A lightweight container correction method and system based on machine vision

Abstract

The present invention relates to the field of image processing and discloses a lightweight container correction method and system based on machine vision, comprising: constructing a detection model to detect corner pieces by enhancing the characteristic response of the corner piece area, calculating the disparity map of the multispectral image, reconstructing the three-dimensional coordinates of the corner pieces and performing posture conversion; online identifying time-varying parameters to construct a state transfer matrix, calculating the predicted vibration offset, projecting the vibration displacement to the world coordinate system, and calculating the innovation sequence variance to dynamically adjust the noise parameter; constructing an inner and outer loop control architecture, wherein the outer loop vision controller generates a target force by calculating the posture error in the world coordinate system; an inner loop force control compensator combines the predicted vibration offset with force sensor feedback, corrects the robot arm control instruction, schedules the control gain through the task stage, and analyzes the vibration frequency to adaptively adjust the parameters; performs force control compensation based on the target force, discretizes the continuous impedance model to obtain a recursive control law to resist external disturbances.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a lightweight container correction method and system based on machine vision. Background Art

[0002] The development of automated container loading and unloading systems has evolved through manual operation, semi-automation, and finally full automation. The current mainstream technology relies on vision-guided robotic arms to grasp corner pieces. Traditional solutions are primarily based on single-modal or binocular vision, combined with classic control algorithms for pose detection and grasping control. With the increase in industrial assembly line speeds, increasing environmental complexity, and vibration interference caused by high-speed robotic arm movement, existing technologies face significant challenges in detection robustness, dynamic adaptability, and control accuracy.

[0003] Existing technologies lack environmental adaptability for single-modal vision. Traditional target detection algorithms have weak small target detection capabilities, poor rotational invariance, and weak anti-fouling capabilities. Stereo vision registration and calibration errors are large. Existing state-space models mostly use fixed parameters and cannot adapt to load changes when the robot arm grasps containers of different weights. Traditional visual servo control mostly uses open-loop or single closed-loop control without force feedback, which cannot resist external disturbances. Gain parameters are fixed and cannot be adjusted according to the task stage. Direct discretization of continuous impedance models has high computational complexity, making it difficult to achieve high update frequencies in embedded platforms, resulting in force control delays and the inability to track high-frequency vibrations.

[0004] In view of this, it is necessary to provide a lightweight container correction method and system based on machine vision. Summary of the Invention

[0005] The purpose of the present invention is to provide a lightweight container correction method and system based on machine vision. To solve the above-mentioned existing technical problems, the present invention is achieved through the following technical solutions:

[0006] In a first aspect, an embodiment of the present invention provides a lightweight container correction method based on machine vision, which specifically includes the following steps:

[0007] The container image is collected and preprocessed in a dynamic environment using multispectral images. A detection model is constructed by enhancing the feature response of the corner parts area to detect corner parts. The disparity map of the multispectral image is calculated, and the 3D coordinates of the corner parts are reconstructed and the pose transformation is performed.

[0008] The state space model of the corner piece is constructed based on its three-dimensional coordinates. The time-varying parameters are identified online to construct the state transfer matrix. The vibration offset is calculated and predicted, the vibration displacement is projected into the world coordinate system, and the innovation sequence variance is calculated to dynamically adjust the noise parameters.

[0009] Construct an inner- and outer-loop control architecture. The outer-loop vision controller generates the target force by calculating the pose error in the world coordinate system. The inner-loop force control compensator combines the predicted vibration offset with the force sensor feedback to correct the manipulator control instructions, schedule the control gain through the task phase, and analyze the vibration frequency to adaptively adjust the parameters.

[0010] Force control compensation is performed based on the target force, and the continuous impedance model is discretized to obtain a recursive control law to resist external disturbances.

