A greenhouse moving robot autonomous navigation method, device, equipment and medium

Abstract

The present application relates to the field of agricultural robot autonomous navigation, and discloses a greenhouse carrying robot autonomous navigation method, device, equipment and medium, which can fuse and filter the original three-dimensional point clouds generated by front and rear laser radars, generate a pure fusion point cloud set and input into SLAM for local scan matching and odometer calculation, obtain local odometer and local trajectory, detect and effectively observe and screen the image stream shot by the vehicle-mounted camera to obtain effective visual landmark observation data, perform three-dimensional pose solution and cross-dimension constraint on the effective visual landmark observation data to obtain visual landmark constraint nodes, and control the global pose and motion trajectory of the carrying robot based on the local odometer, local trajectory, visual landmark constraint nodes and each visual landmark. The present application can break the environmental geometric symmetry and realize robust continuous positioning and navigation of the carrying robot in the greenhouse under high load and high similar environment.

Description

Technical Field

[0001] This invention relates to the field of autonomous navigation for agricultural robots, and more particularly to an autonomous navigation method, apparatus, equipment, and medium for a greenhouse transport robot. Background Technology

[0002] With the development of science and technology, robot navigation technology is constantly improving.

[0003] In recent years, mobile robots have been widely used for handling agricultural products and transporting materials in greenhouses. The core objective of these robots is material transportation, so the top of their vehicles is typically designed as the main cargo platform. Conventional indoor autonomous navigation robots often employ a top-mounted single-LiDAR simultaneous localization and mapping (SLAM) system.

[0004] Among them, in order to meet the safety obstacle avoidance requirements of the chassis of the handling robot with 360 degrees without blind spots, the relevant technology generally adopts a spatial splicing architecture with a single radar at the top.

[0005] However, the related technologies suffer from laser positioning divergence caused by the greenhouse corridor effect, making stable and reliable continuous navigation impossible. Specifically, modern greenhouses often adopt a highly standardized construction model, with equally spaced load-bearing columns and planting furrows of uniform height. This results in strong geometric self-similarity in the point cloud view of lidar. This repetitive structure prevents algorithms based on purely geometric feature matching from distinguishing the current specific location through similar point clouds, leading to severe longitudinal divergence in pose calculations, causing positioning errors, and hindering stable and reliable continuous navigation. Summary of the Invention

[0006] This invention provides an autonomous navigation method, device, equipment, and medium for greenhouse transport robots, which solves the laser positioning divergence problem caused by the greenhouse corridor effect and the inability to achieve stable and reliable continuous navigation in related technologies. It breaks the geometric symmetry of the environment and realizes robust continuous positioning and navigation of transport robots in high-load and highly similar environments in greenhouses.

[0007] In a first aspect, the present invention provides an autonomous navigation method for a greenhouse transport robot. The transport robot includes Simultaneous Localization and Mapping (SLAM), a vehicle-mounted camera mounted on its top, and a front-mounted LiDAR and a rear-mounted LiDAR mounted on its chassis. Multiple visual landmarks are provided in the top space of the greenhouse. The method includes: The original 3D point clouds generated by the front-end lidar and the rear-end lidar are fused and filtered to generate a clean fused point cloud set; The pure fused point cloud set is input into the SLAM for local scan matching and odometry calculation to obtain local odometry and local trajectory. Visual landmark detection and effective observation screening are performed on the image stream captured by the vehicle-mounted camera to obtain effective visual landmark observation data; The effective visual landmark observation data is subjected to three-dimensional pose calculation and cross-dimensional constraints to obtain visual landmark constraint nodes; Based on the local odometry, the local trajectory, the visual landmark constraint nodes, and each visual landmark, the global pose and motion trajectory of the transport robot are controlled.

[0008] Optionally, the step of fusing and filtering the original 3D point clouds generated by the front-end LiDAR and the rear-end LiDAR to generate a clean fused point cloud set includes: Using the created extrinsic transformation matrix, the original 3D point clouds generated by the front lidar and the rear lidar are mapped to the chassis coordinate system to obtain the first mapped point cloud and the second mapped point cloud. The first mapped point cloud and the second mapped point cloud are concatenated and fused to generate an initial point cloud set containing a 360-degree omnidirectional view around the chassis. The statistical outlier filter is invoked to identify and remove isolated noise points in the initial point cloud set, resulting in the clean fused point cloud set.

[0009] Optionally, the step of inputting the clean fused point cloud set into the SLAM for local scan matching and odometry calculation to obtain local odometry and local trajectory includes: The pure fused point cloud set is input into the front end of the SLAM, so that the front end of the SLAM: scans and matches the pure fused point cloud set with the local sub-image at the current time, and solves the optimal two-dimensional pose transformation of the pure fused point cloud set relative to the local sub-image at the current time through nonlinear optimization, so as to obtain the two-dimensional pose transformation at the current time. Based on the two-dimensional pose transformation at the current moment and the two-dimensional pose transformation at the previous moment, the pose increment at adjacent moments is determined. The pose increment at adjacent moments is used as the local odometry at the current moment. The local odometry at consecutive moments is accumulated in chronological order to obtain the local trajectory.

[0010] Optionally, the step of performing visual landmark detection and effective observation filtering on the image stream captured by the vehicle-mounted camera to obtain effective visual landmark observation data includes: For any frame image in the image stream, visual landmark detection is performed on the frame image. When a target visual landmark is detected in the frame image, the distance between the target visual landmark and the transport robot is determined. If the distance is not greater than a set distance threshold, the unique identifier and corner coordinates of the target visual landmark are extracted. If the correction identifier stored in the preset storage space is inconsistent with the unique identifier of the target visual landmark, the current observation result is determined to be a valid observation. The correction identifier in the preset storage space is deleted, and the unique identifier of the target visual landmark is stored as a new correction identifier in the preset storage space. The frame image, the timestamp of the frame image, the unique identifier of the target visual landmark, and the corner coordinates are taken as the valid visual landmark observation data.

