a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to perform the method described above.

[0023] Fourthly, this application provides a computer device, the computer device including a memory, a processor, and a program stored in the memory and executable thereon, the program being executed by the processor to perform the steps of the method described above.

[0024] In summary, this application has the following beneficial effects: 1. The three-dimensional point cloud data of the cantilever of the bracket is collected by a 3D camera and the circular contour data of the inner hole of the I-beam wheel is scanned by a 2D laser sensor. After preprocessing and coordinate system transformation to the same vehicle coordinate system, the dual-target pose collaborative matching algorithm is executed to determine the relative pose deviation. Finally, the deviation is calibrated and adjusted to achieve precise matching of the pose of the cantilever of the bracket and the I-beam wheel, thereby improving the accuracy of loading and unloading. 2. A targetless calibration method combining offline calibration and online dynamic tracking is adopted. By designing the vehicle's running excitation trajectory, defining calibration parameters, and constructing a graph optimization model, the pose and extrinsic parameters are updated in real time using EKF extended Kalman filtering, which accurately converts the coordinate system of the 3D camera and 2D laser sensor into the vehicle coordinate system, thereby improving the accuracy of coordinate transformation. 3. Preprocess the 3D point cloud data and ring contour data, such as adding timestamps, using IMU data compensation, multi-threshold filtering, spatial clustering verification, and repairing point cloud holes, to improve data quality and provide a more accurate data foundation for subsequent pose matching and calibration, thereby improving the reliability of loading and unloading pose calibration. Attached Figure Description

[0025] Figure 1 This is a flowchart of the AGV loading and unloading posture calibration method based on multiple sensors described in a specific embodiment; Figure 2 This is a schematic diagram of the multi-sensor-based AGV loading and unloading posture calibration system described in a specific embodiment. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] like Figure 1 As shown in the figure, this application discloses a multi-sensor-based AGV I-beam wheel loading and unloading posture calibration, including steps such as data acquisition, data preprocessing and coordinate transformation, dual-target posture collaborative matching, real-time dynamic deviation calibration, and adjustment of the vehicle body or gears. By performing these steps, high-precision posture matching between the cantilever and the I-beam wheel is achieved, thereby improving the positioning accuracy and reliability of the AGV's automatic loading and unloading of the I-beam wheel. The following is a further detailed description of this application.

[0028] S1. Collect three-dimensional point cloud data of the cantilever bracket using a 3D camera.

[0029] Specifically, a structured light 3D camera is used, mounted on the upper front of the AGV body, with the lens facing the cantilever arm. The 3D camera acquires 3D point cloud data of the cantilever arm, focusing on extracting the circular contour features of the cantilever arm's holes (such as center coordinates, hole diameter, and spatial normal direction) and the overall 3D pose parameters of the cantilever arm, including pitch and roll angles. It features a high frame rate and point cloud resolution, with a frame rate of no less than 30fps and a high point cloud resolution. With a resolution of 0.1mm, it enables the capture of cantilever structure features under complex lighting conditions in a workshop.

[0030] S2. Scan the annular contour data of the inner hole of the I-beam wheel using a 2D laser sensor.

[0031] Specifically, the 2D laser sensor uses a high-precision line laser sensor, installed on the side of the AGV gear-shaping bracket. The laser beam is parallel to the gear axis and perpendicular to the end face of the I-beam. By scanning the annular contour of the I-beam's inner hole, it obtains the radial offset, axial tilt angle, and inner hole diameter of the I-beam in the vehicle coordinate system, achieving high laser scanning accuracy. 0.05mm, scanning frequency 50Hz, capable of real-time tracking of I-beam wheel attitude changes. Replaceable sensor types include other types of laser rangefinders.

[0032] S3. Preprocess the 3D point cloud data and ring contour data and transform them to the same vehicle coordinate system using coordinate system one.

[0033] Specifically, data preprocessing includes adding timestamps to the data, compensating for data errors, filtering, and contour extraction.

[0034] In this study, considering that different timestamps correspond to different positions during AGV movement, asynchronous data collection can lead to false pose deviations. To ensure the time consistency of heterogeneous data, hardware-level timestamps are added to the data. This is achieved by embedding time stamps during data acquisition, facilitating subsequent data synchronization and processing. Furthermore, interpolation compensation is performed to migrate data from all sensors to the same virtual synchronization moment (e.g., the sampling center moment of a particular sensor), compensating for dynamic target position deviations caused by differences in sampling times. Interpolation compensation includes selecting a reference moment, acquiring the motion trajectory, calculating relative transformation, and applying compensation. Specifically, it utilizes IMU sensors installed on the vehicle body to provide vehicle operating status data, constructs a motion trajectory within a sliding window, and performs interpolation and numerical compensation on the corresponding 3D point cloud data or ring contour data for each frame.