[0011] In a second aspect, an embodiment of the present invention provides a lightweight container correction system based on machine vision, which specifically includes the following modules:

[0012] Visual perception module: This module collects and preprocesses multispectral images of containers in dynamic environments. It then builds a detection model to detect corner parts by enhancing the feature response of the corner parts area. It then calculates the disparity map of the multispectral images, reconstructs the 3D coordinates of the corner parts, and performs pose transformation.

[0013] State Estimation Module: Builds a state space model of the corner piece based on its three-dimensional coordinates, identifies time-varying parameters online to build a state transfer matrix, calculates and predicts vibration offsets, projects the vibration displacements to the world coordinate system, and calculates the innovation sequence variance to dynamically adjust noise parameters.

[0014] Control strategy module: This module constructs an inner- and outer-loop control architecture. The outer-loop vision controller generates the target force by calculating the pose error in the world coordinate system. The inner-loop force control compensator combines the predicted vibration offset with the force sensor feedback to correct the manipulator control instructions, schedules the control gain through the task phase, and analyzes the vibration frequency to adaptively adjust the parameters.

[0015] Safety execution module: performs force control compensation based on the target force, discretizes the continuous impedance model to obtain a recursive control law to resist external disturbances.

[0016] Beneficial effects of the present invention:

[0017] 1. A co-optical axis design for visible light and short-wave infrared cameras is achieved through a beam splitter prism, ensuring that the optical centers coincide with the field of view, eliminating image registration errors at the hardware level. Compared with traditional dual-camera solutions, the spatial consistency of stereo vision is improved. Channel and spatial attention mechanisms are inserted, and the spectral characteristics and geometric contour response of metal corner parts are enhanced through global / local pooling. Deeply separable convolution is combined to reduce the number of parameters and adapt to embedded computing platforms. The combined optimization of CIoU and FocalLoss is introduced to improve the detection accuracy of small and rotating targets. The six-axis IMU and three-axis force sensor are physically co-located on a customized PCB to eliminate vibration measurement deviations caused by the lever arm effect. Sub-millisecond time alignment of visual-inertial data is achieved to resolve state estimation errors caused by asynchronous sampling of multiple sensors. The state space model parameters are dynamically updated based on the recursive least squares method to adapt to load changes and environmental vibrations during the robotic arm grasping process. The ARM NEON instruction set is used to optimize matrix operations to meet real-time requirements.

[0018] 2. The outer loop generates target force based on visual posture error, and the inner loop combines vibration prediction and force feedback to achieve dynamic compensation. Compared with single visual control, it improves the force control response speed and resists sudden external forces. In the approach phase, the visual control gain is increased to quickly converge the posture. In the contact phase, it switches to large force feedback gain to avoid overshoot. A smooth transition of the control strategy is achieved through a finite state machine, and the IMU spectrum is analyzed in real time. The differential gain is increased to suppress jitter during high-frequency vibration, and the proportional gain is enhanced to improve tracking accuracy during low-frequency vibration. It can adapt to different working conditions without pre-setting a vibration model. The control cycle is monitored by a watchdog timer at the hardware level, and a force sensor over-range interrupt is set at the software level. Combined with state limiting to prevent numerical divergence, a full-link fault protection system is built. ARM is responsible for strategy planning and parameter configuration, and FPGA calculates joint torque in real time and generates PWM signals. The control delay is adjusted through heterogeneous computing architecture to adapt to the rhythm of high-speed industrial assembly lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0020] Figure 1 This is a flowchart of a method and system for correcting a lightweight container based on machine vision provided in Example 1 of the present invention;

[0021] Figure 2 This is a structural diagram of a lightweight container correction method and system based on machine vision provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0022] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of the present invention.

[0023] Example 1: Figure 1 As shown, an embodiment of the present invention provides a lightweight container correction method based on machine vision, which specifically includes the following steps:

[0024] Step 1: Collect and preprocess multispectral images of the container in a dynamic environment. Build a detection model to detect corner parts by enhancing the feature response of the corner parts area. Calculate the disparity map of the multispectral image, reconstruct the 3D coordinates of the corner parts, and perform pose transformation.

