AGV pallet intelligent identification and self-adaptive forking control method and system based on 3D laser radar

Abstract

The invention discloses an AGV pallet intelligent identification and adaptive forking control method and system based on a 3D laser radar, and relates to the technical field of image data processing. According to the method, the side features of the pallet are extracted through combination of multi-stage point cloud processing and the point cloud convolutional neural network; a historical feature vector is combined with a grey wolf algorithm to obtain an optimized SVM pallet side classification model, so that the pallet identification accuracy in a complex scene is improved; a weighted RANSAC algorithm and PCA are adopted to realize accurate positioning of pallet center coordinates and error control of attitude angles; through Euclidean clustering segmentation and external cuboid fitting, accurate detection of the cargo size is realized; a path conforming to AGV kinematics is generated based on an improved A * algorithm, PID transverse deviation closed-loop correction is combined, and the pallet fork calibration success rate is increased; according to the method, the autonomous operation capability of the AGV in a complex industrial scene is remarkably improved, the method is compatible with mainstream standard pallets and multi-sensor configuration, and an efficient and safe technical solution is provided for unmanned storage.

Description

Technical Field

[0001] The present invention belongs to the field of image data processing technology, and more specifically, relates to an AGV pallet intelligent recognition and adaptive fork-taking control method and system based on 3D laser radar. Background Art

[0002] In recent years, with the rapid development of intelligent manufacturing, unmanned warehousing, and Industry 4.0, the demand for logistics automation has exploded. Automated Guided Vehicles (AGVs), as core equipment in logistics automation, are widely used in scenarios such as warehousing and handling, material transfer on production lines, and container dispatching at ports. However, traditional AGV systems face significant challenges in high-precision operations and adaptability to complex environments.

[0003] Under the current circumstances, storage space tends to be highly densely laid out (such as narrow aisles and multi-layer shelves), requiring AGVs to complete millimeter-level precise forking within a limited space; scenes such as construction sites and outdoor logistics have problems such as uneven ground, changing lighting, and dust interference, and traditional navigation technology is not stable enough; existing AGVs rely on 2D lidar or visual SLAM, with low positioning accuracy, and cannot meet the forking requirements of precision equipment and fragile goods. Manual assistance is required for pallet posture detection and cargo size verification, resulting in interruptions to the operation process and reduced efficiency; cargo overturning and collision accidents occur frequently, causing losses to enterprises; traditional AGVs mostly use open-loop control, lack dynamic error compensation capabilities, and have difficulty coping with disturbances such as cargo offset and road bumps; emerging 3D vision and AI algorithms have not yet been fully integrated into the AGV control system, and the technological dividend has not been effectively released.

[0004] Existing technologies have problems such as complex scenarios, contradictions between accuracy and efficiency, security and cost pressures, and lack of technological integration. Summary of the Invention

[0005] (1) Technical problems solved In response to the problems in the related art, the present invention provides an AGV pallet intelligent identification and adaptive fork control method based on 3D laser radar to overcome the above-mentioned technical problems existing in the existing related art.

[0006] (2) Technical solution To solve the above technical problems, the present invention is achieved through the following technical solutions: S1. Collect and process the point cloud data of the pallet to obtain the final pallet point cloud data; use the point cloud convolutional neural network to extract the features of the final pallet point cloud data to obtain the side feature vector of the pallet point cloud; S2. Construct an initial SVM classification model, use the historical feature vector combined with the optimization algorithm to train and optimize the initial SVM classification model, and obtain an optimized SVM pallet side classification model; input the pallet side point cloud feature vector into the optimized SVM pallet side classification model to obtain a classification result; S3. When the classification result is a candidate pallet area, the RANSAC algorithm is used to perform plane fitting on the final pallet point cloud data to obtain a pallet cross-section point cloud; the pallet cross-section point cloud is analyzed to obtain the pallet pose; S4. Collect the point cloud of the cargo and perform Euclidean clustering to obtain cargo point cloud clusters; fit the minimum circumscribed cuboid of the cargo point cloud clusters to obtain the cargo size; S5. When the cargo size is within the cargo size threshold, the AGV moves according to the pallet posture, the current AGV posture and the pickup path generated by the improved A* algorithm. During the movement, PID correction is performed until calibration is completed, and the fork operation is started; Through the deep integration of 3D perception, intelligent algorithms and closed-loop control, this invention realizes the full-process autonomy and high-precision operation of AGV from pallet detection, path planning to safe forking, solving the core pain points of traditional technologies such as insufficient precision, high dependence on manual labor, poor dynamic adaptability and low safety, and providing reliable technical support for unmanned warehousing and industrial logistics.

[0007] Preferably, the S1 comprises the following steps: S11. Based on the standard pallet structure, set a standard pallet structure geometric feature point cloud template set; Constructing a three-dimensional side feature model library based on the standard pallet structure geometric feature point cloud template set; S12: The AGV autonomously navigates and moves until the fork is facing the pallet and the distance from the pallet is within 0.4-1 meters; the scanning cycle is set; Scan the real-time pallet using 3D laser radar combined with a scanning cycle to obtain the original pallet point cloud data; S13, using voxel filtering to reduce noise on the original stack point cloud data to obtain reduced noise stack point cloud data; S14, setting an AGV fork operating range constraint, constructing a valid point cloud screening formula based on the AGV fork operating range constraint, and screening valid point clouds in the noise-reduced pallet point cloud data from the AGV center based on the valid point cloud screening formula to obtain valid pallet point cloud data; S15, intercepting point data within a height range from the valid pallet point cloud data through straight-through filtering to obtain final pallet point cloud data; S16. Extract features from the final pallet point cloud data based on a standard pallet 3D feature model library and in combination with a point cloud convolutional neural network to obtain a pallet side point cloud feature vector; The above steps establish a point cloud template set based on the standard pallet geometric features, guide the AGV to locate within a scanning range of 0.4-1 meters, and screen valid point clouds through voxel filtering for noise reduction and fork operating range constraints. Then, a straight-through filter is used to intercept the point cloud in the valid height range. Finally, the template features are fused through the point cloud convolutional neural network for deep feature extraction, forming a feature vector that represents the side structure of the pallet, providing a robust data foundation for subsequent classification and posture calculation.

