Data annotation methods, equipment, media and products

Abstract

The embodiment of the present application provides a data annotation method, device, medium and product. The method includes: constructing a perception point cloud map based on the data to be annotated; obtaining a reference point cloud map corresponding to the perception point cloud map; annotating the data to be annotated according to the position distribution characteristics of the first vector data in the perception point cloud map and the second vector data in the reference point cloud map. The method is used to achieve the effect of improving the annotation accuracy.

Description

Technical Field

[0001] The present application relates to the field of autonomous driving technology, and in particular to a data labeling method, device, medium and product. Background Art

[0002] As the intelligentization process of the automotive industry accelerates, many original equipment manufacturers (OEMs) and autonomous driving technology companies are committed to the research and development of intelligent driving algorithms in order to enhance the autonomy of vehicles. In order to reduce the dependence on high-precision maps and improve the flexibility and adaptability of the system, the bird's eye view (BEV) perception technology has gradually become a research hotspot. By constructing a bird's-eye view of the vehicle's surroundings, this technology provides more intuitive and comprehensive information support for perception, decision-making, planning and control, thereby promoting the further development of intelligent driving technology.

[0003] In the development and training phase of BEV perception algorithms, a large amount of real-value data is usually required for algorithm training and optimization. The acquisition process of these real-value data includes collecting raw data using high-precision sensors, and then annotating these data through automated pre-annotation algorithms to generate labels containing the location and dynamic information of key elements such as vehicles, pedestrians, and obstacles.

[0004] However, existing labeling methods still have the problem of low labeling accuracy. Summary of the invention

[0005] The embodiments of the present application provide a data labeling method, device, medium and product to achieve the effect of improving the labeling accuracy.

[0006] In a first aspect, an embodiment of the present application provides a data labeling method, including:

[0007] Construct a perception point cloud map based on the data to be annotated;

[0008] Obtain a reference point cloud image corresponding to the perception point cloud image;

[0009] The data to be labeled are labeled according to the position distribution characteristics of the first vector data in the perception point cloud image and the second vector data in the comparison point cloud image.

[0010] In a possible implementation, obtaining a reference point cloud image corresponding to the perception point cloud image includes:

[0011] Relocating the perception point cloud image and the preset reference point cloud image, and determining the reference point cloud image corresponding to the perception point cloud image from the reference point cloud image; or,

[0012] A reference point cloud map is constructed based on the target point cloud data, and the collection range of the target point cloud data is the same as the collection range of the data to be annotated.

[0013] In a possible implementation, relocating the perception point cloud image and the preset reference point cloud image, and determining a reference point cloud image corresponding to the perception point cloud image from the reference point cloud image, includes:

[0014] Relocate the perception point cloud image and the reference point cloud image, and determine the sub-image in the reference point cloud image that overlaps with the perception point cloud image;

[0015] The pose of the sub-image corresponding to the overlapping area is optimized to obtain a control point cloud image corresponding to the perceived point cloud image.

[0016] In a possible implementation, labeling the data to be labeled according to the position distribution characteristics of the first vector data in the perception point cloud image and the second vector data in the comparison point cloud image includes:

[0017] The data to be labeled is labeled according to the projection offset between the first vector data and the second vector data in the target projection space.

[0018] In a possible implementation, the data to be annotated includes an image and / or a point cloud; and the target projection space includes a two-dimensional plane and / or a three-dimensional space determined based on the data to be annotated.

[0019] In a possible implementation, labeling the data to be labeled according to the projection offset of the first vector data and the second vector data in the target projection space includes:

[0020] Determine the offset of the vector at the target position according to the geometric accuracy of the vector at the target position in the target projection space of the first vector data and the second vector data;

[0021] The data to be labeled is labeled according to the offset of the vector at the target position.

[0022] In a possible implementation, labeling the data to be labeled according to the offset of the vector at the target position includes:

[0023] If the offset of the vector at the target position meets the preset offset requirement, the data to be labeled is labeled according to the vector at the target position;

[0024] and / or,

[0025] If the data amount of all vectors that do not meet the preset offset requirement within the preset position range where the target position is located is greater than the preset data amount, determining the overall offset of the vectors within the preset position range according to the first vector data and the second vector data;

[0026] The data to be marked is marked according to the overall offset of the first vector data and the second vector data within a preset range.

[0027] In a possible implementation, labeling the data to be labeled according to the projection offset of the first vector data and the second vector data in the target projection space includes:

[0028] Determining target vector data in the second vector data according to a projection position of the second vector data in the target projection space, wherein the target vector data represents vector data in the second vector data that is not aligned with the target projection space;

[0029] According to the target projection space, the target vector data in the second vector data is aligned and adjusted to obtain adjusted second vector data;

[0030] The data to be labeled is labeled according to the projection offset of the first vector data and the adjusted second vector data in the target projection space.

[0031] In a second aspect, an embodiment of the present application provides a data labeling device, including:

[0032] A construction module, used to construct a perception point cloud map based on the data to be annotated;

[0033] An acquisition module, used to acquire a reference point cloud image corresponding to the perception point cloud image;

[0034] The annotation module is used to annotate the data to be annotated according to the position distribution characteristics of the first vector data in the perception point cloud image and the second vector data in the comparison point cloud image.

[0035] In a third aspect, an embodiment of the present application provides a data labeling device, including: a memory, a processor;

[0036] Memory stores computer-executable instructions;

[0037] The processor executes the computer-executable instructions stored in the memory, so that the processor executes the above first aspect and / or various possible implementations of the first aspect.

[0038] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, they are used to implement the first aspect above and / or various possible implementations of the first aspect.

[0039] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the above first aspect and / or various possible implementation methods of the first aspect.

