A lightweight semantic map construction method and system for lawn mowing robots

Abstract

A lightweight semantic map construction method for lawnmower robots includes: designing a lightweight semantic SLAM method, using an RGB camera to identify targets and a depth camera to measure depth information, thereby performing semantic extraction and 3D position estimation of the targets; establishing a semantic map and a topological map, fusing the information extracted from both, with the fused map providing target location information for target search, and continuously updating the map using information collected by the camera; fusing location information provided by GPS and pose data acquired by IMU based on Kalman filtering to accurately determine the lawnmower robot's own position; exploring the lawnmower robot's working area using a random exploration tree; dividing the lawnmower robot's working area into multiple sub-regions using an unsupervised region partitioning method; delineating the environment and assigning targets to a specific sub-region based on their distance from the center of the target region, thereby improving the performance of the target search task.

Description

Technical Field

[0001] This invention belongs to the field of garden machinery and artificial intelligence technology, specifically relating to a lightweight semantic map construction method and system for lawn mowing robots. Background Technology

[0002] In the fields of garden machinery and artificial intelligence, traditional lawnmower robots typically operate along pre-defined fixed paths when performing tasks such as mowing, trimming lawns, and weeding. This method requires prior lawn maps and human intervention to adapt to static obstacles on the lawn, and its perception capabilities are limited and it cannot dynamically adapt to environmental changes. To enable lawnmower robots to better adapt to environmental changes during operation, semantic maps are widely used in lawnmower robot navigation and path planning. Semantic maps are a representation of the environment that can include information such as objects, locations, boundaries, and landmarks. However, if the lawnmower robot is operating in an environment with many dynamic obstacles and targets, static semantic mapping may result in poor performance. Furthermore, the sensing system is a crucial component of the lawnmower robot's environmental perception, primarily used to acquire physical quantities and information about the environment. Therefore, how to combine the sensing system and semantic maps to enable the lawnmower robot to dynamically update the map when the environment changes is a significant challenge in the field of lawnmower robots. Therefore, researchers have proposed a semantic map construction method based on vision and depth sensors to enhance the adaptability of lawnmower robots in semi-structured grass environments.

[0003] However, these methods typically require high-performance computing resources and large amounts of supervised data. Therefore, a lightweight semantic map construction method is needed to improve the adaptability and performance of lawnmower robots. This invention addresses the problem of dynamic semantic map updates for lawnmower robots by proposing a lightweight semantic map construction method and system. This system fuses GPS-provided location information with IMU-acquired pose information using a Kalman filter, further improving the lawnmower robot's positioning accuracy and stability. Furthermore, the lightweight semantic map construction algorithm significantly reduces resource consumption and computation time. Simultaneously, the system introduces a dynamic update mechanism, enabling real-time updates of the semantic map to adapt to environmental changes. Specifically, the method includes the following steps: First, an RGB camera is used to identify the target, and a depth camera is used to measure depth information, thereby performing semantic extraction and 3D position estimation on the target. Second, a semantic map and a topological map are established, and the information extracted from both is fused together. The fused map provides target location information for target search, and the map is continuously updated using information collected by the cameras. Next, information fusion from multiple sensors is performed based on Kalman filtering to accurately determine the lawnmower robot's own position. Finally, by analyzing the robot's motion trajectory and dividing and updating the fused map, the map accuracy is optimized, improving the robot's navigation and path planning capabilities, and enabling rapid target search. This method, by fusing information from multiple sensors and image information, better adapts to environmental changes and further improves the robot's navigation and path planning capabilities, showing broad application prospects. Summary of the Invention

[0004] This invention aims to overcome the aforementioned shortcomings of existing technologies and provides a lightweight semantic map construction method and system for lawnmower robots. This invention uses an unsupervised region partitioning method to establish the attribution of targets to environmental sub-regions, enabling lawnmower robots to better adapt to dynamically updated semi-structured grassland environments.

[0005] A lightweight semantic map construction method for lawnmower robots, with the following specific steps:

[0006] S1. Design a lightweight semantic SLAM method that uses an RGB camera to identify targets and a depth camera to measure depth information, thereby performing semantic extraction and 3D position estimation of the targets.

[0007] S2. Establish semantic maps and topological maps, merge the information extracted from the two together, and use the merged map to provide target location information for target search, and continuously update the map using information collected by the camera;

[0008] S3. Based on Kalman filtering, the location information provided by GPS and the pose data obtained by IMU are fused for positioning to accurately determine the position of the lawnmower robot.