[0008] Preferably, said S2 comprises the following steps: S21, collecting a large amount of historical pallet point cloud feature vector data and flying pallet point cloud feature vector data to obtain historical feature vectors; S22, constructing an initial SVM classification model, setting a learning rate of the initial SVM classification model; setting a training accuracy threshold and training accuracy of the initial SVM classification model; S23, using the historical feature vector to train the initial SVM classification model, combining the optimization algorithm to find the learning rate of the initial SVM classification model during the training process to obtain an optimal solution; using the optimal solution as the learning rate of the initial SVM classification model to obtain an optimized SVM stack side classification model; The above steps construct an initial SVM model by collecting stack / non-stack feature vectors and dynamically adjust the learning rate parameters using an optimization algorithm. This simultaneously optimizes classification accuracy and generalization capability during training, significantly improving recognition accuracy in complex scenarios such as tilt and occlusion, and reducing the false detection rate.

[0009] Preferably, in the S23 training process, finding the learning rate of the initial SVM classification model in combination with an optimization algorithm to obtain the optimal solution includes the following steps: S231. Construct a gray wolf population, set the size of the gray wolf population, and set the maximum number of optimization iterations; S232. Randomly set the initial position of the gray wolf population according to the learning rate of the initial SVM classification model to obtain an initial position set of the gray wolf population; S233. Defining a fitness function for the position of gray wolves in the gray wolf population according to the training accuracy threshold and the training accuracy; S234, performing an iterative operation on the gray wolf population initial position set, calculating the fitness value of each position in the gray wolf population initial position set according to the fitness function during each round of iteration, updating the position of each gray wolf in the gray wolf population initial position set in descending order of fitness value, and obtaining the best individual gray wolf position in the gray wolf population and the global best gray wolf position during each round of iteration; S235, repeat S234, when the maximum number of optimization iterations is reached, stop the iteration, and take the global best gray wolf position as the optimal solution; The above steps initialize the gray wolf population and set the number of iterations. The individual positions are evaluated using a fitness function that integrates the training accuracy. The gray wolf positions are updated according to the fitness value in each iteration. Finally, the global optimal solution is output as the model learning rate, achieving global optimization of the classification model parameters and avoiding local optimal traps.

[0010] Preferably, the step S3 includes the following steps: S31, calculating the feature matching degree of each point in the candidate pallet point cloud by using the Hausdorff distance to obtain a feature matching degree set; S32, setting weights for points in the final pallet point cloud data according to the feature matching degree set to obtain a final pallet point weight set; iterating the RANSAC algorithm on the points in the final pallet point cloud data according to the final pallet point weight set to obtain a pallet side point cloud; S33. Analyze the covariance matrix of the point cloud of the side surface of the pallet using PCA to obtain the direction of the long axis of the pallet; collect the inliers in the point cloud of the side surface of the pallet to obtain an inlier set; Calculate the mean of the three-dimensional coordinates of each interior point in the interior point set to obtain the coordinates of the center of the pallet; S34. Obtain a plane normal vector based on the plane equation of the point cloud on the side of the pallet; The yaw angle and pitch angle are calculated based on the plane normal vector, and the yaw angle and pitch angle together constitute the attitude angle, and the attitude angle and the center coordinate of the pallet together constitute the pallet posture; Set the yaw angle threshold and pitch angle threshold. When the yaw angle is greater than the yaw angle threshold or the pitch angle is greater than the pitch angle threshold, a level 1 warning is triggered. The above steps improve the robustness of plane detection by dynamically calculating the matching degree of point cloud features based on Hausdorff distance and preferentially fitting high-confidence point clouds through the weighted RANSAC algorithm. The PCA decomposition of the stack point cloud covariance matrix determines the direction of the long axis, and the center of the stack is located by combining the inlier mean calculation. The yaw and pitch angles are analyzed through plane normal vector projection. The dual-threshold warning mechanism monitors posture anomalies in real time, and exceeding the threshold triggers a first-level warning, comprehensively solving the problem of pose estimation deviation caused by point cloud noise or local deformation in traditional methods.

[0011] Preferably, in S32, performing RANSAC algorithm iteration on the points in the final pallet point cloud data according to the final pallet point weight set to obtain the pallet side point cloud comprises the following steps: S321, calculating the maximum number of iterations of the RANSAC algorithm on the final stack point cloud data to obtain the maximum RA iteration number; S322. Randomly select three points from the final pallet point cloud data according to the final pallet point weight set to obtain selected points; iterate the selected points using the RANSAC algorithm, and update the number of inliers in each iteration to obtain an updated cross-sectional point cloud; S323, repeat S322, and when the maximum number of RA iterations is reached, stop the iteration to obtain the pallet side point cloud data and the plane equation of the pallet side point cloud; The above steps improve the accuracy of plane detection by improving the RANSAC algorithm. The maximum number of iterations is dynamically calculated based on the confidence level and the ratio of inliers. High-confidence points are preferentially extracted based on the feature matching weight. The number of inliers is updated and the plane equation is optimized in each iteration. The optimal inlier set is cyclically screened to effectively overcome the interference of uneven point cloud density. Finally, the precise pallet side point cloud and plane equation are output, which improves the success rate of plane fitting.

