Submerged fork type mobile robot self-adaptive tray obtaining method and device

Abstract

The invention provides a self-adaptive tray obtaining method and device for a submerged fork type mobile robot, and belongs to the technical field of visual recognition. The method comprises the steps that a panoramic camera is adopted to shoot a full-view image within a preset circumferential distance of the current position of the submerged fork type mobile robot, the full-view image is analyzed, a target operation tray within the preset circumferential distance is recognized, the real posture is judged, and a navigation path is generated according to the real posture and the current position. And the submerging fork type mobile robot moves according to the navigation path, continuously receives multi-sensor data when the submerging fork type mobile robot moves, analyzes the multi-sensor data according to the real posture, dynamically corrects the mobile positioning of the submerging fork type mobile robot, and inserts or forks the target operation tray by the submerging fork type mobile robot. A double-layer sensing framework of cloud sensing and edge guiding is constructed, and the complete process of the submerged fork type mobile robot from long-distance recognition to short-distance accurate butt joint is achieved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of visual recognition, and relates to a method and device for adaptively acquiring a pallet by a fork-entering mobile robot. BACKGROUND

[0002] At present, the fork-entering mobile robot mainly relies on 2D vision, Lidar or single sensor technology in pallet recognition and positioning. These methods have obvious defects: poor environmental adaptability, low recognition rate and robustness of the 2D vision system which is easily disturbed by light changes, complex background and shadows; the Lidar is sensitive to the structure of the bottom of the pallet and is prone to misjudgment; three-dimensional information is missing: the 2D technology is difficult to obtain the depth and attitude angle information (such as azimuth) of the pallet, resulting in insufficient positioning accuracy and affecting the success rate of fork taking; the contradiction between precision and speed: the traditional method often needs complex algorithms and a large amount of calculation to improve the accuracy, which is difficult to meet the real-time requirement (such as responding within 700ms). These defects lead to positioning errors, low efficiency and even operation failure of the robot in complex scenes. SUMMARY

[0003] The application aims at the above-mentioned problems existing in the prior art and provides a method for adaptively acquiring a pallet by a fork-entering mobile robot.

[0004] The object of the application can be achieved by the following technical scheme: a method for adaptively acquiring a pallet by a fork-entering mobile robot, comprising: adopting a panoramic camera to take a full-view image within a preset circumferential distance of the current position of the fork-entering mobile robot; analyzing the full-view image, identifying a target work pallet within the preset circumferential distance and judging a real attitude; generating a navigation path according to the real attitude and the current position, and moving the fork-entering mobile robot according to the navigation path; continuously receiving multi-sensor data of the fork-entering mobile robot during movement; analyzing the multi-sensor data according to the real attitude, dynamically correcting the movement and positioning of the fork-entering mobile robot, and inserting or forking the target work pallet.

[0005] As an optional embodiment of the application, the analysis of the full-view image, the identification of the target work pallet within the preset circumferential distance and the judgment of the real attitude of the pallet comprise: adopting a YOLO model to identify the target work pallet within the full-view image and outputting a two-dimensional bounding box position; performing attitude regression branch analysis on the target work pallet within the two-dimensional bounding box position to obtain a preliminary attitude estimation; The preliminary pose estimation is judged by matching a pre-established tray point cloud template and a minimum Euclidean error to obtain a real pose of the target work tray, wherein the real pose comprises six degrees of freedom information of the target work tray.

[0006] As an optional embodiment of the present application, before continuously receiving the multi-sensor data during movement of the submerged fork mobile robot, the following steps are further included: selecting multi-sensors according to the application environment of the submerged fork mobile robot; establishing a corresponding mathematical model for each type of sensor data; integrating the mathematical models based on state space to uniformly describe each type of sensor data.

[0007] As an optional embodiment of the present application, continuously receiving the multi-sensor data during movement of the submerged fork mobile robot includes: selecting sensors to construct a measurement model, wherein the sensors are odometers, IMUs and laser radars; analyzing the measurement model to obtain a fused positioning value of the sensors.

