Orchard plant protection robot cluster operation system

Abstract

The invention relates to the technical field of automatic agricultural equipment, in particular to an orchard plant protection robot cluster operation system which comprises a central scheduling module, a cluster communication module, an environment modeling module and an updating module. The central scheduling module communicates with a robot through a communication technology, and the cluster communication module constructs a robot cluster communication network by using a communication network configuration strategy; the environment modeling module evaluates the complexity of a local environment through a sensor assembly, and generates a local map containing orchard terrain, vegetation density and obstacle information. The updating module splices the local maps to generate an operation map, updates a communication network according to the environmental complexity, and optimizes task allocation and path planning of the robot cluster; the system has the functions of task scheduling, communication network optimization and map updating, and efficient collaborative operation of a robot cluster in a complex orchard environment is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of automated agricultural equipment, and in particular to an orchard plant protection robot cluster operation system. Background Art

[0002] With the rapid development of agricultural automation, plant protection operations in orchard management are gradually moving towards intelligent and unmanned operations. Traditional orchard plant protection operations mostly rely on manual operations, which are labor-intensive and inefficient. It is difficult to achieve precise application of pesticides and optimal resource allocation, and it is impossible to effectively deal with the complex terrain, vegetation density and diverse obstacles in the orchard. In addition, a single plant protection robot is limited by its operating range and environmental perception capabilities, making it difficult to cope with the complex environment and changing task requirements of large-scale orchards. In the prior art, although there are some plant protection systems based on robotics technology, these systems usually lack the capabilities of cluster collaboration, real-time communication and dynamic task scheduling, and cannot dynamically adjust the operation path and task allocation according to actual environmental changes, resulting in insufficient operation efficiency and effectiveness.

[0003] For example, the patent publication number CN117930841A discloses an orchard modeling and picking robot control method based on three-dimensional point cloud reconstruction, which includes the following steps: Step 1. According to the geometric features, physical properties and kinematic model of the picking robot, a robot digital twin model is established to obtain the three-dimensional point cloud information of the litchi orchard, and the point cloud information is converted into a visual three-dimensional map through the three-dimensional point cloud reconstruction technology to build a virtual orchard platform; Step 2. Based on the virtual orchard platform and the robot digital twin model, a deep reinforcement learning model is established; Step 3. The robot digital twin model and the deep reinforcement learning model are integrated into a reverse control system; Step 4. Based on the digital twin virtual orchard platform, a multi-agent collaborative scheduling system for cluster robot operations is established, and the cluster operation algorithm is scheduled and simulated under the real-time collected orchard environmental constraints to realize collaborative operation control between multiple picking robots or transport robots and realize autonomous collision avoidance picking.

[0004] The above patents all have the problems raised by this background technology: the orchard environment is complex, the terrain is changeable, and the vegetation density is high. The existing communication network is easily interfered by obstacles and terrain, resulting in unstable communication, affecting the coordinated operation and information transmission between robots. To solve the above problems, this application designs an orchard plant protection robot cluster operation system. Summary of the invention

[0005] The technical problem to be solved by the present invention is to address the deficiencies of the prior art and provide an orchard plant protection robot cluster operation system, including a central scheduling module, a cluster communication module, an environmental modeling module and an update module. The central scheduling module communicates with the robot through communication technology, and the cluster communication module uses the communication network configuration strategy to build a robot cluster communication network. The environmental modeling module evaluates the complexity of the local environment through sensor components and generates a local map containing orchard terrain, vegetation density and obstacle information. The update module splices the local maps to generate an operation map, and updates the communication network according to the complexity of the environment to optimize the task allocation and path planning of the robot cluster. The system has task scheduling, communication network optimization and map update functions, and realizes efficient collaborative operation of robot clusters in complex orchard environments.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] An orchard plant protection robot cluster operation system, the system comprising a central scheduling module, a cluster communication module, an environment modeling module and an update module;

[0008] The central dispatch module is used to establish communication with the robot through communication technology;

[0009] The cluster communication module is configured with a communication network configuration strategy, and the communication network configuration strategy is used to build a cluster communication network between robots;

[0010] The environment modeling module is configured with an environment perception strategy, which evaluates the complexity of the local environment and generates a local map based on the sensor components configured on the robot. The local environment complexity includes orchard terrain, vegetation density and obstacle information;

[0011] The update module is connected to the central scheduling module, splices the local maps to generate a work map, updates the cluster communication network according to the complexity of the local environment, and updates the task allocation and path planning of the robot cluster through the updated cluster communication network.

[0012] The cluster communication module comprises:

[0013] A network construction unit, used to construct a self-organizing communication network among robot clusters;

[0014] A network monitoring unit, used to monitor the status of the cluster communication network in real time, and generate an alarm and report to the central dispatch module when the communication is abnormal;

[0015] The communication network configuration strategy includes a network construction logic, and the network construction logic is used to establish a distributed cluster communication network through communication equipment. The network construction logic is configured in the network construction unit.

[0016] The network construction logic includes:

[0017] The robot cluster randomly selects a leader robot in the central area according to the initial placement position, and randomly selects a relay robot from the remaining robots in the central area;

[0018] The remaining robots create blank messages and send them to each relay robot. The relay robot counts the arrival delay of each robot transmitting the blank message, adds the arrival delay as the frame tail information to the blank message, and forwards the blank message to the leader robot. The blank message includes the number of transmissions and the remaining energy of the robot.

[0019] The leading robot receives the blank message and counts the arrival delay of the blank message transmitted by each relay robot, and calculates the total arrival delay based on the arrival delay of the blank message transmitted by the relay robot and the arrival delay of the blank message transmitted by the robot;

[0020] Compare the total arrival delay with the constraint conditions to determine whether the constraint conditions are met. If not, iterate the leader robot and relay robots to recalculate the total arrival delay.

[0021] If the constraints are met, the iteration is stopped, the latest positions of the leader robot and relay robots are determined, and the cluster communication network is constructed.

[0022] The environment modeling module comprises:

[0023] A data processing unit, used to obtain data from the sensor component and process the data, wherein the data includes three-dimensional point cloud data, spatial image data, IMU data and positioning sensor data;

[0024] The map generation unit is used to fuse the sensor data to generate a high-precision local map with semantic information. The multimodal sensor data fusion algorithm is used in the map generation process, and the SLAM technology is combined to generate an operation map with positioning information and semantic labels.

[0025] A complexity evaluation unit, used to evaluate the complexity of the local environment and calculate the local complexity;

[0026] A perception data transmission unit, used for transmitting the local map and the local complexity to a central scheduling module;

[0027] The environmental perception strategy includes sensor data processing logic and complexity calculation logic;

[0028] The sensor data processing logic is configured in the data processing unit, and the sensor data processing logic is used to process the data acquired from the sensor component to extract the front point cloud features and the front entity features;

[0029] The complexity calculation logic is configured in the complexity evaluation unit, and the complexity calculation logic is used to calculate the obstacle complexity and terrain complexity according to the output of the data processing unit, and calculate the local complexity according to the obstacle complexity and terrain complexity.

