Stacker fork delivery correction method and system

Abstract

The present application relates to the technical field of stack deviation correction, and particularly relates to a stacker fork goods delivery deviation correction method and system, the method comprising establishing coordinates with the geometric center of the fork, obtaining a goods attitude vector and a goods entrance geometric parameter vector on the fork; based on the obtained goods attitude vector and goods entrance geometric parameter vector, calculating a coupling deviation degree of goods attitude and entrance geometric parameter; comparing and analyzing the deviation degree threshold value with the obtained coupling deviation degree, and generating a deviation correction instruction. The present application can realize the coordinated analysis of the overall goods attitude and the geometric relationship of the goods location by establishing coordinates with the geometric center of the fork, obtaining a goods attitude vector and a goods entrance geometric parameter vector, and calculating a coupling deviation degree based on the two, thereby generating a targeted deviation correction instruction, avoiding the limitations of the existing single-direction offset correction, and reducing the friction or jamming of the fork with the goods edge during the goods delivery process.

Description

Technical Field

[0001] This invention relates to the field of stacking correction technology, specifically to a method and system for correcting the delivery of goods by stacker crane forks. Background Technology

[0002] Existing stacker crane fork correction control strategies are generally quite simple, mostly relying on mechanical limit fine-tuning or single-variable PID control. These strategies are clearly insufficient in complex operating environments. For example, when goods tilt within the forks during handling, existing methods can only correct based on the offset in one direction, failing to coordinate control with the overall posture of the goods and the geometric relationship of the position. This can easily lead to friction or jamming at the edges when the forks deliver goods, reducing the success rate of operations and resulting in low overall correction efficiency. In complex scenarios, correction may even fail. Summary of the Invention

[0003] To address the aforementioned problems, this invention provides a method and system for correcting the deviation of stacker crane forks during delivery.

[0004] This invention adopts the following technical solution: a method for correcting the delivery deviation of a stacker crane forks, comprising:

[0005] Establish coordinates with the geometric center of the fork to obtain the cargo posture vector on the fork. The cargo posture vector includes the coordinates of the cargo center position and the normal vector of the cargo orientation.

[0006] Establish coordinates with the geometric center of the fork to obtain the geometric parameter vector of the cargo inlet, which includes the center position coordinates of the cargo inlet and the normal vector of the inlet;

[0007] Based on the acquired cargo attitude vector and cargo inlet geometric parameter vector, the coupling deviation metric between cargo attitude and inlet geometric parameters is calculated.

[0008] A preset deviation threshold is set, and the deviation threshold is compared and analyzed with the acquired coupling deviation metric value to generate a correction command.

[0009] When generating correction commands, the cargo attitude vector, cargo entry geometric parameter vector, and coupling deviation metric are input into the adaptive correction control model, which outputs the fork motion control quantity. Based on the output fork motion control quantity, the fork is coordinated and corrected in multiple degrees of freedom.

[0010] As a further description of the above technical solution: the method for obtaining the coordinates of the cargo center position includes:

[0011] Establish a fork coordinate system with the geometric center of the fork, and install at least one depth camera at the end of the fork to obtain a three-dimensional point cloud of the goods in front of the fork;

[0012] The acquired 3D point cloud is denoised and downsampled, and the processed point cloud is transformed into the forklift coordinate system;

[0013] Euclidean clustering is performed on the point cloud transformed to the forklift coordinate system to obtain multiple point cloud clusters. The feature parameters of the multiple point cloud clusters are obtained and sequentially input into the pre-built cargo point cloud cluster prediction model. The predicted point cloud clusters are output and marked as the target point cloud clusters.

[0014] Calculate the centroid of the target point cloud cluster as the center location of the cargo. .

[0015] As a further description of the above technical solution: the method for obtaining the normal vector of the cargo orientation includes:

[0016] Based on the obtained cargo center location The point cloud within the target point cloud cluster is decentralized, and the covariance matrix is ​​calculated.

[0017] Perform eigenvalue decomposition on the covariance matrix, take the eigenvector corresponding to the smallest eigenvalue as the approximate surface normal vector, and perform normalization to obtain the normal vector of the cargo orientation.

[0018] As a further description of the above technical solution: the characteristic parameters of the point cloud cluster include volume value, shape ratio, and centroid distance;

[0019] Training methods for cargo point cloud cluster prediction models include:

[0020] In cargo loading, unloading, warehousing, or transportation scenarios, a depth camera is used to collect 3D point cloud data of the cargo. Outliers and background noise are removed, and the 3D point cloud data is segmented according to individual cargo items to obtain multiple point cloud clusters. Feature parameters, including volume value, shape ratio, and centroid distance, are extracted for each point cloud cluster. Each point cloud cluster is labeled as the corresponding target point cloud cluster category by manual or automatic labeling algorithms. The target point cloud cluster category label is stored in the form of integer ID or one-hot encoding, and corresponds one-to-one with the feature parameters to construct a dataset.

