A method and system for collecting and making high-precision maps by an autonomous driving cleaning robot

Abstract

Disclosed are a method and system for an autonomous driving cleaning robot to collect and produce high-precision maps, the method comprising: S1, synchronizing a laser radar sensor and a camera with a global satellite time system in real time through a controller, and collecting camera data and laser radar point cloud data; S2, extracting semantic information from the laser radar point cloud data and the camera data to obtain dynamic point information and road semantic information; S3, performing motion compensation and global pose optimization on the laser radar point cloud data, and converting the laser radar point cloud data into a point cloud set in a global coordinate system; S4, filtering out dynamic points in the point cloud set based on the dynamic point information, extracting statistical features of the point cloud set based on the road semantic information, and constructing a high-precision map based on the statistical features.

Description

Technical Field

[0001] The present invention relates to the technical field of high-precision map production, and specifically to a method and system for an autonomous driving cleaning robot to collect and produce high-precision maps. Background Art

[0002] The self-driving cleaning robot can realize functions such as automatic cruising, autonomous obstacle avoidance, and path planning, which greatly improves the working efficiency. For the application of the self-driving cleaning robot, the robot's autonomous positioning has always been a basic task in system development.

[0003] In the prior art, autonomous driving cleaning robots are usually equipped with a variety of sensors (such as lidar, camera, GPS, etc.) to perceive the environment. These sensors each have an independent clock source, and the clock source deviation during data fusion will cause data synchronization problems, affecting the robot's positioning and environmental perception accuracy. The traditional method uses a carrier phase differential technology called Real Time Kinematic (RTK) to achieve centimeter-level positioning accuracy when the GPS signal is good. In addition, there are also methods based on environmental feature positioning, such as the high-precision positioning method provided by Lidar (laser sensor), which uses a data acquisition device based on combined navigation, laser and other multi-sensors to pre-collect original positioning information and point cloud information, and post-process and build maps based on the point cloud reflection intensity (intensity) characteristics to complete the construction of offline point cloud maps, and complete vehicle positioning based on the real-time online point cloud and offline point cloud map matching method. This method can also provide centimeter-level positioning accuracy for autonomous driving in areas with rich and obvious environmental intensity characteristics.

[0004] Although existing technologies can generate high-precision maps, in certain application scenarios, such as when cleaning robots use high-precision maps to accurately mark roadside elements, there is still a need to improve accuracy. This places higher demands on map collection, processing, and updating technologies.

[0005] The autonomous driving system that uses RTK positioning technology cannot provide accurate positioning information in weak signal environments such as those blocked by trees, surrounded by tall buildings, under elevated roads, in tunnels and underground parking lots. The autonomous driving system that uses environmental feature-based positioning cannot complete matching convergence in areas with sparse structural features and cannot obtain usable positioning information.

[0006] In addition, the urban environment is full of dynamic obstacles, such as walking pedestrians, moving vehicles, etc. These dynamic obstacles will introduce map deviations and affect the real-time update and accuracy of the map.

[0007] Therefore, during the data collection process, the position and posture measurement of the robot may be affected by many factors, such as sensor accuracy, environmental interference, etc., resulting in position and posture deviation. This deviation will directly affect the accuracy of the high-precision map, and then affect the execution effect of the cleaning task. The point cloud matching accuracy of existing technologies in dynamic environments needs to be improved. Summary of the invention

[0008] In view of the technical problems in the prior art, this application proposes a method for collecting and producing high-precision maps by an autonomous driving cleaning robot, comprising the following steps:

[0009] S1. The laser radar sensor and camera are synchronized with the global satellite time system in real time through the controller. The controller combines the pulse second PPS signal and the minimum specific transmission information recommended by the global positioning system GPRMC for satellite timing. The laser radar is synchronized with the controller in real time through the precision time protocol PTP. The camera is synchronized with the controller in real time through the pulse trigger timing mechanism, and the camera data and the laser radar point cloud data are collected;

[0010] S2, extracting semantic information from the laser radar point cloud data and the camera data to obtain dynamic point information and road semantic information;

[0011] S3, performing motion compensation and global pose optimization on the laser radar point cloud data, and converting the laser radar point cloud data into a point cloud set in a global coordinate system;

[0012] S4. Filter out dynamic points in the point cloud set based on the dynamic point information, extract statistical features of the point cloud set based on the road semantic information, and construct a high-precision map based on the statistical features.