[0040] The data annotation method, device, medium and product provided in the embodiments of the present application construct a perception point cloud map through the data to be annotated, and obtain a reference point cloud map corresponding to the perception point cloud map, and combine and utilize the position distribution characteristics of the first vector data in the perception point cloud map and the second vector data in the reference point cloud map to annotate the data to be annotated, so that when annotating the data to be annotated, the first vector data can be used for annotation with reference to the second vector data in the reference point cloud map, thereby achieving the effect of annotation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0042] Figure 1 A schematic diagram of a scenario for the data annotation method provided in this application;

[0043] Figure 2 Schematic diagram of the data annotation method provided for this application Figure 1 ;

[0044] Figure 3 Schematic diagram of the data annotation method provided for this application Figure 2 ;

[0045] Figure 4 Schematic diagram of the data annotation method provided for this application Figure 3 ;

[0046] Figure 5a A flowchart of another data annotation method provided for this application;

[0047] Figure 5b A schematic diagram of preprocessing in another data annotation method provided in this application;

[0048] Figure 5c A schematic diagram of the posture and positioning information corresponding to the sub-image in another data annotation method provided in this application;

[0049] Figure 5d A schematic diagram of pose results in another data annotation method provided in this application;

[0050] Figure 6 A schematic diagram of the structure of the data annotation device provided in this application;

[0051] Figure 7 A schematic diagram of the structure of the data annotation device provided in this application.

[0052] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0053] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0054] First, the terms involved in this application are explained:

[0055] Data labeling refers to the process of adding labels or annotations to raw data (such as images, point clouds, etc.) through manual or automated means in the field of machine learning and artificial intelligence to ensure the accuracy and reliability of the data when training and verifying algorithm models.

[0056] True data can refer to vector information that reflects the characteristics of the real environment at a specific point in time and space, obtained through high-precision measurement or annotation. It usually includes parameters such as position, direction, and speed, and is used to verify and calibrate the accuracy of sensor data.

[0057] With the development of vehicle intelligence, various OEMs and autonomous driving technology companies are actively developing intelligent driving algorithms, reducing the reliance on high-precision maps by introducing bird's-eye view technology, thereby improving the flexibility and adaptability of the system. BEV technology can build a bird's-eye view of the vehicle's surroundings, providing more intuitive and comprehensive information support for perception, decision-making and path planning, and is one of the key factors in promoting the advancement of intelligent driving technology.

[0058] However, there are some challenges in the development of BEV perception algorithms:

[0059] First, the existing automatic pre-labeling methods have the problem of low accuracy and recall of perception elements when processing BEV4D data, which directly affects the quality of the true value data.

[0060] Secondly, in order to train and optimize perception algorithms, a large amount of high-quality true-value data is required. However, the cost of manual annotation of BEV true-value data on the market is much higher than that of ordinary 2D image annotation. This not only increases the cost of algorithm development, but also limits the availability of data and the iteration speed of the algorithm.

[0061] Therefore, the market urgently needs a technical solution that can effectively improve the automatic pre-labeling capabilities of BEV4D data, so as to reduce costs and improve the efficiency of true value data labeling, and help intelligent driving algorithms achieve mass production faster and better.

[0062] The data annotation method, equipment, medium and product provided by the present application take into account that the same position may have been annotated multiple times during the data annotation process. For this reason, the present application cross-compares and verifies the position distribution characteristics between the second vector data in the constructed control point cloud map and the first vector data in the perception point cloud map constructed based on the data to be annotated, so as to obtain relatively accurate vector true value data results. Data annotation based on these relatively accurate vector true value data results can make full use of the existing annotation information and significantly improve the annotation accuracy. At the same time, partial automation of the annotation process is also achieved, which effectively avoids the common mislabeling and missing labeling phenomena in manual annotation, reduces the need for manual intervention, greatly reduces the amount of repetitive work, and improves the annotation quality and efficiency. High-quality annotation data not only provides solid data support for the stable mass production of intelligent driving algorithms, but also accelerates the commercialization process of intelligent driving technology.

[0063] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0064] Figure 1 The present application provides a scenario structure diagram of a data labeling system, wherein the data labeling system includes a data acquisition device 1100 and a data labeling device 1200, wherein the data acquisition device 1100 and the data labeling device 1200 may be the same device or different devices; in the data labeling system, the data acquisition device 1100 is used to acquire data to be labeled, and the data labeling device 1200 is used to execute the data labeling method provided in the embodiment of the present application; wherein the data acquisition device 1100 may be any device in a server, a server cluster (such as including but not limited to a cloud, a distributed system or other clusters, etc.), a vehicle networking system, an Internet of Things system, a terminal device (such as including but not limited to a vehicle, an on-board processor, a computer, a mobile phone, a tablet or an Internet of Things device, a vehicle networking device, etc.) or other computer devices; the present application embodiment does not impose any special restrictions on the specific configuration of the data acquisition device 1200, and those skilled in the art may set it according to actual needs.

[0065] Figure 2 Schematic diagram of the data annotation method provided for this application Figure 1 , the method may be executed by the aforementioned data annotation device 1200, such as Figure 2 As shown, the method includes:

[0066] S201. Construct a perception point cloud map based on the data to be labeled.

[0067] The data to be labeled may refer to the initial data that needs to be labeled. The present application embodiment does not limit the data type and collection method of the data to be labeled, and those skilled in the art can set it according to actual needs. The data to be labeled may come from various sensors, such as LiDAR, camera, millimeter wave radar, etc.

[0068] In the embodiment of the present application, the data to be annotated may include any one or any combination of image file data (i.e., image / picture), point cloud file data (i.e., point cloud), calibration file data, engineering coordinate origin data, inertial data, positioning file data, wheel speed file data, etc. Before obtaining the data to be annotated, the data to be annotated may be processed for specification adjustment, so as to facilitate the subsequent environment to perform adaptation processing of the same specification. Among them:

[0069] Image file data may refer to environmental visual information captured by a camera and recorded at a specific time point, usually including attributes such as file format, resolution, color mode and timestamp. Image file data that meets preset specification requirements may refer to images whose image timestamps are between the smallest timestamp and the largest timestamp of the positioning file, wherein the images may also be output in ascending order according to the needs of the image timestamps. Image file data may be used for the output of the first vector data and the input of the vector projection reference.