[0012] Preferably, the S4 comprises the following steps: S41: The intelligent main control host first controls the fork to descend to ensure that the fork can normally fork under the pallet when picking up the goods; and obtains the original point cloud data of the goods through 3D laser radar scanning; S42, setting a clustering condition, wherein if two points in the original point cloud data of the goods satisfy a clustering formula, they are classified into the same cluster; Based on the pallet cross-section point cloud and the clustering conditions, Euclidean clustering is performed on the original point cloud data of the cargo above to obtain cargo point cloud clusters; S43. Use the AABB algorithm to combine the cargo point cloud cluster and fit the minimum circumscribed cuboid of the cargo to obtain the cargo size; the cargo size includes the maximum cargo width and the maximum cargo height; Set cargo height threshold and cargo width threshold; when the maximum cargo height > cargo height threshold or the maximum cargo width > cargo width threshold, it indicates that the cargo size is too large and triggers a level 1 alarm; The above steps ensure that the 3D lidar fully captures the cargo point cloud by pre-positioning the forks. The point cloud above the pallet is clustered and segmented based on a dynamic Euclidean distance threshold. The AABB algorithm is used to fit the minimum circumscribed cuboid to accurately measure the cargo width and height. When the size exceeds the preset threshold, a level 1 alarm is triggered, effectively identifying oversized cargo, reducing detection errors, and avoiding fork pickup failures or equipment damage due to abnormal cargo size.

[0013] Preferably, the S5 comprises the following steps: S51, obtaining the pallet posture according to the pallet center coordinates and posture angle; collecting the AGV current posture, wherein the current posture includes the AGV current position and orientation angle; S52, setting the current posture of the AGV as the starting node and the target node as the pallet posture; Set a node expansion rule, wherein the node expansion rule is a fixed moving step length and a fixed steering angle for the steering action; for each steering action, the position of the AGV is updated; the fixed steering angles are a fixed angle for left turning, straight going, and a fixed angle for right turning; S53, simulating the AGV to move from the starting node to the target node according to the node expansion rule; S54, setting weights H1 and H2 of the balance distance and direction priority, and constructing a heuristic function according to the weights of the balance distance and direction priority; At each node, the AGV generates a set of child nodes, which includes multiple child nodes that the AGV will move to next. It uses a heuristic function to calculate the comprehensive cost of each child node in the set, and selects the child node with the smallest comprehensive cost as the next node to be expanded. S55. Set the target distance threshold and the vehicle head deviation threshold; repeat S54. When the distance between the AGV and the target node is less than or equal to the target distance threshold or the vehicle head deviation is less than or equal to the vehicle head deviation threshold, stop the calculation and obtain the pickup path; and execute the pickup according to the pickup path. S56. Set an adjustment cycle and a PID control distance threshold. When the distance between the AGV and the center coordinate of the pallet is less than the PID control distance threshold, collect the fork center coordinate data in each adjustment cycle to obtain the fork center coordinate. Calculate the lateral offset between the fork center coordinate and the pallet center coordinate, and obtain the lateral offset calculation result for each cycle; Set a lateral offset threshold; when the lateral offset calculation result of each cycle is greater than or equal to the lateral offset threshold, calculate the steering angular velocity based on the lateral offset calculation result of each cycle and through the proportional term, integral term, and differential term; drive the differential wheel to adjust the direction according to the steering angular velocity; otherwise, no adjustment is performed; S57. Set the maximum number of repetitions; repeat S56. If the lateral offset is less than the lateral offset threshold for o consecutive cycles, the calibration is considered complete, the correction is stopped, and the fork operation is started. If the lateral offset is still less than the lateral offset threshold after the maximum number of repetitions is reached, a timeout alarm is triggered. The fork operation is that after the AGV reaches the target point on the pallet and the pickup switch is triggered, the intelligent control host controls the fork lifting mechanism to rise; the AGV autonomously navigates to the target point according to the set unloading point to unload the goods. The above steps achieve precise forking by improving path planning and closed-loop control; an improved A* algorithm model is constructed based on the pallet posture and the current position of the AGV, and a fixed steering angle and step-size constraint node expansion rule is set. The comprehensive cost value is dynamically calculated in combination with the dual-weight heuristic function to generate a smooth path that conforms to the AGV kinematic characteristics, thereby improving planning efficiency; when approaching the target, it switches to the PID closed-loop control stage, and calculates the steering angular velocity in real time based on the lateral offset (proportional-integral-differential term coordination), and dynamically adjusts the heading through the differential wheel to improve the lateral positioning accuracy; if the continuous o-cycle offset is less than the threshold, the fork is triggered to lift, and if the maximum number of iterations is exceeded, an alarm is issued; the whole process integrates path optimization and real-time correction to solve the problems of frequent steering in narrow spaces and large cumulative positioning errors of traditional methods, thereby improving the success rate of one-time calibration of the fork and adapting to complex road conditions.

[0014] An AGV pallet intelligent recognition and adaptive fork-picking control system based on 3D laser radar is used to implement the above-mentioned AGV pallet intelligent recognition and adaptive fork-picking control method based on 3D laser radar, including a point cloud data acquisition and processing module, a pallet posture recognition and analysis module, a cargo size detection module, and a path planning and picking module; The point cloud data acquisition and processing module acquires the original point cloud data of the pallet through 3D laser radar scanning, uses voxel filtering for noise reduction, effective range screening, and through-filtering, and combines it with standard pallet feature templates to build a 3D feature model library. The point cloud convolutional neural network extracts the side feature vectors of the pallet, providing a high-dimensional feature representation for subsequent classification, and achieving accurate point cloud feature extraction in complex industrial scenarios. The classification model construction and optimization module constructs an initial SVM classification model based on historical stack feature data and introduces the Gray Wolf Optimization Algorithm to dynamically adjust the learning rate parameters. By defining a fitness function, the global optimization of the model parameters is achieved during the iteration process, significantly improving the classification accuracy and generalization ability of stack / non-stack areas. The pallet posture recognition and analysis module uses a weighted RANSAC algorithm combined with Hausdorff distance feature matching to achieve high-precision plane fitting; PCA principal component analysis is used to calculate the long axis direction of the pallet, and the center coordinates and posture angle are obtained by combining covariance matrix decomposition; a dual-threshold warning mechanism is set to trigger a real-time alarm when the yaw angle or pitch angle exceeds the safe range, ensuring the reliability of posture detection; The cargo size detection module segments cargo point cloud clusters based on Euclidean clustering and uses the AABB algorithm to fit a minimum circumscribed cuboid. It dynamically calculates the maximum width / height of cargo and compares it with a preset threshold to detect size compliance. It also introduces 3D coordinate projection technology and combines it with the pallet cross-section point cloud to establish an accurate size calculation model, effectively solving the problem of cargo size estimation in obscured environments. The path planning and picking module is used to generate a picking path that conforms to the AGV's motion characteristics by improving the A* algorithm, integrating steering constraints and dual-weight heuristic functions; adopting multi-cycle PID closed-loop control, achieving precise positioning through dynamic calculation of lateral offset and differential wheel steering correction; setting multi-level stopping conditions and timeout alarm mechanisms to ensure the safety of fork operation and the integrity of the operation process.