[0008] As an optional embodiment of the present application, analyzing the measurement model to obtain a fused positioning value of the sensors includes: based on a maximum entropy criterion, deriving an expression form of a filtering gain according to the measurement model; based on the expression form, calculating local estimation values under different sensor combinations by using innovation statistical distance; weighting and fusing the local estimation values to obtain a fused positioning value.

[0009] As an optional embodiment of the present application, analyzing the multi-sensor data according to the real pose of the tray and dynamically correcting the movement positioning of the submerged fork mobile robot includes: obtaining a grid map of the full-view image constructed based on SLAM; performing global positioning correction according to the fused positioning value and the grid map to update the local position of the submerged fork mobile robot.

[0010] As an optional embodiment of the present application, the submerged fork mobile robot inserting or forking the target work tray includes: when the submerged fork mobile robot moves to a preset distance in front of the target work tray, entering an edge guiding stage; in the edge guiding stage, scanning surface feature point position information of the target work tray and adjusting the navigation path according to the surface feature point position information.

[0011] The application further provides a tray adaptive acquisition device for a forklift mobile robot, comprising: An image acquisition module is configured to acquire a full-view image within a preset circumferential distance of a current position of the forklift mobile robot by using a panoramic camera. An attitude analysis module is configured to analyze the full-view image, identify a target work tray within the preset circumferential distance, and determine a real attitude. A movement module is configured to generate a navigation path according to the real attitude of the tray and the current position, and the forklift mobile robot moves according to the navigation path. A data receiving module is configured to continuously receive multi-sensor data when the forklift mobile robot moves. A dynamic correction module is configured to analyze the multi-sensor data according to the real attitude of the tray, dynamically correct the movement and positioning of the forklift mobile robot, and insert or fork the target work tray.

[0012] As an optional embodiment of the application, the attitude analysis module comprises: A YOLO recognition module is configured to identify the target work tray in the full-view image by using a YOLO model, and output a two-dimensional bounding box position. A preliminary attitude estimation module is configured to analyze the target work tray in the two-dimensional bounding box position by using an attitude regression branch, and obtain a preliminary attitude estimation. A real attitude acquisition module is configured to match a pre-established tray point cloud template, judge the preliminary attitude estimation by using a minimum Euclidean error, and obtain a real attitude of the target work tray, wherein the real attitude comprises six degrees of freedom information of the target work tray.

[0013] The application further provides an electronic device, comprising: A processor; A memory for storing processor-executable instructions; Wherein, the processor is configured to execute the executable instructions to implement the above-mentioned tray adaptive acquisition method for a forklift mobile robot.

[0014] Compared with the prior art, the forklift mobile robot of the application effectively overcomes the defects of the conventional scheme, and achieves significant performance improvement. The visual recognition can accurately acquire the depth information and three-dimensional attitude of the target work tray, greatly improves the positioning accuracy and forking success rate, and is not affected by light changes and background interference. The multi-sensor fusion enhances the robustness, and can adapt to various tray types such as wood and plastic, and complex stacking scenes. Based on visual recognition and multi-sensor fusion, a double-layer perception architecture of "cloud perception + edge guidance" is constructed, and the complete process of the forklift mobile robot from long-distance recognition to close-range accurate docking is realized. Attached Figure Description

[0015] Figure 1 This is a flowchart of the adaptive pallet acquisition method of the infiltrating forklift mobile robot according to an embodiment of the present invention; Figure 2 This is a block diagram of an adaptive pallet-acquiring device for a diving forklift mobile robot according to an embodiment of the present invention. Detailed Implementation

[0016] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0017] Example 1

[0018] To enhance the automated operation capabilities of dive-in forklift mobile robots in warehousing and sorting environments, an adaptive pallet acquisition method for dive-in forklift mobile robots is proposed, primarily addressing the docking of the robot when the target pallet's pose is uncertain. This method constructs a two-layer perception architecture of "cloud perception + edge guidance," achieving a complete process from long-range identification to precise close-range docking. Figure 1 As shown, the adaptive pallet acquisition method for the dive-in forklift mobile robot includes the following steps: S1 uses a panoramic camera to capture a full field-of-view image of the current position of the forklift mobile robot within a preset circular distance; S2, Analyze the full-view image, identify the target work tray within the preset circumferential distance and determine its true posture; S3, a navigation path is generated based on the actual posture and the current position, and the infiltrating forklift mobile robot moves according to the navigation path; S4, continuously receive multi-sensor data during the movement of the diving fork mobile robot; S5, based on the analysis of the multi-sensor data of the actual posture, dynamically correct the movement and positioning of the diving fork mobile robot, and the diving fork mobile robot inserts into or picks up the target work tray.