[0021] A random forest classifier is used as the prediction model. The dataset is divided into training set, validation set and test set according to a preset ratio. The training set is used for training.

[0022] During training, Gini impurity is used as the node splitting criterion. A training subset is generated for each decision tree through bootstrap sampling. When splitting a node in each decision tree, some features are randomly selected. Finally, the prediction results are integrated by majority voting of the outputs of all decision trees.

[0023] Key hyperparameters are optimized using Bayesian optimization. A validation set is used for validation. When the preset standard is reached, training is stopped, and the model parameters with the best performance on the validation set are retained. The final model is evaluated using a test set, and the overall accuracy is calculated. When the model performance reaches the preset standard, the model is deployed to the actual application system for real-time prediction of newly collected cargo point cloud cluster categories.

[0024] As a further description of the above technical solution: the method for obtaining the coupling deviation metric includes:

[0025] Step S01: Obtain the distance from the center coordinates of the cargo to the center coordinates of the cargo entrance, and mark it as the distance deviation;

[0026] Step S02: Obtain the normal vector of the cargo orientation, and obtain the angle between the cargo orientation and the entrance orientation based on the normal vector of the cargo orientation and the normal vector of the entrance.

[0027] Step S03: Perform a weighted summation operation based on the obtained distance deviation and angle to obtain the coupling deviation metric.

[0028] As a further description of the above technical solution: the method for generating correction instructions includes:

[0029] When the coupling deviation metric is greater than or equal to the deviation threshold, a correction command is generated; otherwise, when the coupling deviation metric is less than the deviation threshold, no correction command is generated.

[0030] As a further description of the above technical solution: the training method of the adaptive correction control model includes:

[0031] Obtain the historical correction dataset for fork offset, and train a pre-built adaptive offset control model based on the historical correction dataset;

[0032] The historical correction dataset includes P sets of cargo attitude vectors, cargo inlet geometric parameter vectors, coupling deviation metrics, and corresponding fork motion control quantities, where P is a positive integer greater than 0. The historical correction dataset is divided into a training set and a validation set, where the training set is used to train the adaptive correction control model and the validation set is used to evaluate the generalization performance of the adaptive correction control model.

[0033] During the training of the adaptive correction control model, a deep neural network structure based on a multilayer perceptron is adopted. The cargo attitude vector, cargo inlet geometric parameter vector, and coupling deviation metric are converted into feature vectors as input. Nonlinear features are extracted through multiple hidden layers, and finally, the probability distribution of the fork motion control quantity is generated by the softmax activation function in the output layer. The fork motion control quantity corresponding to the highest probability is output as the final prediction result. The training process aims to minimize the cross-entropy loss function. At the same time, an early stop strategy is introduced to monitor the performance of the validation set. When the prediction accuracy on the validation set reaches a preset threshold, the adaptive correction control model is considered to have converged, and the training stops.

[0034] As a further description of the above technical solution: a stacker crane fork delivery correction method also includes:

[0035] Based on fork motion control The motion trajectory sequence of the forks is calculated by discrete-time integration. ;in, This represents the translational control amount of the forks in the X-axis direction. This represents the translation control amount of the forks in the Y-axis direction. To represent the attitude angle correction of the forks, including yaw, pitch correction and horizontal angle correction around the Z-axis;

[0036] Obtain the set of obstacle locations within the cargo entry area;

[0037] Based on the set of obstacle locations and the movement trajectory of the forks, the fork trajectory is corrected in real time.

[0038] The method for obtaining the set of obstacle locations includes:

[0039] Deploy depth cameras or 3D LiDAR in the cargo entrance area to collect point cloud data of the entrance area in real time.

[0040] Using the point cloud on the entrance border or entrance wall, the entrance plane is fitted by the least squares method to obtain the equation of the entrance reference plane.

[0041] The distance from the point cloud data of the cargo entrance area to the entrance plane is obtained. When the distance is less than the preset distance threshold, it means that the point cloud data belongs to the entrance reference and is not counted as an obstacle point. For the remaining point cloud data of the entrance area, Euclidean clustering is used to obtain several obstacle point cloud clusters.

[0042] For each obstacle point cloud cluster, calculate its geometric center to obtain the position of the obstacle center point, and combine all the obstacle center point positions to obtain the obstacle position set.

[0043] As a further description of the above technical solution: the method for real-time correction of the fork trajectory based on the obstacle location set and the fork's motion trajectory includes:

[0044] Obtain motion trajectory sequence The collision avoidance distance to the center point of each obstacle in the obstacle location set is preset with a collision avoidance distance threshold. When the collision avoidance distance is less than the preset collision avoidance distance threshold, an offset correction is applied along the direction of the obstacle connection. This operation is repeated for all obstacles to obtain the corrected trajectory.