[0013] Preferably, the controller combines the pulse second (PPS) signal and the global positioning system recommended minimum specific transmission information (GPRMC) for satellite timing, the laser radar is synchronized with the controller in real time through the precision time protocol (PTP), and the camera is synchronized with the controller in real time through a pulse trigger timing mechanism. A unified timestamp protocol is implemented on all sensors through hard real-time synchronization to ensure the temporal consistency of data streams from different sensors.

[0014] Preferably, the step S2 specifically includes using the Pix2PixHD network model to process the camera data to obtain 2D semantic information, establishing a correspondence between the camera data and the lidar point cloud data, and using the PointNet++ network model to process the lidar point cloud data in combination with the 2D semantic information to obtain the dynamic point information and the road semantic information. Through training and a specialized neural network model, the system is able to obtain dynamic point information representing dynamic obstacles (such as pedestrians, vehicles, animals, etc.), and then identify and filter them in subsequent processing, thereby reducing noise interference and improving map accuracy. This step is crucial for building a static base map in a complex and changing urban environment. The road semantic information will subsequently assist in the construction of high-precision maps.

[0015] Preferably, the odometer device is selected for motion compensation and global pose optimization according to the satellite positioning status of the global satellite time system. When the satellite positioning status of the global satellite time system is normal, satellite positioning is used as the odometer device. When the satellite positioning status of the global satellite time system is abnormal, the lidar sensor is used as the odometer device. Due to the movement of the device during the scanning process, the original point cloud data may be deformed, which will affect the subsequent data processing and analysis. Therefore, these deformations can be corrected by the motion compensation algorithm to restore the true geometric structure of the scene. The purpose of global optimization is to ensure that the position and direction of the point cloud data in the global coordinate system are as close to the actual value as possible, and to reduce the cumulative error caused by factors such as equipment errors and environmental changes.

[0016] Preferably, step S1 also includes accessing the integrated navigation data, and when the satellite positioning state of the global satellite time system is abnormal, providing observation values ​​for the laser radar sensor as an odometer device, and performing global pose optimization based on the graph optimization method. By accessing the integrated navigation data, the quality of the real-time data of the laser sensor data is monitored to avoid problems such as data unavailability and low map construction accuracy.

[0017] Further preferably, the point cloud set is projected to form a grid map after filtering out the dynamic points based on the dynamic point information, and statistical features of the point cloud set are extracted based on the position of each grid point, the reflection intensity and the road semantic information, and the statistical features include verticality features, intensity features and semantic features respectively.

[0018] More preferably, a verticality map, an intensity map and road element data are generated respectively based on the verticality features, intensity features and semantic features, and the high-precision map is constructed by comparing the road element data with the verticality map and the intensity map for matches, performing annotation and adjustments.

[0019] Further preferably, extracting the verticality feature specifically includes:

[0020] S61, calculating the mean of the point cloud set P , decentralize the point cloud set and get ,in, , , ,in represents the coordinates of the i-th point cloud data in the point cloud set, , N is the number of points in a frame of point cloud;

[0021] S62, calculate the decentralized point cloud set The covariance matrix C of Perform eigenvalue decomposition of the covariance matrix, where and Respectively represent the i-th eigenvalue and the corresponding eigenvector;

[0022] S63, sorting the eigenvalues ​​in descending order, calculating the angle between the eigenvector corresponding to the minimum eigenvalue and the Z axis, and obtaining the verticality feature.

[0023] This method can effectively extract vertical structures such as curbs and walls from point cloud data.

[0024] Further preferably, the semantic features are extracted based on the normal distribution of the road semantic information.