[0030] The sensor data processing logic includes:

[0031] The three-dimensional point cloud data is randomly sampled by random consistency sampling to obtain the observed point cloud data, the neighborhood points of the observed point cloud data are calculated according to the Lagrange multiplier method, the observed point cloud data and the neighborhood points are discretely determined, and the discrete degree of the observed point cloud data is calculated;

[0032] The coplanarity of the observed point cloud data is judged according to the degree of discreteness, the observed point cloud data that meets the coplanarity condition is merged and the point set space is allocated, and the front point cloud features are calculated;

[0033] Traverse the pixels of the spatial image, calculate the rigid transformation vector and the change amount in the iterative process of two adjacent pixels, and calculate the entity feature value of the pixel point according to the dynamic adjustment factor, the rigid transformation vector and the change amount;

[0034] Compare the entity feature value of the pixel with the entity threshold. If it is greater than the entity threshold, mark the pixel as a sampling point, and confirm the neighborhood of the sampling point according to the local shape distribution function;

[0035] The feature description vector of the sampling point is calculated based on the normal vector angle between the sampling point and the pixel points in the neighborhood of the sampling point. The feature description vector is aggregated through the maximum pooling operation, and then the aggregated features are further abstracted based on sparse convolution to obtain the front entity features through the RPN network.

[0036] The complexity calculation logic includes:

[0037] Constructing an obstacle recognition network, taking the front point cloud features and the front entity features as input parameters of the obstacle recognition network, training the input parameters through the obstacle recognition network, and outputting obstacle complexity;

[0038] The three-dimensional point cloud data set is converted into DEM data, where the DEM data represents the grid data of the earth's surface terrain, the elevation value around each grid cell is calculated based on the DEM data, the height analysis is performed through GIS software, the inclination angle of the local terrain is calculated through the difference method, and the local surface of the orchard terrain is fitted through the least square method according to the inclination angle to calculate the ruggedness, and the terrain complexity is calculated according to the ruggedness;

[0039] The vegetation coverage rate is obtained by calculating the point cloud density within a unit area, and the vegetation height distribution is obtained by counting the height differences between adjacent point cloud points. The vegetation density is calculated based on the vegetation coverage rate and the vegetation height distribution;

[0040] The local complexity is calculated by weighted average method according to the obstacle complexity, terrain complexity and vegetation density.

[0041] The update module comprises:

[0042] A map updating unit, used for receiving the local maps generated by the environment modeling module and splicing the local maps;

[0043] The communication network optimization unit is used to re-optimize the communication network by adjusting the positions of the leader robot and the relay robot when the environment or operating conditions change;

[0044] The task scheduling unit adjusts the task allocation and path planning of the robot cluster according to the spliced ​​global operation map and the network status updated by the cluster communication module;

[0045] The task issuing unit adjusts the map resolution according to the task requirements of the robot, slices the map according to the requirements, and transmits it to the corresponding robot through the central scheduling module.

[0046] The map updating unit is provided with a splicing logic, and the splicing logic includes:

[0047] Performing a preprocessing operation on the local map input by each robot, wherein the preprocessing operation includes denoising, filtering and resampling;

[0048] The point cloud features of the preprocessed local map are extracted through a deep convolutional neural network to obtain geometric point cloud features;

[0049] By calculating the geometric transformation relationship between the geometric point cloud features, the geometric point cloud features are aligned to a unified coordinate system;

[0050] The registered point clouds are fused to generate a global point cloud map.

[0051] The communication network optimization unit is configured with a network optimization logic, and the network optimization logic includes:

[0052] According to the local complexity, the local environment is divided into complex environment and simple environment, and the environment type is determined;

[0053] Adjust the redundancy of the communication network according to the environment type, select multiple relay robots for communication in complex environments, and select relay robots for communication through the shortest path algorithm in simple environments;

[0054] Adjust the transmission power of the communication network according to the environment type, increase the transmission power in complex environments, and reduce the transmission power in simple environments;

[0055] Adjust the communication frequency of the communication network according to the environment type, increase the communication frequency in complex environments, and reduce the number of frequency switching and signal transmission density in simple environments.

[0056] The task scheduling unit is configured with a task allocation logic, and the task allocation logic includes:

[0057] Evaluate the execution capability of each robot and calculate the robot score;

[0058] Sort the tasks to be performed in the local environment by complexity and match them according to the complexity and robot score;

[0059] Based on the allocation results, each robot receives and executes its independent task, while feeding back the task status to the leader robot in real time.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. The present invention constructs a cluster communication module, multiple robots can work together, and dynamically build a self-organizing cluster communication network, which avoids the problem of limited operating range of a single robot and improves the coverage and overall operating efficiency of orchard plant protection operations;

[0062] 2. The present invention uses multimodal sensors (such as lidar, camera, IMU, etc.) to evaluate the local environmental complexity of the orchard in real time, generate high-precision local maps, and accurately identify terrain, vegetation density and obstacle information through complexity calculation logic, providing a more intelligent and accurate path planning basis for robot operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0064] Figure 1 This is a flow chart of a cluster operation method of orchard plant protection robots according to Embodiment 1 of the present invention;

[0065] Figure 2 This is a schematic diagram of a cluster communication network transmission path according to Embodiment 1 of the present invention;

[0066] Figure 3 This is a diagram of the obstacle recognition network structure of Example 1 of the present invention;

[0067] Figure 4This is a structural diagram of an orchard plant protection robot cluster operation system according to Embodiment 2 of the present invention;

[0068] Figure 5 This is a structural diagram of the robot-mounted sensor composition of Example 2 of the present invention;

[0069] Figure 6 This is a flow chart of a cluster operation method of an orchard plant protection robot according to Embodiment 2 of the present invention;

[0070] Figure 7 This is a block diagram of the robot node mapping principle in Example 2 of the present invention;

[0071] Figure 8 This is a schematic diagram of the map stitching process according to Embodiment 2 of the present invention. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0073] Example 1

[0074] See also Figure 1 The present invention provides an embodiment: a cluster operation method of an orchard plant protection robot, the specific steps of the method are as follows:

[0075] S1: Arrange a robot cluster in the orchard area. After the initial deployment, the robot cluster builds a cluster communication network and communicates with the central dispatching unit.

[0076] In this step, multiple robots are first deployed to form a collaborative robot cluster. Each robot is equipped with a sensor component to provide multimodal data for fusion perception and positioning. The robot not only has the ability to build maps and orchard operations, but is also designed with adaptive task planning and collaborative operation modes. Then, a cluster communication network is established through communication equipment.