[0025] In a specific embodiment, a visible light camera and a short-wave infrared camera are used, and a beam splitter prism is used to achieve co-axial integration. The beam splitter prism separates the incident light by wavelength, reflecting the visible light band to the visible light camera and transmitting the short-wave infrared band to the infrared camera, ensuring that the optical centers of the two cameras coincide and the fields of view are consistent, thus avoiding errors in subsequent image registration.

[0026] The dual cameras are controlled to expose simultaneously through hardware synchronization signal lines, and the sampling frequency matches the industrial production line rhythm to ensure that images in motion scenes are not time-shifted.

[0027] Aiming at the fixed pattern noise of short-wave infrared camera, a two-point correction method is used to obtain the corrected pixel value;

[0028] Combining the color vision Retinex theory, which decomposes the image into reflected light and illumination components, with local histogram equalization, the contrast of container corner fittings under complex lighting conditions is enhanced.

[0029] It should be noted that the Retinex theory of color vision explains how the human visual system perceives the true color and brightness of objects, which is not affected by changes in ambient lighting;

[0030] Specifically, multi-scale Retinex decomposition is performed on visible light images to suppress the influence of uneven illumination;

[0031] The infrared image is divided into 8×8 pixel blocks, the local histogram is calculated and stretched to [0, 255] to highlight the edge features of metal corners;

[0032] The bimodal image is filtered using a 3×3 Gaussian kernel with σ=1.5 to suppress high-frequency noise and retain the low-frequency features of the corner piece contours;

[0033] Build a detection model based on the YOLO-Robust model to insert channel and spatial attention;

[0034] It should be noted that the YOLO-Robust model is an improved framework that addresses the performance degradation of the YOLO series object detection models in complex environments. It improves detection accuracy and stability by enhancing the model's robustness to lighting changes, occlusions, small objects, and adversarial attack scenarios.

[0035] Global average pooling and maximum pooling are performed on the feature map, and the channel weights are generated by MLP to emphasize the spectral characteristics of metal corners.

[0036] The feature map is max-pooled and average-pooled along the channel dimension, and convolution is performed to generate spatial weights after splicing, focusing on the rectangular contour area of ​​the corner piece;

[0037] Depthwise separable convolution replacement: splits the standard convolution into depthwise convolution and pointwise convolution, reducing the number of parameters while maintaining feature extraction capabilities;

[0038] Add rust textures, scratches, and oil stains to clean container images to generate degraded samples and obtain synthetic contamination data;

[0039] Collect container images under different lighting and angles, annotate the bounding boxes of six corner fittings, including four top corner fittings and two bottom corner fittings, and obtain real scene data;

[0040] Cluster the annotation box sizes through K-means clustering to generate anchor box priors suitable for corner parts;

[0041] Combining CIoU loss and FocalLoss to construct the loss function, through the formula:

[0042] ;

[0043] Get the loss function ,in, is the balance parameter, is the predicted probability, is the focusing parameter, is the cross entropy loss, is the fusion weight, To consider the intersection-over-union ratio of the bounding box aspect ratio, is the intersection-over-union ratio without considering the aspect ratio;

[0044] The model outputs a five-tuple for each corner piece: the center pixel coordinates , bounding box size , rotation angle , confidence ; The rotation angle is obtained by regressing the aspect ratio and the long side direction of the bounding box;

[0045] Take 20 sets of checkerboard images at different angles and extract corner points;

[0046] Solve the camera intrinsic parameter matrix , the distortion coefficient is:

[0047] ;

[0048] Perform stereo calibration on the dual cameras to obtain the baseline distance B, that is, the horizontal distance between the optical centers of the two cameras, and calculate the three-dimensional coordinates of the calibration plate;

[0049] For the corrected left and right images, a block matching algorithm is used to set the window size and disparity range, using the formula:

[0050] ;

[0051] Calculate the disparity map ;