[0025] According to the second aspect of the present invention, a system for collecting and producing high-precision maps by an autonomous driving cleaning robot is proposed, comprising the following modules:

[0026] Data acquisition module: The laser radar sensor and camera are synchronized with the global satellite time system in real time through the controller. The controller combines the pulse second PPS signal and the global positioning system recommended minimum specific transmission information GPRMC for satellite timing. The laser radar is synchronized with the controller in real time through the precision time protocol PTP. The camera is synchronized with the controller in real time through the pulse trigger timing mechanism, and the camera data and laser radar point cloud data are collected;

[0027] Semantic extraction module: extracting semantic information from the laser radar point cloud data and the camera data to obtain dynamic point information and road semantic information;

[0028] Data processing module: performs motion compensation and global pose optimization on the laser radar point cloud data, and converts the laser radar point cloud data into a point cloud set in a global coordinate system;

[0029] Map making module: filtering out dynamic points in the point cloud set based on the dynamic point information, extracting statistical features of the point cloud set based on the road semantic information, and constructing a high-precision map based on the statistical features.

[0030] The present invention is beneficial in that:

[0031] 1. To address the failure of single sensor positioning based on RTK or feature matching, the inconsistency problem of multi-sensor data caused by clock source deviation is solved by introducing multiple sensors and corresponding efficient real-time synchronization technology, thus laying the foundation for high-precision data fusion and analysis.

[0032] 2. Through semantic information extraction technology, the map deviation introduced by dynamic obstacles is effectively eliminated. This enables the map to more accurately reflect the actual environment, especially in complex urban scenes, improving the operating efficiency and safety of the cleaning robot. At the same time, the introduction of semantic information significantly improves the annotation accuracy of key elements in the map (such as lane lines, curbs, etc.). This is of great significance for the path planning and obstacle recognition of the cleaning robot when performing tasks, and improves the accuracy of the map generation process.

[0033] 3. Through motion compensation algorithm and global closed-loop optimization, the problem of position and posture deviation in the data collection process is solved, and the accuracy of the map generation process is significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated into and constitute a part of this specification. The accompanying drawings illustrate the embodiments and are used together with the description to explain the principles of the present invention. It will be easy to recognize other embodiments and many expected advantages of the embodiments because they become better understood by reference to the following detailed description. The elements of the drawings are not necessarily to scale with each other. The same reference numerals refer to corresponding similar parts.

[0035] Figure 1 A schematic diagram of a process for collecting and making high-precision maps by an autonomous driving cleaning robot according to the present invention is shown;

[0036] Figure 2 Unoptimized surface roughness feature maps are stitched for pure poses;

[0037] Figure 3 It is the surface roughness feature map after global optimization;

[0038] Figure 4 A structural schematic diagram of a system for collecting and producing high-precision maps by an autonomous driving cleaning robot according to the present invention is shown. DETAILED DESCRIPTION

[0039] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It should also be noted that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.

[0040] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0041] Figure 1 A method for collecting and making high-precision maps by an autonomous driving cleaning robot according to the present invention is shown, comprising the following steps:

[0042] S1. The laser radar sensor and camera are synchronized with the global satellite time system in real time through the controller. The controller combines the pulse second PPS signal and the minimum specific transmission information recommended by the global positioning system GPRMC for satellite timing. The laser radar is synchronized with the controller in real time through the precision time protocol PTP. The camera is synchronized with the controller in real time through the pulse trigger timing mechanism, and the camera data and the laser radar point cloud data are collected;

[0043] S2, extracting semantic information from the laser radar point cloud data and the camera data to obtain dynamic point information and road semantic information;

[0044] S3, performing motion compensation and global pose optimization on the laser radar point cloud data, and converting the laser radar point cloud data into a point cloud set in a global coordinate system;

[0045] S4. Filter out dynamic points in the point cloud set based on the dynamic point information, extract statistical features of the point cloud set based on the road semantic information, and construct a high-precision map based on the statistical features.

[0046] In a specific embodiment, the controller combines the pulse second (PPS) signal and the global positioning system recommended minimum specific transmission information (GPRMC) for satellite timing. This method not only utilizes the high-precision characteristics of the PPS signal to ensure the accuracy of time synchronization, but also obtains the global unified time standard through the GPRMC information, realizing the efficient synchronization of the collector terminal and the global satellite time system. This synchronization mechanism ensures that the controller can obtain accurate time tags at any location in the world, providing a solid foundation for the time consistency of data collection.