[0077] S2: Each robot collects environmental data of the local area through the configured sensor components, generates a local map according to the environmental data, analyzes the environmental data, and calculates the local environmental complexity of the current area;

[0078] In this step, the sensor components equipped on each robot can collect various data of the local environment in real time. The sensor components include laser radar, image sensor, IMU sensor and wheel speed meter to ensure comprehensive perception of the environment and generate a point cloud map based on the three-dimensional point cloud data of the laser radar.

[0079] S3: transmitting the local map and the local environment complexity to the central dispatching unit, and the central dispatching unit splices the local maps through a point cloud registration algorithm to generate a global map;

[0080] S4: Feeding back the local environment complexity to the cluster communication network, and performing topology adjustment on the cluster communication network;

[0081] In this step, the cluster communication network will make corresponding topological adjustments based on the local environment complexity assessed by the robot. For example, in areas with weak communication signals or dense obstacles, the network may need to reconfigure the positions of communication nodes to optimize the signal transmission path and avoid communication interruption;

[0082] S5: Redistribute the tasks of the robot cluster through the adjusted cluster communication network, calculate the one-way tasks of the robots, and plan the paths of the robots according to the one-way tasks;

[0083] In this step, task allocation considers the current position, battery status and actual environment of each robot to optimize resource utilization and task execution efficiency. Based on environmental data and communication network information, a safe and efficient path is selected to avoid obstacles, dangerous areas or communication blind spots. During task execution, the robot continuously adjusts itself to ensure that the robot always moves along the optimal path in a dynamic environment.

[0084] S6: According to the adjusted task, the spliced ​​global point cloud map is adjusted through dynamic map resolution and sliced ​​on demand;

[0085] See also Figure 2 , Schematic diagram of the transmission path of the cluster communication network of the embodiment of the present invention, the distributed cluster communication network construction method randomly selects the leader robot and the relay robot, and optimizes based on the blank message transmission delay between the robots, and finally determines the optimal network structure. Through multiple iterations, it is ensured that the communication network can operate efficiently in a complex environment. The specific construction method of the distributed cluster communication network is as follows:

[0086] S1.1: The robot cluster randomly selects a leader robot in the central area based on the initial deployment position. The leader robot is responsible for collecting data from other robots and making global decisions. A relay robot is randomly selected from the remaining robots in the central area. The relay robot will act as an intermediate node for transmitting data to ensure that information can be reliably transmitted to the leader robot.

[0087] S1.2: The remaining robots create blank messages and send them to each relay robot. The relay robot counts the arrival delay of each robot transmitting the blank message, adds the arrival delay as the frame tail information to the blank message, and forwards the blank message to the leader robot. The blank message includes the number of transmissions and the remaining energy of the robot.

[0088] S1.3: The leading robot receives the blank message and counts the arrival delay of the blank message transmitted by each relay robot, and calculates the total arrival delay based on the arrival delay of the blank message transmitted by the relay robot and the arrival delay of the blank message transmitted by the robot;

[0089] S1.4: Compare the total arrival delay with the constraint conditions to determine whether the constraint conditions are met. If not, iterate the leader robot and relay robots to recalculate the total arrival delay;

[0090] S1.5: If the constraints are met, stop the iteration, determine the latest positions of the leader robot and relay robots, and build a distributed cluster communication network;

[0091] The total arrival delay is defined as the sum of the time from when a robot sends a message to when the leader robot receives the message. The total arrival delay involves multiple factors, including the message transmission rate, the number of failed retransmissions, and the choice of transmission path. The entire delay is divided into the delay of the blank message reaching the relay robot and the delay of forwarding to the leader robot.

[0092] In the actual transmission process, message transmission failure is inevitable, especially in complex environments. Therefore, the impact of failed retransmission on the total delay needs to be considered. The exponential correction term is introduced to make corrections. When the number of failures increases, the correction term will have a stronger impact on the total delay, thereby increasing the total delay.

[0093] The calculation formula for the total arrival delay is:

[0094]

[0095] Among them, T d represents the total arrival delay, e represents the exponential function, μ represents the self-similarity coefficient, S r represents the average transmission rate of the robot sending blank messages, N represents the number of failures of the robot to send blank messages, H represents the standard fractal Brownian motion of the robot transmitting blank messages, m represents a single relay robot, M represents the total number of relay robots, α m represents the transmission rate from the robot to the mth relay robot, t m represents the transmission delay of the blank message to the mth relay robot, Q m represents the transmission delay from the mth relay robot to the leader robot, t Qmrepresents the number of retransmissions of the blank message transmitted by the mth relay robot, X m Represents the standard fractal Brownian motion of the mth relay robot receiving a blank message;

[0096] The constraints include maximum delay constraint, network stability and energy consumption constraint. The iteration of the leader robot and relay robot includes:

[0097] Encode the robot's position and relay node selection into antibodies, each antibody represents a complete network configuration solution;

[0098] The fitness function is calculated according to the total arrival delay, the antibody fitness is calculated according to the fitness function, and the antibody with high fitness is selected for cloning;

[0099] During the cloning process, some of the antibody genes (i.e., the robot position or relay node selection) are mutated to generate a new generation of antibody populations;

[0100] The antibodies with the highest fitness are retained in the memory set, and in each iteration, the antibodies in the memory set are introduced into the generation process of the next generation of antibody groups, gradually approaching the global optimal solution;

[0101] The specific steps for analyzing the environmental data are as follows:

[0102] S2.1: The robot scans the space ahead through the laser radar to obtain a 3D point cloud data set of the space ahead. The laser radar measures the target distance by emitting laser pulses and receiving reflected signals to generate high-precision 3D point cloud data. The features of the point cloud ahead are extracted based on the 3D point cloud data set. The robot can detect suspected obstacles and their positions in the environment in real time and with high precision, quickly filter out irrelevant environmental information, focus on obstacles that may affect the robot's movement, and improve the accuracy of subsequent recognition.

[0103] The laser radar is used to scan the space in front of the robot. When the pulse emitted by the laser radar encounters an obstacle, it will be reflected back to the receiver. By measuring the time difference between the emission and reception of the laser pulse, the target distance is calculated, thereby generating three-dimensional point cloud data of the environment.

[0104] By analyzing the 3D point cloud data of the front space in real time, it can quickly identify obstacles in the environment and eliminate irrelevant information, effectively improving the robot's detection accuracy of obstacles in front, reducing invalid calculations, reducing computing resource consumption, and improving the accuracy and efficiency of the robot's perception of the environment;

[0105] S2.2: Jointly calibrate the LiDAR and image sensor. Through the joint calibration process, the three-dimensional point cloud data of the LiDAR is spatially aligned with the two-dimensional image obtained by the image sensor. The purpose of joint calibration is to ensure that the data of the two sensors can be compared and fused in the same coordinate system. The spatial image of the coordinate position of the point cloud feature in front is obtained by the image sensor, and the physical features in front are obtained based on the spatial image. It can provide more texture and color information for the suspected obstacles in the point cloud data, helping to further confirm and identify the specific type and attributes of the obstacle;

[0106] Through joint calibration, the obstacle positions in the LiDAR point cloud data can be accurately mapped to the pixels in the two-dimensional image. The calibration algorithm obtains images and point cloud data from multiple perspectives, calculates the camera's intrinsic and extrinsic parameters, and generates a transformation matrix to accurately align the LiDAR point cloud data with the image sensor's coordinate system.

