Material cage taking and placing method based on double single-line laser radars

Abstract

The invention belongs to the technical field of AGVs, and particularly relates to a method for cooperatively controlling two radars installed on the upper portion and the lower portion of a fork arm to selectively obtain two-dimensional point cloud data of a near-end material cage or a far-end material cage according to goods taking, goods placing or stacking instructions based on the method. According to the method, point clouds are preprocessed and clustered, pairing verification is carried out by utilizing a fixed distance between two positioning parts of the material cage, key feature points of the positioning parts are identified, and then the pose of the material cage in a radar coordinate system is calculated. And after the information is unified to a global map coordinate system through coordinate transformation, the target pose of the AGV is calculated by adopting a corresponding strategy according to a task mode, and finally the AGV is controlled to move in place and a fork arm is operated to complete operation. In the stacking process, closed-loop dynamic calibration can be carried out according to real-time data of the double radars. Through cooperative sensing of double radars and accurate recognition based on the positioning part, the problems that coverage of a single sensor is insufficient and vision is interfered by a grating are solved, and high-precision taking and placing and safe stacking of the material cages are achieved.

Description

Technical Field

[0001] This invention relates to the field of AGV technology, and mainly to a method for picking up and placing materials in a cage based on dual single-line lidar. Background Technology

[0002] As a standardized carrier widely used in warehousing and logistics, the unique perforated grid structure of material cages poses a fundamental challenge to automated identification technology. Currently, automated operation of material cages mainly relies on manual guidance or a single sensor (such as a vision camera or a single LiDAR). These methods have significant shortcomings: when facing the grid structure, the reflection and shadow effects generated by the metal frame under complex lighting conditions severely interfere with edge detection, and the background information seen through the grid gaps is more likely to be visually confused with the main body of the material cage, making it difficult for the system to stably extract the key structural feature of the positioning part, let alone obtain accurate three-dimensional posture information. In stacking operations, the positioning parts of the upper and lower material cages need to be strictly vertically aligned to ensure stability, but vision systems lack depth perception capabilities and cannot provide reliable vertical calibration basis, exhibiting obvious technical limitations.

[0003] Meanwhile, while existing solutions using single-line lidar can overcome the influence of lighting, they face new challenges due to the physical characteristics of the cage structure. When a single-layer laser scanning line passes through the grid, part of the laser beam is reflected by the metal frame to form an effective point cloud, while the other part penetrates the gaps and hits background objects, creating interference data. This mixed point cloud pattern makes it difficult for the system to accurately distinguish between the actual structure of the cage and background noise, easily misjudging grid intersections as positioning points, leading to the failure of the positioning reference. More importantly, single-line lidar can only capture data from a single horizontal cross-section and cannot simultaneously acquire the complete spatial relationship between the nearby forklift cage and the distant ground cage. When the cage tilts slightly due to handling, the projected position of its positioning point on the scanning plane changes. The system cannot determine whether this is a positional shift or an attitude rotation, which can easily cause structural interference during forklift handling or stacking, posing safety hazards.

[0004] In summary, existing cage operation solutions using single-line LiDAR have significant limitations in terms of recognition accuracy, operational safety, and scene adaptability due to their limited scanning dimension, inability to overcome grid interference, and inability to simultaneously perceive multiple target layers. Further improvements are needed to achieve stable, accurate, and real-time perception and pose control of cages (especially the upper and lower cages in stacked scenarios). Summary of the Invention

[0005] The purpose of this invention is to provide a material cage picking and placing method based on dual single-line lidar, so as to overcome the problems of inaccurate material cage identification and unsafe stacking operation in the prior art.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A material cage picking and placing method based on dual single-line lidar, applied to an automated guided vehicle (AGV) with a lifting forklift, includes the following steps: receiving and responding to a work instruction for picking, placing, or stacking goods; coordinating the control of a first lidar and a second lidar deployed at different spatial locations on the robot platform to acquire perception data of a near-end first material cage and a far-end second material cage, respectively; processing the perception data to parse the feature information of the near-end first material cage and the far-end second material cage; determining the local pose of the near-end first material cage and the far-end second material cage relative to the corresponding lidar based on the feature information; fusing the local pose with the robot platform's own state information to calculate the adjustment target pose of the robot platform corresponding to the work instruction in a unified work reference frame; controlling the robot platform to move to the adjustment target pose, and operating the carrying mechanism to perform the material cage picking and placing operation corresponding to the work instruction.