[0015] (3) Beneficial effects The present invention has the following beneficial effects: Through the deep integration of 3D perception, intelligent algorithms and closed-loop control, this invention realizes the full-process autonomy and high-precision operation of AGV from pallet detection, path planning to safe forking, solving the core pain points of traditional technologies such as insufficient precision, high dependence on manual labor, poor dynamic adaptability and low safety, and providing reliable technical support for unmanned warehousing and industrial logistics.

[0016] The present invention introduces the Grey Wolf Optimization Algorithm into the SVM classifier parameter optimization process, dynamically adjusts the learning rate through the fitness function, avoids local optimality while ensuring classification accuracy, and obtains an optimized SVM pallet side classification model specifically suitable for classifying pallet point clouds; improves the recognition accuracy of the model in tilted and semi-occluded pallet scenarios, reduces the false detection rate, and significantly enhances the robustness under complex working conditions.

[0017] The present invention combines 3D lidar with a weighted RANSAC algorithm and PCA main direction analysis to reduce the detection error of the pallet center coordinate and the attitude angle error, significantly improving the detection accuracy compared with traditional lidar and visual SLAM. It also dynamically allocates point cloud weights based on the Hausdorff distance and prioritizes fitting key features such as pallet grooves, thereby improving the detection's anti-interference ability.

[0018] The present invention reduces cargo size measurement errors and improves the accuracy of over-limit determination by adopting Euclidean clustering segmentation and AABB circumscribed cuboid fitting algorithms. It also effectively distinguishes cargo from background noise through height constraints and dynamic thresholds, improving the success rate of non-standard cargo identification. Combined with a first-level alarm mechanism, it avoids operational risks caused by forking tilted pallets.

[0019] The present invention introduces an angular deviation term by improving the A algorithm, balancing distance and direction priorities, thereby improving path planning time. Combined with PID lateral deviation correction, the invention realizes precise fork centering calibration through an integral anti-saturation mechanism and a control cycle, thereby improving dynamic response speed. It is particularly suitable for narrow spaces and uneven road scenarios, reducing safety risks. A three-dimensional feature model library of pallets is constructed based on the ISO6780 specification, supporting plug-and-play recognition of mainstream standard pallets. It is compatible with a variety of alternative solutions such as ToF cameras and stereo vision, reducing hardware costs. The algorithm supports modular replacement and adapts to different AGV models and scenario requirements.

[0020] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the invention. For ordinary technicians in this field, they can also obtain drawings based on these drawings without paying any creative work.

[0022] Figure 1 Schematic diagram of the process of the AGV pallet intelligent recognition and adaptive fork control method based on 3D laser radar of the present invention; Figure 2 This is a module schematic diagram of the AGV pallet intelligent recognition and adaptive fork control system based on 3D laser radar in the present invention. DETAILED DESCRIPTION

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] In the description of the present invention, it should be understood that the terms "opening", "upper", "lower", "top", "middle", "inside" and the like indicating orientation or positional relationship are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the components or elements referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as limiting the invention.