[0107] S2.3: constructing an obstacle recognition network, taking the front point cloud features and the front entity features as input parameters of the obstacle recognition network, training the input parameters through the obstacle recognition network, and outputting obstacle complexity, which can reflect the size and shape of the obstacle. The obstacle recognition network can accurately determine the type, size and position of the obstacle, and calculate its complexity based on this information;

[0108] S2.4: Convert the three-dimensional point cloud data set into DEM data. DEM data represents the grid data of the earth's surface terrain. Calculate the elevation value around each grid cell based on the DEM data. Perform height analysis through GIS software. Calculate the inclination angle of the local terrain through the difference method. Fit the local surface of the orchard terrain through the least square method according to the inclination angle, calculate the ruggedness, and calculate the terrain complexity according to the ruggedness. The terrain ruggedness is calculated by analyzing the change of the elevation value of each grid cell. The greater the height difference around the grid cell, the more complex the terrain. The terrain complexity can reflect the terrain undulation and unevenness of the robot's driving area, provide more accurate terrain information for the robot's path planning, and ensure the stable operation of the robot in complex terrain.

[0109] S2.6: Obtain vegetation coverage by calculating the point cloud density within a unit area, and obtain vegetation height distribution by counting the height differences between adjacent point cloud points. Calculate vegetation density based on the vegetation coverage and the vegetation height distribution;

[0110] Analyze the coverage rate and height distribution of vegetation based on point cloud data. By calculating the density of point cloud data per unit area, the density of vegetation in the orchard is evaluated. The vertical distribution of vegetation is analyzed by the height difference of point cloud data. The vegetation coverage rate reflects the degree to which the ground is covered by vegetation, and the height distribution is used to distinguish vegetation of different heights (such as low grass and tall trees).

[0111] Vegetation density can provide the robot with vegetation distribution information in the operating area, helping the robot determine whether the operating path is blocked and adjust the operating strategy according to the density;

[0112] S2.5: Calculate the local complexity by weighted average method based on the obstacle complexity, terrain complexity and vegetation density. Local complexity is a comprehensive indicator that reflects the difficulty of the robot's action in the corresponding area. It is used to adjust the coordination strategy and path planning of the robot cluster to ensure its efficient operation in a complex environment;

[0113] The specific steps for extracting the front point cloud features based on the 3D point cloud dataset are as follows:

[0114] S2.1.1: Randomly sample the three-dimensional point cloud data through random consistency sampling to obtain the observed point cloud data, calculate the neighborhood points of the observed point cloud data according to the Lagrange multiplier method, make discrete determinations on the observed point cloud data and the neighborhood points, and calculate the discrete degree of the observed point cloud data;

[0115] Random Sample Consensus (RANSAC) is an algorithm for extracting features from a data set, especially for geometric feature extraction in point clouds. This step uses RANSAC to randomly sample from the preprocessed point cloud data set, extract key observation point cloud data, and use the Lagrange multiplier method to calculate the neighborhood points of each observation point. The Lagrange multiplier method is used to optimize the geometric relationship between the observation point and its neighborhood points to ensure the optimal distance distribution between the observation point and the neighborhood points.

[0116] By randomly selecting points from the point cloud data and constructing a geometric model (such as a plane, a curved surface, etc.), the fit between the model and the remaining data is evaluated, and outliers that do not conform to the model are removed;

[0117] S2.1.2: Traverse the three-dimensional point cloud data set, obtain the discrete degree of all observed point cloud data, and compare the discrete degree with the ground point cloud threshold. If it is less than or equal to the ground point cloud threshold, delete the observed point cloud data and eliminate useless ground point clouds;

[0118] In this step, the discrete degree of the observed point cloud data is calculated, and by comparing the discrete degree with the set ground point cloud threshold, useless ground point cloud data is eliminated, and point cloud data related to obstacles is retained. The local neighborhood density of each point is calculated using a discreteness determination algorithm (such as a statistical method based on point cloud density). If the density is low, the point is identified as a ground point cloud or other useless data;

[0119] It can effectively reduce the redundant information of the point cloud data set, eliminate ground points irrelevant to robot navigation, improve the efficiency of feature extraction, and reduce the subsequent calculation burden;

[0120] S2.1.3: Perform coplanarity judgment on the remaining observation point cloud data, merge the observation point cloud data that meet the coplanarity condition and allocate the point set space to obtain the front point cloud features;

[0121] Coplanarity judgment is used to identify point cloud data distributed on the same plane. Point clouds that meet the coplanarity condition will be merged into a group and assigned to the point set space for subsequent processing as part of the obstacle feature. By calculating the angle between the normal vector of each point and the normal vector of the adjacent point, if the angle is less than the set threshold, it is judged that these points are located on the same plane and meet the coplanarity. The coplanar point cloud data are aggregated together to form a complete plane feature, which is convenient for subsequent obstacle identification and classification;

[0122] The specific steps to obtain the entity features of suspected obstacles based on spatial images are as follows:

[0123] S2.2.1: Preprocess the spatial image, where the preprocessing includes regional network joint adjustment and multi-view image matching;

[0124] Regional network joint adjustment is to use joint adjustment technology to optimize and adjust the image errors of multi-view images in an overall manner when multi-view images are acquired. The adjustment process can reduce the geometric errors between images and ensure the accurate registration of each image in three-dimensional space.

[0125] Using stereo matching algorithms (such as SIFT, SURF and other feature matching algorithms), matching feature points are found in images from different perspectives to generate disparity maps or depth maps. This step can reconstruct the three-dimensional information in the image through stereo vision methods;

[0126] The processed spatial images are highly consistent, all pixels have been calibrated and can be compared and calculated in a unified coordinate system, providing high-precision data support for subsequent entity feature extraction;

[0127] S2.2.2: Based on the preprocessed spatial image, traverse the pixels of the spatial image, calculate the rigid transformation vector and the change amount in the iteration process between two adjacent pixels, and calculate the entity feature value of the pixel point based on the dynamic adjustment factor, the rigid transformation vector and the change amount. The rigid transformation vector reflects the rigid change of the object under different viewing angles, and the change amount captures the difference between pixels.