[0008] This invention uses two single-line LiDARs deployed at different spatial positions on the forklift of an Automated Guided Vehicle (AGV) to acquire point cloud data of the near and far end cages, respectively. After filtering and clustering the point clouds, key feature points corresponding to the cage positioning parts are identified. The pose of the cage in the radar coordinate system is calculated based on the coordinates of the feature points, and then unified to the global map coordinate system through a coordinate transformation matrix. Depending on the task of picking up, placing, or stacking goods, the target pose to be reached by the AGV is calculated based on the inverse solution of the global pose of the cage on the ground and the standard loading posture, the compensation for the pose offset of the cage on the forklift, or the alignment relationship of the poses of the upper and lower cages. Finally, the AGV is controlled to move to the pose and operate the forklift to complete the operation. During the stacking process, the target pose can be dynamically corrected based on real-time sensing data. Compared with existing technologies, this invention can synchronously acquire spatial information of multi-layer cages through a special arrangement of dual radars, and achieve precise pose alignment of the cages during the picking, placing and stacking process based on feature recognition of the positioning unit and coordinate system transformation. This reduces dependence on ambient light and improves the robustness of identification and operational safety under the cage grid structure.

[0009] Furthermore, the robot platform is an automated guided vehicle (AGV), and the material carrying mechanism is a liftable forklift. Both the first and second lidars are mounted on the lower part of the forklift, with the second lidar located below the first lidar. The first lidar scans the material cages on its carrying surface; the second lidar scans the material cages in the ground area in front of the AGV. In the picking task, the lower lidar identifies the material cages on the ground; in the placing task, the upper lidar identifies the material cages on the forklift; and in the stacking task, both lidars are simultaneously used to identify the upper and lower layers of material cages. After filtering and clustering the collected point cloud, based on the fixed geometric relationship between the two positioning parts of the material cage, key feature points representing the positions of these two positioning parts are identified and paired, thereby calculating the pose of the material cage in the lidar coordinate system. Subsequently, the pose is unified to the global map coordinate system through a coordinate transformation matrix, and based on the matrix operation strategies corresponding to different task modes (picking, placing, stacking), the target pose required for the AGV to complete precise operations is derived. During the stacking process, the system can perform dynamic closed-loop calibration of the target pose based on the real-time scanning data of the dual radars to ensure that the upper and lower material cages are aligned before placement is performed.

[0010] Furthermore, acquiring perception data includes the following steps: when the operation instruction is to pick up goods, the second LiDAR is activated to scan the second cage on the ground; when the operation instruction is to release goods, the first LiDAR is activated to scan the first cage carried on the forklift; when the operation instruction is to stack, the first and second LiDARs are activated simultaneously to scan the first cage carried on the forklift and the second cage on the ground, respectively. This solution dynamically schedules radar resources at corresponding spatial locations for targeted scanning based on the operation instruction (picking up, releasing, or stacking), achieving precise matching between perception and task. Specifically, during picking up, only the second LiDAR is activated to scan the second cage on the ground; during releasing, only the first LiDAR is activated to scan the cage on the forklift to monitor its attitude deviation; during stacking, both upper and lower LiDARs are activated simultaneously to scan the upper and lower cages respectively, to simultaneously acquire their spatial relative relationship. This design achieves on-demand perception, ensuring complete coverage of key targets while avoiding redundant data acquisition and processing, thus improving the system's efficiency and adaptability in complex operation scenarios.