[0019] First, the infiltrating forklift mobile robot is equipped with a panoramic camera, which can capture images of the work area from different angles during initial mission deployment. In this embodiment, the captured image is a full field-of-view image with a radius of 3-5 meters centered on the robot. Using a panoramic camera can capture the environment in all directions (up, down, left, and right) centered on the robot at once, generating a complete 360°*180° spherical field-of-view photo, eliminating concerns about composition errors.

[0020] After the full-view image is taken, it is uploaded to a cloud server. A YOLOv5 deep learning model deployed in the cloud identifies the target work tray in the image in real time, extracts the real pose of the target work tray, i.e., six degrees of freedom information, by combining a pose regression branch. The cloud sends the real pose back to the robot through a task scheduling system and generates a preliminary navigation path from the current position of the robot to the nearest point of the target work tray. During the movement of the robot, cumulative errors are also generated, including: errors in the movement of the robot itself, such as wheel slippage and uneven ground; the tray itself may be slightly moved; the accuracy of the initial visual detection is limited, especially when the distance is far. Therefore, the robot needs a method for real-time feedback and correction during movement. Therefore, in this embodiment, during the movement of the robot, multi-sensor data is continuously received, and a robust cubature Kalman filtering algorithm based on the maximum entropy criterion is proposed through fusion analysis of the multi-sensor data, and the local position of the robot is updated according to the multi-sensor fusion positioning value, realizing the navigation and positioning of the submerged fork mobile robot in a dynamic interference environment. When the robot is closer and closer to the target work tray, the sensor data is also processed in real time by an edge computing unit to realize millimeter-level alignment correction of the robot and the target work tray, dynamically adjust the direction and angle of travel, realize high-precision approach, and the robot realizes accurate insertion or forking of the target work tray according to the feedback of the multi-sensor data.

[0021] Preferably, the full-view image is analyzed to identify the target work tray within the preset circumferential distance and determine the real pose of the tray, comprising: A YOLO model is used to identify the target work tray in the full-view image and output a two-dimensional bounding box position. A pose regression branch analysis is performed on the target work tray in the two-dimensional bounding box position to obtain a preliminary pose estimate. A pre-established tray point cloud template is matched to determine the preliminary pose estimate using the minimum Euclidean error to obtain the real pose of the target work tray, wherein the real pose includes six degrees of freedom information of the target work tray.

[0022] In this embodiment, the YOLOv5 model in the YOLO series is used to perform real-time identification on the tray target in the full-view image. In order to improve the understanding of the tray pose by the robot, the YOLOv5 model output includes not only the two-dimensional bounding box position of the target work tray, but also the preliminary pose estimation of the target work tray in the bounding box combined with the pose regression branch. Then, the preliminary pose estimation of the identified target work tray is searched and matched with the pre-established tray point cloud template in the offline stage, and the preliminary pose estimation is iterated finely through the minimum Euclidean error to align the point cloud of the template and the point cloud of the target work tray more accurately, so as to obtain the final high-precision pose and obtain the six-degree-of-freedom information of the target work tray, in which three degrees of freedom describe the position of the target work tray, and the other three degrees of freedom describe the pose or equivalent other forms. Thus, the robot can recognize the position of the target work tray in the full-view image and the real pose of the tray.

[0023] Preferably, before continuously receiving the multi-sensor data of the submerged fork mobile robot during movement, the method further comprises: selecting a multi-sensor according to the application environment of the submerged fork mobile robot; establishing a corresponding mathematical model for each kind of sensor data; integrating the mathematical models based on state space to uniformly describe each kind of sensor data.