[0052] in, To match windows, pixel similarity is measured by the sum of absolute differences;

[0053] Center of each corner piece , according to parallax Calculating Depth With three-dimensional coordinates ;

[0054] Through quaternions Avoid Euler angle gimbal lock;

[0055] The rotation matrix R and translation vector T from the world coordinate system to the camera coordinate system are solved by the least squares method to satisfy:

[0056] ;

[0057] in, is the coordinate of the container corner fitting in the world coordinate system, with the center of the container bottom as the origin and the long side as the X-axis;

[0058] Finally solve the six-degree-of-freedom pose of the container: three-dimensional translation and the quaternion rotation q, used for path planning during grasping control;

[0059] Step 2: Construct a state space model of the corner piece based on its 3D coordinates, identify the time-varying parameters online to construct a state transfer matrix, calculate and predict the vibration offset, project the vibration displacement to the world coordinate system, and calculate the innovation sequence variance to dynamically adjust the noise parameters;

[0060] In a specific embodiment, a six-axis IMU and a three-axis force sensor are integrated on the back of the end effector of the robotic arm, and physically co-point mounting is achieved through a customized PCB to reduce vibration artifacts caused by the measurement lever arm effect;

[0061] The IMU and force sensor are connected to the embedded controller via the SPI / I2C bus;

[0062] The controller has a built-in constant temperature crystal oscillator as the clock source to ensure stable sampling frequency;

[0063] The dual-camera module and controller achieve sub-millisecond clock synchronization through the IEEE1588PTP protocol;

[0064] Use the PPS signal as the global synchronization reference to generate multiple synchronous trigger pulses through FPGA:

[0065] The rising edge of PPS triggers the visual camera exposure; after a fixed delay, the IMU and force sensor sampling is triggered to compensate for the camera shutter delay;

[0066] Add hardware timestamps with microsecond accuracy to raw sensor data;

[0067] The IMU data is resampled to the visual frame rate using a cubic spline interpolation algorithm, and the time alignment error is controlled within the preset standard value.

[0068] Constructing a state space model by discretizing the dynamic equations: ; converted to discrete state space, where is the second-order differential, is the first-order differential, is the stiffness coefficient, is the external force, It is an external disturbance;

[0069] Set the system parameters Mapping to a regression model:

[0070] ;

[0071] in, for , for ;

[0072] initialization: ;

[0073] Iterative updates:

[0074] ;

[0075] in, is the forgetting factor, balancing tracking speed and noise robustness, is the second-order differential, is the first-order differential, is the state vector, is the gain matrix, is the model residual, To update the parameters, is the regression vector, is the regression vector matrix;

[0076] Using the ARMNEON instruction set and Vectorize the 3×3 matrix operation;

[0077] Physical constraints are imposed on the estimated parameters, and projection corrections are performed when they exceed the range;

[0078] System parameters based on identification Construct the state transition matrix:

[0079] ;

[0080] in, Using the Pade approximation, is the state vector at time t, is the discretization interval, is the system matrix, is the input matrix, is the input signal;

[0081] Covariance prediction: ;

[0082] Among them, the process noise covariance ,Will Divide into 5 sub-steps, iteratively perform state prediction, combine the angular rate and acceleration data of the IMU, and project the vibration offset into the world coordinate system through coordinate transformation:

[0083] ;

[0084] in, is the rotation matrix calculated by the IMU quaternion;

[0085] Real-time computing innovation sequence The variance of , dynamically adjusts the measurement noise covariance, through the formula:

[0086] Calculate the noise covariance ;

[0087] in, is the preset adjustment factor, The default value is 0.95;

[0088] Step 3: Build an inner- and outer-loop control architecture. The outer-loop vision controller generates the target force by calculating the pose error in the world coordinate system. The inner-loop force control compensator combines the predicted vibration offset with the force sensor feedback to correct the manipulator control instructions, schedules the control gain through the task phase, and analyzes the vibration frequency to adaptively adjust the parameters.