[0035] Taking the compensation of laser point cloud data to the camera exposure center time as an example, the specific steps include: assuming the continuous motion trajectory of the AGV body in the world coordinate system has been obtained, and the extrinsic parameters from the 2D laser sensor to the vehicle coordinate system and the 3D camera to the vehicle coordinate system have been calibrated, the camera exposure center time is selected as a unified reference time; the laser points to be compensated and their measurement times are obtained; the relative motion of the vehicle body between the measurement time and the reference time is calculated; laser point compensation: the original laser point is located in the laser coordinate system, it is first transformed to the vehicle coordinate system at the measurement time, and then transformed to the vehicle coordinate system at the reference time using relative motion transformation. If it is necessary to transform the point to the camera coordinate system, the relative transformation is applied again; the compensated point represents the position of the laser point in the camera coordinate system at the reference time; the above process is repeated for each laser point. If it is necessary to compensate the camera data to the laser time, the camera features are compensated to the laser frame center time, and the process is completely symmetrical.

[0036] Considering that most I-beam wheels are made of metal, laser light often scatters at the rim edge, causing feature point extraction to fail. To avoid noise caused by reflection, multi-threshold filtering and contour extraction are performed on the collected annular contour data. Different filtering algorithms can be used to remove noise, and a specific contour extraction algorithm is used to extract useful contour information. Based on the spatial distribution characteristics of the reflection mechanism, distance-based spatial clustering is performed to screen out and remove outlier noise points. Clustering algorithms can be used to separate outlier points. An attention-based point cloud hole information repair algorithm (e.g., constructing a deep learning-based point cloud repair model) is used to detect and remove abnormal reflective stripe data in the annular contour data. In the stripe center (feature) extraction stage, the center position of the laser line is obtained and the center line is repaired.

[0037] In this embodiment, after the 3D camera acquires the three-dimensional point cloud data of the cantilever bracket, a statistical filtering algorithm is used to remove noise points caused by workshop dust and light interference. Then, the RANSAC circle fitting algorithm is used to extract the circular contour of the cantilever hole from the point cloud data, and the three-dimensional coordinates of the hole in the camera coordinate system are calculated. , , ), aperture and space normal vector A 2D laser sensor scans the inner hole of the I-beam wheel to obtain multiple sets of annular contour point coordinates. The coordinates of the center of the inner hole of the I-beam wheel in the laser coordinate system are obtained by fitting using the least squares method. Aperture and the inclination angle of the inner hole axis .

[0038] Specifically, data processing also includes coordinate transformation. This transformation uses a pre-established calibration matrix to convert the 3D camera coordinate system and the 2D laser sensor coordinate system into the vehicle coordinate system.

[0039] For example, by using a hand-eye calibration algorithm to pre-establish the transformation relationship between the 3D camera coordinate system and the AGV vehicle coordinate system, the calibration matrix Tcam can be obtained. Simultaneously, through sensor installation and calibration, a fixed transformation matrix Tlaser is established between the 2D laser coordinate system and the vehicle coordinate system. The algorithm uses the cantilever hole pose parameters extracted by the 3D camera (car). , , , ) via Tcam Car is converted to its pose in the vehicle coordinate system. , , , This enables the registration of data from both sensors in the same coordinate system, laying the foundation for subsequent pose matching. Simultaneously, by combining the installation and calibration relationship between the laser sensor and the vehicle body, the center coordinates of the I-beam wheel's inner hole in the laser coordinate system are converted into pose parameters in the vehicle body coordinate system. , ).

[0040] S4. Based on the converted data, execute the dual-target pose collaborative matching algorithm to determine the relative pose deviation between the cantilever and the I-beam wheel.

[0041] Specifically, the main steps of the dual-target pose collaborative matching algorithm include: First, feature extraction and feature constraint matching are performed on the transformed data. The feature point matching constraint includes: using the center and normal vector of the cantilever hole as a reference, combined with the center and tilt angle of the inner hole of the I-beam wheel, the spatial distance between the center of the inner hole of the I-beam wheel and the center of the cantilever hole is required. 0.3mm, the angle between the axis of the inner hole of the I-beam wheel and the normal to the cantilever hole position. 0.2°. The specific matching process may include: for the area near the hole in the cantilever point cloud, using RANSAC to fit a plane to obtain the hole end face; then projecting the point cloud onto the end face, fitting a circle to obtain the center and normal vector; for the I-beam point cloud data, using cylindrical fitting (or annular fitting) to obtain the inner hole axis direction and the projection of a point (center) on the axis; matching the center and normal vector, ensuring that the matching meets the constraints, and if not, refitting to obtain the final matched feature points.