[0011] Furthermore, both the first and second cages have a top surface and a bottom surface, each with at least two positioning parts. The sensing data is two-dimensional point cloud data generated by the laser radar scanning the reflections of the positioning parts. The feature information includes position information of key feature points representing the positioning parts, identified based on cluster analysis of the two-dimensional point cloud data. During stacking operations, the system's control objective is to precisely align the positioning parts on the bottom surface of the upper cage with the positioning parts on the top surface of the lower cage. This solution does not rely on the easily disturbed overall outline recognition of the cage, but rather focuses on the precise positioning of the stable and rigid positioning parts at the bottom and top of the cage. In the two-dimensional point cloud generated after laser radar scanning, these positioning parts, as solid reflection sources, form local dense point cloud areas. By performing distance- and density-based cluster analysis on the point cloud data, independent point cloud clusters corresponding to each positioning part can be separated from background noise and other structural reflection points. Furthermore, the centroid of each cluster is calculated, and this centroid is defined as a key feature point representing the precise planar projection position of the corresponding positioning part. This principle transforms the complex problem of cage identification into the problem of detecting multiple stable, discrete entity features, thereby significantly improving the robustness of identification and positioning accuracy of cages with grid structures.

[0012] Furthermore, the positioning parts are located at the four ends of the upper and lower sides of the cage; the key feature point is the position information of the projection center of the positioning part on the lidar scanning plane. In this solution, four positioning parts are provided on both the top and bottom surfaces of the cage, of which two positioning parts located on the same side are used for identification and pose calculation by lidar scanning. In addition, the support feet of the cage are used as positioning parts on its bottom surface.

[0013] Further, the steps for processing the perceived data specifically include: S61: Filtering the original two-dimensional point cloud data based on preset distance and angle thresholds to remove background noise points; S62: Performing spatial clustering based on Euclidean distance on the filtered point cloud to obtain multiple candidate point cloud clusters; S63: Calculating the number of points in each candidate point cloud cluster, and filtering out candidate point cloud clusters that conform to the physical size characteristics of the positioning unit based on a preset point count threshold; S64: Calculating the centroid of each candidate point cloud cluster after filtering in step S63; S65: Based on the known spacing of the positioning unit and combined with a preset spacing matching tolerance threshold, pairing and calculating the distance between each centroid obtained in step S64; From all centroid pairs, selecting a pair of centroids whose centroid spacing matches the known spacing, and defining them as the first key feature point and the second key feature point respectively. First, the original point cloud is initially screened using distance and angle thresholds to filter out background points and noise that obviously do not conform to the location range of the positioning unit, thus achieving data purification. Next, spatial clustering algorithms are used to group physically adjacent points into clusters, thereby separating potential object reflection zones and obtaining candidate point cloud clusters. Then, based on the point cloud density characteristics corresponding to the actual physical size of the positioning unit, invalid clusters that are too large or too small are eliminated through point count threshold filtering, initially locking in candidate clusters that conform to the shape of the positioning unit. Finally, the core geometric constraint verification is introduced: using the fixed and known distance between the first and second positioning units, the filtered candidate clusters are paired and distances are calculated. Only candidate clusters whose centroid distance matches the known positioning unit distance and whose spatial position conforms to the general orientation of the cage are ultimately determined as target point cloud clusters representing the two positioning units, respectively, and their centroid coordinates are confirmed as the precise locations of the first and second key feature points. This principle, through multi-layered filtering from coarse to fine and combined with prior knowledge for final judgment, ensures the accuracy and anti-interference ability of feature point extraction in complex point cloud environments.

[0014] Furthermore, the step of determining the local pose of the first and second cages relative to the corresponding radar is as follows: Let the coordinates of the first key feature point be P. 0(x0,y0) The coordinates of the second key feature point are P. 1(x1,y1) ; Calculate the center position P of the cage (x,y) = (P0+P1) / 2; calculate the orientation angle θ of the cage = atan2(y0-y1,x0-x1); obtain the local pose (x,y,θ).

[0015] Furthermore, the unified operational reference system is a global map coordinate system with a fixed point in the storage area as its origin; the self-state information is the real-time pose of the robot platform in the global map coordinate system. This design establishes the global map coordinate system as a unified spatial reference. Its principle lies in converting and unifying the local cage poses perceived from different locations and times, as well as the robot platform's own motion state, into this global, static reference system for representation and computation. This allows the spatial relationship between the near and far cages, and the relative position of the robot and the cage, to be accurately quantified and compared within the same coordinate framework. The effect is the effective fusion and decoupling of multi-source heterogeneous spatial information, eliminating transformation errors and logical complexity caused by coordinate system inconsistencies. This provides a stable and consistent mathematical foundation for subsequent accurate calculation of the robot's target pose (especially for stacking tasks requiring alignment operations), and is a core prerequisite for the entire system to achieve high-precision spatial perception and coordinated control.