[0011] Optionally, the step of performing 3D pose calculation and cross-dimensional constraints on the effective visual landmark observation data to obtain visual landmark constraint nodes includes: The frame images in the effective visual landmark observation data are preprocessed to obtain the corresponding binarized image data. Connectivity analysis and edge extraction are performed on the binarized image data to identify the closed contours of the binarized image data. Polygon fitting is performed on the closed contours to remove noise points that do not conform to the geometric features of the label, and candidate quadrilateral contour data is obtained. The four vertices of the candidate quadrilateral contour data are extracted, and a sub-pixel interpolation algorithm is called to perform gradient search in the pixel neighborhood of the four vertices to eliminate digitization sampling error and obtain the corresponding four two-dimensional pixel corner coordinate data. The four two-dimensional pixel corner coordinate data and the corner coordinates in the effective visual landmark observation data are input into the perspective n-point pose PnP solver to solve the pose and obtain the initial three-dimensional relative pose data of the label coordinate system relative to the camera coordinate system. The initial three-dimensional relative pose data is mapped to the chassis coordinate system to obtain the mapped three-dimensional relative pose data. Based on the created rotation matrix, the mapped three-dimensional relative pose data is flattened in two dimensions to generate the two-dimensional translation coordinates, pure two-dimensional pose quaternion data, translation weights and rotation weights of the visual landmark relative to the chassis. The two-dimensional translation coordinates, the pure two-dimensional attitude quaternion data, the translation weight, the rotation weight, the timestamp of the frame image in the effective visual landmark observation data, and the unique identifier of the target visual landmark are used as a whole as the visual landmark constraint node.

[0012] Optionally, controlling the global pose and motion trajectory of the transport robot based on the local odometry, the local trajectory, the visual landmark constraint nodes, and each of the visual landmarks includes: Obtain the global pose of each visual landmark in the global map coordinate system; The local odometry, the local trajectory, the visual landmark constraint nodes, and the global pose of each visual landmark in the global map coordinate system are input into the backend of the SLAM so that the SLAM can construct visual landmark residual terms. The visual landmark residual term is inserted into the sparse pose adjustment function to obtain the target function; Solving the objective function yields the corrected global pose and optimized trajectory of the robot. The global pose and motion trajectory of the transport robot are controlled based on the corrected global pose and optimized trajectory.

[0013] Optionally, the visual landmark residual term is: ; in, For the visual landmark residual term, This indicates the index number of the visual landmark participating in the optimization at the current moment; and These represent pose composition and inverse composition operations, respectively. Indicates the first The actual observed relative pose of the visual landmarks with respect to the chassis coordinate system; This indicates the current optimized absolute pose of the robot chassis; This is an abbreviation for the visual landmark. This represents the minimum covariance information matrix specifically assigned to the visual landmark observation item.

[0014] Secondly, the present invention provides an autonomous navigation device for a greenhouse transport robot. The transport robot includes Simultaneous Localization and Mapping (SLAM), a vehicle-mounted camera mounted on its top, a front-mounted LiDAR and a rear-mounted LiDAR mounted on its chassis, and multiple visual landmarks are set in the top space of the greenhouse. The device includes: The fusion unit is used to fuse and filter the original 3D point clouds generated by the front lidar and the rear lidar to generate a clean fused point cloud set. The matching unit is used to input the clean fused point cloud set into the SLAM to perform local scan matching and odometry calculation to obtain local odometry and local trajectory; The filtering unit is used to perform visual landmark detection and effective observation filtering on the image stream captured by the vehicle-mounted camera in order to obtain effective visual landmark observation data. The calculation unit is used to perform three-dimensional pose calculation and cross-dimensional constraints on the effective visual landmark observation data to obtain visual landmark constraint nodes; The control unit is used to control the global pose and motion trajectory of the transport robot based on the local odometry, the local trajectory, the visual landmark constraint nodes, and each of the visual landmarks.

[0015] Thirdly, the present invention provides a computer device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the autonomous navigation method for greenhouse transport robots described in the first aspect or any corresponding embodiment.

[0016] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the autonomous navigation method for a greenhouse transport robot described in the first aspect or any corresponding embodiment.

[0017] This invention provides an autonomous navigation method, device, equipment, and medium for a greenhouse transport robot. The transport robot includes Simultaneous Localization and Mapping (SLAM), a top-mounted vehicle camera, and front and rear LiDARs mounted on the chassis. Multiple visual landmarks are located in the top space of the greenhouse. This invention fuses and filters the original 3D point clouds generated by the front and rear LiDARs to generate a clean fused point cloud set. This clean fused point cloud set is input into SLAM for local scan matching and odometry calculation to obtain local odometry and local trajectory. Visual landmark detection and effective observation filtering are performed on the image stream captured by the vehicle camera to obtain effective visual landmark observation data. 3D pose calculation and cross-dimensional constraints are performed on the effective visual landmark observation data to obtain visual landmark constraint nodes. Based on the local odometry, local trajectory, visual landmark constraint nodes, and each visual landmark, the global pose and motion trajectory of the transport robot are controlled.

[0018] Compared with related technologies, this invention has the following beneficial effects on the problem of laser positioning divergence caused by the corridor effect in greenhouses: Freeing up cargo space and resolving the deadlock of physical interference: The physical conflict between obstacle avoidance perception and cargo space is completely eliminated. The dual radars sinking down ensure 360-degree obstacle avoidance at the bottom, without taking up loading and unloading space, and the radar line of sight will not be blocked by the cargo carried, which is extremely suitable for the engineering and practical needs of handling robots; Overcoming the challenge of perception aliasing in standardized environments: Addressing the inherent limitation of dual LiDAR systems easily getting lost in repetitive greenhouse structures, visual landmarks with non-geometrically unique ID attributes are introduced. Absolute positioning constraints with unique identifiers are provided to eliminate positional ambiguity caused by repetitive structures. Heterogeneous perception complementarity enhances system robustness: the underlying radar is used for high-frequency local positioning and obstacle avoidance, while the vehicle-mounted camera continuously detects visual landmarks; once the robot observes a visual landmark that meets the validity conditions, it generates visual landmark constraints and corrects the laser SLAM pose, thereby periodically eliminating the cumulative positioning error generated in repetitive structural environments and improving the accuracy and stability of continuous navigation. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating an autonomous navigation method for a greenhouse transport robot provided in an embodiment of the present invention; Figure 2 A top view of a handling robot provided in an embodiment of the present invention; Figure 3 A side view of a handling robot provided in an embodiment of the present invention; Figure 4 A front view of a transport robot provided in an embodiment of the present invention; Figure 5 This is a rear view of a transport robot provided in an embodiment of the present invention; Figure 6 A schematic diagram of scanning data from a front-mounted and a rear-mounted lidar, provided as an embodiment of the present invention; Figure 7 This invention provides an embodiment of the visual landmark AprilTag for positioning and hanging on a wall. Figure 8 A landmark marked in a Cartographer map using the visual landmark AprilTag, as provided in this embodiment of the invention; Figure 9 This is a schematic diagram of the structure of an autonomous navigation device for a greenhouse transport robot provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0022] The following is combined Figures 1-8 The present invention describes an autonomous navigation method for a greenhouse transport robot.