[0042] Second, the relative pose deviation between the cantilever and the I-beam wheel is calculated based on matching features. Specifically, this is achieved by establishing a spatial pose deviation function. , , , The required translation compensation for the AGV body is calculated. , , ) and the angle compensation required for the gear cutting This ensures that the poses of the two targets after compensation meet the above constraints. The spatial pose deviation function can be calculated using an averaging algorithm, which obtains multiple matching feature points (center and direction) that meet the constraints and then averages them.

[0043] S5. Perform dynamic deviation calibration in real time based on relative posture deviation, and adjust the vehicle body position or gear angle.

[0044] Specifically, real-time dynamic deviation calibration includes real-time data fusion, updating pose parameters, and correcting compensation amounts.

[0045] The algorithm uses a Kalman filter to fuse 3D point cloud data and ring contour data in real time. The 3D camera collects 30 frames of cantilever pose data per second, and the 2D laser collects 50 frames of I-beam pose data per second. After synchronizing the data in time, the algorithm dynamically updates the pose deviation between the cantilever and the I-beam and corrects the compensation amount in real time according to the new pose deviation.

[0046] Specifically, define the state variables: The real-time pose deviation between the cantilever and the I-beam at time k; the state variable is expressed by the formula: ; Define the observable: The pose deviation observation obtained by the dual-sensor collaborative matching at time k, which has the same dimension as the state variable, is defined as: Perform Kalman filtering prediction-update, using dynamic state variables and covariance, to finally obtain the updated pose deviation.

[0047] Among them, based on the optimal deviation at the previous moment, the deviation at the current moment is predicted. The state prediction formula (predicting the pose deviation at the current moment) is as follows: In the formula, The predicted value of the pose deviation at time k; This is the optimal estimate of the pose deviation at time k-1; The state transition matrix is ​​assumed to have a gradual change in pose deviation between adjacent time steps (with weak dynamic fluctuations). The identity matrix is ​​used to fit the AGV fine-tuning and calibration scenario. The noise is a process noise that follows a Gaussian distribution. , The observation noise covariance matrix; where the covariance prediction formula is: In the formula, Let k be the predicted value of the state covariance at time k; The optimal estimate of the state covariance at time k-1 is obtained; based on the observation bias of the two sensors at time k, the prediction bias is corrected to obtain the optimal pose bias at the current time, including steps such as calculating Kalman gain, state update, and covariance update.

[0048] The AGV's control unit can send real-time calibration commands (translation compensation and angle compensation) to the AGV's motion mechanism and gear-shaping drive mechanism via the EtherCAT protocol. The AGV adjusts its position based on the translation compensation, while the gear-shaping drive mechanism fine-tunes the angle of the I-beams based on the angle compensation. When the position reaches a preset threshold, a "calibration complete" signal is sent, the algorithm stops calibration, and loading / unloading actions are triggered.

[0049] A specific embodiment, differing from the above embodiments, utilizes a targetless calibration method combining offline calibration and online dynamic tracking. This method simultaneously acquires multi-sensor data based on the vehicle's running excitation trajectory, and solves for the optimal calibration parameters using a graph optimization model. This achieves accurate conversion from the 3D camera coordinate system and the 2D laser sensor coordinate system to the vehicle coordinate system, avoiding the limitations of relying on specific targets and improving the flexibility and accuracy of coordinate transformation. The method includes: firstly, using a targetless calibration method combining offline calibration and online dynamic tracking to convert the 3D camera coordinate system and the 2D laser sensor coordinate system into the vehicle coordinate system.

[0050] First, offline calibration is performed. To address issues such as inaccurate initial extrinsic parameter calibration and time synchronization deviations, and to provide foundational extrinsic parameter data for online operation, this process can be repeated periodically to correct long-term drift in the extrinsic parameters. Specific steps include: Specifically, considering the AGV's loading and unloading scenario using I-beam wheels, an excitation trajectory for the vehicle's movement is designed, with pre-calibration preparations made. Equipment status: The AGV is in an unloaded state, and the 3D camera, 2D laser sensor, and IMU are all running normally. Standard posture arrangement: The cantilever arm and I-beam wheels are placed in the standard initial loading and unloading posture as a calibration reference. The excitation trajectory design includes: translational motion (e.g., the AGV moving at a constant speed along the X, Y, and Z axes (forward, lateral, and longitudinal directions), stationary, and moving at a constant speed in the opposite direction), and rotational motion (the AGV rotating clockwise around the Z-axis (perpendicular to the ground). , stationary, counterclockwise rotation ) and compound motion (combining translation and rotation to simulate the actual motion trajectory of the I-beam wheel during loading and unloading); during the process of the AGV executing the motion excitation trajectory, the 3D point cloud data collected by the 3D camera, the 2D contour data collected by the 2D laser sensor, and the inertial data collected by the IMU sensor installed on the vehicle body are acquired simultaneously.