[0009] S4. Use a random exploration tree to explore the working area of ​​the lawnmower robot;

[0010] S5. The working area of ​​the lawnmower robot is divided into multiple sub-regions using an unsupervised region division method.

[0011] S6. Delineate the environment and assign the target ion region to a sub-region based on its distance from the center, thereby improving the performance of the target search task.

[0012] Further, step S1 specifically includes:

[0013] S11: Use an RGB camera to identify image targets, capture RGB data of the current environment, and use a target detection algorithm to identify them;

[0014] S12: Use a depth camera to determine the target depth information, and the depth information is captured for 3D positioning;

[0015] S13: The steps to obtain the 3D coordinates of the 2D semantic target in the environment model are as follows: Map the pixel plane coordinates to the imaging plane, and use the following coordinate transformation formula to obtain:

[0016]

[0017] In the formula, (u,v,1) represents the aligned coordinates of the target in the pixel reference system, u0 and v0 represent the translations from the origin of the pixel coordinate system to the optical axis, and Z represents the distance from the target to the optical axis. c For depth information measured by a depth camera, f x and f y Let X be the focal length of the camera in the x and y directions, respectively. c and Y c The two-dimensional coordinates of the target in the camera coordinate system;

[0018] S14: The steps to convert the target's three-dimensional coordinates to coordinates in the robot's coordinate system through coordinate transformation are as follows:

[0019]

[0020] Where the rotation matrix R and the translation matrix T represent the rotation and translation transformation relationship between the camera coordinates and the robot coordinates, X r Y r and Z r Let be the three-dimensional coordinates of the target in the robot coordinate system.

[0021] Furthermore, step S2 specifically includes:

[0022] S21: Integrate the acquired important semantic information with the spatial information of the target area;

[0023] S22: When marking frequently occurring areas, the robot can learn from historical data to identify these areas and automatically mark them on the map;

[0024] S23: Automatically update the fused image using environmental information collected by the camera.

[0025] Furthermore, step S3 specifically includes:

[0026] S31: GPS location information may be affected by a variety of factors, so data preprocessing and correction are required to remove errors and biases;

[0027] S32: The IMU acquires position information such as acceleration and angular velocity, which needs to be preprocessed and calibrated to remove noise and bias.

[0028] S33: By using the Kalman filter algorithm, the position and pose information measured by GPS and IMU and preprocessed is fused to obtain more accurate and reliable robot position and state information.

[0029] The lawnmower robot has a position L(k) at time k, a velocity V(k), an acceleration A(k), a position measured by GPS Z1(k), a pose information measured by IMU Z2(k), a state vector X(k)=[L(k),V(k),A(k)], and an observation vector Z(k)=[Z1(k),Z2(k)].

[0030] The state equation and the observation equation can be expressed as follows:

[0031] X(k+1)=AX(k)+Bu(k)+w(k) (3)

[0032] Z(k)=HX(k)+v(k) (4)

[0033] Where A, B, and H are coefficient matrices, and w and v are noise vectors. It is assumed that the noise follows a Gaussian distribution, and the covariance matrices are Q and R, respectively.

[0034] In the Kalman filter algorithm, the prediction step is performed first, predicting the state vector and covariance matrix of the lawnmower robot based on the state equation and observation equation:

[0035] X(l+1|k)=AX(k|k)+Bu(k) (5)

[0036] P(k+1|k)=AP(k|k)A T +Q (6)

[0037] Where X(k|k) represents the posterior estimate of the state vector at time k, and P(k|k) represents the posterior estimate of the covariance matrix at time k.

[0038] Then, an update step is performed: based on the observation equation and the predicted values, the Kalman gain is calculated, and the predicted values ​​are corrected.

[0039]

[0040] X(k+1|k+1)=X(k+1|k)+K(k+1)(Z(k+1)-HX(k+1|k)) (8)

[0041] P(k+1|k+1)=(IK(k+1)H(k+1))P(k+1|k) (9)

[0042] Where K is the Kalman gain, X(k+1|k+1) represents the posterior estimate of the state vector at time k+1, and P(k+1|k+1) represents the posterior estimate of the covariance matrix at time k+1.

[0043] Through the above fusion steps, more accurate information on the location and status of the lawnmower robot is obtained, which can be used to accurately determine the robot's own position.