[0025] Example 1: See also Figure 1 The present invention discloses an AGV pallet intelligent recognition and adaptive fork control method based on 3D laser radar, comprising the following steps: S1. Collect the point cloud data of the pallet, obtain the initial pallet point cloud data, and process it to obtain the final pallet point cloud data; use the point cloud convolutional neural network to extract the features of the midpoint cloud of the final pallet point cloud data to obtain the side feature vector of the pallet point cloud; Said S1 comprises the following steps: S11. Based on the standard pallet structure, set the standard pallet structure geometric feature point cloud template set ,in, a i Represents the first i A standard pallet structure geometric feature point cloud template, b Indicates the total number of standard pallet structure geometric feature point cloud templates; Constructing a three-dimensional side feature model library based on the standard pallet structure geometric feature point cloud template set; S12: The AGV autonomously navigates and moves until the fork is facing the pallet and the distance from the pallet is within 0.4-1 meters; the scanning cycle is set; Scan the real-time pallet through 3D laser radar and combine the scanning cycle to obtain the original pallet point cloud data ,in, c i Represents the first i The three-dimensional coordinates of the points, d Indicates the total number of points in the scanning cycle; ,in c ix 、 c iy and c iz Respectively represent the first i points x Coordinate values, y Coordinate values and z Coordinate value; S13, using voxel filtering to reduce noise on the original stack point cloud data to obtain reduced noise stack point cloud data; S14, set the AGV fork operating range constraint, construct an effective point cloud screening formula based on the AGV fork operating range constraint, and screen the effective point cloud in the noise-reduced pallet point cloud data from the AGV center according to the effective point cloud screening formula to obtain the effective pallet point cloud data. ,in, e i Indicates the first i The three-dimensional coordinates of the points, f Represents the total number of points in the valid pallet point cloud data, where e ix 、e iy and e iz Respectively represent the first i points x Coordinate values, y Coordinate values and z Coordinate value; The effective point cloud screening formula is as follows: ; in, a valid Represents valid point cloud data, AGV x and AGV y Respectively represent the AGV center x Coordinate values and y Coordinate value, α, β Indicates the effective range of the pallet constrained by the AGV fork operating range; S15, intercepting point data within a height range from the valid pallet point cloud data through straight-through filtering to obtain final pallet point cloud data; the straight-through filtering is as follows: ; in, e valid Represents the final stack point cloud data, λ, d Indicates the effective height range of the pallet constrained by the AGV fork operating range; S16. Extract features from the final pallet point cloud data based on a standard pallet 3D feature model library and in combination with a point cloud convolutional neural network to obtain a pallet side point cloud feature vector; S2. Construct an initial SVM classification model, use the historical feature vector combined with the optimization algorithm to train and optimize the initial SVM classification model, and obtain an optimized SVM pallet side classification model; input the pallet side point cloud feature vector into the optimized SVM pallet side classification model to obtain a classification result; The S2 comprises the following steps: S21, collecting a large amount of historical pallet point cloud feature vector data and flying pallet point cloud feature vector data to obtain historical feature vectors; S22, build the initial SVM classification model, set the learning rate of the initial SVM classification model; set the training accuracy threshold of the initial SVM classification model to r 1. The training accuracy is r 2; S23, using the historical feature vector to train the initial SVM classification model, combining the optimization algorithm to find the learning rate of the initial SVM classification model during the training process to obtain an optimal solution; using the optimal solution as the learning rate of the initial SVM classification model to obtain an optimized SVM stack side classification model; In the S23 training process, the optimization algorithm is combined to find the learning rate of the initial SVM classification model to obtain the optimal solution, which includes the following steps: S231. Construct a gray wolf population and set the size of the gray wolf population to m , then the gray wolf population is represented by ,in, u i Indicates the gray wolf population i Gray wolves; set the maximum number of optimization iterations; S232, randomly setting the initial position of the gray wolf population according to the learning rate of the initial SVM classification model, and obtaining the initial position set of the gray wolf population ;in v i Represents the first i The position of the gray wolf indicates the distance to the prey; S233, according to the training accuracy threshold p 1 and training accuracy p 2. Define the fitness function of the gray wolf position in the gray wolf population. The fitness function formula is as follows: ; Z represents the fitness function, cs Indicates the offset; S234, performing an iterative operation on the initial position set of the gray wolf population, wherein the higher the fitness value, the better the position; during each round of iteration, calculating the fitness value of each position in the initial position set of the gray wolf population according to the fitness function, and updating the position of each gray wolf in the initial position set of the gray wolf population in descending order of fitness value, and obtaining the best individual gray wolf position in the gray wolf population and the global best gray wolf position during each round of iteration; S235, repeat S234, when the maximum number of optimization iterations is reached, stop the iteration, and take the global best gray wolf position as the optimal solution; S3. When the classification result is a candidate pallet area, the RANSAC algorithm is used to perform plane fitting on the final pallet point cloud data to obtain a pallet cross-section point cloud; the pallet cross-section point cloud is analyzed to obtain the pallet center coordinates and attitude angle; The S3 includes the following steps: S31. Calculate the feature matching degree of each point in the candidate pallet point cloud by using the Hausdorff distance to obtain a feature matching degree set. The feature matching degree calculation formula is as follows: ; in w i represents the feature matching degree, d i Indicates the first i The Euclidean distance between each point and the nearest neighbor feature point in the standard pallet 3D feature model library, σ represents the scale parameter; σ is dynamically adjusted according to the pallet feature size, and is usually set to 0.05 times the pallet width; S32, setting weights for points in the final pallet point cloud data according to the feature matching degree set to obtain a final pallet point weight set; iterating the RANSAC algorithm on the points in the final pallet point cloud data according to the final pallet point weight set to obtain a pallet side point cloud; In the step S32, the RANSAC algorithm is iterated on the points in the final pallet point cloud data according to the final pallet point weight set to obtain the pallet side point cloud, which includes the following steps: S321. Calculate the maximum number of iterations of the RANSAC algorithm for the final stack point cloud data to obtain the maximum number of RA iterations. The calculation formula is as follows: ; in K Indicates the maximum number of RA iterations, or Represents the confidence of the RANSAC algorithm, e Indicates the initial estimate of the proportion of all points in the final pallet point cloud data that are internal points; S322. Randomly select three points from the final pallet point cloud data according to the final pallet point weight set to obtain selected points; iterate the selected points using the RANSAC algorithm, and update the number of inliers in each iteration to obtain an updated cross-sectional point cloud; S323, repeat S322, when the maximum number of RA iterations is reached, stop the iteration, and obtain the stack side point cloud data And the plane equation of the point cloud on the side of the pallet; where, p i Indicates the first i The three-dimensional coordinates of the points, q Indicates the total number of points in the valid pallet point cloud data; ,in p ix 、 p iy and p iz Respectively represent the