[0047] The laser radar is synchronized with the controller in real time through the precision time protocol (PTP). The PTP protocol can transmit time information in the network and achieve millisecond-level or even higher precision time synchronization between the laser radar and the controller by accurately controlling the propagation delay of information. This synchronization method effectively ensures that the time tag of the laser radar point cloud data is strictly consistent with the controller's time system, thereby providing an accurate time reference for subsequent data processing and fusion.

[0048] The camera is synchronized with the controller in real time through a pulse-triggered timing mechanism. The pulse-triggered timing mechanism triggers the camera's shooting by sending a synchronization pulse signal, ensuring that each shot is synchronized with the controller's time system. In this way, each photo taken will be given a time tag consistent with the controller's time system, greatly improving the accuracy and reliability of data fusion from different sensors.

[0049] In a specific embodiment, the step S2 specifically includes using the Pix2PixHD network model to process the camera data to obtain 2D semantic information, establishing a correspondence between the camera data and the lidar point cloud data, and using the PointNet++ network model to process the lidar point cloud data in combination with the 2D semantic information to obtain the dynamic point information and the road semantic information.

[0050] Pix2PixHD is a high-performance image-to-image conversion network. The present invention can accurately extract 2D semantic information such as lane lines and curbs from 2D images of urban scenes by processing the camera data using the Pix2PixHD network model. By using the pre-calibrated camera-to-lidar external parameters and the camera's internal parameters, the correspondence between the laser point cloud and the camera image data and the 2D semantic information corresponding to the camera image data can be established, providing a basis for the subsequent PointNet++ to extract 3D semantic information from the point cloud. The calibration method can use traditional camera calibration methods such as Tsai's two-step method, Zhang's calibration method, etc.

[0051] PointNet++ is an efficient point cloud processing network. The present invention uses the Pix2PixHD network model to learn the local structural features in the point cloud to calculate the dynamic point information of dynamic points including pedestrians, non-motor vehicles, and motor vehicles, as well as the road semantic information such as roads and green belts, thereby improving the recognition and segmentation accuracy of complex scenes.

[0052] In a specific embodiment, an odometer device is selected for motion compensation and global posture optimization according to the satellite positioning status of the global satellite time system.

[0053] When the satellite positioning status of the global satellite time system is normal, the satellite positioning is used as the odometer device.

[0054] When the satellite positioning status of the global satellite time system is abnormal, the laser radar sensor is used as the odometer device, and the integrated navigation data accessed provides observation values ​​for the laser radar sensor used as the odometer device. By real-time monitoring of the signal standard deviation value of the integrated navigation data accessed, the accuracy of the current integrated navigation data is judged, and then the current laser sensor data is screened for mapping, avoiding problems such as data unavailability and low mapping accuracy.

[0055] When performing motion compensation on the laser radar point cloud data, after selecting the odometer device, the motion information data of the odometer device is aligned with the laser radar point cloud data through the timestamp, and then the laser radar point cloud data is divided into smaller segments according to actual needs, each of which corresponds to the motion information data of the odometer device in a shorter time. Because in a very short time interval, the movement of the device can be approximated as linear or limited rotation, thereby simplifying the compensation calculation.

[0056] Then, the difference calculation is performed for each point cloud data segment, and the specific motion state of each point in the segment is calculated based on the motion information data (position and posture) at the start and end times. While the integrated navigation data provided observation values ​​for the laser odometer, it also uses the integrated navigation posture corresponding to each frame of laser after posture interpolation as the vertex and the odometer device data obtained by laser matching as the edge to construct an optimization factor graph. The optimization goal is to adjust the posture of each node to minimize the error function on all edges. This optimization process usually needs to take into account various constraints, such as relative motion constraints between consecutive frames, position constraints of known landmarks, and constraints generated during closed-loop detection. Iterative optimization can be performed using formulas such as Taylor expansion. Global optimization by accessing integrated navigation data can effectively improve posture accuracy.