[0044] Furthermore, step S4 specifically includes:

[0045] S41: After knowing the position and state information of the lawnmower robot, the continuous workspace of the lawnmower robot is converted into a discrete workspace by random sampling. That is, the continuous space is divided into several discrete values ​​in each dimension, and then the discrete values ​​of each dimension are combined to obtain a state in the discrete space. This discrete state corresponds to a small region in the continuous space, so that the corresponding nodes can be accessed.

[0046] S42: Since randomly sampled points in the workspace cannot be guaranteed to be interconnected, the robot cannot reach these isolated points. By generating a random tree covering the entire workspace, and using a Rapid-exploration Random Tree (RRT) to sample the points, and because each node of the RRT performs collision detection during growth, it can cover the entire workspace with the random tree regardless of whether the environment is known or unknown. This ensures that the sampled access points are connected and accessible to the lawnmower robot.

[0047] Furthermore, step S5 specifically includes:

[0048] S51: After ensuring that the robot can search in unknown environments, the above tree nodes are divided according to Euclidean distance using an unsupervised clustering algorithm, and the working area is divided into many sub-regions.

[0049] Furthermore, the specific process of performing unsupervised sub-region partitioning is as follows:

[0050] Based on the distance, sample points {x1,…,x} are selected. n The clusters are divided into K classes. Through iterative processing, suitable cluster centers are found to minimize the loss function j. The loss function is calculated as follows:

[0051]

[0052] Where, μ k Let r be the cluster center of the k-th class. nk This indicates whether the nth sampling point belongs to the kth class, where k = 1, ..., k, and its values ​​are:

[0053]

[0054] Cluster center μ k for

[0055]

[0056] The obtained cluster center μ k It is the center of each sub-region after the division.

[0057] S52: Based on the distance between the target and the center point of the sub-region, establish the relationship between the target and the sub-region to locate the target more accurately.

[0058] Furthermore, step S6 specifically includes:

[0059] S61: When performing a target search task, determine the sub-region where the target is located based on the distance of the target ion region's center point;

[0060] S62: Use the area where the target is located as the target area for the search, and quickly search the area where the target is located by calculating the pre-planned route.

[0061] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: It proposes a lightweight semantic map construction method and system for lawnmower robots, fusing topological maps and semantic maps. Unlike topological maps, which only simplify the environmental structure, the fused map provides precise location and functional attribute information for target search. Compared to 3D point cloud maps, which require significant computational resources and storage space, this semantic map exhibits excellent real-time performance and adaptability. This novel semantic map can be automatically updated in dynamic environments, employing a random exploration tree to explore the lawnmower robot's working area and using an unsupervised region partitioning method to establish the target's affiliation to environmental sub-regions. Compared to traversal searches in dynamic environments, the target search time based on the fused mapping is significantly reduced, improving robot recognition accuracy. This is crucial for maintaining the robot's autonomy under long-term continuous working conditions. Attached Figure Description

[0062] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0063] Appendix Figure 1 This document presents a lightweight semantic map construction method for lawn mowing robots and a system overall framework diagram.

[0064] Appendix Figure 2 This is a schematic diagram illustrating the semantic extraction and 3D position estimation process of the target according to the present invention;

[0065] Appendix Figure 3 This is a schematic diagram of the process of fusing GPS-provided location information and IMU-acquired pose data based on Kalman filtering according to the present invention.

[0066] Appendix Figure 4 This is a schematic diagram illustrating the process of exploring the robot's work area using a random exploration tree, as described in this invention.

[0067] Appendix Figure 5 This is a schematic diagram illustrating the process of dividing the robot's working area using unsupervised region partitioning in this invention;

[0068] Appendix Figure 6 This is a schematic diagram of a lightweight semantic map construction system for a lawnmower robot, provided in Embodiment 2 of the present invention. Detailed Implementation

[0069] To more clearly describe the technical solution of the present invention, this document will provide a detailed description of the technical solution provided by the present invention in conjunction with specific embodiments, and further illustrate the present invention with reference to the accompanying drawings. It should be noted that these specific embodiments are only some embodiments of the present invention, used to explain the present invention, and not to limit the present invention.

[0070] Example 1

[0071] This invention provides a lightweight semantic map construction method for lawnmower robots, comprising the following steps:

[0072] S1. Design a lightweight semantic SLAM method that uses an RGB camera to identify targets and a depth camera to measure depth information, thereby performing semantic extraction and 3D position estimation of the targets.