[0023] like Figure 1 As shown, this embodiment proposes a first autonomous navigation method for a greenhouse transport robot. The transport robot includes Simultaneous Localization and Mapping (SLAM), a vehicle-mounted camera on its top, and front and rear LiDAR sensors mounted on its chassis. Multiple visual landmarks are located in the top space of the greenhouse. The method may include the following steps: S101. The original 3D point clouds generated by the front-end LiDAR and the rear-end LiDAR are fused and filtered to generate a clean fused point cloud set.

[0024] Specifically, the front-mounted and rear-mounted LiDAR sensors can be horizontally installed at the bottom front and rear of the chassis of the handling robot, respectively. These sensors are used to collect omnidirectional point clouds of the underlying environment around the chassis for collision avoidance and positioning, while avoiding spatial interference from the cargo platform. The horizontal scanning planes of the front-mounted and rear-mounted LiDAR sensors are coplanar, and their installation height is below the safety limits of the cargo platform and the drooping crop canopies inside greenhouses. The point cloud data is stitched together to form a 360-degree blind-spot-free bottom-level collision avoidance profile at the bottom of the chassis.

[0025] Specifically, this embodiment can install at least one forward-facing observation camera, i.e., a vehicle-mounted camera, positioned on top of the cargo platform. Its field of view extends beyond the cargo platform to capture in real-time a global visual landmark network deployed on the vertical facade structure of the greenhouse. Specifically, the vehicle-mounted camera can be installed at the front or side edge of the cargo platform, with its lens optical axis approximately parallel to the horizontal plane and facing the robot's forward travel area. This allows it to capture visual landmarks on the vertical facade structure directly in front or to the side. This installation position ensures that the camera's field of view avoids blind spots caused by the loaded cargo on the cargo platform.

[0026] like Figures 2 to 5 The transport robot shown includes a top-mounted camera 1, a front-mounted LiDAR 2, and a rear-mounted LiDAR 3 mounted on the chassis. The top-mounted camera 1 can be a color vision camera.

[0027] Specifically, when the transport robot performs cargo-carrying tasks and moves within the greenhouse, this embodiment can autonomously navigate the transport robot based on onboard cameras, front-mounted LiDAR, rear-mounted LiDAR, and visual landmarks set in the top space of the greenhouse.

[0028] The original 3D point cloud can be generated from the laser scanning data performed by the front and rear lidar sensors in the greenhouse during the handling robot's cargo-carrying tasks and movements. For example... Figure 6 As shown, the green box represents the scanning data from the front-mounted LiDAR, and the red box represents the scanning data from the rear-mounted LiDAR.

[0029] like Figure 7 As shown, visual landmarks can be AprilTag tags used for positioning when hung on a wall.

[0030] Optionally, step S101 includes: Using the created extrinsic transformation matrix, the original 3D point clouds generated by the front and rear lidars are mapped to the chassis coordinate system to obtain the first and second mapped point clouds. The first and second mapped point clouds are concatenated and fused to generate an initial point cloud set containing a 360-degree omnidirectional view around the chassis. The statistical outlier filter is invoked to identify and remove isolated noise points in the initial point cloud set, resulting in a clean fused point cloud set.

[0031] Specifically, this embodiment can perform dual-radar point cloud fusion and dynamic filtering. Taking the original 3D point clouds output by the robot's front and rear lidars as the processing objects, the two sets of point clouds are unified to the robot chassis tracking coordinate system through time synchronization, homogeneous extrinsic parameter coordinate transformation, and point cloud cascading stitching. Then, statistical outlier filtering is used to remove discrete noise points, resulting in a clean fused point cloud set covering a 360-degree field of view around the robot chassis.

[0032] Specifically, this embodiment can subscribe to the raw 3D point cloud data of both the front and rear LiDARs. Since the front and rear LiDARs have independent physical installation poses, low-level spatial registration of the point cloud coordinate systems is required. This embodiment can extract the rigid extrinsic transformation matrix from the front LiDAR coordinate system to the chassis tracking coordinate system (base_link). This matrix contains Rotation matrix and Translation vector of Homogeneous transformation matrix: ; Let the coordinates of any single point in the original point cloud of the forward radar be represented as a homogeneous column vector. The system calculates its absolute mapped coordinates in the chassis coordinate system through cross-coordinate system matrix multiplication. : ; Similarly, based on the homogeneous extrinsic matrix of the backward radar The system synchronously projects the rearward radar point cloud onto the chassis coordinate system. After completing the matrix projection, the system performs cascaded fusion on the two sets of spatial point sets. Landmarks are used to generate an initial point cloud set containing a 360-degree omnidirectional view around the chassis. .

[0033] Because the unstructured environment of a greenhouse contains not only ground clutter but also dust and random discrete noise from plant leaf edges, this embodiment can utilize a statistical outlier filter to process the fused point cloud. Filtering is performed. This embodiment can calculate the set. Each point in and its neighboring areas Nearest neighbor points (preferred setting) The average distance is obtained by using the spatial Euclidean distance. Let the mean distance of the global point cloud be... The standard deviation is Filtering and truncation are performed using a dynamic threshold formula: ; in, This is the standard deviation multiplier threshold. This embodiment can forcibly remove isolated noise points whose spatial distance distribution exceeds the dynamic adaptive threshold, resulting in a high-quality, clean, fused point cloud set. .

[0034] S102. Input the clean fused point cloud set into SLAM for local scan matching and odometry calculation to obtain local odometry and local trajectory.

[0035] Specifically, in this embodiment, after obtaining a clean fused point cloud set, the clean fused point cloud set can be input into SLAM for local scan matching and odometry calculation to obtain the local odometry and local trajectory generated and output by SLAM.