[0089] In the world coordinate system, define the six-dimensional pose error vector e;

[0090] The visual inspection corner piece pose is converted from the camera coordinate system to the robot arm base coordinate system, and the hand-eye calibration matrix is ​​used accomplish;

[0091] Calculate target force using proportional-derivative control : ;

[0092] in, is the proportional gain matrix, is the differential gain matrix, is the error differential, , To control the cycle;

[0093] According to the predicted vibration deviation , calculate the compensation force :

[0094] ;

[0095] in, To compensate for the proportional gain, To compensate for the differential gain, is the offset differential;

[0096] Introducing force sensor data , and modify the control instructions through the impedance control principle:

[0097] ;

[0098] Gain ultimate control ,in is the force feedback gain, is the expected contact force;

[0099] Gain scheduling based on task phase, approach phase: increase visual control gain , speed up the response and reduce the force feedback gain , reduce the influence of force control;

[0100] Contact phase: Reduce visual control gain , to avoid overshoot, increase , enhanced force feedback correction;

[0101] Adaptive adjustment of vibration frequency:

[0102] Real-time analysis of the spectrum of IMU data, when the dominant frequency is detected When, if If the vibration is higher than 10Hz, the compensation differential gain should be increased. ;like If the vibration frequency is less than 5Hz, the compensation proportional gain should be increased. ;

[0103] Step 4: Perform force control compensation based on the target force, discretize the continuous impedance model to obtain a recursive control law to resist external disturbances;

[0104] The interaction between the end of the robotic arm and the environment is equivalent to a spring-mass-damper system:

[0105] ;

[0106] in, is the expected trajectory, is the external force;

[0107] For continuous systems, the sampling period Discretize it and get the difference equation:

[0108] ;

[0109] The recursive formula is: ;

[0110] Inverse the matrix Using Cholesky decomposition preprocessing, The complexity is reduced to ;

[0111] Introducing a state limiting mechanism to prevent numerical divergence caused by abnormal force input;

[0112] The force control compensation and execution process is as follows:

[0113] ARM processor configures FPGA parameters: 、 、 、 ;

[0114] Perform hand-eye calibration to obtain the camera and robotic arm base coordinate system transformation matrix ;

[0115] FPGA real-time calculation of joint torque instructions , converted into joint angles through inverse kinematics;

[0116] The PWM module outputs the motor control signal to drive the robotic arm to track the desired trajectory;

[0117] When the vibration frequency changes, ARM dynamically adjusts parameter;

[0118] When the contact force suddenly changes, the impedance parameters are quickly switched to increase To resist shock, the PGA has a built-in watchdog timer, which automatically triggers safety braking if the control cycle times out;

[0119] When the force sensor data exceeds the range, the PL-PS interrupt notifies the ARM to perform an emergency stop.

[0120] Example 2: Figure 2 As shown, an embodiment of the present invention provides a lightweight container correction system based on machine vision, which specifically includes the following modules:

[0121] Visual perception module: This module collects and preprocesses multispectral images of containers in dynamic environments. It then builds a detection model to detect corner parts by enhancing the feature response of the corner parts area. It then calculates the disparity map of the multispectral images, reconstructs the 3D coordinates of the corner parts, and performs pose transformation.

[0122] State Estimation Module: Builds a state space model of the corner piece based on its three-dimensional coordinates, identifies time-varying parameters online to build a state transfer matrix, calculates and predicts vibration offsets, projects the vibration displacements to the world coordinate system, and calculates the innovation sequence variance to dynamically adjust noise parameters.

[0123] Control strategy module: This module constructs an inner- and outer-loop control architecture. The outer-loop vision controller generates the target force by calculating the pose error in the world coordinate system. The inner-loop force control compensator combines the predicted vibration offset with the force sensor feedback to correct the manipulator control instructions, schedules the control gain through the task phase, and analyzes the vibration frequency to adaptively adjust the parameters.