[0057] Specifically, the process of calculating the specific motion state of each point in the segment through interpolation is as follows. The end time of each frame of point cloud is used as the timestamp of the frame data. It is known that the transformation relationship from the maximum time lidar coordinate system to the world coordinate system in a frame of point cloud is , the transformation relationship from the minimum time lidar coordinate system to the world coordinate system is: and the corresponding time points and So for any point in the point cloud data segment The coordinates below The transformation relationship to the world coordinate system can be expressed as , assuming that the point is The coordinates of the pose are ,but:

[0058]

[0059]

[0060] Represents matrix multiplication operation, let the transformation matrix ,but

[0061] because , so that both sides of the equation are multiplied from the left , then . Because the matrix The inverse of the matrix = 1, so ,so ,at the same time , so we have: , .therefore, . And because , .

[0062]

[0063]

[0064] As can be seen from the above formula, the transformation matrix is ​​divided into the rotation part and pan part part.

[0065] For rotating parts , expressed in quaternions, namely: ,in express time The posture of , according to the spherical interpolation formula:

[0066]

[0067] Where t represents the time interpolation scale factor, that is .

[0068] but

[0069] For translation part , expressed in quaternions, namely: ,in .

[0070] Using the obtained motion information, adjust the spatial position of the points in each point cloud data segment to compensate for the deformation caused by the movement of the device. According to the above formula, each point in the point cloud corresponds to the rotation part and translation part of the standard timestamp of the frame point cloud data. Each point is added with the corresponding rotation part and translation part according to the vector calculation method to complete the adjustment and compensation. All motion-compensated point cloud segments are recombined to form a whole, motion-compensated point cloud data.

[0071] like Figure 2 and Figure 3 As shown, Figure 2 Unoptimized surface roughness feature maps are stitched for pure poses; Figure 3 This is the surface roughness feature map after global optimization. It is obvious that the authenticity of the point cloud data compensated by the optimized splicing point cloud clarity algorithm and the mapping data collected by the combined navigation has increased by more than 10%.

[0072] In a specific embodiment, after global optimization using a graph optimization method, motion compensation is performed again, and the global pose information is combined to calculate the point cloud set of the lidar data in the global coordinate system, and the local observation data is converted into a point cloud set in the global coordinate system, thereby achieving a unified representation.

[0073] In a specific embodiment, the point cloud set is projected to form a grid map after filtering out the dynamic points based on the dynamic point information obtained in the previous semantic recognition, and the statistical features of the point cloud set are extracted based on the position of each grid point, the reflection intensity and the road semantic information, and the statistical features include verticality features, intensity features and semantic features.

[0074] In a specific embodiment, extracting the verticality feature specifically includes:

[0075] S61, calculating the mean of the point cloud set P , decentralize the point cloud set and get ,in, , , ,in represents the coordinates of the i-th point cloud data in the point cloud set, , N is the number of points in a frame of point cloud;

[0076] S62, calculate the decentralized point cloud set The covariance matrix of ,based on Perform eigenvalue decomposition of the covariance matrix, where and Respectively represent the i-th eigenvalue and the corresponding eigenvector;

[0077] S63, sorting the eigenvalues ​​in descending order, calculating the angle between the eigenvector corresponding to the minimum eigenvalue and the Z axis, and obtaining the verticality feature. The closer the angle is to 0 degrees, the higher the verticality; the closer the angle is to 90 degrees, the lower the verticality.

[0078] In a specific embodiment, the semantic feature is extracted based on the normal distribution of the road semantic information, and the intensity feature is directly obtained according to the intensity attribute of the point cloud collected by the laser radar.

[0079] In a specific embodiment, a verticality map, an intensity map and road element data are generated respectively based on the verticality feature, the intensity feature and the semantic feature.

[0080] Import the verticality map, the intensity map and the road element data into a map editing tool. The verticality map can show vertical structures such as curbs, walls, and curbstones. These structures mark the boundaries of the road. Road element data includes coordinate data such as lane lines, curbs, and other road elements obtained through semantic information. Based on the vertical structure of the road surface displayed on the verticality map, the actual road boundary can be determined. With the help of the visualization information of the curb data in the road element data, the location of the curb data in the road element data can be compared with the verticality map to see if it matches. Use tools to mark and adjust the road boundary lines and intersection areas and save them.