[0045] A stacker crane fork delivery correction system is provided for implementing the aforementioned stacker crane fork delivery correction method. The system includes:

[0046] The cargo parameter acquisition module establishes coordinates with the geometric center of the fork and obtains the cargo posture vector on the fork. The cargo posture vector includes the coordinates of the cargo center position and the normal vector of the cargo orientation.

[0047] The inlet parameter acquisition module establishes coordinates with the geometric center of the fork and obtains the cargo inlet geometric parameter vector, which includes the center position coordinates of the cargo inlet and the normal vector of the inlet;

[0048] The data processing module calculates the coupling deviation metric between the cargo attitude and the inlet geometric parameters based on the acquired cargo attitude vector and cargo inlet geometric parameter vector.

[0049] The instruction generation module presets a deviation threshold, compares and analyzes the deviation threshold with the acquired coupling deviation metric value, and generates a correction instruction.

[0050] When a correction command is generated, the correction module inputs the cargo attitude vector, cargo entry geometric parameter vector, and coupling deviation metric into the adaptive correction control model, outputs the fork motion control quantity, and performs multi-degree-of-freedom coordinated correction of the fork based on the output fork motion control quantity.

[0051] The beneficial effects of this invention are as follows:

[0052] This invention provides a method for correcting the deviation of stacker crane forks during delivery, which improves the correction efficiency and success rate of stacker crane forks in complex operating environments. By establishing coordinates with the geometric center of the forks, the method obtains the cargo posture vector and the cargo entry geometric parameter vector, and calculates the coupling deviation metric based on the two. This enables the coordinated analysis of the overall posture of the cargo and the geometric relationship of the cargo position, thereby generating targeted correction instructions. This avoids the limitations of existing unidirectional offset correction methods, reduces friction or jamming between the forks and the cargo edges during cargo delivery, and improves the stability and safety of cargo handling.

[0053] Furthermore, this invention introduces an adaptive correction control model to coordinate the cargo posture, entry geometric parameters, and coupling deviation metric values ​​with multiple degrees of freedom, thereby realizing the intelligent generation of fork motion control quantities. Combined with a point cloud cluster prediction model and a deep neural network trained on historical correction data, this technical solution can learn and optimize correction strategies in real time, improving the adaptability of the forks in diverse cargo shapes and complex warehousing scenarios, and significantly enhancing the efficiency, reliability, and safety of cargo loading and unloading operations. Attached Figure Description

[0054] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0055] Figure 1 This is a flowchart of a stacker crane fork delivery correction method provided in Embodiment 1 of the present invention;

[0056] Figure 2 This is a flowchart of a method for obtaining the center position coordinates of cargo according to Embodiment 1 of the present invention;

[0057] Figure 3 This is a flowchart of a method for obtaining the normal vector of cargo orientation provided in Embodiment 1 of the present invention;

[0058] Figure 4 This is a flowchart of a stacker crane fork delivery correction method provided in Embodiment 2 of the present invention;

[0059] Figure 5 This is a module connection diagram of a stacker crane fork delivery correction system provided in Embodiment 3 of the present invention. Detailed Implementation

[0060] To make the technical means, creative features, objectives, and effects of this invention readily understandable, the invention is further described below with reference to specific illustrations. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0061] Example 1

[0062] Please see Figures 1-3 This invention provides a technical solution: a method for correcting the delivery deviation of a stacker crane forks, comprising:

[0063] Establish coordinates with the geometric center of the fork to obtain the posture vector of the cargo on the fork. The posture vector of the cargo includes the coordinates of the center position of the cargo and the normal vector of the cargo orientation.

[0064] The cargo attitude vector is represented as follows: In the formula, Let be the cargo attitude vector. , and The coordinates of the cargo's center position. This is the normal vector to the orientation of the goods;

[0065] The method for obtaining the coordinates of the cargo center position includes:

[0066] Establish a fork coordinate system with the geometric center of the fork, and install at least one depth camera at the fork tip to obtain a 3D point cloud of the goods in front of the forks, represented as: ;

[0067] The acquired 3D point cloud is denoised and downsampled, and the processed point cloud is transformed into the forklift coordinate system, as follows: ;

[0068] It should be noted that the denoising method uses statistical outlier removal to eliminate isolated points, and the downsampling method uses voxel mesh filtering to reduce the point cloud resolution to a range of voxel side length of 5–20 mm. It should also be noted that the denoising, downsampling, and coordinate transformation are all existing technologies and will not be elaborated on here.