[0124] Safety execution module: performs force control compensation based on the target force, discretizes the continuous impedance model to obtain a recursive control law to resist external disturbances.

[0125] An embodiment of the present invention is described in detail above, but the content described is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention; the above formulas are all dimensionless and numerical calculations, and the formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field based on actual conditions and historical experience, and can be adjusted according to actual conditions; the above description is only a preferred embodiment of the present invention and is not used to limit the present invention. All equal changes and improvements made according to the scope of application of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A lightweight container correction method based on machine vision, characterized in that: The following steps are involved: The container image is collected and preprocessed in a dynamic environment using multispectral images. A detection model is constructed by enhancing the feature response of the corner parts area to detect corner parts. The disparity map of the multispectral image is calculated, and the 3D coordinates of the corner parts are reconstructed and the pose transformation is performed. The method for constructing the detection model is: Cluster the annotation box sizes through K-means clustering to generate anchor box priors suitable for corner parts; Combining CIoU loss and FocalLoss to construct the loss function, through the formula: ； Get the loss function ,in, is the balance parameter, is the predicted probability, is the focusing parameter, is the cross entropy loss, is the fusion weight, To consider the intersection-over-union ratio of the bounding box aspect ratio, is the intersection-over-union ratio without considering the aspect ratio; The state space model of the corner piece is constructed based on its three-dimensional coordinates. The time-varying parameters are identified online to construct the state transfer matrix. The vibration offset is calculated and predicted, the vibration displacement is projected into the world coordinate system, and the innovation sequence variance is calculated to dynamically adjust the noise parameters. Construct an inner- and outer-loop control architecture. The outer-loop vision controller generates the target force by calculating the pose error in the world coordinate system. The inner-loop force control compensator combines the predicted vibration offset with the force sensor feedback to correct the manipulator control instructions, schedule the control gain through the task phase, and analyze the vibration frequency to adaptively adjust the parameters. Force control compensation is performed based on the target force, and the continuous impedance model is discretized to obtain a recursive control law to resist external disturbances.

2. The method for correcting lightweight containers based on machine vision according to claim 1, characterized in that: The method for reconstructing the three-dimensional coordinates of the corner piece is: Perform stereo calibration on the dual cameras to obtain the baseline distance B, that is, the horizontal distance between the optical centers of the two cameras, and calculate the three-dimensional coordinates of the calibration plate; For the corrected left and right images, a block matching algorithm is used to set the window size and disparity range, and the disparity map is calculated by the formula ; in, To match windows, pixel similarity is measured by the sum of absolute differences; Center of each corner piece , according to parallax Calculating Depth With three-dimensional coordinates ; Through quaternions Avoid Euler angle gimbals; The rotation matrix R and translation vector T from the world coordinate system to the camera coordinate system are solved by the least squares method to satisfy: ； in, is the coordinate of the container corner fitting in the world coordinate system, with the center of the container bottom as the origin and the long side as the X-axis; Finally solve the six-degree-of-freedom pose of the container: three-dimensional translation The quaternion rotation q is used for path planning during grasping control.

3. The method for correcting lightweight containers based on machine vision according to claim 1, characterized in that: The method for performing posture conversion is: The IMU data is resampled to the visual frame rate using a cubic spline interpolation algorithm, and the time alignment error is controlled within the preset standard value. Constructing a state space model by discretizing the dynamic equations: Converted to discrete state space, where is the second-order differential, is the first-order differential, is the stiffness coefficient, is the external force, It is an external disturbance; Set the system parameters Mapping to a regression model: ； in, for , for , is the model residual.

4. The method for correcting lightweight containers based on machine vision according to claim 1, characterized in that: The method for reconstructing the three-dimensional coordinates of the corner piece is: ； ； in, is the forgetting factor, balancing tracking speed and noise robustness, is the second-order differential, is the first-order differential, is the state vector, is the gain matrix, is the model residual, To update the parameters, is the regression vector, is the regression vector matrix; Using the ARMNEON instruction set and Vectorize the 3×3 matrix operation; Physical constraints are imposed on the estimated parameters, and projection corrections are performed when they are out of range.