[0016] Furthermore, the fusion calculation step is implemented through coordinate transformation, including the following steps: using the pre-calibrated radar installation pose, the local pose (x, y, θ) is converted into a pose in the robot platform's base coordinate system; combined with the robot platform's real-time global pose, the pose in the base coordinate system is further transformed into the global map coordinate system to obtain the global pose information of the cage. This scheme integrates the cage position observed locally by the radar into the global map coordinate system through a coordinate transformation chain. This provides a unique and consistent absolute position reference for all subsequent task planning and precise control based on global position.

[0017] Furthermore, let the moments formed by the local poses be T. (laser-shelf) Let the radar installation pose matrix be T. (base-laser) Let T be the real-time global pose matrix of the robot platform. (map-base) The global pose matrix of the cage is T. (map-shelf) The coordinate transformation is achieved through the product of two-dimensional homogeneous transformation matrices, and the transformation formula is: T (map-shelf) =T (map-base) *T (base-laser) *T (laser-shelf) . Attached Figure Description

[0018] Figure 1 This is the process of the present invention. Figure 1 ;

[0019] Figure 2 This is the process of the present invention. Figure 2 . Detailed Implementation

[0020] The following describes a preferred embodiment of the present invention in conjunction with the accompanying drawings.

[0021] See Figure 1-2 This invention discloses a method for picking up and placing material cages based on dual single-line lidar, applied to an automated guided vehicle (AGV) with a lifting fork arm. The method includes the following steps: receiving and responding to a work instruction for picking up, placing, or stacking goods; coordinating the control of a first lidar and a second lidar deployed at different spatial locations on the robot platform to acquire perception data of a near-end first material cage and a far-end second material cage, respectively; processing the perception data to parse feature information of the near-end first material cage and the far-end second material cage; determining the local pose of the near-end first material cage and the far-end second material cage relative to the corresponding lidar based on the feature information; fusing the local pose with the robot platform's own state information to calculate the target pose of the robot platform corresponding to the work instruction in a unified work reference frame; controlling the robot platform to move to the target pose and operating the carrying mechanism to perform the material cage picking and placing operation corresponding to the work instruction.

[0022] In one embodiment, all radars are mounted on the lower part of the fork arm, and the second lidar is located below the first lidar. The first lidar is used to scan the cages on its bearing surface; the second lidar is used to scan the cages on the ground in front of the AGV. This solution uses two single-line lidars, installed vertically and vertically on the lower part of the AGV fork arm, to collect point cloud data for the first cage carried by the fork arm nearby and the cage to be operated on the ground further away. In the picking task, the second lidar is used to identify the cage on the ground; in the placing task, the first lidar is used to identify the cage on the fork arm; and in the stacking task, both radars are used simultaneously to identify the upper and lower layers of cages respectively. After filtering and clustering the collected point clouds, based on the fixed geometric relationship of at least two positioning parts (e.g., positioning parts) on the bottom or top surface of the cage, key feature points characterizing the positions of these positioning parts are identified and paired, thereby calculating the pose of the cage in the radar coordinate system. Subsequently, the pose is unified to the global map coordinate system through a coordinate transformation matrix. Based on the matrix operation strategies corresponding to different task modes (picking, placing, and stacking), the target pose required for the automated guided vehicle to complete the precise operation is derived. For stacking tasks, this target pose aims to precisely align the positioning parts on the bottom surface of the upper cage with the positioning parts on the top surface of the lower cage. During stacking, the system can perform dynamic closed-loop calibration of the target pose based on real-time scanning data from dual radars to ensure that the positioning parts of the upper and lower cages meet the preset alignment accuracy conditions before placement.