[0069] The expression for coordinate transformation is: ;

[0070] In the formula, Indicates the first A 3D point cloud of a point in the camera coordinate system This is a homogeneous transformation from the depth camera to the forklift coordinate system, i.e., a 4×4 homogeneous transformation matrix. For the first The coordinate vectors of a point in the forklift coordinate system;

[0071] Euclidean clustering is performed on the point cloud transformed to the forklift coordinate system to obtain multiple point cloud clusters. The feature parameters of the multiple point cloud clusters are obtained and sequentially input into the pre-constructed cargo point cloud cluster prediction model. The predicted point cloud clusters are output and marked as target point cloud clusters.

[0072] It should be noted that the obtained Euclidean clustering to obtain multiple point cloud clusters is an existing technique and will not be elaborated here.

[0073] The centroid of the target point cloud cluster is calculated and used as the center location of the cargo, denoted as: ; , and The coordinates of the cargo's center position;

[0074] The formula for calculating the coordinates of the cargo's center position is:

[0075] ;

[0076] ;

[0077] ;

[0078] In the formula, For the target point cloud cluster, , and The coordinates of the cargo's center position. This indicates the number of points in the target point cloud cluster. , and Points The three components.

[0079] The method for obtaining the normal vector of the cargo orientation includes:

[0080] Based on the obtained cargo center location The point cloud within the target point cloud cluster is decentralized, and the covariance matrix is ​​calculated.

[0081] The expression for decentralization is: ; The decentralized coordinate vector within the target point cloud cluster. Let T be the coordinate vector within the target point cloud cluster, and T be the transpose of the matrix.

[0082] The expression for the covariance matrix is: ; The outer product of all decentralized points is summed to construct the covariance matrix.

[0083] Perform eigenvalue decomposition on the covariance matrix, take the eigenvector corresponding to the smallest eigenvalue as the approximate surface normal vector, and then normalize it.

[0084] The expression for eigenvalue decomposition is: ;in, and The first 1 eigenvalue and eigenvector;

[0085] The expression for normalization is: , The eigenvector corresponding to the smallest eigenvalue. Let be the normal vector of the cargo's orientation. Let be the magnitude of the vector.

[0086] The method for obtaining feature parameters of multiple point cloud clusters, sequentially inputting them into a pre-built cargo point cloud cluster prediction model, and outputting the predicted point cloud clusters, which are then marked as target point cloud clusters, includes:

[0087] The point cloud cluster The characteristic parameters include volume value, aspect ratio, and centroid distance;

[0088] The volume value is obtained by using the convex hull method for calculation, and the formula for calculating the volume value is: In the formula, For point cloud clusters The volume value, For point cloud clusters The three-dimensional convex hull, convex hull Volume;

[0089] The method for calculating the shape ratio is as follows: In the formula, The shape ratio measures how compact the volume is relative to the bounding box volume. For point cloud clusters The bounding box volume, To prevent decimals with a denominator of zero, For point cloud clusters The volume value.

[0090] The method for calculating the centroid distance is as follows: ; The distance between the centroids, For point cloud clusters The coordinates of the centroid, This is the origin of the forklift coordinate system.

[0091] Training methods for cargo point cloud cluster prediction models include:

[0092] In cargo loading, unloading, warehousing, or transportation scenarios, a depth camera is used to collect 3D point cloud data of the cargo. Outliers and background noise are removed, and the point cloud is segmented according to individual cargo items to obtain multiple point cloud clusters. The following feature parameters are extracted for each point cloud cluster, including volume value, shape ratio, and centroid distance. Each point cloud cluster is labeled with the corresponding target point cloud cluster category, such as cargo, pallet, or shelf, by manual or automatic labeling algorithms. The labels are stored in the form of integer IDs or one-hot encodings, corresponding one-to-one with the feature parameters, and a dataset is constructed.

[0093] A random forest classifier is used as the prediction model. The dataset is divided into training set, validation set and test set according to a preset ratio. The training set is used for training.

[0094] During training, Gini impurity is used as the node splitting criterion. A training subset is generated for each tree through bootstrap sampling. When splitting a node in each tree, some features are randomly selected. Finally, the prediction results are integrated by majority voting from the outputs of all decision trees.

[0095] Key hyperparameters were optimized using Bayesian optimization methods.

[0096] The key hyperparameters include: number of decision trees: 50-200; maximum depth of each tree: 5-15 layers; minimum number of samples for node splitting: 2-10 nodes; maximum number of features considered during splitting: 1-3.

[0097] Validation is performed using a validation set. When the accuracy of the validation set improves by less than 0.1% for 15 consecutive rounds, training is stopped, and the model parameters with the best performance on the validation set are retained.

[0098] The final model is evaluated using a test set, and the overall accuracy is calculated.

[0099] When the model performance reaches the preset standard (e.g., overall accuracy ≥ 95%), the model will be deployed to the actual application system for real-time prediction of the category of newly collected cargo point cloud clusters.