[0024] This embodiment selects sensors suitable for the localization of a stealth forklift mobile robot based on the application environment. Analysis and mathematical modeling are performed on the selected sensors. A multi-sensor fusion robot localization method is proposed by combining sensor data and multi-sensor fusion methods. Specifically, in actual operations, stealth robots may face various typical application environments, including: indoor warehousing scenarios: complex spatial structures, GNSS (geometry system) unavailable; semi-outdoor loading and unloading areas: partial metal reflection and obstruction; narrow passages or high-shelf areas: severe LiDAR obstruction, significant IMU (inertial measurement unit) drift. To address these environmental characteristics, the following sensors are selected: wheeled odometer: used for short-term, high-precision incremental displacement estimation, low cost, and high real-time performance; IMU: provides attitude angular velocity information, supplementing short-term dynamic response; 2D / 3D LiDAR: enables large-area map matching and localization; panoramic camera: used for visual-assisted localization and target pallet identification; UWB (Ultra-Wideband) (optional): used for global localization calibration in open environments. By combining the above sensors, a multi-source heterogeneous data stream is constructed. For each type of sensor data, the following mathematical models are established: Odometry: State transition model, representing position increment; IMU: Attitude angle integral model, represented by quaternions / Euler angles; LiDAR: Observation model based on ICP / NDT point cloud matching; Camera: PnP visual positioning model. Based on state space modeling, the observation behavior of the above different sensors is uniformly described, and a unified interface is provided for subsequent filtering processing.

[0025] Preferably, the continuous reception of multi-sensor data during the movement of the dive-in forklift mobile robot includes: A measurement model is constructed by selecting sensors, namely, an odometer, an IMU, and a lidar. The measurement model is analyzed to obtain the fused positioning value of the sensor.

[0026] To address the impact of dynamic interference such as measurement censoring, measurement delay, and maximum outliers on navigation and positioning accuracy, this embodiment proposes a robust capacitive Kalman filter algorithm based on the maximum entropy criterion. First, a set of Bernoulli random variables β is defined. k ∈{0,1}, used to indicate whether the sensor's k-th measurement was received normally. The following measurement model is constructed using three types of sensors: odometer, IMU, and lidar: z k =β k h(x k )+v k .

[0027] Where h(x) k Let v be the nonlinear mapping function from state to observation. k It is Gaussian noise. If β k=0 indicates that the sensor is currently experiencing a measurement failure or severe delay.

[0028] Based on this measurement model, the measurement censoring and delay scenarios are described, and the multi-sensor fusion positioning value is further calculated.

[0029] Preferably, analyzing the measurement model to obtain the fused positioning value of the sensor includes: Based on the maximum entropy criterion, the expression form of the filter gain is derived according to the measurement model. Based on the aforementioned expression, local estimates for different sensor combinations are calculated using innovation statistical distance. The local estimates are weighted and fused to obtain the fused location value.

[0030] Based on the maximum entropy criterion, the posterior entropy of the state is defined as H(p(x)). k ∣z 1:k The goal is to maximize this entropy to obtain the most informative estimate. In RCKF, the filter gain K... k The expression is: K k =P xz P zz -1 .

[0031] Where P xz Let P be the covariance matrix of the state and the observations. zz To observe the covariance matrix, the expression for the filter gain is derived based on the maximum entropy criterion to reduce the impact of maximum outliers on navigation and positioning accuracy, thereby obtaining locally optimal navigation and positioning estimates. Secondly, a fusion rule for local navigation and positioning estimates is designed based on innovation statistical distance to obtain multi-sensor fused positioning values ​​from the odometry / IMU and lidar / IMU local estimates.

[0032] Specifically, using the new information to statistically determine the distance d k =(z k -z^ k ) T P zz -1 (z k -z^ k To evaluate the reliability of the current sensor observations, local estimates are obtained for different sensor combinations (e.g., odometer + IMU as one group, laser + IMU as another). The local estimates x^ for each group are then evaluated. k (1) ,x^ k (2) Weighted fusion is used: x^ k =w1x^ k (1) +w2x^ k (2), where w i ∝1 / dk (i) .

[0033] Preferably, the movement and positioning of the forklift robot are dynamically corrected based on the multi-sensor data analyzed according to the actual posture of the pallet, including: Obtain a raster map of the full-view image constructed based on SLAM; Global positioning correction is performed based on the fused positioning value and the grid map to update the local position of the diving forklift mobile robot.

