Ground static element automatic labeling method and device and computer program product

Abstract

The invention provides a ground static element automatic labeling method and device and a computer program product. The method comprises the steps of obtaining a ground static element point cloud map of a target scene; segmenting the ground static element point cloud map to obtain a plurality of point cloud map labeling units; performing aerial view angle projection on the plurality of point cloud map labeling units to obtain a plurality of point cloud map aerial view images; the color of each pixel point in the aerial view image of the point cloud map is determined according to the corresponding point cloud intensity; segmenting the ground static elements in the aerial view images of the plurality of point cloud maps based on a preset image segmentation large model to obtain a plurality of segmentation masks; and for each segmentation mask, obtaining point cloud data of each ground static element according to the corresponding point cloud map labeling unit, and generating vectorization representation of each ground static element according to the point cloud data of each ground static element. According to the method, the accuracy and the robustness of automatic marking of the three-dimensional point cloud ground static elements are improved, and any manual marking is not needed.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a method and device for automatically labeling static ground elements, and a computer program product. Background Art

[0002] Traditional autonomous driving technology relies heavily on high-precision maps, which require pre-production and constant updating, resulting in high costs. Consequently, the current technological trend in the autonomous driving field is clearly leaning towards a "perception-heavy, map-light" approach. This approach uses a variety of sensors, including cameras, radar, and lidar, to enable autonomous vehicles to perceive their surroundings in real time, identifying roads, obstacles, traffic signs, pedestrians, and other vehicles. This reduces reliance on pre-produced high-precision maps, allowing autonomous vehicles to operate without them.

[0003] The "heavy perception, light map" solution places more stringent requirements on the perception algorithm of ground static elements (including but not limited to lane lines, curbs, zebra crossings, stop lines, and road signs, etc.). In order to meet the coverage and perception accuracy of a wider range of scenarios, the ground static element perception algorithm requires larger-scale and higher-precision perception truth annotation data as input to train the ground static element perception model. Traditional ground static element annotation mainly relies on manual annotation, which requires a lot of manpower and material resources, has a slow annotation speed, low production capacity, and is difficult to meet the needs of large-scale annotation. Due to the complexity and tediousness of the annotation work, a lot of management costs, time costs, and labor costs are also required. In response to this, some technicians have proposed an automatic annotation method for ground static elements in three-dimensional point cloud maps, but this method still has a problem: in the early stage, it is still necessary to use a large amount of manual annotation data of ground static elements to train and optimize the large point cloud automatic annotation model in order to obtain more accurate three-dimensional point cloud map annotation results.

[0004] In summary, although there are many methods and technologies, there is no method that can completely get rid of the dependence on a large amount of manually labeled data of static ground elements in three-dimensional point clouds. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and device for automatic labeling of ground static elements, and a computer program product to improve the accuracy and robustness of automatic labeling of ground static elements in three-dimensional point clouds without any manual labeling.

[0006] To achieve the above object, according to a first aspect of the present invention, a method for automatically labeling static elements on the ground is provided, the method comprising:

[0007] Obtain a point cloud map of static ground elements of the target scene;

[0008] Segmenting the ground static element point cloud map to obtain a plurality of point cloud map annotation units;

[0009] Projecting the multiple point cloud map annotation units from a bird's-eye view perspective to obtain multiple point cloud map bird's-eye view images; the color of each pixel in the point cloud map bird's-eye view image is determined according to its corresponding point cloud intensity;

[0010] Based on a preset image segmentation model, ground static elements in the multiple point cloud map bird's-eye view images are segmented to obtain multiple segmentation masks;

[0011] For each segmentation mask, point cloud data of each ground static element is obtained according to its corresponding point cloud map annotation unit, and a vectorized representation of each ground static element is generated according to the point cloud data of each ground static element.

[0012] According to a second aspect of the present invention, there is provided an automatic labeling device for static ground elements, comprising a module for executing the above method.

[0013] According to a third aspect of the present invention, there is provided an automatic labeling device for static elements on the ground, comprising:

[0014] Communication interface, used for communicating with other electronic devices;

[0015] a memory for storing computer program instructions;

[0016] A processor is configured to execute the computer program instructions to support the apparatus in implementing the method as described in the first aspect.