[0036] Optionally, step S102 includes: Using the created extrinsic transformation matrix, the original 3D point clouds generated by the front and rear lidars are mapped to the chassis coordinate system to obtain the first and second mapped point clouds. The first and second mapped point clouds are concatenated and fused to generate an initial point cloud set containing a 360-degree omnidirectional view around the chassis. The statistical outlier filter is invoked to identify and remove isolated noise points in the initial point cloud set, resulting in a clean fused point cloud set.

[0037] Specifically, in this embodiment, the clean fused point cloud set can be scanned and matched with the local sub-image of the laser SLAM front end, and the optimal two-dimensional pose transformation of the current point cloud frame relative to the local sub-image can be obtained through nonlinear optimization. Local odometry and local trajectory are obtained based on the two-dimensional pose transformation of continuous point cloud frames.

[0038] Specifically, this embodiment first performs local scan matching, inputting the clean fused point cloud set at the current moment into the SLAM front end and performing scan matching with the local subgraph at the current moment. The optimal 2D pose transformation of the current clean fused point cloud set relative to the local subgraph is solved through nonlinear optimization to obtain the 2D pose transformation at the current moment. Then, odometry and local trajectory generation are performed. Based on the 2D pose transformation at the current moment and the 2D pose transformation at the previous moment, the pose increment between adjacent moments is calculated to obtain the local odometry at the current moment. The continuous local odometry is accumulated in chronological order to obtain the local trajectory of the transport robot.

[0039] S103. Visual landmark detection and effective observation screening are performed on the image stream captured by the vehicle-mounted camera to obtain effective visual landmark observation data.

[0040] Specifically, this embodiment can acquire an image stream captured by a vehicle-mounted camera, perform visual landmark detection and effective observation filtering on the image stream, and obtain effective visual landmark observation data.

[0041] Optionally, step S103 includes: For any frame in the image stream, visual landmark detection is performed on the frame. When a target visual landmark is detected in the frame, the distance between the target visual landmark and the handling robot is determined. If the distance is not greater than a set distance threshold, the unique identifier and corner coordinates of the target visual landmark are extracted. If the correction identifier stored in the preset storage space is inconsistent with the unique identifier of the target visual landmark, the current observation result is determined to be a valid observation. The correction identifier in the preset storage space is deleted, and the unique identifier of the target visual landmark is stored as a new correction identifier in the preset storage space. The frame image, the timestamp of the frame image, the unique identifier of the target visual landmark, and the corner coordinates are taken as the whole as valid visual landmark observation data.

[0042] Specifically, this embodiment can perform visible visual landmark detection and effective observation screening. Using the image data continuously output by the vehicle-mounted camera as the processing object, visual landmarks in the image are identified in real time. When a visual landmark is detected within the field of view, its validity is determined based on the distance threshold from the visual landmark to the handling robot and the unique ID status lock of the tag, resulting in effective visual landmark observation data containing the unique tag ID, image timestamp, and corner coordinates.

[0043] In this embodiment, the continuously output color image stream from the vehicle-mounted camera is used as the processing object. Visual landmark detection is continuously performed during robot operation. When a visual landmark is detected within the camera's field of view, a mechanism combining physical distance determination and ID state locking is used for effective observation and filtering.

[0044] Specifically, this embodiment can perform distance determination and ID resolution. Visual landmark detection is continuously performed on the color image stream captured by the vehicle-mounted camera. When a visual landmark is detected within the field of view, and the physical straight-line distance from the visual landmark to the handling robot calculated based on the initial pose is less than or equal to a set distance threshold, the unique identifier ID of the visual landmark and its corner coordinates are extracted.

[0045] Specifically, this embodiment can perform single-state lock interception based on ID. A recently corrected ID register is maintained in memory. If the current tag ID is inconsistent with the ID recorded in the register, the current observation is marked as the first valid observation, a single-correction allow signal is output, and the register is simultaneously overwritten with the current tag ID; if the current tag ID is the same as the ID recorded in the register, it is determined that they are still in the same observation period, duplicate observations are intercepted, and the correction allow signal is not output repeatedly.

[0046] Optionally, in other autonomous navigation methods for greenhouse handling robots proposed in this embodiment, the method may further include the following after step S103: If the correction identifier stored in the preset storage space matches the unique identifier of the target visual landmark, the current observation result is determined to be an invalid observation, and it is prohibited to delete the correction identifier in the preset storage space and to generate valid visual landmark observation data. If no visual landmarks are detected in N consecutive frames, the correction identifier stored in the preset storage space is deleted.

[0047] Specifically, this embodiment can perform an observation cycle reset (state release). When a visual landmark leaves the camera's field of view, or when the camera fails to detect any tag for a set number of consecutive frames, the ID stored in the most recently corrected ID register is automatically cleared. This release mechanism ensures that the correction process can be retried when the robot returns after leaving, or when it reverses and observes the same tag again.

[0048] S104. Perform three-dimensional pose calculation and cross-dimensional constraints on the effective visual landmark observation data to obtain visual landmark constraint nodes.

[0049] Specifically, in this embodiment, after obtaining valid visual landmark observation data, three-dimensional pose calculation and cross-dimensional constraints are performed on the valid visual landmark observation data to generate visual landmark constraint nodes.

[0050] Optionally, step S104 includes: The frame images in the effective visual landmark observation data are preprocessed to obtain the corresponding binarized image data. Connectivity analysis and edge extraction are performed on the binarized image data to identify the closed contours of the binarized image data. Polygon fitting is performed on the closed contours to remove noise points that do not conform to the geometric features of the label, and candidate quadrilateral contour data are obtained. Extract the four vertices of the candidate quadrilateral contour data, call the sub-pixel level interpolation algorithm to perform gradient search in the pixel neighborhood of the four vertices to eliminate digital sampling error, and obtain the corresponding four two-dimensional pixel corner coordinate data; The corner coordinates of four two-dimensional pixels and the corner coordinates in the effective visual landmark observation data are input into the perspective n-point pose PnP solver to solve the pose and obtain the initial three-dimensional relative pose data of the label coordinate system relative to the camera coordinate system. The initial 3D relative pose data is mapped to the chassis coordinate system to obtain the mapped 3D relative pose data. Based on the created rotation matrix, the mapped 3D relative pose data is flattened in two dimensions to generate the 2D translation coordinates of the visual landmark relative to the chassis, pure 2D pose quaternary data, translation weights and rotation weights. The two-dimensional translation coordinates, pure two-dimensional attitude four-dimensional data, translation weights, rotation weights, timestamps of frame images in the effective visual landmark observation data, and unique identifiers of target visual landmarks are used as a whole as visual landmark constraint nodes.