[0051] Define calibration parameters and use them as optimization variables in graph optimization. These calibration parameters include: camera-IMU extrinsic parameters, laser-IMU extrinsic parameters, and time offset parameters; camera-IMU extrinsic parameters: Assume the IMU coordinate system is... The 3D camera coordinate system is The extrinsic parameters include: rotation matrix (Camera rotation relative to IMU) and translation vector (The translation of the camera relative to the IMU), that is This describes the position and attitude relationship between the camera and the IMU; Laser-IMU extrinsic parameters: assuming a 2D laser coordinate system... The extrinsic parameters include: rotation matrix (Rotation and translation vector of the laser relative to the IMU) (The translation of the laser relative to the IMU), that is This describes the position and orientation relationship between the laser and the IMU; time offset parameter: assuming the time offset of the camera relative to the IMU is... (i.e., camera data timestamp = IMU data timestamp + ...) The time offset of the laser relative to the IMU is (That is, laser data timestamp = IMU data timestamp + ...) ).

[0052] The design involves an optimization model for the graph. In this model, the nodes represent calibration parameters (extrinsic parameters and time offset), and the edges represent the constraints between data collected by each sensor. The optimal calibration parameters are found by minimizing the weighted sum of squared residuals from all constraints. These constraints include: IMU pre-integration constraints, camera-IMU extrinsic parameter constraints, and laser-IMU extrinsic parameter constraints.

[0053] Among them, the IMU pre-integration constraint refers to the calculation of the AGV's pose change at adjacent time points based on the angular velocity and acceleration of the IMU, forming a constraint with the pose sequence of the nodes in the graph. The corresponding constraint residual is the deviation between the IMU pre-integration result and the pose sequence of the graph nodes, and the specific formula is as follows: ; ; In the formula, This represents the rotational change of the IMU from time k to time k+1; This represents the change in the IMU's position from time k to time k+1; This represents the change in IMU velocity from time k to time k+1; , This represents the angular velocity and acceleration measurements of the IMU at time t; , This indicates the angular velocity deviation and acceleration deviation of the IMU; , Represents the rotation matrix and velocity of the IMU at time k; The sampling time interval is represented by the constraint parameter formula: Camera-IMU extrinsic constraints refer to the transformation of feature points of the cantilever arm acquired by a 3D camera from the camera coordinate system to the IMU coordinate system, and then to the world coordinate system, based on the camera-IMU extrinsic parameters. This transformation forms a constraint with the world coordinates of the cantilever arm feature points in the graph nodes. The corresponding constraint residual is the deviation from the world coordinates of the cantilever arm feature points in the graph nodes, and the specific formula is as follows: In the formula, The world coordinates of the cantilever feature point in the graph node; The center of the aperture of the cantilever bracket; , Considering the time offset, the rotation matrix and translation vector of the IMU in the world coordinate system at time k; corresponding constraint residual formula: Laser-IMU extrinsic constraint refers to fitting the contour data acquired by 2D laser to obtain the I-shaped wheel circle, and then, using laser-IMU extrinsic parameters, transforming the center of the circle from the laser coordinate system to the IMU coordinate system, and then to the world coordinate system. This forms a constraint with the world coordinates of the I-shaped wheel circle center in the graph node. The corresponding constraint residual is the deviation from the world coordinates of the inner hole center of the I-shaped wheel in the graph node. The specific formula is: Corresponding constraint residual formula: The formula for minimizing the weighted sum of squares of all constrained residuals is: In the formula, Let N represent the set of all optimization variables, and N be the total number of samples. , , This represents the covariance matrix of the residuals for each constraint.

[0054] The optimal calibration parameters are obtained by using a graph optimization model. The optimized extrinsic parameters are saved to the AGV control system as initial extrinsic parameters. The 3D camera coordinate system and 2D laser sensor coordinate system are converted into the vehicle coordinate system.

[0055] Secondly, dynamic tracking is performed. Considering the slight deviation of external parameters caused by factors such as vibration, temperature changes, and sensor loosening during the real-time tracking of the AGV's online operation, the external parameters are dynamically corrected. Moreover, since the deviation of external parameters is slight during online operation (i.e., the deviation between the true value of the external parameters and the optimal value calibrated offline is small), linearization can be used.