[0023] In one embodiment, acquiring sensing data includes the following steps: when the operation instruction is to pick up goods, the second lidar is activated to scan the second cage on the ground; when the operation instruction is to release goods, the first lidar is activated to scan the first cage carried on the forklift; when the operation instruction is to stack, the first lidar and the second lidar are activated simultaneously to scan the first cage carried on the forklift and the second cage on the ground, respectively. The principle of this solution in the sensing data acquisition stage is: dynamically scheduling radar resources at corresponding spatial locations for targeted scanning according to the operation instruction (picking up, releasing, or stacking), achieving precise matching between sensing and tasks. Specifically, when picking up goods, only the second lidar is activated to scan the second cage on the ground; when releasing goods, only the first lidar is activated to scan the cage on the forklift to monitor its attitude deviation; when stacking, both upper and lower lidars are activated simultaneously to scan the upper and lower cages respectively, to simultaneously acquire their spatial relative relationship. This design achieves on-demand sensing, ensuring complete coverage of key targets while avoiding redundant data acquisition and processing, improving the system's efficiency and adaptability in complex operation scenarios.

[0024] In one embodiment, the cage has a top surface and a bottom surface, each with at least two positioning parts located at the four ends of the cage's upper and lower sides. The sensing data is two-dimensional point cloud data generated by the lidar scanning the reflections of the positioning parts. The feature information includes the positional information of key feature points, each representing at least two positioning parts, identified through cluster analysis of the two-dimensional point cloud data. This solution does not rely on the easily disturbed overall outline recognition of the cage, but rather focuses on the precise positioning of the positioning parts, which are structurally stable and rigid on the top and bottom surfaces of the cage. In the two-dimensional point cloud generated after lidar scanning, these positioning parts, as solid reflection sources, form local dense point cloud areas. By performing distance- and density-based cluster analysis on the point cloud data, independent point cloud clusters corresponding to each positioning part can be separated from background noise and other structural reflection points. Furthermore, the centroid of each cluster is calculated, and this centroid is defined as a key feature point representing the precise planar projection position of the corresponding positioning part. This principle transforms the complex problem of cage identification into the problem of detecting multiple stable, discrete entity features, thereby significantly improving the robustness of identification and positioning accuracy of cages with grid structures. In stacking operations, by identifying and matching the positioning parts of the upper and lower cages, precise alignment of the bottom surface of the upper cage with the top surface of the lower cage can be achieved.

[0025] In one embodiment, the positioning part is located at the four ends of the upper and lower sides of the cage; the key feature point is the position information of the projection center of the positioning part on the lidar scanning plane.

[0026] In one embodiment, the steps for processing the perceived data specifically include: S61: filtering the original two-dimensional point cloud data based on preset distance and angle thresholds to remove background noise points; S62: performing spatial clustering based on Euclidean distance on the filtered point cloud to obtain multiple candidate point cloud clusters; S63: calculating the number of points in each candidate point cloud cluster, and filtering out candidate point cloud clusters that conform to the physical size characteristics of the positioning part based on a preset point count threshold; S64: calculating the centroid of each candidate point cloud cluster after filtering in step S63; S65: pairing and calculating the distance between each centroid obtained in step S64 based on the known spacing of the positioning part and a preset spacing matching tolerance threshold; selecting a pair of centroids whose centroid spacing matches the known spacing from all centroid pairs, and defining them as the first key feature point and the second key feature point, respectively. First, the original point cloud is initially screened using distance and angle thresholds to remove background points and noise that obviously do not conform to the location range of the positioning part, thereby achieving data purification. Next, spatial clustering algorithms are used to group physically adjacent points into clusters, thereby separating potential object reflection zones and obtaining candidate point cloud clusters. Then, based on the point cloud density characteristics corresponding to the actual physical size of the positioning unit, invalid clusters that are too large or too small are eliminated through point count threshold filtering, initially locking in candidate clusters that conform to the shape of the positioning unit. Finally, the core geometric constraint verification is introduced: using the fixed and known distance between the first and second positioning units, the filtered candidate clusters are paired and distances are calculated. Only candidate clusters whose centroid distance matches the known positioning unit distance and whose spatial position conforms to the general orientation of the cage are ultimately determined as target point cloud clusters representing the two positioning units, respectively, and their centroid coordinates are confirmed as the precise locations of the first and second key feature points. This principle, through multi-layered filtering from coarse to fine and combined with prior knowledge for final judgment, ensures the accuracy and anti-interference ability of feature point extraction in complex point cloud environments.