[0073] S2. Establish semantic maps and topological maps, and merge the information extracted from the two together. The merged map provides target location information and functional attribute information for target search.

[0074] S3. Based on Kalman filtering, the location information provided by GPS and the pose data obtained by IMU are fused for positioning to accurately determine the position of the lawnmower robot.

[0075] S4. Use a random exploration tree to explore the working area of ​​the lawnmower robot;

[0076] S5. The working area of ​​the lawnmower robot is divided into multiple sub-regions using an unsupervised region division method.

[0077] S6. Delineate the environment and assign the target ion region to a sub-region based on its distance from the center, thereby improving the performance of the target search task.

[0078] See attached document Figure 2 The RGB camera mentioned in step S1 is used to identify the target. The specific identification process is as follows:

[0079] S1-1: Use an RGB camera to acquire images of the target, capturing information such as the color and shape of the target object;

[0080] S1-2: Preprocess the acquired images by performing image enhancement, noise reduction, filtering, etc., to remove noise and enhance target features;

[0081] S1-3: Extract edge features of the target from the preprocessed image and use a convolutional neural network (CNN) to extract various features in the image;

[0082] S1-4: Match the extracted features with the known target features;

[0083] S1-5: Determine the target object in the image based on the feature matching results.

[0084] Furthermore, after determining the target object as described in step S1, the depth information of the target is measured using a depth camera. The specific measurement process is as follows:

[0085] S1-1: Capture the depth and texture information of the target object to generate a depth image and a color image;

[0086] S1-2: Calibrate the depth camera and camera coordinates to determine the camera's internal and external parameters;

[0087] S1-3: Use depth images to generate point clouds to represent the shape and location information of the target;

[0088] S1-4: Filter the generated point cloud to remove noise and outliers, and extract target features from the filtered point cloud;

[0089] S1-6: Match the extracted features with known targets to determine the target's location information;

[0090] S1-7: Calculate the depth information of the target based on the point cloud matching results.

[0091] The specific process of combining the semantic map and the topological map in step S2 is as follows:

[0092] S2-1: The semantic map is constructed by using SLAM technology to perceive the environment through an RGB camera and model the topological relationships of different regions in the environment to obtain a topological map.

[0093] S2-2: Use the ICP algorithm to match the semantic map and the topological map in terms of location relationships in order to perform information fusion;

[0094] S2-3: Add semantic information to the topology map for fusion;

[0095] S2-4: Utilize the environmental information continuously collected by sensors to update the semantic map and topological map, and then perform map alignment and information fusion.

[0096] See attached document Figure 3 The specific fusion process of the location information provided by GPS and the pose data acquired by IMU based on Kalman filtering in step S3 is as follows:

[0097] The S3-1 GPS provides position observations, and the IMU provides attitude and acceleration observations, which serve as inputs to the Kalman filter.

[0098] The initial state of S3-2 is provided by GPS, including position and attitude information. The state transition matrix and state model are updated using motion state information provided by the IMU.

[0099] S3-3 Based on the observations and the state model, the Kalman filter calculates the optimal estimates of position and attitude, and updates the state variables and covariance matrix;

[0100] When the S3-4 GPS signal is blocked or unavailable, the IMU provides continuous motion state information and updates the state variables according to the state transition matrix and state model.

[0101] After multiple iterations of calculation, S3-5 can be fused to obtain accurate position and attitude estimates of the lawnmower robot. Furthermore, the estimation accuracy and stability can be improved by correcting and updating the Kalman filter.

[0102] See attached document Figure 4 Step S4, which involves exploring the robot's work area using a random exploration tree, is described in the following steps:

[0103] S4-1: Create a two-dimensional working area map using image information acquired by an RGB camera;

[0104] S4-2: Set the robot's start and end points. The start point is the robot's current position, and the end point is all the areas that need to be mowed.

[0105] S4-3: Using the RRT algorithm, a random exploration tree is built on the map. Starting from the starting point, nodes are randomly generated to expand the tree to the target area until the tree reaches the destination area.