first i pointsx Coordinate values, y Coordinate values and z Coordinate value; S33. Analyze the covariance matrix of the point cloud of the side surface of the pallet using PCA to obtain the direction of the long axis of the pallet; collect the inliers in the point cloud of the side surface of the pallet to obtain an inlier set; Calculate the mean of the three-dimensional coordinates of each interior point in the interior point set to obtain the center coordinates of the pallet , F x 、 F y and F z Respectively represent the center of the pallet x Coordinate values, y Coordinate values and z Coordinate value; S34. According to the plane equation of the point cloud on the side of the pallet, the plane normal vector is obtained ;in, n 1. n 2 and n 3 represents the normal vector x 、 y 、 z Components on the axis; The yaw angle and pitch angle are calculated based on the plane normal vector, and the yaw angle and pitch angle together constitute the attitude angle; the yaw angle calculation formula is as follows: ; Among them, it means i represents the yaw angle; The pitch angle calculation formula is as follows: ; in, P Indicates the pitch angle; Set the yaw angle threshold and pitch angle threshold. When the yaw angle is greater than the yaw angle threshold or the pitch angle is greater than the pitch angle threshold, a level 1 warning is triggered. S4. Collect the point cloud of the cargo and perform Euclidean clustering to obtain cargo point cloud clusters; fit the minimum circumscribed cuboid of the cargo point cloud clusters to obtain the cargo size; The S4 comprises the following steps: S41, the intelligent main control host first controls the fork to descend to ensure that the fork can be normally inserted under the pallet when picking up the goods; through 3D laser radar scanning, the original point cloud data of the goods is obtained ,in, f i Indicates the first i The three-dimensional coordinates of the points, g Indicates the total number of points in the original point cloud data of the cargo; ,inf ix 、 f iy and f iz Respectively represent the first i points x Coordinate values, y Coordinate values and z Coordinate value; S42, set the clustering condition, the clustering condition is that if two points in the original point cloud data of the goods f i 、 f j Satisfy the clustering formula ; are classified into the same cluster; among them, f jx 、 f jx and f jx Respectively represent the first j points x Coordinate values, y Coordinate values and z Coordinate value; Based on the pallet cross-section point cloud, the original point cloud data of the cargo above is subjected to Euclidean clustering according to the clustering conditions to obtain the cargo point cloud clusters. ,in, k i Indicates the first i The three-dimensional coordinates of the points, k Indicates the total number of clusters in the cargo point cloud cluster; S43. Use the AABB algorithm to combine the cargo point cloud cluster to fit the minimum circumscribed cuboid of the cargo to obtain the cargo size. The cargo size includes the maximum width and maximum height. The maximum width is calculated as follows: ; in, W goods is the maximum width of the cargo observed from the AGV’s perspective, max ( p ix ) represents the maximum value of all points in the pallet cross-section point cloud on the x-coordinate; minutes ( p ix ) means that all points in the pallet cross-section point cloud are x The minimum value on the coordinate; The maximum height calculation formula is as follows, ; in, H goods It is the maximum height of the cargo observed from the AGV’s perspective. max ( p iz ) represents the maximum value of all points in the pallet cross-section point cloud on the z coordinate; minutes ( p iz ) means that all points in the pallet cross-section point cloud are z The minimum value on the coordinate; Set cargo height threshold and cargo width threshold; when W goods >Cargo height threshold or H goods When the cargo width exceeds the threshold, it indicates that the cargo size is too large and triggers a level 1 alarm. S5. When the cargo size is within the threshold, the pallet position is obtained based on the pallet center coordinates and attitude angle. Based on the pallet position and the current position of the AGV, an improved A* algorithm is used to generate a pickup path. The AGV moves according to the pickup path, performing PID correction during movement until calibration is complete, and then starting the fork operation. The S5 comprises the following steps: S51, obtaining the pallet posture according to the pallet center coordinates and posture angle; collecting the AGV current posture, wherein the current posture includes the AGV current position and orientation angle; S52, setting the current posture of the AGV as the starting node and the target node as the pallet posture; Set a node expansion rule, wherein the node expansion rule is a fixed moving step length and a fixed steering angle for the steering action; for each steering action, the position of the AGV is updated; the fixed steering angles are a fixed angle for left turning, straight going, and a fixed angle for right turning; S53, simulating the AGV to move from the starting node to the target node according to the node expansion rule; S54. Set the weight of balancing distance and direction priority H 1. H 2. Constructing a heuristic function based on the weight of the balance distance and direction priority; At each node, the AGV generates a subnode set, which contains multiple subnodes that the AGV will go to next. The heuristic function is used to calculate the comprehensive cost value of each subnode in the subnode set, and the subnode with the smallest comprehensive cost value is selected as the next node to be expanded. The heuristic function formula is as follows: ; in, D Represents the Euclidean distance from the child node to the target node, N x 、 N y Represents the child nodes x Axis coordinate values and y Axis coordinate values; Indicates the absolute difference between the node orientation of the child node and the target orientation, H Indicates the comprehensive cost value of the child node; S55. Set the target distance threshold and the vehicle head deviation threshold; repeat S54. When the distance between the AGV and the target node is less than or equal to the target distance threshold or the vehicle head deviation is less than or equal to the vehicle head deviation threshold, stop the calculation and obtain the pickup path; and execute the pickup according to the pickup path. S56, set the adjustment cycle, set the PID control distance threshold, when the distance between the AGV and the pallet center coordinate is less than the PID control distance threshold, collect the fork center coordinate data in each adjustment cycle to obtain the fork center coordinate , O x 、 O y and O z Respectively represent the center of the pallet x Coordinate values, y Coordinate values and z Coordinate value; Calculate the lateral offset between the fork center coordinate and the pallet center coordinate to obtain the lateral offset calculation result for each cycle; the lateral offset calculation formula is as follows: ; Among them, W E t Indicates time t The calculation results of the lateral offset when F xt Indicates time t The target lateral position of the AGV at O xt Indicates time t The target lateral position of the AGV at time ; Set the lateral offset threshold; when the lateral offset calculation result of each cycle is greater than or equal to the lateral offset threshold, the steering angular velocity is obtained based on the lateral offset calculation result of each cycle and through the proportional term, integral term, and differential term. The differential wheel is driven to adjust the direction according to the steering angular velocity; otherwise, no adjustment is performed. The comprehensive calculation formula is as follows: ; in, P t represents the steering angular velocity, K p 、 K i and K d They represent the proportional coefficient, integral coefficient and differential coefficient of PID control respectively; KpEt represents the current deviation, Represents the historical cumulative deviation, Indicates the trend of forecast deviation; S57, set the maximum number of repetitions; repeat S56, if continuous o If the lateral deviation of each cycle is less than the lateral deviation threshold, the calibration is considered complete, the correction is stopped and the fork operation is started; if the lateral deviation is still less than the lateral deviation threshold when the maximum number of repetitions is reached, a timeout alarm is triggered; the fork operation is that after the AGV reaches the target point of the pallet and the pickup switch is triggered, the intelligent control host controls the fork lifting mechanism to rise; the AGV autonomously navigates to the target point according to the set unloading point to unload.