[0027] In one embodiment, the step of determining the local pose of the first and second cages relative to the corresponding radar is as follows: Let the coordinates of the first key feature point be P. 0(x0,y0) The coordinates of the second key feature point are P. 1(x1,y1) ; Calculate the center position P of the cage (x,y) = (P0+P1) / 2; calculate the orientation angle θ of the cage = atan2(y0-y1,x0-x1); obtain the local pose (x,y,θ).

[0028] In one embodiment, the unified operation reference system is a global map coordinate system with a fixed point in the warehouse area as the origin; the self-state information is the real-time pose of the robot platform in the global map coordinate system.

[0029] In one embodiment, the fusion calculation step is implemented through coordinate transformation, including the following steps: using the pre-calibrated radar installation pose, the local pose (x,y,θ) is converted into the pose in the robot platform base coordinate system; combined with the real-time global pose of the robot platform, the pose in the base coordinate system is further converted to the global map coordinate system to obtain the global pose information of the cage.

[0030] In one embodiment, the coordinate transformation is achieved by a series of multiplications of two-dimensional homogeneous transformation matrices, and the transformation formula is: T (map-shelf) =T (map-base) *T (base-shelf) *T (laser-laser) Among them, T (laser-shelf) Let T be a matrix composed of the local poses. (base-laser) To install the pose matrix for the radar, T (map-base) T is the real-time global pose matrix for the robot platform. (map-shelf) Let be the global pose matrix of the cage.

[0031] In one embodiment, the step of calculating and adjusting the target pose is distinguished according to the type of operation instruction as follows: For a picking instruction, the target pose of the robot platform is calculated by inverse kinematics based on the global pose of the target cage on the ground and the standard loading posture of the cage on the robot platform; for a placing instruction, the compensated target pose of the robot platform is calculated based on the global pose of the cage on the forklift (or its offset relative to the standard posture) and the global pose of the preset target placement point; for a stacking instruction, the alignment target pose of the robot platform that aligns the lower cage on the ground and the upper cage on the forklift is calculated based on the global poses of the lower cage on the ground and the upper cage on the forklift.

[0032] In one embodiment, T target (map-base) is calculated using the following formula: T target (map-base) = T down (map-shelf)×T up (base-shelf).inverse; where T down (map-shelf) represents the pose matrix of the lower-level material cage identified by the second lidar in the global map coordinate system; T up (base-shelf) represents the pose matrix of the upper cage of the forklift identified by the first lidar in the base coordinate system of the automated guided vehicle, and .inverse represents the inverse matrix.

[0033] Furthermore, during the stacking operation, the control steps include dynamic closed-loop calibration: as the robot platform moves towards the target pose, the target pose is recalculated and updated in real time based on the latest acquired perception data until a preset alignment accuracy condition is met, and then the placement action is performed. This dynamic closed-loop calibration, through continuous synchronous scanning and real-time calculation by dual radars during movement, forms a "perception-calculation-adjustment" closed-loop control loop to continuously compensate for the cumulative deviations caused by AGV movement errors, environmental disturbances, and changes in cage pose. This transforms traditional one-time open-loop positioning into progressive, precise alignment, effectively overcoming the risk of stack instability or tipping due to uncalibrated errors, and significantly improving the success rate and safety of stacking operations.

[0034] In one embodiment, a safety verification step is included before performing the picking or placing operation: the robot platform is paused, and the corresponding radar is reactivated for verification scanning. The verification results are compared with the predetermined plan, and the operation is only performed after safety is confirmed. This safety verification step establishes an active safety barrier before performing critical operations through a "pause-re-scan-compare" process. Its principle is that at the critical moment of the final action execution, the target state is independently verified one last time by radar, comparing real-time perception data with the theoretical expectations from the action planning stage. This mechanism can effectively detect and intercept operational risks that may be caused by sudden environmental changes, unexpected movement of the target object, or minor deviations that were not detected in previous identification stages. Its effect is to significantly improve the robustness and operational safety of the system, preventing misoperations, collisions, or cargo damage caused by information lag or sudden state changes. It adds a crucial AI-based "double-check" guarantee to the automated process, and is particularly suitable for high-value goods or industrial scenarios with stringent operational precision requirements.