[0056] The specific steps include: using initial extrinsic parameters as initialization input, and updating the extrinsic parameters in real time through an Extended Kalman Filter (EKF). The initialization parameters of the online EKF are entirely based on the optimal calibration results from offline calibration. The real-time update step includes: Define state variables: Use the offsets of extrinsic parameters as state variables in the EKF. The formula is: In the formula, : The minute translational deviation of the camera-IMU extrinsic parameters (X, Y, Z axis directions, unit: m) relative to the optimal extrinsic parameters calibrated offline. A slight shift; : The minute rotational deviation of the camera-IMU extrinsic parameters (around the X, Y, and Z axes, in rad) relative to the optimal extrinsic parameters calibrated offline. A slight shift; Small translational deviations of the laser-IMU extrinsic parameters (X, Y, Z axes, unit: m) relative to the optimal extrinsic parameters calibrated offline. A slight shift; Small rotational deviations of the laser-IMU extrinsic parameters (around the X, Y, and Z axes, in rad) relative to the optimal extrinsic parameters calibrated offline. A slight shift.

[0057] To further simplify the method, the extrinsic parameters R and T can be directly incorporated as state variables to be optimized during the real-time calibration process, and optimized online together with the relative pose calibration. Simultaneously, the Kalman filter tracking algorithm used in real-time calibration is replaced with an EFK Kalman filter, thereby achieving simultaneous optimization of extrinsic parameters and pose deviations. The corresponding state variable formula is: In the formula, , , , These represent the translational and rotational pose deviations of the cantilever arm and the I-beam wheel in the X, Y, and Z axes, respectively. Initial values ​​of the state variables. The optimal extrinsic parameters and initial pose deviation are determined by offline calibration.

[0058] Definition of Observation Measurement: The external parameter offsets derived from the world coordinate deviations of the cantilever feature points observed by the 3D camera and the world coordinate deviations of the inner hole center of the I-beam wheel observed by the 2D laser are used as the observation measurement of EKF. Specifically, the formula for the observation value is: In the formula, the observed external parameter offset is calculated by comparing the real-time data acquired by the dual sensors with the offline calibration benchmark. The core logic is the difference between the online measured external parameter and the offline optimal external parameter. The specific steps are: Camera-IMU external parameter measurement: feature point coordinates acquired by the 3D camera. Combined with online EKF real-time pose Reverse engineering online measured camera-IMU external parameters , With offline optimal extrinsic parameters , By comparison, the observed values ​​of the extrinsic parameter offset are obtained: (Rotation offset, (for mapping from antisymmetric matrices to vectors) (Translation and offset); Laser-IMU external measurement: Coordinates of the center of the I-shaped wheel acquired by 2D laser. Combined with online EKF real-time pose Inverse calculation of online measured laser-IMU extrinsic parameters , With offline optimal extrinsic parameters , By comparison, the observed values ​​of the extrinsic parameter offset are obtained: , .

[0059] The real-time calibrated Kalman filter tracking algorithm is replaced with the EFK Kalman filter, thereby optimizing both extrinsic parameters and pose deviations. The corresponding formula is: In the formula, As the observed value of the pose deviation between the cantilever and the I-beam wheel, it is obtained by coordinate registration and pose matching of data acquired by 3D camera and 2D laser. That is, the difference between the world coordinates of the feature point of the cantilever and the world coordinates of the center of the inner hole of the I-beam wheel, which are measured online.

[0060] Execution of prediction and observation update: During the online operation of the AGV, based on the dynamic characteristics of the slight deviation of the external parameters, the state variables and covariance are predicted; based on the three-dimensional point cloud data collected by the 3D camera and the two-dimensional contour data collected by the 2D laser sensor, the observation residuals are calculated, the Kalman gain is calculated, and the optimal estimates of the state variables and the optimal estimates of the state covariance are updated.

[0061] Based on the dynamic characteristics of the slight shift in extrinsic parameters, the state variables and covariance are predicted. The state prediction formula includes: In the formula, Let the state transition matrix be an example of a state transition matrix, assuming "slight shifts and slow changes". The design includes: the transition blocks for extrinsic parameter offsets are converted into identity matrices. (One translation and one rotation offset each) The specific structure is as follows: In the formula, the first four... These correspond to the four components of the extrinsic parameter offset. , , , Substituting into the prediction formula, the prediction process for the extrinsic parameter offset includes: Camera-IMU translation extrinsic parameter offset prediction: Camera-IMU rotation extrinsic parameter offset prediction: Laser-IMU translational extrinsic parameter migration prediction: Laser-IMU rotational extrinsic parameter migration prediction: ; , The process noise is represented by the external parameter offset (following a Gaussian distribution with a mean of 0 and a very small variance). Further consideration can be given to adding relative pose deviations, with the transfer blocks for these deviations designed based on the AGV's motion characteristics.