[0026] Example 2

[0027] See also Figure 2 , an AGV pallet intelligent recognition and adaptive fork-picking control system based on 3D laser radar, which is used to implement the above-mentioned AGV pallet intelligent recognition and adaptive fork-picking control method based on 3D laser radar, including a point cloud data acquisition and processing module, a classification model construction and optimization module, a pallet posture recognition and analysis module, a cargo size detection module, and a path planning and picking module; The point cloud data acquisition and processing module acquires the original point cloud data of the pallet through 3D laser radar scanning, uses voxel filtering for noise reduction, effective range screening, and through-filtering, and combines it with standard pallet feature templates to build a 3D feature model library. The point cloud convolutional neural network extracts the side feature vectors of the pallet, providing a high-dimensional feature representation for subsequent classification, and achieving accurate point cloud feature extraction in complex industrial scenarios. The classification model construction and optimization module constructs an initial SVM classification model based on historical stack feature data and introduces the Gray Wolf Optimization Algorithm to dynamically adjust the learning rate parameters. By defining a fitness function, the global optimization of the model parameters is achieved during the iteration process, significantly improving the classification accuracy and generalization ability of stack / non-stack areas. The pallet posture recognition and analysis module uses a weighted RANSAC algorithm combined with Hausdorff distance feature matching to achieve high-precision plane fitting; PCA principal component analysis is used to calculate the long axis direction of the pallet, and the center coordinates and posture angle are obtained by combining covariance matrix decomposition; a dual-threshold warning mechanism is set to trigger a real-time alarm when the yaw angle or pitch angle exceeds the safe range, ensuring the reliability of posture detection; The cargo size detection module segments cargo point cloud clusters based on Euclidean clustering and uses the AABB algorithm to fit a minimum circumscribed cuboid. It dynamically calculates the maximum width / height of cargo and compares it with a preset threshold to detect size compliance. It also introduces 3D coordinate projection technology and combines it with the pallet cross-section point cloud to establish an accurate size calculation model, effectively solving the problem of cargo size estimation in obscured environments. The path planning and picking module is used to generate a picking path that conforms to the AGV's motion characteristics by improving the A* algorithm, integrating steering constraints and dual-weight heuristic functions; adopting multi-cycle PID closed-loop control, achieving precise positioning through dynamic calculation of lateral offset and differential wheel steering correction; setting multi-level stopping conditions and timeout alarm mechanisms to ensure the safety of fork operation and the integrity of the operation process.

[0028] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0029] The preferred embodiments of the invention disclosed above are intended only to help illustrate the invention. These preferred embodiments do not exhaust all details, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. AGV pallet intelligent recognition and adaptive fork control method based on 3D laser radar, characterized by: The following steps are involved: S1. Collect and process the point cloud data of the pallet to obtain the final pallet point cloud data; use the point cloud convolutional neural network to extract the features of the final pallet point cloud data to obtain the side feature vector of the pallet point cloud; S2. Build an initial SVM classification model, use the historical feature vector combined with the optimization algorithm to train and optimize the initial SVM classification model, and obtain an optimized SVM stack side classification model; Input the feature vector of the pallet side point cloud into the optimized SVM pallet side classification model to obtain the classification result; S3. When the classification result is a candidate pallet area, the RANSAC algorithm is used to perform plane fitting on the final pallet point cloud data to obtain the pallet cross-section point cloud; Analyze the pallet cross-section point cloud to obtain the pallet pose; S4. Collect the point cloud of the cargo and perform Euclidean clustering to obtain cargo point cloud clusters; fit the minimum circumscribed cuboid of the cargo point cloud clusters to obtain the cargo size; S5. When the cargo size is within the cargo size threshold, the AGV moves according to the pallet posture, the current AGV posture and the pickup path generated by the improved A* algorithm. During the movement, PID correction is performed until the calibration is completed, and the fork operation is started.

2. The AGV pallet intelligent identification and adaptive fork control method based on 3D laser radar according to claim 1 is characterized in that: Said S1 comprises the following steps: S11. Based on the standard pallet structure, set a standard pallet structure geometric feature point cloud template set; Constructing a three-dimensional side feature model library based on the standard pallet structure geometric feature point cloud template set; S12: The AGV autonomously navigates and moves until the fork is facing the pallet and the distance from the pallet is within 0.4-1 meters; the scanning cycle is set; Scan the real-time pallet using 3D laser radar combined with a scanning cycle to obtain the original pallet point cloud data; S13, using voxel filtering to reduce noise on the original stack point cloud data to obtain reduced noise stack point cloud data; S14, setting an AGV fork operating range constraint, and filtering valid point clouds in the denoised pallet point cloud data using a valid point cloud screening formula constructed using the AGV fork operating range constraint to obtain valid pallet point cloud data; S15, intercepting point data within a height range from the valid pallet point cloud data through straight-through filtering to obtain final pallet point cloud data; S16. Based on the standard pallet 3D feature model library and in combination with the point cloud convolutional neural network, feature extraction is performed on the final pallet point cloud data to obtain the pallet side point cloud feature vector.

3. The AGV pallet intelligent identification and adaptive fork control method based on 3D laser radar according to claim 1 is characterized in that: The S2 comprises the following steps: S21, collecting a large amount of historical pallet point cloud feature vector data and flying pallet point cloud feature vector data to obtain historical feature vectors; S22, constructing an initial SVM classification model, setting a learning rate, a training accuracy threshold, and a training accuracy of the initial SVM classification model; S23. Use the historical feature vector to train the initial SVM classification model. During the training process, combine the optimization algorithm to find the learning rate of the initial SVM classification model to obtain the optimal solution; use the optimal solution as the learning rate of the initial SVM classification model to obtain an optimized SVM stack side classification model.