[0081] The intensity map can intuitively represent the reflective characteristics of different objects such as the ground, buildings, and vegetation. Lane lines have high reflectivity, regularity, and continuity, and can be distinguished from the intensity map. Assisted by the lane line data in the road elements, accurate guide lines are generated using tool annotations to guide the autonomous vehicle along the correct path. Through this series of comprehensive analyses, editors can accurately mark lane lines on the intensity map and provide accurate navigation information for autonomous vehicles. Other road information such as stop lines, parking spaces, speed bumps, crosswalks, etc. can be verified by matching the intensity map and road element data, so that they can be accurately marked on the map.

[0082] After all road elements are labeled, the topology and associations between lanes and road elements need to be set. This step ensures that the elements in the map are not only spatially accurate, but also logically interconnected, providing the necessary information for path planning and vehicle decision-making. Finally, all the information and annotations are integrated to build a complete high-precision map. This map will contain all necessary navigation data such as road geometry, lane information, traffic signs and lights, and the relationships between them.

[0083] According to a second aspect of the present invention, a system for collecting and producing high-precision maps by an autonomous driving cleaning robot is proposed. Figure 4 As shown, it includes the following modules:

[0084] Data acquisition module 401: The lidar sensor and camera are synchronized with the global satellite time system in real time through the controller. The controller combines the pulse second PPS signal and the global positioning system recommended minimum specific transmission information GPRMC for satellite timing. The lidar is synchronized with the controller in real time through the precision time protocol PTP. The camera is synchronized with the controller in real time through the pulse trigger timing mechanism, and the camera data and lidar point cloud data are collected.

[0085] Semantic extraction module 402: extracting semantic information from the laser radar point cloud data and the camera data to obtain dynamic point information and road semantic information.

[0086] Data processing module 403: performs motion compensation and global pose optimization on the laser radar point cloud data, and converts the laser radar point cloud data into a point cloud set in a global coordinate system.

[0087] Map making module 404: filtering out dynamic points in the point cloud set based on the dynamic point information, extracting statistical features of the point cloud set based on the road semantic information, and constructing a high-precision map based on the statistical features.

[0088] The above describes the specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0089] In the description of the present application, it should be understood that the terms "upper", "lower", "inside", "outside", etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, which are only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. The wording 'comprising' does not exclude the presence of elements or steps not listed in the claims. The wording 'one' or 'an' preceding an element does not exclude the presence of a plurality of such elements. The simple fact that certain measures are recorded in mutually different dependent claims does not indicate that a combination of these measures cannot be used for improvement. Any reference symbols in the claims should not be interpreted as limiting the scope.

Claims

1. A method for collecting and producing high-precision maps by an autonomous driving cleaning robot, characterized in that: The following steps are involved: S1. The laser radar sensor and camera are synchronized with the global satellite time system in real time through the controller. The controller combines the pulse second PPS signal and the minimum specific transmission information recommended by the global positioning system GPRMC for satellite timing. The laser radar is synchronized with the controller in real time through the precision time protocol PTP. The camera is synchronized with the controller in real time through the pulse trigger timing mechanism, and the camera data and the laser radar point cloud data are collected; S2, extracting semantic information from the laser radar point cloud data and the camera data, processing the camera data using a network model to obtain 2D semantic information, establishing a corresponding relationship between the camera data and the laser radar point cloud data, and processing the laser radar point cloud data using the network model in combination with the 2D semantic information to obtain dynamic point information and road semantic information; S3, performing motion compensation and global pose optimization on the laser radar point cloud data, and converting the laser radar point cloud data into a point cloud set in a global coordinate system; S4. Filter out dynamic points in the point cloud set based on the dynamic point information, extract statistical features of the point cloud set based on the position of each grid point in the grid map of the point cloud set, the reflection intensity and the road semantic information, the statistical features respectively include verticality features, intensity features and semantic features, generate a verticality map, an intensity map and road element data based on the verticality features, the intensity features and the semantic features respectively, and construct the high-precision map by comparing the road element data with the verticality map and the intensity map for matching, annotating and adjusting.