[0034] Relying solely on a grid map constructed using SLAM to schedule robot movement is prone to errors. This embodiment, based on the fusion positioning values ​​of multiple sensors, performs global positioning correction through a grid map within the robot's field of vision constructed using SLAM, realizing a navigation scheme of "local tracking + global calibration". This can effectively cope with data failure, delay and abnormal interference in complex dynamic environments, ensuring the navigation robustness and accuracy of the infiltrating forklift mobile robot.

[0035] Preferably, the submersible forklift mobile robot inserts into or picks up the target work tray, comprising: When the forklift mobile robot moves to a preset distance in front of the target work tray, it enters the edge guidance stage; During the edge guidance phase, the location information of surface feature points on the target work tray is scanned, and the navigation path is adjusted based on the location information of the surface feature points.

[0036] In this embodiment, when the robot approaches within 1 meter of the tray, it enters the edge guidance phase. During this phase, the edge computing unit (NVIDIA Jetson platform) processes LiDAR and visual servo sensor data in real time. The 2D / 3D LiDAR scans the bottom edge and side contours of the tray and performs ICP matching with the built-in tray point cloud model to achieve millimeter-level alignment correction. The visual servo controller dynamically adjusts its direction and angle of travel by detecting the deviation between the tray's surface feature points (such as tray edge lines and corners) and the target trajectory, achieving high-precision approach.

[0037] Through the above methods, the stealth forklift mobile robot effectively overcomes the shortcomings of traditional solutions and achieves significant performance improvements. Visual recognition can accurately acquire the depth information and 3D posture of the target pallet, greatly improving positioning accuracy and forklift success rate, unaffected by changes in lighting or background interference. Multi-sensor fusion enhances robustness, enabling it to adapt to various pallet types such as wood and plastic, as well as complex stacking scenarios. Based on visual recognition and multi-sensor fusion, a two-layer perception architecture of "cloud perception + edge guidance" is constructed, realizing the complete process of stealth forklift mobile robot from long-distance recognition to close-range precise docking.

[0038] Example 2

[0039] Based on the same principle as the aforementioned method, an adaptive pallet-acquiring device 100 for a forklift mobile robot is also proposed, such as... Figure 2 As shown, it includes: Image acquisition module 110 is used to capture a full field-of-view image within a preset circumferential distance of the current position of the dive fork mobile robot using a panoramic camera; The attitude analysis module 120 is used to analyze the full-view image, identify the target work tray within the preset circumferential distance, and determine the true attitude. The moving module 130 is used to generate a navigation path based on the actual posture of the pallet and the current position, and the forklift mobile robot moves according to the navigation path; The data receiving module 140 is used to continuously receive multi-sensor data during the movement of the diving fork mobile robot. The dynamic correction module 150 is used to analyze the multi-sensor data based on the actual posture of the pallet and dynamically correct the movement and positioning of the submersible forklift mobile robot, which inserts into or picks up the target work pallet.

[0040] Preferably, the attitude analysis module 120 includes: The YOLO recognition module is used to identify the target work tray within the full-view image using the YOLO model and output the position of the two-dimensional bounding box. The preliminary attitude estimation module is used to perform attitude regression branch analysis on the target work tray within the two-dimensional bounding box to obtain a preliminary attitude estimate. A true attitude acquisition module is used to match a pre-established pallet point cloud template, use the minimum Euclidean error to determine the preliminary attitude estimate, and obtain the true attitude of the target work pallet, wherein the true attitude includes the six degrees of freedom information of the target work pallet.

[0041] Example 3

[0042] Furthermore, an electronic device is proposed, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to implement the adaptive pallet acquisition method for the dive-fork mobile robot described in Embodiment 1 when executing the executable instructions.