[0128] The rigid transformation vector is calculated by calculating the relative position change between each pixel and its neighboring points, usually described by a rotation matrix and a displacement vector;

[0129] The dynamic adjustment factor is combined with the rigid transformation vector to adjust the nonlinear changes that occur during the transformation process, which is used to more accurately calculate the entity feature value of the pixel point;

[0130] Through the rigid transformation vector and the change amount, the geometric changes of the object under different viewing angles can be reflected, thereby accurately capturing the structural information of the object. This calculation result provides more accurate geometric information for subsequent feature extraction;

[0131] S2.2.3: Based on the entity feature values ​​calculated from the rigid transformation vector and the change amount, traverse all pixels and compare the feature value of each pixel with the preset entity threshold. If the feature value is greater than the entity threshold, the pixel is marked as a sampling point, indicating that it belongs to a certain entity surface;

[0132] The calculation of entity eigenvalues ​​reflects the geometric stability of pixel points in space. If the eigenvalue is greater than the entity threshold, the point is likely to be an obstacle or an important entity surface;

[0133] S2.2.4: For each pixel marked as a sampling point, calculate the angle between its normal vector and other pixels in the neighborhood. Based on these normal vector angles, generate a feature description vector to describe the geometric features of the sampling point in its local neighborhood. The calculation formula for the sampling point feature description vector is:

[0134]

[0135] Among them, V f represents the sampling point feature description vector, t c represents the entity feature value of the sampling point, i represents a single pixel in the neighborhood of the sampling point, N represents the total number of pixels in the neighborhood of the sampling point, and θ i Represents the normal vector angle between the sampling point and the i-th pixel in the sampling point neighborhood, C i represents the curvature of the i-th pixel in the neighborhood of the sampling point, δ represents the normalization factor, e represents the exponential function, a and b represent the horizontal and vertical matrix coordinates of the sampling point, a i and b irepresents the horizontal and vertical matrix coordinates of the i-th pixel in the neighborhood of the sampling point, and c represents the convolution factor;

[0136] By calculating the angle between the normal vector of each sampling point and the pixels in its neighborhood, the surface shape of the point is described, and a vector is generated to represent the local geometric structure of the sampling point by combining multiple angle information. These vectors can be used for subsequent feature aggregation and classification. The feature description vector accurately describes the local geometric characteristics of each sampling point, allowing the robot to identify the shape of objects in the environment through this information, and further use it for path planning and obstacle detection;

[0137] S2.2.5: Aggregate the feature description vectors through the maximum pooling operation to simplify and centralize the local geometric features of each sampling point. Subsequently, use a sparse convolutional network to further abstract these features and simplify the complex point cloud data structure. Finally, generate the front entity features through the region proposal network (RPN);

[0138] The maximum value of the local area is extracted from the feature description vector to retain the most representative features. By performing convolution operations on sparse point cloud data, the geometric features in the data are further compressed and abstracted, enabling the network to process large-scale point cloud data at a low computational cost.

[0139] The maximum pooling operation can effectively reduce redundant features and retain important geometric information; the sparse convolutional network can extract key features from large-scale point cloud data, and then generate accurate entity features through the RPN network to help the robot identify obstacles or key objects in the front environment;

[0140] See also Figure 3 , a structural diagram of an obstacle recognition network according to an embodiment of the present invention, wherein the obstacle recognition network comprises an input layer, a feature extraction layer and a feature recognition layer;

[0141] The input layer establishes a feature sequence according to the input parameters, wherein the feature sequence includes a suspected obstacle laser point cloud feature sequence and a suspected obstacle entity point cloud feature sequence;

[0142] The feature extraction layer is used to extract feature vectors of feature sequences. The feature extraction layer includes a geometric feature extraction module and a spatial feature extraction module. The feature vectors include spatial features and geometric features.

[0143] The feature recognition layer is used to deeply fuse the spatial features and the geometric features to generate obstacle fusion point cloud data, and output obstacle information according to the obstacle fusion point cloud data;

[0144] The input layer is used to accumulate input parameters, establish a grey differential equation, calculate the least squares parameters of the grey differential equation, calculate the fitting value of the feature sequence by discretizing the least squares parameter solution, and obtain the suspected obstacle point cloud feature sequence and the suspected obstacle entity point cloud feature sequence;

[0145] The geometric feature extraction module is used to obtain the center point position of the suspected obstacle laser point cloud feature sequence and the suspected obstacle entity point cloud feature sequence through the attention mechanism, calculate the neighboring points of the center point according to the K nearest neighbor algorithm, use the center point and the neighboring points as the source of the Gaussian kernel function, output the local geometric features of the center point through the Gaussian kernel function, and finally map all local geometric features to a high-dimensional space through a residual connection network, perform a pooling operation, and output the geometric features;

[0146] The spatial feature extraction module includes a multi-layer perceptron network. Each layer first performs symmetrical operations on the suspected obstacle laser point cloud feature sequence and the suspected obstacle entity point cloud feature sequence through average pooling, and then spatially describes the suspected obstacle point cloud feature sequence and the suspected obstacle entity point cloud feature sequence through one-dimensional convolution and normalization. Finally, they are activated according to the ReLU activation function to obtain spatial features. The feature recognition layer maps the spatial features and geometric features into a tuple sequence [(Q1, K1), (Q2, K2), ..., (Q m ,K m )], where Q m represents the corresponding input of the mth feature in the geometric feature, K m Represent the corresponding input of the mth feature in the spatial feature, perform dot product on the corresponding input features in each candidate sequence in the tuple sequence, calculate the correlation score after the dot product of the candidate sequence by the Pearson correlation coefficient, compress the correlation score according to the sigmoid function, and compare it with the correlation threshold. If it is greater than the correlation threshold, retain the candidate sequence. If it is less than or equal to the correlation threshold, filter the candidate sequence, collect the local features of the candidate sequence retained in the tuple sequence, and enhance them. Splice the enhanced candidate sequence along the head sequence of the tuple sequence, fuse the local features through residual connection, and perform pooling operation on the feature vector after residual connection fusion to obtain obstacle fusion point cloud data;

[0147] The specific steps of S3 are as follows:

[0148] S3.1: performing a preprocessing operation on the local map input by each robot, wherein the preprocessing operation includes denoising, filtering and resampling;

[0149] After the local map is input, a series of preprocessing operations are first performed. These preprocessing operations are aimed at improving the accuracy and consistency of the local map and ensuring the stability and accuracy of subsequent feature extraction. The preprocessing steps include the following key operations:

[0150] Use statistical filtering algorithms (such as noise point filtering based on average distance or standard deviation) to remove abnormal noise points in the point cloud. These noise points are usually introduced by sensor errors or environmental factors;

[0151] Use Gaussian filtering or voxel grid filtering to smooth the point cloud to reduce random noise in the point cloud data while maintaining key geometric details;

[0152] Use voxel gridding to downsample the point cloud, reduce the amount of point cloud data and evenly distribute the point cloud data. The points in each voxel are averaged to generate a representative point to reduce redundant data;

[0153] After preprocessing, the quality of the local map is significantly improved, the noise points are effectively removed, the point cloud data is smoother and more consistent, and the complexity of subsequent calculations is reduced. At the same time, resampling reduces redundant point cloud data and significantly improves the efficiency of point cloud processing;

[0154] S3.2: The preprocessed local map is used for point cloud feature extraction through a deep convolutional neural network (CNN). The deep convolutional neural network extracts local and global features of point cloud data layer by layer. Usually, network architectures such as PointNet or PointCNN are used to extract features from point clouds, automatically capturing local shapes and global spatial distribution information in geometric point clouds;

[0155] Use 3D convolutional neural network to perform convolution operation on the local neighborhood of point cloud data to extract the geometric features of each point, such as normal vector, curvature, etc.