[0051] Specifically, this embodiment can perform visual pose calculation and cross-dimensional constraint generation. Using effective visual landmark observation data as the processing object, the three-dimensional relative pose of the visual landmark with respect to the camera is calculated using camera calibration parameters and the Perspective-n-Point (PnP) algorithm. After extrinsic parameter transformation from camera to chassis, the three-dimensional relative pose is flattened in two dimensions based on rotation matrix deconstruction to generate the two-dimensional relative pose of the visual landmark with respect to the robot chassis and visual landmark constraint nodes.

[0052] In this embodiment, visual landmark observation data can be effectively used as the processing object. The high-precision camera intrinsic parameter matrix calibrated offline is loaded, the corner coordinates of the visual landmarks are extracted, and the initial three-dimensional relative pose of the label coordinate system with respect to the camera coordinate system is calculated.

[0053] Specifically, in this embodiment, after receiving valid visual landmark observation data, the relevant logical steps and mathematical models are executed to achieve a precise conversion from the original image to a three-dimensional pose.

[0054] Specifically, this embodiment can perform image preprocessing. Frame images captured by a color camera are acquired from valid visual landmark observation data. Through grayscale conversion and adaptive threshold filtering, interference from complex light and shadow within the greenhouse is eliminated, converting the images into high-contrast binarized image data.

[0055] Specifically, this embodiment can perform feature detection and initial screening. Connectivity analysis and edge extraction are performed on the binarized image data to identify closed contours and perform polygon fitting, eliminating noise points that do not conform to the label's geometric features, and outputting candidate quadrilateral contour data. Sub-pixel corner refinement is performed, extracting the four vertices of the candidate quadrilateral contour data, and a sub-pixel interpolation algorithm is introduced to perform gradient search within the pixel neighborhood to eliminate digitization sampling errors, obtaining four high-precision two-dimensional pixel corner coordinate data. , .

[0056] Specifically, this embodiment can perform pose calculation based on the PnP algorithm, and convert the above-mentioned two-dimensional pixel corner coordinate data The input is fed into the PnP algorithm solver to establish a mapping relationship from the image plane to the physical space.

[0057] In this embodiment, data mapping can be established, and the extracted two-dimensional pixels can be... Compared with a predefined real physical model (i.e., the three-dimensional spatial point coordinate data of visual landmarks in their own local coordinate system) A one-to-one correspondence is made to form 2D-3D feature point pairs.

[0058] This embodiment can perform projection error model construction and call the overloaded high-precision camera intrinsic parameter matrix: ; in, For camera focal length, The primary point coordinates are used. Utilizing camera imaging geometry principles, a three-dimensional spatial coordinate system is constructed. The transformation formula from theoretical projection to the virtual imaging plane is used to calculate the theoretical projection pixel coordinates. : ; in, s For depth scaling factor, R Let be the rotation matrix to be found. t Let be the translation vector to be determined.

[0059] This embodiment can perform iterative optimization to find the solution. The solver constructs a nonlinear least squares objective function and calculates the actual extracted corner points. The geometric distance between the theoretical projection point and the actual projection point, i.e., the reprojection error. E : ; The solver uses the Gauss-Newton iterative method to continuously adjust the assumed pose matrix $[R,t]$ of the camera in mathematical space until the aforementioned reprojection error is reached. It drops to a minimum value.

[0060] This embodiment can output pose data. When the error function converges, the optimal pose state is extracted, that is, the initial 3D relative pose data of the label coordinate system relative to the camera coordinate system is calculated. (This data is the rotation matrix obtained through the above iterations.) R With translation vector t The combination of .

[0061] Since chassis-driven laser SLAM operates in a two-dimensional planar space, to prevent minor observational perturbations in the visual 3D pose at height, roll angle, and pitch angle from disrupting the smoothness of the 2D pose map, this embodiment designs a 3D-to-2D cross-dimensional pose forced flattening bridging algorithm based on rotation matrix deconstruction: First, based on the static extrinsic parameters from the camera to the chassis, calculate the 3D rotation quaternion of the visual landmark AprilTag in the robot chassis tracking coordinate system. Convert it to a standard three-dimensional rotation matrix. : ; In the mathematical geometry definition, the third column vector of this rotation matrix This represents the direction vector of the Z-axis of the AprilTag coordinate system in the chassis coordinate system (i.e., perpendicularly pointing inwards from the wall). Extracting this column vector and inverting it yields the planar normal vector of the wall perpendicularly outwards (pointing towards the robot itself). : ; Discard vertical components that interfere with 2D optimization. The yaw angle in a pure 2D plane is calculated using only the horizontal plane component: ; Subsequently, the altitude Z=0, roll angle Roll=0, and pitch angle Pitch=0 are forcibly set, and a dimension-reduced two-dimensional quaternion is regenerated based on the two-dimensional yaw angle. Finally, the pure two-dimensional relative pose, after matrix deconstruction and flattening, along with the unique label ID, is encapsulated into a visual landmark constraint node. At this point, this embodiment can eliminate the altitude, roll, and pitch disturbances in the visual three-dimensional pose that interfere with the optimization of the two-dimensional pose graph.

[0062] It should be noted that in this embodiment, the pure 2D pose, after rigorous matrix deconstruction and flattening, along with its unique label ID, is encapsulated into a LandmarkList constraint node natively supported by Cartographer, such as... Figure 8 The image shows a landmark marked on a Cartographer map using the visual landmark AprilTag. Specifically, the data format and content contained in this constraint node are as follows: tracking_time (timestamp data): Records the high-precision system time of capturing the current valid visual observation frame; landmark_id (identifier code data): Stores the extracted AprilTag unique identifier string encoding; Translation data: Stores the two-dimensional translation coordinates [x, y, 0] output by the dimensionality reduction calculation above; rotation (rotation data): Stores pure two-dimensional attitude quaternions ; weight (confidence weight data): includes translation weight and rotation weight. Since the visual landmark AprilTag provides the absolute physical anchor point, the system assigns a very large weight coefficient value to this observation, i.e., a very small covariance matrix.