[0070] Point cloud file data may refer to a large number of points collected by a laser radar or other equipment at a specific time point, representing the three-dimensional spatial information of the environment. Each point may contain location coordinates and other attribute information such as reflection intensity. Point cloud file data that meets the preset specification requirements may refer to a point cloud file whose timestamp is between the minimum timestamp and the maximum timestamp of the positioning file. Point cloud file data may be used for the construction of a point cloud map, the output of the first vector data, and the input of a vector projection reference.

[0071] Calibration file data may refer to data that contains the intrinsic and extrinsic information of a sensor (such as a camera) and is used to convert the data collected by the sensor into a unified coordinate system to ensure accurate alignment and fusion of multi-sensor data. Calibration file data that meets preset specifications may refer to data that converts the camera extrinsic parameters into an FRD coordinate system for use in vector projection.

[0072] The engineering coordinate origin data may refer to the starting reference point of map construction, which is used as the reference point for measuring other positions and coordinate conversion. The engineering coordinate origin data that meets the preset specification requirements may refer to the data used for point cloud map construction and vector coordinate conversion.

[0073] Inertial data may refer to the acceleration and angular velocity information of a device at a specific time point recorded by an inertial measurement unit (IMU), which is used to determine the motion state and attitude change of the device. Inertial data that meets the preset specification requirements may refer to data in the format of timestamp, acceleration, angular velocity, and used for the construction of point cloud maps.

[0074] Positioning file data can refer to the data used to record the precise position and posture of a vehicle or device at a specific time point, as well as the corresponding speed information. Positioning file data can be used for path planning, map construction, and location positioning. Positioning file data that meets the preset specification requirements requires the quaternion corresponding to the timestamp, the enu coordinate component, the three speeds under the flu coordinate, and the data used for point cloud map construction and data alignment and relocation.

[0075] Wheel speed file data may refer to the rotation speed information of the four wheels (front left, front right, rear left, rear right) of the vehicle at a specific time point. Wheel speed file data that meets the preset specification requirements may refer to data in the format of timestamp and four wheel speeds.

[0076] A point cloud may refer to a collection of three-dimensional points, each of which represents a position in space. These points are usually collected by three-dimensional scanning equipment (such as LiDAR) or other three-dimensional sensors. In an embodiment of the present application, the perception point cloud map can be a point cloud map obtained by mapping the data to be annotated. For example, the sub-image point cloud corresponding to the data to be annotated can be mapped starting from the northeast celestial coordinate system with the first frame of the point cloud as the origin, and the corresponding complete spliced ​​point cloud and the posture and positioning information corresponding to each sub-image can be obtained after splicing.

[0077] S202: Obtain a reference point cloud image corresponding to the perception point cloud image.

[0078] Among them, the reference point cloud map refers to a point cloud map with the same geographical location as the perception point cloud map, but with different collection time or method, which can be used for reference comparison and assist in subsequent data annotation. This type of point cloud map must meet certain accuracy requirements, and its type and source are not restricted. For example, it can be an existing point cloud map, a high-definition (HD) map built by other manufacturers, a standard definition (SD) map, or a simplified version of the map. By comparing with these high-precision maps, the accuracy and efficiency of data annotation can be improved.

[0079] S203: label the data to be labeled according to the position distribution characteristics of the first vector data in the perceived point cloud image and the second vector data in the reference point cloud image.

[0080] The vector data may refer to the vector data of each object in the point cloud map, where:

[0081] The object in the point cloud image may refer to an entity layer in the point cloud image. For example, the object in the point cloud image may include any one or any combination of lane separators, pole-shaped objects, road surface objects, traffic lights, traffic signs, and physical isolation objects represented in the point cloud image.

[0082] The vector data of an object may include attribute feature information used to describe the attribute features of the object, and the attribute features include but are not limited to any one or any combination of position, shape, and direction, and the attribute feature information includes but is not limited to any one or any combination of the pose information and positioning information of the object in the point cloud map; the vector data of the object may also include vector data of a virtual layer that is associated with the physical layer, and the vector data of the aforementioned virtual layer may include but is not limited to any one or any combination of lane topology relationships, lane regular speed limit tables, lane construction tables, element spatial relationship tables, object-associated road tables, object-associated lane signs, road line morphology tables, road speed limits, etc.

[0083] In an embodiment of the present application, the first vector data may refer to the vector data of each object in the perception point cloud image, and the second vector data may refer to the vector data of each object in the comparison point cloud image.

[0084] The position distribution feature may refer to any one or any combination of the position, distribution, relative position, etc. of the spatial position of the vector data in the point cloud image.

[0085] In the embodiment of the present application, by determining the position distribution of the first vector data in the perceived point cloud map and the position distribution of the second vector data in the reference point cloud map, and comparing the position of the first vector data in the perceived point cloud map with the position of the second vector data in the reference point cloud map, it is possible to determine which positions of the vector data in the first vector data and the second vector data correspond to each other, so that the vector data corresponding to the position can be marked as true value data. In the embodiment of the present application, the method for determining the first vector data in the perceived point cloud map may include:

[0086] The perception point cloud image is input into a preset recognition model to obtain the first vector data output by the recognition model.

[0087] The preset recognition model may refer to a model for determining objects in a perceived point cloud image and vector data of the objects. The recognition model may be a perception algorithm model for fusion of point cloud and image. For example, the recognition model may be any one or a combination of PointPillars model, MV3D (Multi-View 3D) model, AVOD (Anchor-based Voxel Object Detection) model, SECOND (Sparsity-aware Convolutional Detector) model, etc. In some embodiments, the first vector data may be stored in any format as required. For example, the first vector data may be stored in, but not limited to, geojson format data.

[0088] In an embodiment of the present application, in the aforementioned step S203, the position distribution feature may be a projection offset of the first vector data and the second vector data in the target projection space, and then in step S203, the data to be labeled may be labeled according to the projection offset of the first vector data and the second vector data in the target projection space.