[0035] Working principle:

[0036] (1) Formula Explanation

[0037] T (Parent-Child) :Child: Parent coordinates, Child: Child coordinates, corresponding 3x3 matrix:

[0038]

[0039] (2) Examples:

[0040]

[0041] From T (base-shelf)Extract P (base-shelf) (x bs ,y bs ,θ bs ):

[0042] x bs =x bl +x ls *cosθ bl -y ls *sinθ bl

[0043] y bs =y bl +x ls *sinθ bl +y ls *cosθ bl

[0044] θ bs =θ bl +θ ls

[0045] From P( base-shelf) (x bs ,y bs ,θ bs ) Convert to T (base-shelf) :

[0046]

[0047] The "parent coordinates" mentioned above refer to the target coordinate system that serves as the spatial reference, i.e., the destination of the coordinate transformation; while the "child coordinates" refer to the source coordinate system that needs to be located and described, i.e., the starting point of the transformation. Transformation matrix T (父坐标-子坐标) The mathematical meaning of T is to describe the precise position and orientation (i.e., translation and rotation angle) of the "sub-coordinate system" within the "parent coordinate system". For example, T (base-laser) This represents the installation pose of the lidar coordinate system (child) relative to the AGV base coordinate system (parent); through chain multiplication T (map-shelf) =T (map-base) ×T (base-laser) ×T (laser-shelf) This allows the material cage to be gradually transformed from local radar coordinates (sub) to global map coordinates (final parent coordinates), enabling the fusion and calculation of all spatial data under a unified reference system.

[0048] This invention utilizes two single-line LiDARs deployed at different spatial positions on the forklift of an Automated Guided Vehicle (AGV) to acquire point cloud data of the near and far end cages, respectively. After filtering and clustering the point clouds, key feature points corresponding to the cage positioning section are identified. The pose of the cage in the radar coordinate system is calculated based on the coordinates of the feature points, and then unified to the global map coordinate system through a coordinate transformation matrix. Depending on the task of picking up, placing, or stacking, the target pose to be reached by the AGV is calculated based on the inverse solution of the global pose of the cage on the ground and the standard loading posture, compensation for the pose offset of the cage on the forklift, or the alignment relationship of the poses of the upper and lower cages. Finally, the AGV is controlled to move to the target pose and operate the forklift to complete the operation. During the stacking process, the target pose can be dynamically corrected based on real-time sensing data. Compared with the prior art, this invention can acquire spatial information of multiple cages simultaneously through dual radars, and achieve precise pose alignment of the cages during picking up, placing, and stacking based on feature recognition of the positioning section and coordinate system transformation. This reduces the dependence on ambient light and improves the robustness of identification and operational safety under the cage grid structure.

[0049] Based on the disclosure and teachings of the foregoing specification, those skilled in the art can make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and some modifications and changes to the present invention should also fall within the protection scope of the claims of the present invention. Furthermore, although some specific terms are used in this specification, these terms are only for convenience of explanation and do not constitute any limitation on the present invention.

Claims

1. A method for loading and unloading a material cage based on dual single-line lidar, applied to an automated guided vehicle (AGV) with a lifting forklift, characterized in that, Includes the following steps: Receive and respond to operation instructions for picking up, placing or stacking goods, and coordinate control of the first and second lidars deployed at different spatial locations on the robot platform to acquire perception data of the first cage at the near end and the second cage at the far end, respectively. The sensed data is processed to extract the feature information of the first and second cages; Based on the aforementioned feature information, the local poses of the first and second cages relative to their respective radars are determined. By integrating the local pose with the robot platform's own state information, the target pose of the robot platform corresponding to the operation instruction in a unified operation reference frame is calculated. Control the robot platform to move to the target position, and operate the carrying mechanism to perform the material cage picking and placing operation corresponding to the operation instruction.