[0063] S105, based on local odometry, local trajectory, visual landmark constraint nodes and each visual landmark, controls the global pose and motion trajectory of the handling robot.

[0064] Specifically, this embodiment can control the global pose and motion trajectory of the handling robot based on the aforementioned local odometry, local trajectory, and visual landmark constraint nodes, thereby improving the autonomous navigation accuracy of the handling robot.

[0065] Optionally, step S105 includes: Obtain the global pose of each visual landmark in the global map coordinate system; The local odometry, local trajectory, visual landmark constraint nodes, and the global pose of each visual landmark in the global map coordinate system are input into the backend of SLAM so that SLAM can construct visual landmark residual terms. The visual landmark residual terms are inserted into the sparse pose adjustment function to obtain the target function; The objective function is solved to obtain the corrected global pose and optimized trajectory of the robot. The global pose and motion trajectory of the handling robot are controlled based on the corrected global pose and optimized trajectory.

[0066] Among them, the visual landmark residual term is: ; in, For visual landmark residuals, Indicates the index number of the visual landmark participating in the optimization at the current moment; and These represent pose composition and inverse composition operations, respectively. Indicates the first The actual observed relative pose of a visual landmark with respect to the chassis coordinate system; This indicates the current optimized absolute pose of the robot chassis; It is an abbreviation for visual landmark. This represents the minimum covariance information matrix specifically assigned to visual landmark observations.

[0067] Specifically, this embodiment can take the aforementioned local odometer and its formed local trajectory, visual landmark constraint nodes, and the pre-recorded global pose of each visual landmark in the global map coordinate system as the processing objects.

[0068] When the SLAM backend receives visual landmark constraints, it can dynamically insert high-weight visual landmark residual terms into the original sparse pose adjustment objective function. .

[0069] It should be noted that since the visual landmark AprilTag provides global physical anchor points with unique IDs, the backend optimizer performs joint optimization on the above local trajectories when solving the global objective function that includes the visual landmark residual term. This eliminates the longitudinal cumulative drift caused by the greenhouse repetitive structure, outputs the corrected global pose and optimized trajectory of the robot, and provides the result to the local path planner. The local path planner then generates speed commands to complete the chassis motion control.

[0070] This embodiment is based on the cargo-carrying requirements of the handling robot. It ensures omnidirectional obstacle avoidance through a low-position dual-radar architecture and introduces a visual camera and AprilTag landmark network to address the radar positioning error problem that is easily caused by standardized greenhouses. By complementing the advantages and disadvantages of heterogeneous sensors, the geometric symmetry of the environment is broken, and robust continuous positioning and navigation of the robot in high-load and highly similar environments is achieved.

[0071] The autonomous navigation method for greenhouse handling robots proposed in this embodiment has the following significant advantages over related technologies in addressing the laser positioning divergence problem caused by the greenhouse corridor effect: Freeing up cargo space and resolving the deadlock of physical interference: The physical conflict between obstacle avoidance perception and cargo space is completely eliminated. The dual radars sinking down ensure 360-degree obstacle avoidance at the bottom, without encroaching on loading and unloading space, and the radar line of sight will not be obstructed by the carried goods, which is extremely suitable for the engineering and practical needs of handling robots.

[0072] Overcoming the challenge of perception aliasing in standardized environments: Addressing the inherent limitation of dual LiDAR systems easily getting lost in repetitive greenhouse structures, visual landmarks with non-geometrically unique ID attributes are introduced. Absolute positioning constraints with unique identifiers are provided to eliminate positional ambiguity caused by repetitive structures.

[0073] Heterogeneous perception complementarity enhances system robustness: the underlying radar is used for high-frequency local positioning and obstacle avoidance, while the vehicle-mounted camera continuously detects visual landmarks; once the robot observes a visual landmark that meets the validity conditions, it generates visual landmark constraints and corrects the laser SLAM pose, thereby periodically eliminating the cumulative positioning error generated in repetitive structural environments and improving the accuracy and stability of continuous navigation.

[0074] like Figure 9 As shown, this embodiment proposes an autonomous navigation device for a greenhouse transport robot, applicable to any of the aforementioned autonomous navigation methods for greenhouse transport robots. The transport robot includes Simultaneous Localization and Mapping (SLAM), a top-mounted vehicle-mounted camera, a front-mounted LiDAR and a rear-mounted LiDAR on the chassis, and multiple visual landmarks are set in the top space of the greenhouse. The device may include: The fusion unit 101 is used to fuse and filter the original 3D point clouds generated by the front lidar and the rear lidar to generate a clean fused point cloud set. Matching unit 102 is used to input the clean fused point cloud set into SLAM for local scan matching and odometry calculation to obtain local odometry and local trajectory; The filtering unit 103 is used to perform visual landmark detection and effective observation filtering on the image stream captured by the vehicle-mounted camera in order to obtain effective visual landmark observation data. Solving unit 104 is used to perform three-dimensional pose calculation and cross-dimensional constraints on effective visual landmark observation data to obtain visual landmark constraint nodes; The control unit 105 is used to control the global pose and motion trajectory of the transport robot based on local odometry, local trajectory, visual landmark constraint nodes and each visual landmark.

[0075] It should be noted that the processing procedures and beneficial effects of the fusion unit 101, matching unit 102, filtering unit 103, solving unit 104, and control unit 105 can be referred to respectively. Figure 1 Steps S101 to S105 are not described in detail here.

[0076] Optionally, the fusion unit 101 is also used for: Using the created extrinsic transformation matrix, the original 3D point clouds generated by the front and rear lidars are mapped to the chassis coordinate system to obtain the first and second mapped point clouds. The first and second mapped point clouds are concatenated and fused to generate an initial point cloud set containing a 360-degree omnidirectional view around the chassis. The statistical outlier filter is invoked to identify and remove isolated noise points in the initial point cloud set, resulting in a clean fused point cloud set.