[0062] Among them, state covariance prediction: The formula for calculating the observation residuals is as follows: In the formula, The core block structure of the observation matrix (the linearized matrix) is as follows (only the key blocks related to extrinsic parameter offset and pose offset are shown; the remaining blocks are zero matrices). Includes external parameter offset observations and pose deviation observations. It includes predicted values ​​of extrinsic parameter offset and pose deviation, and the residual reflects the difference between the predicted value and the observed value.

[0063] Among them, the Kalman gain is calculated as follows: ,in Includes external parameter offset observation noise and pose deviation observation noise, balancing the weights of predicted and observed values. Update state variables: Simultaneously update the optimal estimates of extrinsic parameter offset and pose deviation; update the covariance: This provides a basis for predicting the next moment.

[0064] Finally, based on the updated optimal estimate of the state variables, the extrinsic parameters at the current moment are corrected, and the corrected extrinsic parameters are output to perform the subsequent conversion of the 3D camera coordinate system and the 2D laser sensor coordinate system into the vehicle coordinate system.

[0065] A specific embodiment, differing from the above embodiments, utilizes an iterative nearest neighbor algorithm to construct a joint cost function to solve for the relative pose deviation. This allows for a more accurate determination of the final relative pose deviation between the cantilever and the I-beam wheel, further improving positioning accuracy and enhancing the reliability of automatic loading and unloading of the I-beam wheel. The method includes: executing a binocular pose cooperative matching algorithm. The step further includes: Based on the iterative nearest neighbor algorithm, a joint cost function is constructed that includes data point cloud registration terms, circle center alignment constraints, and axis parallelism constraints. The transformation containing relative pose deviation is solved, and the process is iterated until convergence, and the final relative pose deviation is output.

[0066] Specifically, instead of pre-matching features, the deviation is calculated based on the matched feature points. Direct feature constraints are used as both pre- and embedded constraints in ICP (Integrated Point Registration). Fine registration is performed using all point cloud data, ultimately outputting accurate relative pose deviations and matching point pairs. The steps include: First, perform a coarse matching. Using only the center and axis, ignoring the point cloud, calculate the initial transformation using closed-form solutions or least squares to obtain the initial transformation containing relative pose deviations. The specific formula is: In the formula, It can be decomposed into rotations Peaceful relocation Among them, rotation is the calculation arrive Angle in the horizontal plane Translation refers to the circular shape of the rotated H-shaped wheel. Translation .in, , Let the center and inclination angle of the inner hole of the I-beam bevel bevel bevel. , Using the center of the cantilever hole and the normal vector as references, solve for... .

[0067] Second, in the fine registration stage, at each step of the ICP iteration, the point cloud matching error and feature constraint error are simultaneously optimized to form a joint cost function. The matched point pairs are dynamically updated based on the current transformation, and the final transformation is finally output. This refers to the relative pose that needs to be compensated.

[0068] The joint cost function formula is as follows: In the formula, in each ICP iteration, the matching point pairs It depends on the current T. Matching point pairs are established in the ICP iteration. Let the transformation estimate of the current iteration step be... To establish point cloud matching, for each H-shaped point... Perform the transformation: Search for the nearest point: Find the nearest point in the cantilever point cloud. European distance to the nearest point At the same time, record the normal vector of that point. ;like If the match is not found, then the matching pairs are rejected and the set {( )}.

[0069] Assuming the matching point pairs are known, and the circular and axis features are also known, the algorithm flow for solving the problem with T as a variable includes: Let... To initialize the transformation values, iterate k=0,1,2,... until convergence; establish the matching, i.e., as described above for each I-beam point. Calculate the transformed Search for the nearest point; solve for the increment. The formula is: In the formula, yes Corresponding Lie algebra; updated to obtain The iteration ends, and the final transformation is output. This refers to the amount of compensation that the vehicle body needs to perform.

[0070] like Figure 2 As shown in the figure, this application discloses an AGV loading and unloading posture calibration system based on multiple sensors, specifically including: The pose data acquisition module 100 is used to acquire three-dimensional point cloud data of the cantilever arm of the hanger through a 3D camera; and to scan the annular contour data of the inner hole of the I-beam wheel through a 2D laser sensor. The pose processing module 200 is used to preprocess the 3D point cloud data and the ring contour data and transform them to the same vehicle coordinate system through coordinate system one; The relative pose calculation module 300 is used to execute a dual-target pose collaborative matching algorithm based on the converted data to determine the relative pose deviation between the cantilever arm and the I-beam wheel. The steps of executing the dual-target pose collaborative matching algorithm include: performing feature extraction and feature constraint matching on the converted data, and the constraints include center alignment constraints and axis parallel constraints; calculating the relative pose deviation between the cantilever arm and the I-beam wheel based on the matching features; the relative pose deviation includes the translation compensation amount required by the vehicle body and the angle compensation amount required by the gear. The pose calibration and adjustment module 400 is used to perform dynamic deviation real-time calibration based on relative pose deviation, and adjust the vehicle body position or gear angle. The dynamic deviation real-time calibration steps include: fusing three-dimensional point cloud data and ring contour data in real time through Kalman filter tracking algorithm, dynamically updating the pose parameters of the cantilever and I-beam wheel, and correcting the compensation amount in real time based on the calculated new relative pose deviation.