[0089] Among them, the target projection space may refer to the space used to project the first vector data in the perception point cloud map and the second vector data in the comparison point cloud map when performing data annotation. The target projection space may be determined based on a two-dimensional plane or a three-dimensional space determined by the data to be annotated; for example, the aforementioned target projection space may be determined by scene data collected for the target scene; wherein the aforementioned target scene is a scene for collecting the data to be annotated, and the scene data may be other data collected for the target scene in addition to the data to be annotated, and the scene data may also be the data to be annotated itself collected for the target scene; if the data to be annotated is an image or point cloud collected for a certain expressway toll station, the target scene may be the expressway toll station, and the scene data may be other images or point clouds collected for the expressway toll station in addition to the data to be annotated, and the scene data may also be the data to be annotated itself. In the embodiments of the present application, the target scene is not limited, and those skilled in the art may set it according to actual needs.

[0090] When the data to be annotated are photos or pictures collected for the target scene and used to describe the geographic location, the scene data can be other images collected for the target scene. By constructing a two-dimensional coordinate system on other images and using important landmarks or boundary lines in the image as key coordinate points and boundaries, the two-dimensional plane can be determined.

[0091] For example, if the other image is a bird's-eye view of the entire highway toll station captured by drone photography, satellite imagery, or a high-altitude fixed-point camera, a two-dimensional coordinate system can be established on the bird's-eye view. Select the center of the toll station or a prominent landmark as the origin, the x-axis extends along the main lane, and the y-axis is perpendicular to the lane. After constructing the two-dimensional coordinate system, mark the key landmarks on the bird's-eye view, such as lane boundaries, toll booth locations, safety islands, speed bumps, traffic lights, and signs. The key coordinate points and boundaries. Finally, the information of these key coordinate points and boundaries is integrated into a detailed two-dimensional plane representation, so that the two-dimensional plane can be used as the target projection space.

[0092] When the data to be annotated is a point cloud collected for the target scene, the scene data can be other point clouds collected for the target scene. A three-dimensional coordinate system is set on other point clouds, wherein the origin of the three-dimensional coordinate system can be determined according to the significant landmark or center position in other point clouds, and the directions of the x-axis, y-axis and z-axis can be set according to the actual geographical orientation. Then, in the three-dimensional coordinate system, the feature points representing important structural elements in other point clouds are identified and extracted, and the extracted feature points are plane-fitted to determine the significant geometric plane (such as the ground, street lights, etc.). After the geometric plane is determined, a new three-dimensional coordinate system is established based on the geometric plane, wherein the origin of the new coordinate system can be any point on the geometric plane, the directions of the x-axis and y-axis can be determined according to the actual geographical orientation, and the direction of the z-axis can be the direction of the normal vector. Finally, possible geometric distortion is checked and corrected to ensure data consistency and accuracy, and the target projection parameters of the three-dimensional space (such as orthogonal or perspective projection, scale factor, etc.) are clarified. Thus, a three-dimensional space can be obtained, and the three-dimensional space can be used as a target projection space.

[0093] For example, for a point cloud map of a highway toll station, the main toll booth can be selected as the origin, the x-axis extends along the main lane, the y-axis is perpendicular to the lane direction, and the z-axis points to the sky, thereby constructing a three-dimensional coordinate system. Then, in this three-dimensional coordinate system, the Random Sample Consensus Algorithm (RANSAC) can be used to identify the positions of the ground plane and lane boundaries, toll booths, safety islands, speed bumps, traffic lights, and signs. After identifying the positions of the ground plane and lane boundaries, toll booths, safety islands, speed bumps, traffic lights, and signs, the ground plane can be used as a significant geometric plane to establish a new three-dimensional coordinate system, in which a point on the ground plane is selected as the origin of the new coordinate system, and the direction of the normal vector is used as the positive direction of the z-axis to ensure that the x-axis extends along the main lane direction and the y-axis is perpendicular to the lane direction. All extracted key landmarks (such as lane boundaries, toll booth locations, etc.) are marked in the new three-dimensional coordinate system, and the consistency and accuracy of the data are ensured. Finally, the target projection parameters of the three-dimensional space are clarified, including but not limited to the projection type (such as orthogonal projection), scale factor, etc., thereby constructing a three-dimensional space that can be used as the target projection space.

[0094] In some embodiments, when there are multiple other point clouds, a new three-dimensional coordinate system established on each other point cloud can be transformed by matrix, so as to obtain a target projection space in the same three-dimensional coordinate system and composed of multiple other point clouds.

[0095] Projection offset may refer to the offset of the first vector data and the second vector data at the same geographical location in the target projection space, wherein when the first vector data and the second vector data are projected to the same target projection space (whether a two-dimensional plane or a three-dimensional space), the difference in their positions at the same geographical location is the projection offset, wherein the projection offset may include deviations such as position deviation, direction difference, shape change, and scale difference.

[0096] After determining the projection offset, the vector data without deviation is labeled as the true value data.

[0097] In order to avoid inaccurate projection offset judgment caused by different vector data formats, in an embodiment of the present application, the data to be annotated is annotated according to the projection offset of the first vector data and the second vector data in the target projection space, including:

[0098] Determining target vector data in the second vector data according to a projection position of the second vector data in the target projection space, wherein the target vector data represents vector data in the second vector data that is not aligned with the target projection space;

[0099] According to the target projection space, the target vector data in the second vector data is aligned and adjusted to obtain adjusted second vector data;

[0100] The data to be labeled is labeled according to the projection offset of the first vector data and the adjusted second vector data in the target projection space.

[0101] Among them, the target vector data may refer to the vector data of the part that does not fit the second vector data and the target projection space. For example, when the second vector data is projected to the target projection space, the part that exceeds the image boundary may be the target vector data, or the part projected on other erroneous objects or backgrounds may be the target vector data.

[0102] The alignment adjustment process of the target vector data may refer to a process of correcting the target vector data so that the target vector data is completely aligned with the actual object in the image. For example, the target vector data may be interrupted and then the point closest to the edge of the image may be connected.

[0103] The adjusted second vector data may refer to data adjusted to avoid inaccurate projection offset judgment caused by different vector data formats, thereby ensuring that the second vector data is accurately aligned with the image boundary to avoid the exceeding part affecting the accuracy of the annotation.