[0106] S4-4: During the tree expansion process, consider the robot's maximum turning angle and maximum speed to ensure that the tree growth process is feasible;

[0107] S4-5: After the tree is built, it is used to plan the robot's path. Starting from the starting point, the optimal path to the destination region is searched using the node and edge information of the tree;

[0108] S4-6: Once the optimal path is found, the robot moves along the optimal path and completes the lawn mowing task;

[0109] S4-7: Update the map in real time during mowing to avoid hitting obstacles or repeatedly mowing the same areas.

[0110] See attached document Figure 5 To divide the robot's working area using unsupervised region partitioning, the specific process is as follows:

[0111] S5-1: Collect image data as the lawnmower moves within the work area;

[0112] S5-2: Using the K-means algorithm, K cluster centers are randomly initialized. Then, each data point is assigned to the cluster containing the nearest cluster center. The cluster center of each cluster is updated according to the average value of the data points in the cluster. This process is repeated until convergence.

[0113] S5-3: After clustering algorithm calculation, the working area of ​​the lawnmower robot is divided into multiple sub-regions. The extent of each sub-region is determined based on its center point and boundaries, and stored in the robot's memory for later use;

[0114] S5-4: When the lawnmower is performing a task, the defined sub-regions are used as a navigation map to help the robot determine its current location and automatically plan the mowing path.

[0115] The specific process of step S6 is as follows:

[0116] S6-1: Divide the lawn into several areas based on its specific characteristics, and mark each sub-area accordingly;

[0117] S6-2: Determine the vicinity of the sub-region where the target is located based on the distance between the target and the center point of the sub-region;

[0118] S6-3: By using information about the distance between the center of the sub-region and the target ion region, the target can be searched more efficiently.

[0119] Example 2

[0120] Appendix Figure 6 This is a schematic diagram of a lightweight semantic map construction system for a lawnmower robot, provided in Embodiment 2 of the present invention. The system includes: environmental information monitoring equipment, sensor data acquisition equipment, communication equipment, and a main control device.

[0121] The environmental information detection device is used to collect image information in the environment and send the detection data to the data acquisition device.

[0122] The environmental data processing and analysis device is used to acquire environmental detection-related data sent by the detection device, including color, shape, and location data of images within the current monitoring area; and to obtain the category attributes of the sample data by extracting, analyzing, and processing the features of the sample data.

[0123] The sensor data acquisition device is used to collect the current position and pose data of the lawnmower robot. The collected data includes latitude data, acceleration data, angle data, etc. More accurate data is obtained after preprocessing and filtering.

[0124] The communication device enables real-time data transmission and communication between various devices to achieve information synchronization and updating. It features multiple communication methods and protocols to adapt to different scenario requirements, while also ensuring the confidentiality and integrity of data during communication to prevent data leakage.

[0125] The main control device is used to receive the detection data sent by the environmental data processing and analysis device and the data acquired by the sensors, and to visualize the detection data and the image information to analyze the current environment and the position information of the lawnmower robot itself.

[0126] Finally, the above are merely embodiments of the present invention and the technical principles applied. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, as various obvious changes, readjustments, and alternatives are possible without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail above, the present invention is not limited to the above embodiments. Many other equivalent embodiments may be included without departing from the scope of the concept of the present invention. The scope of the present invention is defined by the claims.