2. The method for handling and placing the material cage according to claim 1, characterized in that, The robot platform is an automated guided vehicle (AGV), and the material carrying mechanism is a liftable forklift. The first and second laser radars are both installed on the lower part of the forklift, with the second laser radar located below the first laser radar. The first laser radar is used to scan the material cage on its carrying surface, and the second laser radar is used to scan the material cage in the ground area in front of the AGV.

3. The method for handling and placing the material cage according to claim 1, characterized in that, Acquiring sensory data includes the following steps: When the work instruction is to pick up goods, the second lidar is activated to scan the second material cage on the ground; When the work instruction is to release the goods, the first laser radar is activated to scan the first material cage carried on the fork arm. When the operation instruction is stacking, the first lidar and the second lidar are activated simultaneously to scan the first cage on the forklift and the second cage on the ground, respectively.

4. The method for handling and placing materials in a cage according to claim 1, characterized in that, Both the first and second material cages have a top surface and a bottom surface, and each of the top and bottom surfaces is provided with at least two positioning parts; the sensing data is two-dimensional point cloud data generated by the laser radar scanning the reflection of the positioning parts; the feature information includes the position information of key feature points that respectively characterize at least two positioning parts, identified based on cluster analysis of the two-dimensional point cloud data. During the stacking operation, the positioning part on the bottom surface of the first cage is aligned with the positioning part on the top surface of the second cage by control.

5. The method for handling and placing the material cage according to claim 4, characterized in that, The positioning part is located at the four ends of the upper and lower sides of the cage; the key feature point is the position information of the projection center of the positioning part on the laser radar scanning plane.

6. The method for handling and placing the material cage according to claim 5, characterized in that, The specific steps for processing the sensed data include: S61: Based on preset distance and angle thresholds, filter the original two-dimensional point cloud data to remove background noise points; S62: Perform spatial clustering based on Euclidean distance on the filtered point cloud to obtain multiple candidate point cloud clusters; S63: Calculate the number of points in each candidate point cloud cluster, and select candidate point cloud clusters that meet the physical size characteristics of the positioning part according to the preset point number threshold. S64: Calculate the centroid of each candidate point cloud cluster after screening in step S63; S65: Based on the known spacing of the positioning part and combined with the preset spacing matching tolerance threshold, pair up and calculate the distance between each centroid obtained in step S64; from all centroid pairs, select a pair of centroids whose centroid spacing matches the known spacing, and define them as the first key feature point and the second key feature point respectively.

7. The method for handling and placing the material cage according to claim 6, characterized in that, The steps for determining the local pose of the first and second feed cages relative to their respective radars are as follows: Let the coordinates of the first key feature point be P. 0(x0,y0) The coordinates of the second key feature point are P. 1(x1,y1) ; Calculate the center position P of the cage (x,y) = (P0 + P1) / 2; Calculate the orientation angle θ of the cage = atan2(y0-y1,x0-x1); The local pose (x, y, θ) is obtained.

8. The method for handling and placing the material cage according to claim 7, characterized in that, The unified operation reference system is a global map coordinate system with a fixed point in the warehouse area as the origin; the self-state information is the real-time pose of the robot platform in the global map coordinate system.

9. The method for handling and placing a material cage according to claim 8, characterized in that, The fusion calculation step is achieved through coordinate transformation, including the following steps: Using the pre-calibrated radar installation pose, the local pose (x, y, θ) is converted into a pose in the robot platform base coordinate system; By combining the real-time global pose of the robot platform, the pose in the base coordinate system is further transformed to the global map coordinate system to obtain the global pose information of the cage.

10. The method for handling and placing a material cage according to claim 8, characterized in that, a The moments formed by the local poses are T (laser-shelf) Let the radar installation pose matrix be T. (base-laser) Let T be the real-time global pose matrix of the robot platform. (map-base) The global pose matrix of the cage is T. (map-shelf) The coordinate transformation is achieved through the product of two-dimensional homogeneous transformation matrices, and the transformation formula is: T (map-shelf) =T (map-base) *T (base-laser) *T (laser-shelf) .

Images