[0104] The data annotation method provided in the embodiment of the present application can obtain relatively accurate vector truth data results by referencing the second vector data in the constructed reference point cloud map and cross-comparing and verifying the position distribution characteristics of the first vector data in the perception point cloud map constructed based on the data to be annotated. Among them, since the position distribution characteristics can fully reflect the arrangement pattern and relative position relationship of the object in space. Therefore, data annotation can be performed based on these relatively accurate vector truth data results, thereby making full use of the existing annotation information and significantly improving the annotation accuracy.

[0105] Figure 3 Schematic diagram of the data annotation method provided for this application Figure 2 ,like Figure 3 As shown, in this embodiment Figure 2 Based on the embodiment, a method for obtaining a reference point cloud image corresponding to a perception point cloud image is described in detail. The method includes:

[0106] S301. Relocate the perceived point cloud map and the preset reference point cloud map, and determine the reference point cloud map corresponding to the perceived point cloud map from the reference point cloud map; or, construct a reference point cloud map based on the target point cloud data, and the collection range of the target point cloud data is the same as the collection range of the data to be annotated.

[0107] The control point cloud map can be determined from the reference point cloud map, or can be constructed based on the target point cloud data, wherein:

[0108] When a control point cloud map is determined from a reference point cloud map, the reference point cloud map may refer to a point cloud map that includes the control point cloud map. Since the geographical location range represented by the point cloud map used for reference is generally larger and the formats of the point cloud maps may be different, in order to determine the control point cloud map from the reference point cloud map, the perceived point cloud map and the preset reference point cloud map may be repositioned, thereby determining the control point cloud map from the reference point cloud map.

[0109] Repositioning may refer to the process of comparing and matching the sensor data in the perception point cloud with the data in the reference point cloud, thereby determining the precise position and posture of the perception point cloud in the reference point cloud. In an embodiment of the present application, the perception point cloud and the preset reference point cloud may be repositioned. If the perception point cloud overlaps at the edge of the reference point cloud, the sub-image pose of the overlapping portion may be obtained by jointly optimizing multiple point cloud data, thereby generating an accurate repositioning pose result.

[0110] In some embodiments, the repositioning pose result may include point cloud data in the reference point cloud directly corresponding to the perceived point cloud, and may also include point cloud data in the reference point cloud indirectly corresponding to the perceived point cloud. The directly corresponding point cloud data may refer to point cloud data with the same data type and acquisition method, and the indirectly corresponding point cloud data may refer to point cloud data that needs to be converted before it can correspond due to different sensor types, different acquisition methods, etc.

[0111] When the reference point cloud map is constructed based on the target point cloud data, the target point cloud data may refer to historical data that has been collected and annotated within the same collection range. By mapping the historical data, the reference point cloud map may be obtained.

[0112] In the embodiment of the present application, the perception point cloud image and the preset reference point cloud image are relocated, and a reference point cloud image corresponding to the perception point cloud image is determined from the reference point cloud image, including:

[0113] Relocate the perception point cloud image and the reference point cloud image, and determine the sub-image in the reference point cloud image that overlaps with the perception point cloud image;

[0114] The pose of the sub-image corresponding to the overlapping area is optimized to obtain a control point cloud image corresponding to the perceived point cloud image.

[0115] The sub-image in the perceived point cloud image that overlaps with the reference point cloud image may refer to the portion of the overlapping area of ​​the perceived point cloud image in the reference point cloud image. In an embodiment of the present application, by comparing the position information of the perceived point cloud image and the reference point cloud image, the sub-image in the reference point cloud image that overlaps with the perceived point cloud image can be determined.

[0116] Pose optimization can refer to adjusting the position and posture of the sub-image in the overlapping area of ​​the reference point cloud image and the perception point cloud image through an algorithm to minimize the difference between the two. For example, the ICP algorithm (Iterative Closest Point algorithm) and the error minimization method can be used to ensure that the point cloud data in the overlapping area is accurately aligned in space, thereby generating a more accurate reference point cloud image.

[0117] The data annotation method provided in the embodiment of the present application can reposition the perceived point cloud image and the reference point cloud image, determine the sub-images in the reference point cloud image that overlap with the perceived point cloud image, and optimize the position and posture of the sub-images in the overlapping area, thereby generating a reference point cloud image that is precisely aligned with the perceived point cloud image. This can improve the alignment accuracy and overall consistency of the point cloud data, and provide a basis for subsequent data annotation and analysis.

[0118] As an embodiment, in step S203, during the process of annotating the data to be annotated based on the projection offset of the first vector data and the second vector data on the target projection space, the offset of the vector at the target position can be determined based on the geometric accuracy of the vector at the target position in the aforementioned target projection space of the first vector data and the second vector data, and then the data to be annotated can be annotated based on the offset of the vector at the target position.

[0119] The vector at the target position may refer to the vector at the same position in the first vector data and the second vector data; the geometric accuracy may refer to the degree of consistency between the vector at the same position in the first vector data and the second vector data in terms of spatial position and shape. That is, it can characterize the difference in geometric features (for example, any one or any combination of the starting point, end point, curvature, etc.) of the vector at the same position in the first vector data and the second vector data. The offset of the vector at the target position can be obtained by calculating the difference in spatial distance between corresponding points.

[0120] Furthermore, the data to be labeled can be labeled according to the offset of the vector at the target position by any one or a combination of the following labeling methods A1 and A2, wherein:

[0121] In marking method A1, it is determined whether the offset of the vector at the target position meets the preset offset requirement. If the offset of the vector at the target position meets the preset offset requirement, the data to be marked is marked according to the vector at the target position; if it does not meet the preset offset requirement, the data to be marked can be marked by other marking methods, or by marking method A2.

[0122] In marking method A2, a vector that does not meet the preset offset requirement is determined as an abnormal vector, and then the data amount of all abnormal vectors within the preset position range of the target position is detected. When the data amount of all abnormal vectors within the preset position range of the target position is greater than the preset data amount, the overall offset of the vectors within the preset position range is determined based on the first vector data and the second vector data; the data to be marked is marked based on the overall offset of the vectors within the preset range of the first vector data and the second vector data; when the data amount of all abnormal vectors within the preset position range of the target position is less than or equal to the preset data amount, the data to be marked can be marked by other marking methods.