Claims

1. A lightweight semantic map construction method for a mowing robot, comprising the following steps: S1. Designing a lightweight semantic SLAM method, using an RGB camera to identify targets and using a depth camera to measure depth information, and then performing semantic extraction and three-dimensional position estimation on the targets; S2. Establishing a semantic map and a topological map, fusing the information extracted from the two maps together, providing target position information for target search, and continuously updating the fusion map using information collected by the camera; S3. Based on Kalman filtering, fusing the position information provided by GPS and the pose data obtained by IMU to accurately determine the position of the mowing robot itself; S4. Using a random exploration tree to explore the working area of the mowing robot; S5. Using an unsupervised region division method to divide the working area of the mowing robot into multiple sub-regions; S6. Circumscribing the environment and attributing a target to a sub-region based on the distance between the target and the center of the sub-region, thereby improving the performance of the target search task; Step S1 specifically comprises: S11. Using an RGB camera to identify image targets, capturing RGB data of the current environment, and using a target detection algorithm to identify the targets; S12. Using a depth camera to determine target depth information, which is captured for three-dimensional positioning; S13. Obtaining the three-dimensional coordinates of two-dimensional semantic targets in the environment model, which is done by mapping pixel plane coordinates to imaging plane coordinates, and introducing the following coordinate transformation formula: wherein , , 1) is the flat coordinate of the target in the pixel reference system, and are the translation of the pixel coordinate system origin to the optical axis, is the depth information measured by the depth camera, and are the focal length of the camera in x and y direction, respectively, and are the two-dimensional coordinates of the target in the camera coordinate system; S14. Converting the three-dimensional coordinates of the target to coordinates in the robot coordinate system through coordinate transformation, which is done as follows: where the rotation matrix and the translation matrix represent the rotation and translation transformation relationship between the camera coordinates and the robot coordinates, , and is the three-dimensional coordinate of the target in the robot coordinate system; Step S3 specifically comprises: S31. Preprocessing and correcting the position information provided by GPS to remove errors and biases; S32. IMU obtains acceleration and angular velocity pose information, which is preprocessed and calibrated to remove noise and bias; S33. Through the Kalman filtering algorithm, the preprocessed position information of GPS and the preprocessed pose information measured by IMU are fused to obtain the position information and state information of the robot; Step S33 specifically comprises: The position of the mowing robot at time k , the velocity is , the acceleration is , the position measured by GPS is , the pose information measured by IMU is , the state vector is [ , , ], and the observation vector is [ , ]. The state equation and the observation equation are represented as: wherein , , is a coefficient matrix, , is a noise vector, the noise satisfying a Gaussian distribution; In the Kalman filtering algorithm, first, the prediction step is performed, and the state vector and covariance matrix of the mowing robot are predicted according to the state equation and the observation equation: wherein represents a posteriori estimate of the state vector at time represents a posteriori estimate of the covariance matrix at time represents a posteriori estimate of the covariance matrix of the noise​​ Then, the update step is performed, and the Kalman gain is calculated according to the observation equation and the predicted value to correct the predicted value: wherein is a Kalman gain, denotes a posteriori estimate of the state vector at time denotes a posteriori estimate of the covariance matrix at time is a covariance matrix of the noise .​​ Through the above fusion steps, the position and state information of the mowing robot are obtained; Step S4 specifically comprises: S41. After knowing the position and state information of the mowing robot, the continuous working space of the mowing robot is converted into a discrete working space by random sampling, i.e., the continuous space is divided into several discrete values in each dimension, and then the discrete values in each dimension are combined to obtain a state in the discrete space, which corresponds to a small region in the continuous space; S42. A random tree covering the entire working space is generated, and points are sampled using a Rapid-exploration Random Tree (RRT).

2. The lightweight semantic map construction method for mowing robots of claim 1, wherein: The step S2 specifically comprises: S21: The semantic map is aware of the environment through an RGB camera, is constructed by using a SLAM technology, and models a topological relationship of different regions in the environment to obtain a topological map; S22: An ICP algorithm is used to match the semantic map and the topological map in a positional relationship, so as to perform information fusion; S23: Semantic information is added to the topological map for fusion; the fusion map provides target position information for target search, and when a region frequently appears is marked, the robot identifies the region by learning historical data and automatically marks on the map; S24: The semantic map and the topological map are updated by using environmental information collected by the camera, and then map alignment and information fusion are performed, and the fusion map is automatically updated.

3. The lightweight semantic map construction method for a mowing robot of claim 1, wherein: The step S5 specifically comprises: S51: A tree node is divided according to a Euclidean distance by using an unsupervised clustering algorithm, and a work region is divided into a plurality of sub-regions; S52: A relationship between a target and a sub-region is established according to a distance between the target and a center point of the sub-region, and the target is positioned.

4. The lightweight semantic map construction method for a mowing robot of claim 3, wherein: The step S51 specifically comprises: According to the distance, the sampling points { } are divided into M classes; through continuous iteration, the appropriate clustering center is solved, and the loss function j is minimized, and the loss function is calculated as follows: wherein, is the cluster center of the m-th class, denotes whether the n-th sample belongs to the m-th class, m = 1,..., M, and takes the value: Cluster centers To The obtained cluster centers are the centers of each sub-region after partitioning.

5. The lightweight semantic map construction method for a mowing robot of claim 1, wherein: The step S6 specifically comprises: S61: When a target search task is performed, a sub-region where the target is located is determined according to a distance between the target and a center point of the sub-region; S62: The sub-region where the target is located is taken as a target region for search, and a pre-planned route is calculated to quickly search the region where the target is located.

Images