[0017] According to a fourth aspect of the present invention, a computer program product is provided, comprising computer program instructions, wherein the computer program instructions instruct a computer device to perform operations corresponding to the method according to the first aspect.

[0018] The present invention provides a method, device, and computer program product for automatically labeling static elements on the ground, which have the following beneficial effects:

[0019] A point cloud map of the ground static elements of the target scene is obtained through sensors such as lidar, and the obtained point cloud map data is segmented and processed to form multiple point cloud map annotation units, which decomposes the complex overall point cloud data into smaller units that are easier to process, and converts these point cloud map annotation units into a point cloud map bird's-eye view image. The ground static elements in the point cloud map bird's-eye view image are segmented using a preset image segmentation large model to obtain multiple segmentation masks. The image segmentation large model can identify different elements in the image. Finally, for each segmentation mask, the point cloud data of each ground static element is extracted, and a vectorized representation of each ground static element is generated based on these point cloud data. In summary, the present invention innovatively integrates advanced image segmentation large model technology to realize the automatic annotation of ground static elements in three-dimensional point clouds without any manual annotation, and improves the accuracy and robustness of the automatic annotation of ground static elements in three-dimensional point clouds, significantly improving the intelligence level and work efficiency of point cloud data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0021] Figure 1 The figure is a flow chart of a method for automatically labeling static elements on the ground according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The detailed description of the accompanying drawings is intended as an illustration of the current embodiment of the present invention and is not intended to represent the only form in which the present invention can be implemented. It should be understood that the same or equivalent functions can be accomplished by different embodiments that are intended to be included in the spirit and scope of the present invention.

[0023] See Figure 1 An embodiment of the present invention provides a method for automatically labeling static elements on the ground, the method comprising the following steps:

[0024] Step S10: Obtain a point cloud map of ground static elements of the target scene.

[0025] Specifically, this embodiment uses three-dimensional sensors such as lidar to scan the target scene and collect three-dimensional spatial data about static ground elements in the scene (such as lane markings, curbs, and zebra crossings). A point cloud map is a dataset containing a large number of three-dimensional coordinate points, each representing a specific location in the scene and containing additional information such as reflection intensity. The data collection process for the target scene is crowdsourced and consists of multiple trips, with different vehicles taking different routes to ensure comprehensive coverage of the target scene.

[0026] Step S20: segmenting the ground static element point cloud map to obtain a plurality of point cloud map annotation units.

[0027] Specifically, in this embodiment, the ground static element point cloud map is divided into multiple smaller, more manageable units, called point cloud map annotation units. The segmentation can be based on a variety of criteria, such as spatial area, point cloud density, or specific ground features, in order to improve the efficiency and accuracy of subsequent processing. For example, along the vehicle's trajectory, the vehicle's position P = {p1, p2, ..., pn} is sampled at fixed distance intervals (e.g., 50 meters). With the sampling position P = {p1, p2, ..., pn} as the center, multiple smaller point cloud maps are cropped from the ground static element point cloud map. The entire scene is then divided into a series of point cloud map annotation units.

[0028] Step S30: Project the multiple point cloud map annotation units from a bird's-eye view to obtain multiple point cloud map bird's-eye view images; the color of each pixel in the point cloud map bird's-eye view image is determined according to its corresponding point cloud intensity.

[0029] Specifically, during the conversion process, for each point in the point cloud map annotation unit, its three-dimensional coordinates (x, y, z) are projected onto a two-dimensional plane to obtain its coordinate position (x, y, 0) in the point cloud map bird's-eye view image, and combined with its point cloud intensity to obtain the final point cloud map bird's-eye view image. Each pixel point on the point cloud map bird's-eye view image can be expressed as (x, y, z, i), where x, y, z are the x-axis, y-axis and z-axis coordinates of each point in the point cloud map annotation unit, respectively. The point cloud map bird's-eye view image is a plane map, so z = 0, and i is the normalized value of the point cloud intensity (0 to 255), representing the color of the pixel point.

[0030] Step S40 : Based on a preset large image segmentation model, ground static elements in the plurality of point cloud map bird's-eye view images are segmented to obtain a plurality of segmentation masks.