[0123] Among them, the preset offset requirement in marking method A1 and marking method A2 may refer to the requirement for the offset threshold, and the offset threshold may be set in advance as needed. When the offset of the vector at the target position meets the preset offset requirement, the first vector data and the second vector data at the target position are represented as true value data. Thus, the data to be marked is marked by the first vector data and the second vector data at the target position.

[0124] The preset range in the marking method A2 may refer to a range in the target projection space, which may be determined according to the target position, or the target projection space may be divided and determined in advance; the data volume of all abnormal vectors within the preset range where the target position is located may refer to the number / number of all abnormal vectors within the preset range where the target position is located, and the preset data volume may refer to the limit data of the number of vectors set in advance.

[0125] The overall offset in the marking method A2 may refer to the average offset and / or comprehensive offset of all abnormal vectors within the preset position range (the comprehensive offset refers to the comprehensive offset of all abnormal vectors). In an embodiment of the present application, when the number of abnormal vectors that do not meet the preset offset requirement within the preset range where the target position is located exceeds the preset data amount, the overall offset within the preset range is determined by calculating the average value or sum of the offsets of these abnormal vectors; if the overall offset meets the preset overall offset requirement, the vector data within the preset range can be determined to be true value data, that is, the data to be marked can be marked; if the overall offset does not meet the preset overall offset requirement, the data to be marked is marked according to the offset of each vector data within the preset range.

[0126] In the embodiment of the present application, the combination of the marking method A1 and the marking method A2 can be arbitrary, for example:

[0127] 1. You can first execute the marking method A1 and then execute the marking method A2, that is, first execute the scheme for marking the true value data according to the offset of the vector at the target position and the preset offset requirement, and then execute the scheme for marking the true value data according to the data amount of all vectors that do not meet the preset offset requirement within the preset position range of the target position and the preset data amount.

[0128] 2. You can first execute marking method A2 and then execute marking method A1, that is, first execute the plan for marking the true value data based on the data amount of all vectors that do not meet the preset offset requirements within the preset position range of the target position and the preset data amount, and then execute the plan for marking the true value data based on the offset of the vector at the target position and the preset offset requirements.

[0129] 3. The marking method A1 and the marking method A2 can be executed at the same time, that is, the scheme for marking the true value data is determined according to the data amount of all vectors that do not meet the preset offset requirements within the preset position range of the target position and the preset data amount, and the scheme for marking the true value data is determined according to the offset of the vector at the target position and the preset offset requirements.

[0130] 4. Only execute the marking method A2, that is, only execute the scheme for marking the true value data according to the data amount of all vectors that do not meet the preset offset requirements within the preset position range of the target position and the preset data amount.

[0131] 5. Only execute the marking method A1, that is, only execute the scheme of determining the true value data for marking according to the offset of the vector at the target position and the preset offset requirement.

[0132] See also Figure 4, a specific example of labeling the data to be labeled by combining the above-mentioned labeling method A1 and labeling method A2 is given:

[0133] S401, determining an offset of a vector at a target position according to the geometric accuracy of the vector at a target position in a target projection space based on the first vector data and the second vector data;

[0134] S402: if the offset of the vector at the target position meets the preset offset requirement, annotate the data to be annotated according to the vector at the target position;

[0135] S403, if the offset of the vector at the target position does not meet the preset offset requirement, determine the data amount of all vectors that do not meet the preset offset requirement within the preset range where the target position is located (i.e., determine the data amount of all abnormal vectors within the preset range where the target position is located);

[0136] S404, if the data amount of all vectors that do not meet the preset offset requirement within the preset position range where the target position is located is greater than the preset data amount (that is, the data amount of all abnormal vectors within the preset position range is greater than the preset data amount), determine the overall offset of the vectors within the preset position range according to the first vector data and the second vector data;

[0137] S405: label the data to be labeled according to the overall offset of the first vector data and the second vector data within a preset range.

[0138] Therefore, the data annotation method provided by the embodiment of the present application, in order to perform more detailed comparison and optimization in the local area and ensure the high accuracy and consistency of the vector data, when the offset of the vector at the target position does not meet the preset offset requirement, by determining the data amount of all vectors that do not meet the preset offset requirement within the preset range of the target position, and at the same time comparing the data amount of all vectors that do not meet the preset offset requirement within the preset position range of the target position with the preset data amount, the geometric accuracy of the two types of vector data can be concentratedly compared in a small range, thereby more accurately identifying and correcting local errors and improving the quality of the overall data. At the same time, it avoids misjudgment caused by errors in a single point or a few points, ensuring that the vector result finally output has high confidence and reliability.

[0139] Figure 5a A flowchart of another data annotation method provided for this application is shown below. Figure 5b A schematic diagram of preprocessing in another data annotation method provided in this application; Figure 5c A schematic diagram of the posture and positioning information corresponding to the sub-image in another data annotation method provided in this application; Figure 5d A schematic diagram of pose results in another data annotation method provided in this application;

[0140] Among them, Figure 5a As shown, another data annotation method may include:

[0141] S501, preprocessing: processing the data to be labeled into data that can be directly processed by the downstream.

[0142] Among them, since the data structures and data types of point clouds, images, trajectories, and sensor parameters uploaded by each customer's vehicle are different, it is necessary to store different types of data in a unified format folder for storage, such as:

[0143] For image files: the images need to be screened. The timestamp of the images needs to be greater than the smallest timestamp of the positioning file and less than the largest timestamp of the positioning file. The timestamps of the images need to be output in ascending order. This is used for the subsequent perception algorithm verification input and vector projection reference input.

[0144] For point cloud files: point clouds need to be screened and processed in a similar way to images for subsequent mapping, perception algorithms, and vector projection conversion inputs.

[0145] For calibration files: the camera external parameters need to be uniformly converted to the frd coordinate system for use in the subsequent projection filter vector;

[0146] For the engineering coordinate origin: the first mapping positioning point, used for subsequent mapping and vector coordinate conversion;

[0147] For IMU data: the data format is timestamp, acceleration, and angular velocity, which is used for subsequent mapping;

[0148] For positioning files: the quaternion corresponding to the timestamp, the enu coordinate component, the three velocities under the flu coordinate, and pry are used for subsequent mapping, alignment and repositioning.