[0100] Establish coordinates with the geometric center of the fork to obtain the geometric parameter vector of the cargo inlet, which includes the center position coordinates of the cargo inlet and the normal vector of the inlet;

[0101] The geometric parameter vector of the cargo inlet is represented as follows: In the formula, This is the geometric parameter vector for the cargo inlet. , and These are the coordinates of the center position of the cargo entrance. This is the normal vector facing the entrance;

[0102] The method for obtaining the center position coordinates of the cargo entrance includes: fixing two sets of laser rangefinders on both sides of the entrance, obtaining the distance from each set of rangefinders to the edge of the obstacle, calculating the lateral position of the entrance centerline through the corresponding distance difference on both sides, which is the midpoint, calculating the entrance tilt angle through two sets of height differences or two longitudinal sensors, transforming the measured geometric quantities to the forklift coordinate system, and obtaining the cargo entrance geometric parameter vector.

[0103] Based on the acquired cargo attitude vector and cargo inlet geometric parameter vector, the coupling deviation metric between cargo attitude and inlet geometric parameters is calculated.

[0104] The method for obtaining the coupling deviation metric includes:

[0105] Step S01: Obtain the distance from the center coordinates of the cargo to the center coordinates of the cargo entrance, and mark it as the distance deviation;

[0106] The formula for calculating the distance deviation is as follows:

[0107] In the formula, For distance deviation, , and The coordinates of the cargo's center position. , and These are the coordinates of the center position of the cargo entrance.

[0108] Step S02: Obtain the normal vector of the cargo orientation, and obtain the angle between the cargo orientation and the entrance orientation based on the normal vector of the cargo orientation and the normal vector of the entrance.

[0109] The formula for calculating the included angle is: In the formula, The angle between the direction the goods face and the direction the entrance faces. Let the magnitude of the vector be . Let be the normal vector of the entrance orientation. This is the normal vector of the cargo's orientation.

[0110] Step S03: Perform a weighted summation operation based on the obtained distance deviation and angle to obtain the coupling deviation metric.

[0111] Optionally, the expression for the coupling deviation metric is:

[0112] In the formula, This is a measure of coupling deviation. This is the normalized value of the distance deviation. This is the normalized value of the angle between the cargo orientation and the entrance orientation. and These are the weighting coefficients. and All are greater than 0. + =1;

[0113] It should be noted that the normalization method used is minimum-maximum normalization, which is an existing technique and will not be elaborated upon here.

[0114] The weighting coefficients in the formula and the preset thresholds in the analysis process are set by those skilled in the art based on the actual situation or obtained by simulation of a large amount of data. The size of the weighting coefficient is a specific value obtained by quantifying each parameter to facilitate subsequent comparison. The size of the weighting coefficient depends on the amount of sample data and the processing coefficients initially set by those skilled in the art for each set of sample data. As long as it does not affect the proportional relationship between the parameter and the quantified value, it is acceptable.

[0115] A preset deviation threshold is set, and the deviation threshold is compared and analyzed with the acquired coupling deviation metric value to generate a correction command.

[0116] The method for generating correction instructions includes:

[0117] When the coupling deviation metric is greater than or equal to the deviation threshold, a correction command is generated; otherwise, when the coupling deviation metric is less than the deviation threshold, no correction command is generated.

[0118] When generating a correction command, the cargo attitude vector, cargo entry geometric parameter vector, and coupling deviation metric are input into the adaptive correction control model, and the fork motion control quantity is output. Based on the output fork motion control quantity, the fork is coordinated with multiple degrees of freedom for pitch, roll, and offset along the X, Y, and Z axes.

[0119] The fork motion control quantity is expressed as: ;

[0120] In the formula, This refers to the overall motion control parameters of the forks, encompassing all the action commands that the forks need to adjust in space. This represents the translational control amount of the forks in the X-axis direction. This represents the translation control amount of the forks in the Y-axis direction. To represent the attitude angle correction of the forks, we include yaw about the Z-axis, pitch correction, and horizontal angle correction.

[0121] The training method for the adaptive correction control model includes:

[0122] Obtain the historical correction dataset for fork offset, and train a pre-built adaptive offset control model based on the historical correction dataset;

[0123] The historical correction dataset includes P sets of cargo attitude vectors, cargo inlet geometric parameter vectors, coupling deviation metrics, and corresponding fork motion control quantities, where P is a positive integer greater than 0. The historical correction dataset is divided into a training set and a validation set, where the training set is used to train the adaptive correction control model and the validation set is used to evaluate the generalization performance of the adaptive correction control model.