4. The AGV pallet intelligent identification and adaptive fork control method based on 3D laser radar according to claim 3 is characterized in that: In the S23 training process, the optimization algorithm is combined to find the learning rate of the initial SVM classification model to obtain the optimal solution, which includes the following steps: S231. Construct a gray wolf population and set the maximum number of optimization iterations; S232, randomly setting the initial position of the gray wolf population according to the learning rate of the initial SVM classification model to obtain an initial position set of the gray wolf population; S233. Defining a fitness function for the gray wolf position according to the training accuracy threshold and the training accuracy; S234, iterating the initial position set of the gray wolf population, calculating the fitness value of each position in the initial position set of the gray wolf population, updating the position of each gray wolf in the initial position set of the gray wolf population, and obtaining the best individual gray wolf position and the global best gray wolf position in each round of iteration; S235. Repeat S234. When the maximum number of optimization iterations is reached, stop the iteration and take the global best gray wolf position as the optimal solution.

5. The AGV pallet intelligent identification and adaptive fork control method based on 3D laser radar according to claim 1 is characterized in that: The S3 includes the following steps: S31, calculating the feature matching degree of each point in the candidate pallet point cloud to obtain a feature matching degree set; S32, using the feature matching set combined with the final pallet point cloud data to obtain a final pallet point weight set, and iterating the RANSAC algorithm on the points in the final pallet point cloud data to obtain a pallet side point cloud; S33, using PCA to analyze the covariance matrix of the point cloud of the side of the pallet to obtain the direction of the long axis of the pallet; collecting the inliers in the point cloud of the side of the pallet to obtain an inlier set; Calculate the mean of the three-dimensional coordinates of each interior point in the interior point set to obtain the coordinates of the center of the pallet; S34. Obtaining a plane normal vector based on a plane equation of the point cloud on the side of the pallet; The yaw angle and pitch angle are calculated based on the plane normal vector. The yaw angle and pitch angle together constitute the attitude angle, and the attitude angle and the center coordinate of the pallet together constitute the pallet posture; Set the yaw angle threshold and pitch angle threshold. When the yaw angle is greater than the yaw angle threshold or the pitch angle is greater than the pitch angle threshold, a level 1 warning is triggered.

6. The AGV pallet intelligent identification and adaptive fork control method based on 3D laser radar according to claim 5 is characterized in that: In the step S32, the RANSAC algorithm is iterated on the points in the final pallet point cloud data according to the final pallet point weight set to obtain the pallet side point cloud, which includes the following steps: S321, calculate and obtain the maximum number of RA iterations; S322. Randomly select three points from the final pallet point cloud data according to the final pallet point weight set to obtain selected points; iterate the selected points using the RANSAC algorithm, and update the number of inliers in each iteration to obtain an updated cross-sectional point cloud; S323. Repeat S322. When the maximum number of RA iterations is reached, stop the iteration to obtain the pallet side point cloud data and the plane equation of the pallet side point cloud.

7. The AGV pallet intelligent identification and adaptive fork control method based on 3D laser radar according to claim 1 is characterized in that: The S4 comprises the following steps: S41: The intelligent main control host first controls the fork to descend to ensure that the fork can normally fork under the pallet when picking up the goods; and obtains the original point cloud data of the goods; S42, setting clustering conditions; Based on the pallet cross-section point cloud and the clustering conditions, Euclidean clustering is performed on the original point cloud data of the cargo above to obtain cargo point cloud clusters; S43. Use the AABB algorithm to combine the cargo point cloud cluster and fit the minimum circumscribed cuboid of the cargo to obtain the cargo size; the cargo size includes the maximum cargo width and the maximum cargo height; Set cargo height threshold and cargo width threshold; when the maximum cargo height > cargo height threshold or the maximum cargo width > cargo width threshold, a level 1 alarm is triggered.

8. The AGV pallet intelligent identification and adaptive fork control method based on 3D laser radar according to claim 1 is characterized in that: The S5 comprises the following steps: S51, collecting the current position of the AGV; S52, setting the current posture of the AGV as the starting node and the target node as the pallet posture; Set node expansion rules; S53, simulating the AGV to move from the starting node to the target node according to the node expansion rule; S54, setting the weight of the balance distance and the direction priority, and constructing a heuristic function according to the weight of the balance distance and the direction priority; At each node, the AGV generates a set of child nodes, which includes multiple child nodes that the AGV will move to next. It uses a heuristic function to calculate the comprehensive cost of each child node in the set, and selects the child node with the smallest comprehensive cost as the next node to be expanded. S55. Set the target distance threshold and the vehicle head deviation threshold; repeat S54. When the distance between the AGV and the target node is less than or equal to the target distance threshold or the vehicle head deviation is less than or equal to the vehicle head deviation threshold, stop the calculation and obtain the pickup path; and execute the pickup according to the pickup path. S56. Set an adjustment cycle and a PID control distance threshold. When the distance between the AGV and the center coordinate of the pallet is less than the PID control distance threshold, collect the fork center coordinate data in each adjustment cycle to obtain the fork center coordinate. Calculate the lateral offset between the fork center coordinate and the pallet center coordinate, and obtain the lateral offset calculation result for each cycle; Set the lateral offset threshold; when the lateral offset calculation result of each cycle is greater than or equal to the lateral offset threshold, the steering angular velocity is obtained through PID comprehensive calculation; the differential wheel is driven to adjust the direction according to the steering angular velocity; S57, set the maximum number of repetitions; repeat S56, if the lateral offset is less than the lateral offset threshold for o consecutive cycles, then the calibration is determined to be complete, the correction is stopped and the fork operation is started; If the lateral offset is still less than the lateral offset threshold when the maximum number of repetitions is reached, a timeout alarm is triggered.

9. AGV pallet intelligent recognition and adaptive fork-taking control system based on 3D laser radar, characterized by: Implementing the AGV pallet intelligent recognition and adaptive fork picking control method based on 3D laser radar as described in any one of claims 1-8, the system includes a point cloud data acquisition and processing module, a pallet posture recognition and analysis module, a cargo size detection module, and a path planning and picking module.

10. A storage medium, characterized in that: A program is stored thereon, and when the program is executed by the processor, the AGV pallet intelligent recognition and adaptive forking control method based on 3D laser radar as described in any one of claims 1 to 8 is implemented.

Images