[0071] This application also discloses a computer-readable storage medium.

[0072] Specifically, the computer-readable storage medium stores a computer program that can be loaded and executed by a processor, such as the above-described multi-sensor-based AGV wheel loading and unloading posture calibration method. The computer-readable storage medium includes, for example, 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.

[0073] This application also discloses a computer device.

[0074] Specifically, the computer device includes a memory and a processor. The memory stores a computer program that can be loaded and executed by the processor to perform the above-mentioned multi-sensor-based AGV loading and unloading posture calibration method.

[0075] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for calibrating the loading and unloading posture of an AGV (Automated Guided Vehicle) based on multiple sensors, characterized in that, Includes the following steps: The 3D point cloud data of the cantilever arm of the hanger is acquired using a 3D camera; Scan the annular contour data of the inner hole of the I-beam wheel using a 2D laser sensor; The 3D point cloud data and the ring contour data are preprocessed and transformed to the same vehicle coordinate system using coordinate system one; Based on the transformed data, a dual-target pose collaborative matching algorithm is executed to determine the relative pose deviation between the cantilever and the I-beam wheel; The steps of executing the binocular pose cooperative matching algorithm include: performing feature extraction and feature constraint matching on the converted data, and the constraints include center alignment constraints and axis parallel constraints; calculating the relative pose deviation between the cantilever and the I-beam wheel based on the matching features; the relative pose deviation includes the translation compensation amount required by the vehicle body and the angle compensation amount required by the gear. The dynamic deviation is calibrated in real time based on the relative pose deviation, and the vehicle body position or gear angle is adjusted. The dynamic deviation real-time calibration steps include: fusing three-dimensional point cloud data and ring contour data in real time through Kalman filter tracking algorithm, dynamically updating the pose parameters of the cantilever and the I-beam wheel, and correcting the compensation amount in real time based on the calculated new relative pose deviation.

2. The AGV loading and unloading posture calibration method based on multiple sensors according to claim 1, characterized in that, The preprocessing steps for 3D point cloud data and ring contour data include: adding hardware-level timestamps to the 3D point cloud data and ring contour data; using IMU sensors installed on the vehicle body to provide vehicle operating status data, constructing motion trajectories within a sliding window, and interpolating and numerically compensating the corresponding 3D point cloud data or ring contour data for each frame.

3. The AGV loading and unloading posture calibration method based on multiple sensors according to claim 1, characterized in that, The preprocessing steps for the annular contour data also include: performing multi-threshold filtering and contour extraction on the collected annular contour data; performing distance-based spatial clustering tests based on the spatial distribution characteristics of the reflection mechanism to screen out and remove outlier noise points; using a point cloud hole information repair algorithm based on attention fusion network to detect and remove abnormal reflection stripe data in the annular contour data, and obtaining the center position of the laser line and repairing the center line in the stripe center extraction stage.

4. The AGV loading and unloading posture calibration method based on multiple sensors according to claim 1, characterized in that, The coordinate system is one that converts the 3D camera coordinate system and the 2D laser sensor coordinate system into the vehicle coordinate system through a pre-established calibration matrix.