[0124] During the training of the adaptive correction control model, a deep neural network structure based on a multilayer perceptron is adopted. The cargo attitude vector, cargo inlet geometric parameter vector, and coupling deviation metric are converted into feature vectors as input. Nonlinear features in the data are extracted through multiple hidden layers, and finally, the probability distribution of the fork motion control quantity is generated by the softmax activation function in the output layer. The fork motion control quantity corresponding to the highest probability is output as the final prediction result. The training process aims to minimize the cross-entropy loss function. At the same time, an early stop strategy is introduced to monitor the performance of the validation set. When the prediction accuracy on the validation set reaches a preset threshold, the adaptive correction control model is considered to have converged, and the training stops.

[0125] Example 2

[0126] Please see Figure 4 Based on Embodiment 1, this invention further provides a technical solution, specifically including:

[0127] Based on fork motion control The motion trajectory sequence of the forks is calculated by discrete-time integration. ;in, This represents the translational control amount of the forks in the X-axis direction. This represents the translation control amount of the forks in the Y-axis direction. To represent the attitude angle correction of the forks, including yaw, pitch correction and horizontal angle correction around the Z-axis;

[0128] Specifically, let the state of the forks at time t be... , The coordinates of the geometric center of the forks are [value]. For the fork orientation angle, within a set time step The updated status yields:

[0129]

[0130] By repeating the iteration, the fork is obtained from its initial state. To the target state A continuous sequence of states.

[0131] Obtain the set of obstacle locations within the cargo entry area;

[0132] Based on the set of obstacle locations and the movement trajectory of the forks, the fork trajectory is corrected in real time.

[0133] The method for obtaining the set of obstacle locations includes:

[0134] Deploy depth cameras or 3D LiDAR in the cargo entrance area to collect point cloud data of the entrance area in real time.

[0135] Using the point cloud on the entrance border or entrance wall, the entrance plane is fitted by the least squares method to obtain the equation of the entrance reference plane.

[0136] The equation of the inlet reference plane is expressed as: In the formula, Let be the normal vector of the entrance plane. Located at the center of the entrance. Point cloud data for the entry area;

[0137] The distance from the point cloud data of the entrance area to the entrance plane is obtained. When the distance is less than the preset distance threshold, it means that the point cloud data belongs to the entrance reference and is not counted as an obstacle point. For the remaining point cloud data of the entrance area, Euclidean clustering is used to obtain several obstacle point cloud clusters.

[0138] For each obstacle point cloud cluster, calculate its geometric center to obtain the position of the obstacle center point, and combine all the obstacle center point positions to obtain the obstacle position set.

[0139] Methods for real-time correction of fork trajectory based on obstacle location set and fork movement trajectory include:

[0140] Obtain motion trajectory sequence The collision avoidance distance to the center point of each obstacle in the obstacle location set is preset with a collision avoidance distance threshold. When the collision avoidance distance is less than the preset collision avoidance distance threshold, an offset correction is applied along the direction of the obstacle connection line. This operation is repeated for all obstacles to obtain the corrected trajectory.

[0141] motion trajectory sequence The method for calculating the collision avoidance distance to the center point of each obstacle within the obstacle location set is as follows:

[0142] In the formula, This represents the geometric center of the m-th obstacle. For collision avoidance distance, Let be the pose of the fork at time t;

[0143] The method for applying offset correction along the direction of the obstacle connection is as follows:

[0144] ;

[0145] In the formula, These are the corrected fork trajectory points. Let be the pose of the fork at time t. For collision avoidance distance threshold, The trajectory correction scaling factor. This represents the geometric center of the m-th obstacle.

[0146] In this embodiment, real-time perception and dynamic obstacle avoidance of obstacles can be achieved during fork operation, significantly improving the safety and reliability of forks in complex environments. By performing discrete-time integration on the fork motion control quantity, a continuous fork trajectory sequence is generated, enabling the fork motion planning to accurately reflect actual state changes. By combining the obstacle position set in the entrance area for trajectory correction, proactive avoidance of potential collision risks is achieved, preventing the forks from colliding with goods, shelves or other obstacles during operation, and realizing anti-collision protection during the correction process.

[0147] Furthermore, by real-time acquisition of point clouds in the entrance area, extraction of obstacle point cloud clusters, and calculation of geometric centers, this application can quickly construct a set of obstacle locations. In trajectory correction, an offset correction strategy based on collision distance judgment along the obstacle direction is adopted, enabling the fork movement trajectory to be continuously and smoothly adjusted in a dynamic environment, ensuring the smoothness and safety of cargo loading and unloading. Moreover, this method is compatible with different cargo sizes and cargo entrance geometry, improving the stacker crane's adaptability and overall operational efficiency in diverse operating scenarios.

[0148] Example 3

[0149] Please see Figure 5 This invention provides a technical solution: a stacker crane fork delivery correction system, used to implement the aforementioned stacker crane fork delivery correction method, the system comprising:

[0150] The cargo parameter acquisition module establishes coordinates with the geometric center of the fork and obtains the cargo posture vector on the fork. The cargo posture vector includes the coordinates of the cargo center position and the normal vector of the cargo orientation.