[0149] For wheel speed files: the data format is timestamp and four wheel speeds (fl, fr, rl, rr).

[0150] In the embodiments of the present application, Figure 5b As shown, the preprocessed data to be labeled can be input into the preprocessing model to obtain the preprocessed data to be labeled.

[0151] S502, map building: map and splice the data to be labeled, and obtain the corresponding map building and relocation pose.

[0152] Among them, mapping the data to be labeled may include:

[0153] According to the preprocessing output results, the sub-image point cloud corresponding to the data to be annotated is mapped in the northeast celestial coordinate system with the first frame point cloud as the origin. After splicing, the corresponding complete spliced ​​point cloud and the pose and positioning information of each sub-image are obtained. Among them, the pose and positioning information structure corresponding to each sub-image is as follows Figure 5c shown.

[0154] Repositioning can mean: according to the result of the mapping output, the point cloud data of the data to be annotated is matched and repositioned with the Langge point cloud. If the point cloud data of the data to be annotated overlaps with the Langge point cloud at the edge, the sub-image pose at the overlap is optimized by multiple point clouds, and finally the corresponding repositioning pose result is obtained. Among them, the pose result can be as follows Figure 5d shown.

[0155] S503, vector production: pull or automatically produce corresponding vector results based on the data to be annotated.

[0156] Among them, HD vector query can refer to:

[0157] According to the precise range results output by preprocessing, mapping and repositioning, the corresponding range is used as a parameter to query the data of each layer in the high-precision map (HD). By judging the intersection principle of the corresponding range and the vector coordinates, the relevant HD vectors are extracted. The extracted vector layers include physical layers such as lane dividers, road surface line objects, rod-shaped objects, physical isolation objects, road surface objects, traffic lights, traffic signs, lane centerlines, lane attribute points, intersection functional surfaces, road surface objects and roads. Then, according to the extracted physical layer object ID, the corresponding virtual layer data is obtained, such as lane topology relationship, lane conventional speed limit table, lane construction table, element spatial relationship table, object-associated road table, object-associated lane sign, road line morphology table and road speed limit.

[0158] Automated vector perception + physical library query may refer to:

[0159] The output results of preprocessing, mapping and relocalization are imported into the point cloud + image perception algorithm model. Based on the comparison results of point cloud and image perception vectors, vector results of lane dividers, rod-shaped objects, road surface objects, traffic lights, traffic signs, physical isolation objects and other layers are output. These results are packaged as GeoJSON format data for storage.

[0160] S504, data specification conversion: the vectors within the required range are screened and matched, and the point cloud and image results are processed into the format agreed upon by the customer platform.

[0161] Among them, data specification conversion may refer to:

[0162] According to the above HD vector, the corresponding vector is projected onto the image. According to the proportion of the image projection, it is determined whether the vector and the image are completely aligned. The excess part needs to be interrupted and the point closest to the edge of the image is produced for attachment. After the vector processing is completed, the overall result needs to be output as a 3D point cloud projection result and a 2D image projection result for inspection and judgment in the subsequent visualization link.

[0163] S505, vector conversion: after comparing the vector results, obtain the labeled true value, and convert the corresponding data into the agreed format.

[0164] Among them, vector conversion can refer to:

[0165] In S504, the precise HD vector result after vector processing and the automatic perception vector result are projected and compared within the same position range. The comparison mainly includes two aspects:

[0166] a. Geometric accuracy comparison: Using the current HD vector record as a benchmark, compare the geometric accuracy of the perception vector at the same position. If the offset is small, it is marked as a high-confidence attribute for subsequent result output; otherwise, it is marked as a low-confidence vector and flows into the next processing link.

[0167] b. Regional overall comparison: When a sufficient number of low-confidence vectors appear in a small area, the HD vector results in the current small area are used as a benchmark to compare all perception vector results in the area. If the overall offset is low, the HD vector results of the entire area are uniformly marked as high-confidence vectors for subsequent output; otherwise, they are uniformly recorded as current problem labels.

[0168] After the vector result comparison stage, the overall vector result needs to be filtered to finally obtain an accurate vector result for true value output.

[0169] Therefore, in combination with the above content, the data annotation method provided in the embodiment of the present application can obtain the perception point cloud map of the data to be annotated and its vector matching relationship with the reference point cloud map by performing data preprocessing, mapping and repositioning on the data to be annotated, and then automatically perceive and generate the perception vector result, and finally cross-compare and verify the first vector data representing the perception vector result with the second vector data to obtain relatively accurate vector true value data. This process accelerates manual annotation and improves the annotation efficiency of the overall BEV true value data, thereby shortening the research and development cycle of the intelligent driving perception algorithm and improving the stable mass production capability of the algorithm.

[0170] Figure 6 A schematic diagram of the structure of the data annotation device provided in this application, such as Figure 6 As shown, the data labeling device 60 provided in this embodiment includes:

[0171] A construction module 601 is used to construct a perception point cloud map based on the data to be annotated;

[0172] An acquisition module 602 is used to acquire a reference point cloud image corresponding to the perception point cloud image;

[0173] The labeling module 603 is used to label the data to be labeled according to the position distribution characteristics of the first vector data in the perception point cloud image and the second vector data in the comparison point cloud image.

[0174] In a possible implementation, the acquisition module 602 may also be specifically configured to:

[0175] Relocating the perception point cloud image and the preset reference point cloud image, and determining the reference point cloud image corresponding to the perception point cloud image from the reference point cloud image; or,

[0176] A reference point cloud map is constructed based on the target point cloud data, and the collection range of the target point cloud data is the same as the collection range of the data to be annotated.

[0177] In a possible implementation, the acquisition module 602 may also be specifically configured to:

[0178] Relocate the perception point cloud image and the reference point cloud image, and determine the sub-image in the reference point cloud image that overlaps with the perception point cloud image;

[0179] The pose of the sub-image corresponding to the overlapping area is optimized to obtain a control point cloud image corresponding to the perceived point cloud image.