[0077] Optionally, the matching unit 102 is also used for: The pure fused point cloud set is input into the front end of SLAM so that the front end of SLAM: scans and matches the pure fused point cloud set with the local subgraph at the current time, and solves the optimal two-dimensional pose transformation of the pure fused point cloud set relative to the local subgraph at the current time through nonlinear optimization to obtain the two-dimensional pose transformation at the current time. Based on the two-dimensional pose transformation at the current moment and the two-dimensional pose transformation at the previous moment, the pose increment at adjacent moments is determined. The pose increment at adjacent moments is used as the local odometry at the current moment. The local odometry at consecutive moments is accumulated in chronological order to obtain the local trajectory.

[0078] Optionally, the filtering unit 103 is also used for: For any frame in the image stream, visual landmark detection is performed on the frame. When a target visual landmark is detected in the frame, the distance between the target visual landmark and the handling robot is determined. If the distance is not greater than a set distance threshold, the unique identifier and corner coordinates of the target visual landmark are extracted. If the correction identifier stored in the preset storage space is inconsistent with the unique identifier of the target visual landmark, the current observation result is determined to be a valid observation. The correction identifier in the preset storage space is deleted, and the unique identifier of the target visual landmark is stored as a new correction identifier in the preset storage space. The frame image, the timestamp of the frame image, the unique identifier of the target visual landmark, and the corner coordinates are taken as the whole as valid visual landmark observation data.

[0079] Optionally, the solution unit 104 is also used for: The frame images in the effective visual landmark observation data are preprocessed to obtain the corresponding binarized image data. Connectivity analysis and edge extraction are performed on the binarized image data to identify the closed contours of the binarized image data. Polygon fitting is performed on the closed contours to remove noise points that do not conform to the geometric features of the label, and candidate quadrilateral contour data are obtained. Extract the four vertices of the candidate quadrilateral contour data, call the sub-pixel level interpolation algorithm to perform gradient search in the pixel neighborhood of the four vertices to eliminate digital sampling error, and obtain the corresponding four two-dimensional pixel corner coordinate data; The corner coordinates of four two-dimensional pixels and the corner coordinates in the effective visual landmark observation data are input into the perspective n-point pose PnP solver to solve the pose and obtain the initial three-dimensional relative pose data of the label coordinate system relative to the camera coordinate system. The initial 3D relative pose data is mapped to the chassis coordinate system to obtain the mapped 3D relative pose data. Based on the created rotation matrix, the mapped 3D relative pose data is flattened in two dimensions to generate the 2D translation coordinates of the visual landmark relative to the chassis, pure 2D pose quaternary data, translation weights and rotation weights. The two-dimensional translation coordinates, pure two-dimensional attitude four-dimensional data, translation weights, rotation weights, timestamps of frame images in the effective visual landmark observation data, and unique identifiers of target visual landmarks are used as a whole as visual landmark constraint nodes.

[0080] Optionally, the control unit 105 is also used for: Obtain the global pose of each visual landmark in the global map coordinate system; The local odometry, local trajectory, visual landmark constraint nodes, and the global pose of each visual landmark in the global map coordinate system are input into the backend of SLAM so that SLAM can construct visual landmark residual terms. The visual landmark residual terms are inserted into the sparse pose adjustment function to obtain the target function; The objective function is solved to obtain the corrected global pose and optimized trajectory of the robot. The global pose and motion trajectory of the handling robot are controlled based on the corrected global pose and optimized trajectory.

[0081] Optional, the visual landmark residual term is: ; in, For visual landmark residuals, Indicates the index number of the visual landmark participating in the optimization at the current moment; and These represent pose composition and inverse composition operations, respectively. Indicates the first The actual observed relative pose of a visual landmark with respect to the chassis coordinate system; This indicates the current optimized absolute pose of the robot chassis; It is an abbreviation for visual landmark. This represents the minimum covariance information matrix specifically assigned to visual landmark observations.

[0082] The autonomous navigation device for greenhouse transport robots proposed in this embodiment can solve the problem of laser positioning divergence caused by the greenhouse corridor effect and the inability to achieve stable and reliable continuous navigation in related technologies. It breaks the geometric symmetry of the environment and realizes robust continuous positioning and navigation of transport robots in high-load and highly similar environments inside greenhouses.

[0083] In this embodiment, the autonomous navigation device for the greenhouse transport robot is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit), a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0084] This invention also provides a computer device having the above-described features. Figure 9 The image shows an autonomous navigation device for a greenhouse transport robot.

[0085] Please see Figure 10 The present invention provides a schematic diagram of the structure of a computer device according to an optional embodiment. The computer device includes one or more processors 10, a memory 20, and interfaces for connecting the various components, including high-speed interfaces and low-speed interfaces. The various components are interconnected via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on an external input / output device (such as a display device coupled to the interface). In some optional embodiments, multiple processors and / or multiple buses can be used with multiple memories, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 10 Take a processor 10 as an example.

[0086] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0087] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0088] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function. The data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, which can be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0089] Memory 20 may include volatile memory, such as random access memory. Memory may also include non-volatile memory, such as flash memory, hard disk, or solid-state drive. Memory 20 may also include combinations of the above types of memory.

[0090] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0091] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An autonomous navigation method for a greenhouse transport robot, characterized in that, The handling robot includes Simultaneous Localization and Mapping (SLAM), a top-mounted vehicle-mounted camera, and front and rear LiDAR sensors mounted on the chassis. Multiple visual landmarks are installed in the roof space of the greenhouse. The method includes: The original 3D point clouds generated by the front-end lidar and the rear-end lidar are fused and filtered to generate a clean fused point cloud set; The pure fused point cloud set is input into the SLAM for local scan matching and odometry calculation to obtain local odometry and local trajectory. Visual landmark detection and effective observation screening are performed on the image stream captured by the vehicle-mounted camera to obtain effective visual landmark observation data; The effective visual landmark observation data is subjected to three-dimensional pose calculation and cross-dimensional constraints to obtain visual landmark constraint nodes; Based on the local odometry, the local trajectory, the visual landmark constraint nodes, and each visual landmark, the global pose and motion trajectory of the transport robot are controlled.