[0151] The inlet parameter acquisition module establishes coordinates with the geometric center of the fork and obtains the cargo inlet geometric parameter vector, which includes the center position coordinates of the cargo inlet and the normal vector of the inlet;

[0152] The data processing module calculates the coupling deviation metric between the cargo attitude and the inlet geometric parameters based on the acquired cargo attitude vector and cargo inlet geometric parameter vector.

[0153] The instruction generation module presets a deviation threshold, compares and analyzes the deviation threshold with the acquired coupling deviation metric value, and generates a correction instruction.

[0154] When a correction command is generated, the correction module inputs the cargo attitude vector, cargo entry geometric parameter vector, and coupling deviation metric into the adaptive correction control model, outputs the fork motion control quantity, and performs multi-degree-of-freedom coordinated correction of the fork based on the output fork motion control quantity.

[0155] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for correcting the deviation of stacker crane forks during delivery, characterized in that, include: Establish coordinates with the geometric center of the fork to obtain the cargo posture vector on the fork. The cargo posture vector includes the coordinates of the cargo center position and the normal vector of the cargo orientation. Obtain the cargo inlet geometric parameter vector, which includes the center position coordinates of the cargo inlet and the normal vector of the inlet; Based on the acquired cargo attitude vector and cargo inlet geometric parameter vector, the coupling deviation metric between cargo attitude and inlet geometric parameters is calculated. A preset deviation threshold is set, and the deviation threshold is compared and analyzed with the acquired coupling deviation metric value to generate a correction command. When generating a correction command, the cargo attitude vector, cargo entry geometric parameter vector, and coupling deviation metric are input into the adaptive correction control model, and the fork motion control quantity is output. Based on the output fork motion control quantity, the fork is coordinated and corrected in multiple degrees of freedom. The method for obtaining the coordinates of the cargo center position includes: Install at least one depth camera at the end of the forks to obtain a three-dimensional point cloud of the cargo in front of the forks; The acquired 3D point cloud is denoised and downsampled, and the processed point cloud is transformed into the forklift coordinate system; Euclidean clustering is performed on the point cloud transformed to the forklift coordinate system to obtain multiple point cloud clusters. The feature parameters of the multiple point cloud clusters are obtained and sequentially input into the pre-built cargo point cloud cluster prediction model. The predicted point cloud clusters are output and marked as the target point cloud clusters. Calculate the centroid of the target point cloud cluster as the center location of the cargo. ; The method for obtaining the normal vector of the cargo orientation includes: Based on the obtained cargo center location The point cloud within the target point cloud cluster is decentralized, and the covariance matrix is ​​calculated. Perform eigenvalue decomposition on the covariance matrix, take the eigenvector corresponding to the smallest eigenvalue as the approximate surface normal vector, and perform normalization to obtain the normal vector of the cargo orientation; The characteristic parameters of the point cloud cluster include volume value, shape ratio, and centroid distance; Training methods for cargo point cloud cluster prediction models include: In cargo loading, unloading, warehousing, or transportation scenarios, a depth camera is used to collect 3D point cloud data of the cargo. Outliers and background noise are removed, and the 3D point cloud data is segmented according to individual cargo items to obtain multiple point cloud clusters. Feature parameters, including volume value, shape ratio, and centroid distance, are extracted for each point cloud cluster. Each point cloud cluster is labeled as a corresponding target point cloud cluster category by manual or automatic labeling algorithms. The target point cloud cluster category label is stored in the form of integer ID or one-hot encoding, corresponding one-to-one with the feature parameters, and a dataset is constructed.

2. The stacker crane fork delivery correction method according to claim 1, characterized in that, The training method for the cargo point cloud cluster prediction model also includes: A random forest classifier is used as the prediction model. The dataset is divided into training set, validation set and test set according to a preset ratio. The training set is used for training. During training, Gini impurity is used as the node splitting criterion. A training subset is generated for each decision tree through bootstrap sampling. When splitting a node in each decision tree, some features are randomly selected. Finally, the prediction results are integrated by majority voting of the outputs of all decision trees. Key hyperparameters are optimized using Bayesian optimization. A validation set is used for validation. When the preset standard is reached, training is stopped, and the model parameters with the best performance on the validation set are retained. The final model is evaluated using a test set, and the overall accuracy is calculated. When the model performance reaches the preset standard, the model is deployed to the actual application system for real-time prediction of newly collected cargo point cloud cluster categories.