2. The method for collecting and making high-precision maps by an autonomous driving cleaning robot according to claim 1, characterized in that: Step S2 specifically includes using the Pix2PixHD network model to process the camera data to obtain 2D semantic information, establishing a correspondence between the camera data and the lidar point cloud data, and using the PointNet++ network model to process the lidar point cloud data in combination with the 2D semantic information to obtain the dynamic point information and the road semantic information.

3. The method for collecting and making high-precision maps by an autonomous driving cleaning robot according to claim 1, characterized in that: Step S2 specifically includes: selecting an odometer device for motion compensation and global posture optimization according to the satellite positioning status of the global satellite time system; when the satellite positioning status of the global satellite time system is normal, using satellite positioning as the odometer device; when the satellite positioning status of the global satellite time system is abnormal, using the lidar sensor as the odometer device.

4. The method for collecting and making high-precision maps by an autonomous driving cleaning robot according to claim 3, characterized in that: Step S1 also includes accessing the combined navigation data, and when the satellite positioning status of the global satellite time system is abnormal, providing observation values ​​for the laser radar sensor as an odometer device, and performing global posture optimization based on a graph optimization method.

5. The method for collecting and making high-precision maps by an autonomous driving cleaning robot according to claim 2, characterized in that: The point cloud set is projected to form a grid map after filtering out the dynamic points based on the dynamic point information.

6. The method for collecting and making high-precision maps by an autonomous driving cleaning robot according to claim 1, characterized in that: Extracting the verticality feature specifically includes: S61, calculating the mean of the point cloud set P Decentralize the point cloud set to obtain in, Where (X i , Y i , Z i ) represents the coordinates of the i-th point cloud data of the point cloud set, i=0,…,N, N is the number of points in a frame of point cloud; S62, calculate the decentralized point cloud set P center The covariance matrix C of , based on CV i =λ i V i Perform eigenvalue decomposition of the covariance matrix, where λ i and V i Respectively represent the i-th eigenvalue and the corresponding eigenvector; S63, sorting the eigenvalues ​​in descending order, calculating the angle between the eigenvector corresponding to the minimum eigenvalue and the Z axis, and obtaining the verticality feature.

7. The method for collecting and making high-precision maps by an autonomous driving cleaning robot according to claim 1, characterized in that: The semantic feature is extracted based on the normal distribution of the road semantic information.

8. A system for collecting and producing high-precision maps by an autonomous cleaning robot, characterized in that: Includes the following modules: Data acquisition module: The laser radar sensor and camera are synchronized with the global satellite time system in real time through the controller. The controller combines the pulse second PPS signal and the global positioning system recommended minimum specific transmission information GPRMC for satellite timing. The laser radar is synchronized with the controller in real time through the precision time protocol PTP. The camera is synchronized with the controller in real time through the pulse trigger timing mechanism, and the camera data and laser radar point cloud data are collected; Semantic extraction module: extracting semantic information from the laser radar point cloud data and the camera data, processing the camera data using a network model to obtain 2D semantic information, establishing a corresponding relationship between the camera data and the laser radar point cloud data, and processing the laser radar point cloud data using the network model in combination with the 2D semantic information to obtain dynamic point information and road semantic information; Data processing module: performs motion compensation and global pose optimization on the laser radar point cloud data, and converts the laser radar point cloud data into a point cloud set in a global coordinate system; Map making module: filtering out dynamic points in the point cloud set based on the dynamic point information, extracting statistical features of the point cloud set based on the position of each grid point in the grid map of the point cloud set, the reflection intensity and the road semantic information, wherein the statistical features include verticality features, intensity features and semantic features respectively, and generating a verticality map, an intensity map and road element data based on the verticality features, the intensity features and the semantic features respectively, and constructing the high-precision map by comparing the road element data with the verticality map and the intensity map for matching, annotating and adjusting.

Images