[0156] Through layer-by-layer convolution, local features are integrated into global features to capture the geometric and spatial information of the entire local map;

[0157] The geometric point cloud features extracted by the convolutional neural network can effectively capture the key geometric information and global structure of the local map. These features have extremely high accuracy for subsequent point cloud alignment and fusion;

[0158] S3.3: After point cloud feature extraction, the next step is to align these geometric point cloud features into a unified global coordinate system. The calculation of geometric transformation relationships usually uses the iterative closest point algorithm (ICP) to align different local maps;

[0159] Use the key points in the point cloud features to perform rough alignment through feature matching algorithms to generate a preliminary geometric transformation matrix, including a rotation matrix and a displacement vector;

[0160] Based on the preliminary alignment, the ICP algorithm is used to iterate step by step to further accurately adjust the alignment effect of the local map;

[0161] By calculating the geometric transformation relationship, each local map is accurately aligned to the global coordinate system, laying the foundation for subsequent point cloud fusion;

[0162] S3.4: After the alignment, the local point cloud maps are fused to form a complete global point cloud map. The point cloud fusion process includes integrating all aligned point cloud data into a unified coordinate system and dealing with redundancy and overlap issues of point clouds;

[0163] The global map after point cloud fusion has high accuracy and consistency, which can reflect the overall picture of the entire operation area and provide an accurate environment model for subsequent robot navigation and operation;

[0164] In the orchard environment, in order to ensure the efficient collaborative work of the robot cluster, the topology of the distributed cluster communication network needs to be dynamically adjusted. Specifically, for different types of environments, the communication network will automatically adjust the redundancy, transmission power and communication frequency to ensure the stability, reliability and energy efficiency of the network. The adjustment is based on the real-time evaluation of the complexity of the local environment, and the specific steps for topological adjustment of the distributed cluster communication network are as follows:

[0165] S4.1: Compare the local complexity with the complexity threshold. If it is greater than or equal to the complexity threshold, mark the robot's environment as a complex environment. If it is less than the complexity threshold, mark the robot's environment as a simple environment. This marking operation provides basic data for subsequent network topology adjustments.

[0166] S4.2: Adjust the redundancy of the communication network, select multiple relay robots for communication in complex environments, and select relay robots for communication through the shortest path algorithm in simple environments;

[0167] In complex environments, due to possible signal attenuation, obstacles and other issues, single-path communication may not meet reliability requirements. Therefore, in these environments, the communication network will automatically enable multi-path redundant communication methods, that is, distribute relay robots through distributed routing algorithms, and transmit the same information packet through multiple relay robots at the same time. The network will increase communication redundancy in complex environments, that is, select multiple relay robots to transmit the same information to improve the communication success rate;

[0168] In a simple environment, due to the superior communication conditions, low signal loss, and good signal propagation conditions, the network topology will be adjusted to minimize redundancy, enable the shortest path-based routing selection algorithm, select the most direct relay path for communication, ensure the shortest communication link, and further reduce the complexity of the network;

[0169] In areas marked as simple environments, the communication network will use the shortest path algorithm (such as Dijkstra algorithm or A* algorithm) to select relay robots and build the shortest path for information transmission. This not only reduces communication delay, but also reduces energy consumption, making network resources more efficiently utilized;

[0170] S4.3: Adjust the transmission power of the communication network, increase the transmission power in complex environments, and reduce the transmission power in simple environments;

[0171] Complex environments are usually accompanied by higher energy consumption, so it is necessary to optimize energy management strategies. In complex environments, the transmission power should be appropriately increased to ensure communication quality, but at the same time, power consumption should be controlled to prevent nodes from failing due to energy exhaustion. The transmission power of nodes can be adjusted in real time through power control algorithms. In complex environments, power output can be moderately increased, but the remaining energy and task requirements of the nodes should be comprehensively considered to reasonably allocate energy resources;

[0172] In simple environments, since signal propagation conditions are good and high power is not required to maintain communication quality, the transmission power can be reduced to save energy;

[0173] S4.4: Adjust the communication frequency of the communication network, increase the communication frequency in complex environments, and reduce the number of frequency switching and signal transmission density in simple environments;

[0174] The task reallocation includes:

[0175] Evaluate the execution capability of each robot and calculate the robot score;

[0176] Sort the tasks to be performed in the local environment by complexity and match them according to the complexity and robot score;

[0177] Based on the allocation results, each robot receives and executes its independent task, while feeding back the task status to the leader robot in real time.

[0178] Example 2

[0179] See also Figure 4 The present invention provides an embodiment: an orchard plant protection robot cluster operation system, comprising the following components:

[0180] Central dispatch unit. In this embodiment, the central dispatch unit is a server for computing and map processing in the monitoring room of the orchard manager or the operation site. The server communicates with each robot node through the 4G network, and the communication data is divided into uplink and downlink. The uplink data includes the status data, longitude and latitude, map data, etc. of each robot node; the downlink data includes the issued operation data, the issued local map data, etc. After the map data is uploaded from the robot node, it will enter the multi-stage map splicing module, the segmentation module, the database and the task scheduling optimization module in sequence; among them, the multi-stage map splicing module is responsible for splicing the maps of different robot nodes through feature extraction, point cloud registration, point cloud fusion and other methods; the segmentation module divides the spliced ​​map into fixed-size sub-maps according to the positioning requirements. In this embodiment, the segmentation is based on geographic information, and the size of each sub-map is 20m*20m; the database is responsible for storing the segmented sub-maps, and each sub-map is associated with the global coordinates of its map center coordinates for storage, so that the sub-maps can be matched and issued by the central dispatch unit during the subsequent autonomous movement of the robot.

[0181] Data transmission unit. In this embodiment, the data transmission unit is an existing communication base station, and the robot can directly use the base station for data transmission. The existing 4G network can provide the system with low-latency and high-bandwidth data transmission capabilities.

[0182] Robot node. In this embodiment, the robot is a crawler orchard operation robot with airborne sensors and operation systems. Among them, the airborne sensors, such as Figure 5 As shown, it includes lidar, camera, IMU, wheel speed meter for building high-precision maps and local positioning, and GPS-RTK differential positioning sensor for absolute positioning; the operating system is the spraying system carried at the rear of the vehicle.