3. The stacker crane fork delivery correction method according to claim 1, characterized in that, The method for obtaining the coupling deviation metric includes: Step S01: Obtain the distance from the center coordinates of the cargo to the center coordinates of the cargo entrance, and mark it as the distance deviation; Step S02: Obtain the normal vector of the cargo orientation, and obtain the angle between the cargo orientation and the entrance orientation based on the normal vector of the cargo orientation and the normal vector of the entrance. Step S03: Perform a weighted summation operation based on the obtained distance deviation and angle to obtain the coupling deviation metric.

4. The stacker crane fork delivery correction method according to claim 1, characterized in that, The method for generating correction instructions includes: When the coupling deviation metric is greater than or equal to the deviation threshold, a correction command is generated; otherwise, when the coupling deviation metric is less than the deviation threshold, no correction command is generated.

5. A method for correcting the delivery deviation of a stacker crane forks according to claim 1, characterized in that, The training method for the adaptive correction control model includes: Obtain the historical correction dataset for fork offset, and train a pre-built adaptive offset control model based on the historical correction dataset; The historical correction dataset includes P sets of cargo attitude vectors, cargo inlet geometric parameter vectors, coupling deviation metrics, and corresponding fork motion control quantities, where P is a positive integer greater than 0. The historical correction dataset is divided into a training set and a validation set, where the training set is used to train the adaptive correction control model and the validation set is used to evaluate the generalization performance of the adaptive correction control model. During the training of the adaptive correction control model, a deep neural network structure based on a multilayer perceptron is adopted. The cargo attitude vector, cargo inlet geometric parameter vector, and coupling deviation metric are converted into feature vectors as input. Nonlinear features are extracted through multiple hidden layers, and finally, the probability distribution of the fork motion control quantity is generated by the softmax activation function in the output layer. The fork motion control quantity corresponding to the highest probability is output as the final prediction result. The training process aims to minimize the cross-entropy loss function. At the same time, an early stop strategy is introduced to monitor the performance of the validation set. When the prediction accuracy on the validation set reaches a preset threshold, the adaptive correction control model is considered to have converged, and the training stops.

6. The stacker crane fork delivery correction method according to claim 1, characterized in that, Also includes: Based on fork motion control The motion trajectory sequence of the forks is calculated by discrete-time integration. ;in, This represents the translational control amount of the forks in the X-axis direction. This represents the translation control amount of the forks in the Y-axis direction. To represent the attitude angle correction of the forks, including yaw, pitch correction and horizontal angle correction around the Z-axis; Obtain the set of obstacle locations within the cargo entry area; Based on the set of obstacle locations and the movement trajectory of the forks, the fork trajectory is corrected in real time. The method for obtaining the set of obstacle locations includes: Deploy depth cameras or 3D LiDAR in the cargo entrance area to collect point cloud data of the entrance area in real time. Using the point cloud on the entrance border or entrance wall, the entrance plane is fitted by the least squares method to obtain the equation of the entrance reference plane. The distance from the point cloud data of the cargo entrance area to the entrance plane is obtained. When the distance is less than the preset distance threshold, it means that the point cloud data belongs to the entrance reference and is not counted as an obstacle point. For the remaining point cloud data of the entrance area, Euclidean clustering is used to obtain obstacle point cloud clusters. For each obstacle point cloud cluster, calculate its geometric center to obtain the position of the obstacle center point, and combine all the obstacle center point positions to obtain the obstacle position set.

7. A method for correcting the delivery deviation of a stacker crane forks according to claim 6, characterized in that, The method for real-time correction of the fork trajectory based on the set of obstacle locations and the fork's motion trajectory includes: Obtain motion trajectory sequence The collision avoidance distance to the center point of each obstacle in the obstacle location set is preset with a collision avoidance distance threshold. When the collision avoidance distance is less than the preset collision avoidance distance threshold, an offset correction is applied along the direction of the obstacle connection. This operation is repeated for all obstacles to obtain the corrected trajectory.

8. A stacker crane fork delivery correction system, used to implement the stacker crane fork delivery correction method according to any one of claims 1-7, characterized in that, The system includes: The cargo parameter acquisition module establishes coordinates with the geometric center of the fork and obtains the cargo posture vector on the fork. The cargo posture vector includes the coordinates of the cargo center position and the normal vector of the cargo orientation. The inlet parameter acquisition module obtains the cargo inlet geometric parameter vector, which includes the center position coordinates of the cargo inlet and the normal vector of the inlet; The data processing module calculates the coupling deviation metric between the cargo attitude and the inlet geometric parameters based on the acquired cargo attitude vector and cargo inlet geometric parameter vector. The instruction generation module presets a deviation threshold, compares and analyzes the deviation threshold with the acquired coupling deviation metric value, and generates a correction instruction. When a correction command is generated, the correction module inputs the cargo attitude vector, cargo entry geometric parameter vector, and coupling deviation metric into the adaptive correction control model, outputs the fork motion control quantity, and performs multi-degree-of-freedom coordinated correction of the fork based on the output fork motion control quantity.

Images