2. The method according to claim 1, characterized in that, The process of fusing and filtering the original 3D point clouds generated by the front-end LiDAR and the rear-end LiDAR to generate a clean fused point cloud set includes: Using the created extrinsic transformation matrix, the original 3D point clouds generated by the front lidar and the rear lidar are mapped to the chassis coordinate system to obtain the first mapped point cloud and the second mapped point cloud. The first mapped point cloud and the second mapped point cloud are concatenated and fused to generate an initial point cloud set containing a 360-degree omnidirectional view around the chassis. The statistical outlier filter is invoked to identify and remove isolated noise points in the initial point cloud set, resulting in the clean fused point cloud set.

3. The method according to claim 1, characterized in that, The step of inputting the clean fused point cloud set into the SLAM for local scan matching and odometry calculation to obtain local odometry and local trajectory includes: The pure fused point cloud set is input into the front end of the SLAM, so that the front end of the SLAM: scans and matches the pure fused point cloud set with the local sub-image at the current time, and solves the optimal two-dimensional pose transformation of the pure fused point cloud set relative to the local sub-image at the current time through nonlinear optimization, so as to obtain the two-dimensional pose transformation at the current time. Based on the two-dimensional pose transformation at the current moment and the two-dimensional pose transformation at the previous moment, the pose increment at adjacent moments is determined. The pose increment at adjacent moments is used as the local odometry at the current moment. The local odometry at consecutive moments is accumulated in chronological order to obtain the local trajectory.

4. The method according to claim 1, characterized in that, The step of performing visual landmark detection and effective observation filtering on the image stream captured by the vehicle-mounted camera to obtain effective visual landmark observation data includes: For any frame image in the image stream, visual landmark detection is performed on the frame image. When a target visual landmark is detected in the frame image, the distance between the target visual landmark and the transport robot is determined. If the distance is not greater than a set distance threshold, the unique identifier and corner coordinates of the target visual landmark are extracted. If the correction identifier stored in the preset storage space is inconsistent with the unique identifier of the target visual landmark, the current observation result is determined to be a valid observation. The correction identifier in the preset storage space is deleted, and the unique identifier of the target visual landmark is stored as a new correction identifier in the preset storage space. The frame image, the timestamp of the frame image, the unique identifier of the target visual landmark, and the corner coordinates are taken as the valid visual landmark observation data.

5. The method according to claim 1, characterized in that, The process of performing 3D pose calculation and cross-dimensional constraints on the effective visual landmark observation data to obtain visual landmark constraint nodes includes: The frame images in the effective visual landmark observation data are preprocessed to obtain the corresponding binarized image data. Connectivity analysis and edge extraction are performed on the binarized image data to identify the closed contours of the binarized image data. Polygon fitting is performed on the closed contours to remove noise points that do not conform to the geometric features of the label, and candidate quadrilateral contour data is obtained. The four vertices of the candidate quadrilateral contour data are extracted, and a sub-pixel interpolation algorithm is called to perform gradient search in the pixel neighborhood of the four vertices to eliminate digitization sampling error and obtain the corresponding four two-dimensional pixel corner coordinate data. The four two-dimensional pixel corner coordinate data and the corner coordinates in the effective visual landmark observation data are input into the perspective n-point pose PnP solver to solve the pose and obtain the initial three-dimensional relative pose data of the label coordinate system relative to the camera coordinate system. The initial three-dimensional relative pose data is mapped to the chassis coordinate system to obtain the mapped three-dimensional relative pose data. Based on the created rotation matrix, the mapped three-dimensional relative pose data is flattened in two dimensions to generate the two-dimensional translation coordinates, pure two-dimensional pose quaternion data, translation weights and rotation weights of the visual landmark relative to the chassis. The two-dimensional translation coordinates, the pure two-dimensional attitude quaternion data, the translation weight, the rotation weight, the timestamp of the frame image in the effective visual landmark observation data, and the unique identifier of the target visual landmark are used as a whole as the visual landmark constraint node.

6. The method according to claim 1, characterized in that, The control of the global pose and motion trajectory of the transport robot based on the local odometry, the local trajectory, the visual landmark constraint nodes, and each visual landmark includes: Obtain the global pose of each visual landmark in the global map coordinate system; The local odometry, the local trajectory, the visual landmark constraint nodes, and the global pose of each visual landmark in the global map coordinate system are input into the backend of the SLAM so that the SLAM can construct visual landmark residual terms. The visual landmark residual term is inserted into the sparse pose adjustment function to obtain the target function; Solving the objective function yields the corrected global pose and optimized trajectory of the robot. The global pose and motion trajectory of the transport robot are controlled based on the corrected global pose and optimized trajectory.

7. The method according to claim 6, characterized in that, The visual landmark residual term is: ； in, For the visual landmark residual term, This indicates the index number of the visual landmark participating in the optimization at the current moment; and These represent pose composition and inverse composition operations, respectively. Indicates the first The actual observed relative pose of the visual landmarks with respect to the chassis coordinate system; This indicates the current optimized absolute pose of the robot chassis; This is an abbreviation for the visual landmark. This represents the minimum covariance information matrix specifically assigned to the visual landmark observation item.

8. An autonomous navigation device for a greenhouse transport robot, characterized in that, An autonomous navigation method for a greenhouse transport robot according to any one of claims 1 to 7, wherein the transport robot includes Simultaneous Localization and Mapping (SLAM), a vehicle-mounted camera mounted on the top, a front-mounted LiDAR and a rear-mounted LiDAR mounted on the chassis, and multiple visual landmarks are provided in the top space of the greenhouse; the device includes: The fusion unit is used to fuse and filter the original 3D point clouds generated by the front lidar and the rear lidar to generate a clean fused point cloud set. The matching unit is used to input the clean fused point cloud set into the SLAM to perform local scan matching and odometry calculation to obtain local odometry and local trajectory; The filtering unit is used to perform visual landmark detection and effective observation filtering on the image stream captured by the vehicle-mounted camera in order to obtain effective visual landmark observation data. The calculation unit is used to perform three-dimensional pose calculation and cross-dimensional constraints on the effective visual landmark observation data to obtain visual landmark constraint nodes; The control unit is used to control the global pose and motion trajectory of the transport robot based on the local odometry, the local trajectory, the visual landmark constraint nodes, and each of the visual landmarks.

9. A computer device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the autonomous navigation method for a greenhouse transport robot as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the autonomous navigation method for the greenhouse transport robot according to any one of claims 1 to 7.

Images