[0183] Multi-sensor fusion mapping module: Each robot integrates data from multiple onboard sensors to improve the temporal and spatial consistency of the map, generate a high-precision global map with semantic information, and provide a more accurate reference for robot navigation and operation;

[0184] Adaptive task planning module: Each robot node is equipped with an intelligent task allocation algorithm, which can automatically adjust the task priority according to different operation requirements, real-time operation progress and the current remaining amount of medicine, and report to the central control unit for re-planning, thereby improving the system's operating efficiency and flexibility;

[0185] Collaborative operation module: Robot nodes work together and share local map data to achieve distributed mapping and efficient collaboration, reduce the coverage of repeated operation areas, and improve the overall efficiency of the system.

[0186] This embodiment also includes a method for cluster operation of orchard plant protection robots, wherein the steps are as follows: Figure 6 As shown:

[0187] Step 1: Mapping tasks are issued. The central dispatch unit divides the plots according to the number of operating plots and orchard plant protection robot clusters, and issues the boundary coordinates of each sub-plot to be mapped to each robot node;

[0188] Step 2: On-site personnel remotely build maps. Since the need for map building mainly occurs in areas where GPS-RTK signals are incomplete in the orchard, the remote control process uses the robot's meter-level precision GPS positioning signal to confirm and distinguish boundaries; the on-site personnel remotely control the robot to shuttle back and forth between orchards to ensure the traversal coverage of the drivable paths within the sub-plots; the robot obtains synchronized laser radar, camera, IMU, wheel speed meter and other airborne sensor data, and creates online local maps through the SLAM algorithm. Figure 7 A mapping principle of robot nodes is shown, such as Figure 7 As shown in the figure, the SLAM algorithm input consists of a lidar, an IMU, and a camera, and the algorithm includes two subsystems: a laser odometry (LIO) subsystem and a visual odometry (VIO) subsystem. The LIO subsystem first compensates for motion distortion in the lidar scan through backpropagation, and then calculates the frame mapping point-surface residual. Similarly, the VIO subsystem extracts the visual submap in the current field of view from the visual global map and discards outliers in the submap (points that are occluded or have depth discontinuities). Then, sparse direct visual alignment is performed to calculate the image photometric error from the frame to the map. In the error state iterative Kalman filter, the lidar point-surface residual and the image photometric error are tightly fused with the IMU propagation. The fused posture is used to add new points to the global map.

[0189] Specifically, through the synergy of multiple sensors such as LiDAR, IMU (Inertial Measurement Unit) and camera, high-precision environment map generation is achieved. LiDAR acquires 3D point cloud data of the environment at a frequency of 10-100Hz. After the data is processed by back propagation, it enters the point-surface residual calculation module to estimate the pose information of the laser point cloud; at the same time, the IMU sensor continuously records the motion information of the robot, and after forward propagation, the predicted data is passed to the state estimation module to provide pose prior information, further improving the accuracy of pose prediction. The camera also collects environmental visual data at a frequency of 10-50Hz, and the outlier removal module filters out the data that may interfere to ensure the accuracy of the visual data. After that, the camera data enters the sparse direct visual alignment module for feature matching to help generate accurate visual subgraphs. The state estimation module fuses the data from LiDAR, IMU and camera to generate high-precision pose estimation results. This pose information is then passed to the visual global map and visual subgraph modules, and further associated with the global map to improve the continuity and consistency of the entire map. At the same time, the kd tree structure is used to manage and quickly search for point cloud data, achieving efficient point cloud processing and being able to update map data at a frequency of 20-150Hz. Ultimately, through multi-sensor data fusion and map updates, the robot can build an accurate global map and local submap to support its real-time positioning, navigation, and environmental perception;

[0190] Step 3: Map upload. When each robot node completes the traversal and coverage of the assigned sub-plot, it initiates a map upload request to the central dispatch unit. After confirmation, the central dispatch unit establishes a long connection for data compression and upload;

[0191] Step 4: Map stitching. After the central dispatch unit confirms that the maps of all nodes have been collected, it stitches the incremental point cloud maps. Stitching includes the following processes: data preprocessing, feature extraction, point cloud registration and point cloud fusion to obtain the stitched global point cloud map. Figure 8A flow chart of map stitching is shown. First, the point cloud map is read. In this embodiment, a total of n robot nodes upload n local point cloud maps. Then, the point cloud map needs to be preprocessed, and the data preprocessing includes removing meaningless Nan points, downsampling filtering, etc. Then, the preprocessed point cloud map is subjected to normal estimation, key point detection and FPFH feature descriptor calculation, which are respectively used to describe local surface features, extract key points to reduce the amount of calculation, and generate feature descriptors for each key point to compare the geometric structure between points. Next, feature matching is performed between different point cloud maps to achieve pairwise alignment in three-dimensional space, wherein the SAC coarse registration uses the random sampling consensus (RANSAC) algorithm to roughly align the point clouds in the initial stage to obtain a preliminary transformation matrix; the VGICP transformation matching criterion is based on voxelization and generalized ICP optimization of the preliminary transformation results to further refine the point cloud alignment; the ICP fine registration finally achieves high-precision point cloud registration through iterative nearest neighbor (ICP) matching and error minimization. Finally, point cloud fusion is performed, which mainly combines the registered point cloud data into a set of consistent and non-redundant global point cloud maps.

[0192] Step 5: Map segmentation and storage. The spliced ​​global point cloud map is segmented into multiple small segmented maps according to a certain spatial range. Each small segmented map covers a limited spatial range. In this method, the point cloud blocks are usually stored in the database together with their positions in the global coordinate system for subsequent indexing, matching and other processing.

[0193] Step 6: Autonomous operation of robot nodes. The central dispatch unit divides the operation according to the area to be operated, the number of robots, and the location of the robots, and issues the operation tasks. After receiving the operation tasks, each robot node starts to operate autonomously. Each robot determines its own operation mode based on its real-time GPS-RTK status and completes the autonomous operation according to the corresponding operation mode.

[0194] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. An orchard plant protection robot cluster operation system, characterized in that: The system includes a central dispatch module, a cluster communication module, an environment modeling module and an update module; The central dispatch module is used to establish communication with the robot through communication technology; The cluster communication module is configured with a communication network configuration strategy, and the communication network configuration strategy is used to build a cluster communication network between robots; The environment modeling module is configured with an environment perception strategy, which evaluates the complexity of the local environment and generates a local map based on the sensor components configured on the robot. The local environment complexity includes orchard terrain, vegetation density and obstacle information; The update module is connected to the central scheduling module, splices the local maps to generate a work map, updates the cluster communication network according to the complexity of the local environment, and updates the task allocation and path planning of the robot cluster through the updated cluster communication network.

2. According to claim 1, the orchard plant protection robot cluster operation system is characterized in that: The cluster communication module comprises: A network construction unit, used to construct a self-organizing communication network among robot clusters; A network monitoring unit, used to monitor the status of the cluster communication network in real time, and generate an alarm and report to the central dispatch module when the communication is abnormal; The communication network configuration strategy includes a network construction logic, and the network construction logic is used to establish a distributed cluster communication network through communication equipment. The network construction logic is configured in the network construction unit.