5. The AGV loading and unloading posture calibration method based on multiple sensors according to claim 1, characterized in that, The coordinate system employs a targetless calibration method combining offline calibration and online dynamic tracking to transform the 3D camera coordinate system and the 2D laser sensor coordinate system into a vehicle coordinate system. The offline calibration steps include: The design of the vehicle body running excitation trajectory, and the simultaneous acquisition of three-dimensional point cloud data collected by 3D camera, two-dimensional contour data collected by 2D laser sensor, and inertial data collected by IMU sensor installed on the vehicle body during the execution of the motion excitation trajectory; the excitation running trajectory includes: translational motion, rotational motion, and compound motion; Define calibration parameters and use them as optimization variables for graph optimization; the calibration parameters include: camera-IMU extrinsic parameters, laser-IMU extrinsic parameters, and time offset parameters; The design involves an optimization model for the graph. In this model, the nodes represent calibration parameters, and the edges represent the constraints between data collected by various sensors. The goal is to minimize the weighted sum of squares of all constraint residuals to find the optimal calibration parameters. These constraints include: IMU pre-integration constraints, camera-IMU extrinsic parameter constraints, and laser-IMU extrinsic parameter constraints. IMU pre-integration constraints refer to calculating the pose changes of the AGV at adjacent time points based on the angular velocity and acceleration of the IMU, forming constraints with the pose sequences of the nodes in the graph. Camera-IMU extrinsic parameter constraints refer to using feature points of the cantilever arm collected by the 3D camera, combined with camera-IMU extrinsic parameters, to determine the optimal calibration parameters. The feature points are transformed from the camera coordinate system to the IMU coordinate system, and then to the world coordinate system, forming a constraint with the world coordinates of the cantilever feature points in the graph node; the laser-IMU extrinsic constraint refers to fitting the I-shaped wheel circle based on the contour data acquired by 2D laser, and combining the laser-IMU extrinsic parameters to transform the center of the circle from the laser coordinate system to the IMU coordinate system, and then to the world coordinate system, forming a constraint with the world coordinates of the center of the I-shaped wheel circle in the graph node; the constraint residual refers to: the deviation between the IMU pre-integration result and the pose sequence of the graph node, the deviation from the world coordinates of the cantilever feature points in the graph node, and the deviation from the world coordinates of the center of the inner hole of the I-shaped wheel in the graph node; The optimal calibration parameters are obtained by using a graph optimization model. The optimized extrinsic parameters are saved to the AGV control system as initial extrinsic parameters. The 3D camera coordinate system and 2D laser sensor coordinate system are converted into the vehicle coordinate system.

6. The AGV loading and unloading posture calibration method based on multiple sensors according to claim 5, characterized in that, The online dynamic tracking steps include: Using initial extrinsic parameters as initialization input, the extrinsic parameters are updated in real time through an EKF extended Kalman filter; the real-time update step includes: Define state variables: use the offsets of extrinsic parameters as state variables in the EKF; Definition of observations: The offset of external parameters inferred from the world coordinate deviation of the cantilever feature points observed by the 3D camera and the world coordinate deviation of the inner hole center of the I-beam wheel observed by the 2D laser is used as the observation of EKF. Execution of prediction and observation update: During the online operation of the AGV, based on the dynamic characteristics of the slight deviation of the external parameters, the state variables and covariance are predicted; based on the three-dimensional point cloud data collected by the 3D camera and the two-dimensional contour data collected by the 2D laser sensor, the observation residuals are calculated, the Kalman gain is calculated, and the optimal estimates of the state variables and the optimal estimates of the state covariance are updated. Based on the updated optimal estimate of the state variables, the extrinsic parameters at the current moment are corrected, and the corrected extrinsic parameters are output to perform subsequent conversion of the 3D camera coordinate system and the 2D laser sensor coordinate system into the vehicle coordinate system.

7. The AGV loading and unloading posture calibration method based on multiple sensors according to claim 1, characterized in that, The steps for executing the binocular pose cooperative matching algorithm also include: Based on the iterative nearest neighbor algorithm, a joint cost function is constructed that includes data point cloud registration terms, circle center alignment constraints, and axis parallelism constraints. The transformation containing relative pose deviation is solved, and the process is iterated until convergence, and the final relative pose deviation is output.

8. A multi-sensor-based AGV (Automated Guided Vehicle) loading and unloading posture calibration system, characterized in that, include: The pose data acquisition module is used to acquire three-dimensional point cloud data of the cantilever arm of the hanger through a 3D camera; and to scan the annular contour data of the inner hole of the I-beam wheel through a 2D laser sensor. The pose processing module is used to preprocess 3D point cloud data and ring contour data and transform them to the same vehicle coordinate system through coordinate system one; The relative pose calculation module is used to execute a dual-target pose collaborative matching algorithm based on the converted data to determine the relative pose deviation between the cantilever and the I-beam wheel; The steps of executing the binocular pose cooperative matching algorithm include: performing feature extraction and feature constraint matching on the converted data, and the constraints include center alignment constraints and axis parallel constraints; calculating the relative pose deviation between the cantilever and the I-beam wheel based on the matching features; the relative pose deviation includes the translation compensation amount required by the vehicle body and the angle compensation amount required by the gear. The pose calibration and adjustment module is used to perform dynamic deviation calibration in real time based on relative pose deviation, and adjust the vehicle body position or gear angle. The dynamic deviation real-time calibration steps include: fusing three-dimensional point cloud data and ring contour data in real time through Kalman filter tracking algorithm, dynamically updating the pose parameters of the cantilever and I-beam wheel, and correcting the compensation amount in real time based on the calculated new relative pose deviation.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the method as described in any one of claims 1 to 7.

10. A computer device, characterized in that, The computer device includes a memory, a processor, and a program stored in and executable on the memory, the program being executed by the processor to implement the steps of the method as described in any one of claims 1 to 7.

Images