[0180] In a possible implementation, the marking module 603 may also be specifically used for:

[0181] The data to be labeled are labeled according to the projection offset between the first vector data and the second vector data in the target projection space.

[0182] In a possible implementation, the marking module 603 may also be specifically used for:

[0183] Determine the offset of the vector at the target position according to the geometric accuracy of the vector at the target position in the target projection space of the first vector data and the second vector data;

[0184] The data to be labeled is labeled according to the offset of the vector at the target position.

[0185] In a possible implementation, the marking module 603 may also be specifically used for:

[0186] If the offset of the vector at the target position meets the preset offset requirement, the data to be labeled is labeled according to the vector at the target position;

[0187] and / or,

[0188] If the data amount of all vectors that do not meet the preset offset requirement within the preset position range where the target position is located is greater than the preset data amount, determining the overall offset of the vectors within the preset position range according to the first vector data and the second vector data;

[0189] The data to be marked is marked according to the overall offset of the first vector data and the second vector data within a preset range.

[0190] In a possible implementation, the marking module 603 may also be specifically used for:

[0191] Determining target vector data in the second vector data according to a projection position of the second vector data in the target projection space, wherein the target vector data represents vector data in the second vector data that is not aligned with the target projection space;

[0192] According to the target projection space, the target vector data in the second vector data is aligned and adjusted to obtain adjusted second vector data;

[0193] The data to be labeled is labeled according to the projection offset of the first vector data and the adjusted second vector data in the target projection space.

[0194] The data labeling device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and this embodiment will not be described in detail here.

[0195] Figure 7 This is a schematic diagram of the structure of the data annotation device provided in this application. Figure 7 As shown, the data labeling device 70 provided in this embodiment includes: at least one processor 701 and a memory 702. Optionally, the device 70 also includes a communication component 703. The processor 701, the memory 702 and the communication component 703 are connected via a bus 704.

[0196] In a specific implementation process, at least one processor 701 executes the computer-executable instructions stored in the memory 702, so that at least one processor 701 executes the above method.

[0197] The specific implementation process of the processor 701 can be found in the above method embodiment, and its implementation principle and technical effect are similar, so this embodiment will not be repeated here.

[0198] In the above embodiments, it should be understood that the processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the invention can be directly implemented as a hardware processor, or can be implemented by a combination of hardware and software modules in the processor.

[0199] The memory may include a high-speed memory (Random Access Memory, RAM), and may also include a non-volatile memory (NVM), such as at least one disk storage.

[0200] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus in the drawings of this application is not limited to only one bus or one type of bus.

[0201] The present application also provides a computer program product, including a computer program, which implements the above method when executed by a processor.

[0202] The present application also provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the above method is implemented.

[0203] The above-mentioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The readable storage medium can be any available medium that can be accessed by a general or special-purpose computer.

[0204] An exemplary readable storage medium is coupled to a processor so that the processor can read information from the readable storage medium and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can be located in an application specific integrated circuit (Application Specific Integrated Circuits, referred to as: ASIC). Of course, the processor and the readable storage medium can also exist in the device as discrete components.

[0205] The division of units is only a logical function division, and there may be other divisions in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or unit, which can be electrical, mechanical or other forms.

[0206] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0207] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0208] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0209] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0210] Finally, it should be noted that those skilled in the art will readily conceive of other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses or adaptations of the present invention, which follow the general principles of the present invention and include common knowledge or customary technical means in the art not disclosed by the present invention, are not limited to the precise structure described above and shown in the drawings, and may be modified and changed in various ways without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. A data labeling method, characterized in that: include: Construct a perception point cloud map based on the data to be annotated; Relocating the perceived point cloud image and a preset reference point cloud image, and determining a reference point cloud image corresponding to the perceived point cloud image from the reference point cloud image; Annotating the data to be annotated according to a projection offset of the first vector data in the perception point cloud image and the second vector data in the reference point cloud image in a target projection space; The step of labeling the data to be labeled according to the projection offset of the first vector data in the perception point cloud image and the second vector data in the reference point cloud image in the target projection space includes: Determine the offset of the vector at the target position according to the geometric accuracy of the vector at the target position in the target projection space based on the first vector data and the second vector data; If the offset of the vector at the target position meets the preset offset requirement, the data to be labeled is labeled according to the vector at the target position; If the data amount of all vectors that do not meet the preset offset requirement within the preset position range of the target position is greater than the preset data amount, determine the overall offset of the vectors within the preset position range based on the first vector data and the second vector data; and mark the data to be marked based on the overall offset of the vectors within the preset position range based on the first vector data and the second vector data.

2. The method according to claim 1, characterized in that The repositioning of the perception point cloud image and the preset reference point cloud image, and determining a reference point cloud image corresponding to the perception point cloud image from the reference point cloud image, includes: Relocating the perception point cloud image and the reference point cloud image, and determining a sub-image in the reference point cloud image that overlaps with the perception point cloud image; The overlapping sub-images are subjected to pose optimization to obtain a control point cloud image corresponding to the perception point cloud image.

3. The method according to claim 1, characterized in that The data to be labeled includes an image and / or a point cloud; the target projection space includes a two-dimensional plane and / or a three-dimensional space determined based on the data to be labeled.

4. The method according to claim 1, characterized in that The labeling of the data to be labeled according to the projection offset of the first vector data in the perception point cloud image and the second vector data in the reference point cloud image in the target projection space comprises: Determine target vector data in the second vector data according to a projection position of the second vector data in the target projection space, wherein the target vector data represents vector data in the second vector data that is not aligned with the target projection space; According to the target projection space, aligning and adjusting the target vector data in the second vector data to obtain adjusted second vector data; The data to be labeled is labeled according to a projection offset of the first vector data and the adjusted second vector data in a target projection space.

5. A data annotation device, characterized in that: include: Memory, processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory, so that the processor performs the method according to any one of claims 1 to 4.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 4 when executed by a processor.

7. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 4 when being executed by a processor.

Images