3. According to claim 2, the orchard plant protection robot cluster operation system is characterized in that: The network construction logic includes: The robot cluster randomly selects a leader robot in the central area according to the initial placement position, and randomly selects a relay robot from the remaining robots in the central area; The remaining robots create blank messages and send them to each relay robot. The relay robot counts the arrival delay of each robot transmitting the blank message, adds the arrival delay as the frame tail information to the blank message, and forwards the blank message to the leader robot. The blank message includes the number of transmissions and the remaining energy of the robot. The leading robot receives the blank message and counts the arrival delay of the blank message transmitted by each relay robot, and calculates the total arrival delay based on the arrival delay of the blank message transmitted by the relay robot and the arrival delay of the blank message transmitted by the robot; Compare the total arrival delay with the constraint conditions to determine whether the constraint conditions are met. If not, iterate the leader robot and relay robots to recalculate the total arrival delay. If the constraints are met, the iteration is stopped, the latest positions of the leader robot and relay robots are determined, and the cluster communication network is constructed.

4. The orchard plant protection robot cluster operation system according to claim 3, characterized in that: The environment modeling module comprises: A data processing unit, used to obtain data from the sensor component and process the data, wherein the data includes three-dimensional point cloud data, spatial image data, IMU data and positioning sensor data; The map generation unit is used to fuse the sensor data and generate a local map with semantic information. The multimodal sensor data fusion algorithm is used in the map generation process, and the SLAM technology is combined to generate an operation map with positioning information and semantic labels. A complexity evaluation unit, used to evaluate the complexity of the local environment and calculate the local complexity; A perception data transmission unit, used for transmitting the local map and the local complexity to a central scheduling module; The environmental perception strategy includes sensor data processing logic and complexity calculation logic; The sensor data processing logic is configured in the data processing unit, and the sensor data processing logic is used to process the data acquired from the sensor component to extract the front point cloud features and the front entity features; The complexity calculation logic is configured in the complexity evaluation unit, and the complexity calculation logic is used to calculate the obstacle complexity and terrain complexity according to the output of the data processing unit, and calculate the local complexity according to the obstacle complexity and terrain complexity.

5. The orchard plant protection robot cluster operation system according to claim 4, characterized in that: The sensor data processing logic includes: The three-dimensional point cloud data is randomly sampled by random consistency sampling to obtain the observed point cloud data, the neighborhood points of the observed point cloud data are calculated according to the Lagrange multiplier method, the observed point cloud data and the neighborhood points are discretely determined, and the discrete degree of the observed point cloud data is calculated; The coplanarity of the observed point cloud data is judged according to the degree of discreteness, the observed point cloud data that meets the coplanarity condition is merged and allocated to the point set space, and the front point cloud features are calculated; Traverse the pixels of the spatial image, calculate the rigid transformation vector and the change amount in the iterative process of two adjacent pixels, and calculate the entity feature value of the pixel point according to the dynamic adjustment factor, the rigid transformation vector and the change amount; Compare the entity feature value of the pixel with the entity threshold. If it is greater than the entity threshold, mark the pixel as a sampling point, and confirm the neighborhood of the sampling point according to the local shape distribution function; The feature description vector of the sampling point is calculated based on the normal vector angle between the sampling point and the pixel points in the neighborhood of the sampling point. The feature description vector is aggregated through the maximum pooling operation, and then the aggregated features are further abstracted based on sparse convolution to obtain the front entity features through the RPN network.

6. The orchard plant protection robot cluster operation system according to claim 4, characterized in that: The complexity calculation logic includes: Constructing an obstacle recognition network, taking the front point cloud features and the front entity features as input parameters of the obstacle recognition network, training the input parameters through the obstacle recognition network, and outputting obstacle complexity; The three-dimensional point cloud data set is converted into DEM data, where the DEM data represents the grid data of the earth's surface terrain, the elevation value around each grid cell is calculated based on the DEM data, the height analysis is performed through GIS software, the inclination angle of the local terrain is calculated through the difference method, and the local surface of the orchard terrain is fitted through the least square method according to the inclination angle to calculate the ruggedness, and the terrain complexity is calculated according to the ruggedness; The vegetation coverage rate is obtained by calculating the point cloud density within a unit area, and the vegetation height distribution is obtained by counting the height differences between adjacent point cloud points; the vegetation density is calculated according to the vegetation coverage rate and the vegetation height distribution; The local complexity is calculated by weighted average method according to the obstacle complexity, terrain complexity and vegetation density.

7. The orchard plant protection robot cluster operation system according to claim 4, characterized in that: The update module comprises: A map updating unit, used for receiving the local maps generated by the environment modeling module and splicing the local maps; The communication network optimization unit is used to re-optimize the communication network by adjusting the positions of the leader robot and the relay robot when the environment or operating conditions change; The task scheduling unit adjusts the task allocation and path planning of the robot cluster according to the spliced ​​global operation map and the network status updated by the cluster communication module; The task issuing unit adjusts the map resolution according to the task requirements of the robot, slices the map according to the requirements, and transmits it to the corresponding robot through the central scheduling module.

8. The orchard plant protection robot cluster operation system according to claim 7, characterized in that: The map updating unit is provided with a splicing logic, and the splicing logic includes: Performing a preprocessing operation on the local map input by each robot, wherein the preprocessing operation includes denoising, filtering and resampling; The point cloud features of the preprocessed local map are extracted through a deep convolutional neural network to obtain geometric point cloud features; By calculating the geometric transformation relationship between the geometric point cloud features, the geometric point cloud features are aligned to a unified coordinate system; The registered point clouds are fused to generate a global point cloud map.

9. The orchard plant protection robot cluster operation system according to claim 7, characterized in that: The communication network optimization unit is configured with a network optimization logic, and the network optimization logic includes: According to the local complexity, the local environment is divided into complex environment and simple environment, and the environment type is determined; Adjust the redundancy of the communication network according to the environment type, select multiple relay robots for communication in complex environments, and select relay robots for communication through the shortest path algorithm in simple environments; Adjust the transmission power of the communication network according to the environment type, increase the transmission power in complex environments, and reduce the transmission power in simple environments; Adjust the communication frequency of the communication network according to the environment type, increase the communication frequency in complex environments, and reduce the number of frequency switching and signal transmission density in simple environments.

10. The orchard plant protection robot cluster operation system according to claim 7, characterized in that: The task scheduling unit is configured with a task allocation logic, and the task allocation logic includes: Evaluate the execution capability of each robot and calculate the robot score; Sort the tasks to be performed in the local environment by complexity and match them according to the complexity and robot score; Based on the allocation results, each robot receives and executes its independent task, while feeding back the task status to the leader robot in real time.

Images