[0043] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0044] Furthermore, it should be noted that the use of terms such as "first," "second," and "a" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified. The terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two elements or the interaction between two elements, unless otherwise explicitly specified. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0045] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0046] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for adaptive pallet acquisition by a forked mobile robot, comprising: The method comprises the following steps: taking a panoramic camera to take a full-view image within a preset circumferential distance of a current position of a forklift mobile robot; analyzing the full-view image to identify a target work pallet within the preset circumferential distance and determine a real pose; generating a navigation path according to the real pose and the current position, and moving the forklift mobile robot according to the navigation path; continuously receiving multi-sensor data when the forklift mobile robot moves; analyzing the multi-sensor data according to the real pose, dynamically correcting the moving position of the forklift mobile robot, and inserting or forking the target work pallet.

2. The method of claim 1, wherein, The method of analyzing the full-view image to identify a target work pallet within the preset circumferential distance and determine a real pose comprises: using a YOLO model to identify the target work pallet in the full-view image, and outputting a two-dimensional bounding box position; performing pose regression branch analysis on the target work pallet in the two-dimensional bounding box position to obtain a preliminary pose estimate; matching a pre-established pallet point cloud template, using the minimum Euclidean error to judge the preliminary pose estimate, and obtaining the real pose of the target work pallet, wherein the real pose comprises six degrees of freedom information of the target work pallet.

3. The method of claim 1, wherein, Before continuously receiving multi-sensor data when the forklift mobile robot moves, the method further comprises the following steps: selecting multi-sensors according to the application environment of the forklift mobile robot; establishing a corresponding mathematical model for each type of sensor data; integrating the mathematical models based on state space to uniformly describe each type of sensor data.

4. The method of claim 3, wherein, Continuously receiving multi-sensor data when the forklift mobile robot moves comprises: selecting sensors to construct a measurement model, wherein the sensors are odometers, IMUs, and laser radars; analyzing the measurement model to obtain a fusion positioning value of the sensors.

5. The method of claim 4, wherein, Analyzing the measurement model to obtain a fusion positioning value of the sensors comprises: based on the maximum entropy criterion, deriving an expression form of a filtering gain according to the measurement model; based on the expression form, using a innovation statistical distance to calculate a local estimate value under different sensor combinations; weighting and fusing the local estimate values to obtain a fusion positioning value.

6. The method of claim 5, wherein, Analyzing the multi-sensor data according to the real pose to dynamically correct the moving position of the forklift mobile robot comprises: obtaining a grid map of the full-view image constructed based on SLAM; performing global positioning correction according to the fusion positioning value and the grid map to update the local position of the forklift mobile robot.

7. The method of claim 1, wherein, The forklift mobile robot inserting or forking the target work pallet comprises: when the forklift mobile robot moves to a preset distance in front of the target work pallet, entering an edge guiding stage; in the edge guiding stage, scanning surface feature point position information of the target work pallet, and adjusting the navigation path according to the surface feature point position information.

8. A self-adapting pallet acquisition device for a forked mobile robot, comprising: The method comprises the following steps: an image acquisition module is configured to take a panoramic camera to take a full-view image within a preset circumferential distance of a current position of a forklift mobile robot; An analysis posture module is configured to analyze the full-view image, identify a target work tray within the preset circumferential distance, and determine a real posture; A movement module is configured to generate a navigation path according to the real posture and the current position, and the undercover fork mobile robot moves according to the navigation path; A receiving data module is configured to continuously receive multi-sensor data when the undercover fork mobile robot moves; A dynamic correction module is configured to analyze the multi-sensor data according to the real posture, and dynamically correct the movement and positioning of the undercover fork mobile robot, so that the undercover fork mobile robot inserts or forks the target work tray.

9. The self-adapting pallet acquisition device for a fork mobile robot that dives in according to claim 8, wherein, The analysis posture module comprises: A YOLO identification module is configured to identify the target work tray in the full-view image by using a YOLO model, and output a two-dimensional bounding box position; A preliminary posture estimation module is configured to analyze the posture regression branch of the target work tray in the two-dimensional bounding box position, and obtain a preliminary posture estimation; An obtaining real posture module is configured to match a pre-established tray point cloud template, judge the preliminary posture estimation by using a minimum Euclidean error, and obtain a real posture of the target work tray, wherein the real posture comprises six-degree-of-freedom information of the target work tray.

10. An electronic device, comprising: It comprises: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to implement the method for the undercover fork mobile robot to adaptively obtain a tray according to any one of claims 1-7 when executing the executable instructions.

Images