5. The method for correcting lightweight containers based on machine vision according to claim 1, characterized in that: The method for constructing the state space model of the corner piece is: System parameters based on identification Construct the state transition matrix: ； in, Using the Pade approximation, is the state vector at time t, is the discretization interval, is the system matrix, is the input matrix, is the input signal; Covariance prediction: ; Among them, the process noise covariance ; Will Divide into 5 sub-steps, iteratively perform state prediction, combine the angular rate and acceleration data of the IMU, and project the vibration offset into the world coordinate system through coordinate transformation: ； in, is the rotation matrix calculated by the IMU quaternion; Real-time computing innovation sequence The variance of , dynamically adjusts the measurement noise covariance, through the formula: ； Calculate the noise covariance ; in, is the preset adjustment factor, The default value is 0.

95.

6. The method for correcting lightweight containers based on machine vision according to claim 1, characterized in that: The method for calculating and predicting vibration offset is: In the world coordinate system, define the six-dimensional pose error vector e; The visual inspection corner piece pose is converted from the camera coordinate system to the robot arm base coordinate system, and the hand-eye calibration matrix is ​​used accomplish; Calculate target force using proportional-derivative control : ; in, is the proportional gain matrix, is the differential gain matrix, is the error differential, , To control the cycle; According to the predicted vibration deviation , calculate the compensation force : ； in, To compensate for the proportional gain, To compensate for the differential gain, is the offset differential.

7. The method for correcting lightweight containers based on machine vision according to claim 1, characterized in that: The method for dynamically adjusting noise parameters is: Introducing force sensor data , and modify the control instructions through the impedance control principle: ； Gain ultimate control ,in, is the force feedback gain, is the expected contact force; Gain scheduling based on task phase, approach phase: increase visual control gain , speed up the response and reduce the force feedback gain , reduce the influence of force control; Contact phase: Reduce visual control gain , to avoid overshoot, increase , enhanced force feedback correction; Adaptively adjust the vibration frequency.

8. The method for correcting lightweight containers based on machine vision according to claim 1, characterized in that: The method for performing force control compensation is: The interaction between the end of the robotic arm and the environment is equivalent to a spring-mass-damper system: ； in, is the expected trajectory, is the external force; For continuous systems, the sampling period Discretize it and get the difference equation: ； The recursive formula is: ; Inverse the matrix Using Cholesky decomposition preprocessing, The complexity is reduced to ; A state limiting mechanism is introduced to prevent numerical divergence caused by abnormal force input.

9. A lightweight container correction system based on machine vision, the system is used to execute the correction method according to any one of claims 1 to 8, characterized in that: include: Visual perception module: This module collects and preprocesses multispectral images of containers in dynamic environments. It then builds a detection model to detect corner parts by enhancing the feature response of the corner parts area. It then calculates the disparity map of the multispectral images, reconstructs the 3D coordinates of the corner parts, and performs pose transformation. State Estimation Module: Builds a state space model of the corner piece based on its three-dimensional coordinates, identifies time-varying parameters online to build a state transfer matrix, calculates and predicts vibration offsets, projects the vibration displacements to the world coordinate system, and calculates the innovation sequence variance to dynamically adjust noise parameters. Control strategy module: This module constructs an inner- and outer-loop control architecture. The outer-loop vision controller generates the target force by calculating the pose error in the world coordinate system. The inner-loop force control compensator combines the predicted vibration offset with the force sensor feedback to correct the manipulator control instructions, schedules the control gain through the task phase, and analyzes the vibration frequency to adaptively adjust the parameters. Safety execution module: performs force control compensation based on the target force, discretizes the continuous impedance model to obtain a recursive control law to resist external disturbances.

Images