[0031] Specifically, based on the multiple point cloud map bird's-eye view images, a preset image segmentation model is used to segment several static ground elements in the multiple point cloud map bird's-eye view images, obtaining corresponding multiple segmentation masks. The input of the image segmentation model is the point cloud map bird's-eye view image, and the output is a segmentation mask for each static ground element in the bird's-eye view image. The segmentation mask will mark different static ground elements with corresponding values, for example, background is 0, lane line is 1, and road sign is 2. To eliminate false detections caused by noise, this embodiment sets a confidence threshold of 0.7 to perform confidence filtering on the segmentation results.

[0032] It should be noted that the image segmentation large model is an open-source pre-trained image segmentation large model with strong generalization capabilities, eliminating the need for retraining using manually annotated data. In contrast, the generalization performance of point cloud segmentation models is often limited by numerous factors, such as differences in sensor installation locations and different LiDAR models. Due to the diversity of these specific conditions, a widely applicable general point cloud segmentation large model has not yet been established in academia or even industry. Therefore, training point cloud segmentation models for specific application scenarios still relies heavily on large-scale manually annotated datasets.

[0033] Step S50 : for each segmentation mask, obtaining point cloud data of each ground static element according to its corresponding point cloud map annotation unit, and generating a vectorized representation of each ground static element according to the point cloud data of each ground static element.

[0034] Specifically, the segmented ground static elements can be determined based on the segmentation mask, and the point cloud map annotation unit provides the point cloud data of the segmentation mask, so the point cloud data of these segmented ground static elements can be obtained based on the point cloud map annotation unit; based on the point cloud data of the ground static elements, a vectorized representation of each ground static element is further generated, which involves converting the point cloud data into a parametric description of geometric figures.

[0035] In summary, the method of this embodiment innovatively integrates advanced image segmentation large-scale model technology to realize the automatic labeling of static ground elements in three-dimensional point clouds. This unique combination abandons the traditional need for manual labeling of three-dimensional point cloud maps, indicating that in the absence of any pre-labeled three-dimensional point cloud data, it is still possible to efficiently and accurately complete the automatic identification and labeling of static ground elements, significantly improving the intelligence level and work efficiency of point cloud data processing.

[0036] In some embodiments, step S50 generates a vectorized representation of each ground static element based on the point cloud data of each ground static element, including:

[0037] For ground static elements of line element type, curve fitting is performed based on their point cloud data to obtain a cubic curve equation, a starting point and an end point. According to the cubic curve equation, the starting point and the end point, the point cloud data is downsampled to obtain multiple downsampling points, and the multiple downsampling points are connected in sequence to obtain a broken line for representing the ground static element.

[0038] Specifically, line elements mainly include curbs, lane lines, and parking lines. For ground static elements identified as line elements by the image segmentation model, the curve fitting technology is first used to process their point cloud data. The purpose of curve fitting is to find a mathematical model. In this embodiment, the cubic curve equation is used. The cubic curve equation is a parametric equation that can better fit complex curves because it can represent the starting point, end point, and curvature of the curve. After obtaining the cubic curve equation, it is necessary to determine the starting point and end point of the line element. These points are the actual starting and ending positions of the line element in the point cloud data.

[0039] Downsampling involves selecting a subset of points from the original point cloud data according to specific rules to reduce the data size while preserving the basic shape of the line elements. In this process, multiple downsampling points are selected from the point cloud data according to the cubic curve equation, with more points sampled in areas with greater curvature.

[0040] Finally, these downsampled points are connected in sequence to form a broken line. This broken line is the vectorized representation of the ground static element of the line element type. The advantages of vectorized representation are small data volume, easy calculation and processing, and the ability to clearly describe the shape and position of the line element.

[0041] In some embodiments, step S50 of generating a vectorized representation of each ground static element based on the point cloud data of each ground static element includes:

[0042] For a ground static element of the area element type, the boundary of its point cloud data is uniformly sampled to obtain a plurality of sampling points, and the plurality of sampling points are sequentially connected to obtain a polygon for representing the ground static element.

[0043] Specifically, area elements refer to static ground elements with closed shapes in autonomous driving scenarios, such as sidewalks, diversion areas, road signs, etc. These area elements are usually represented as a group of dense points in point cloud data, which together constitute a closed area.

[0044] For regional elements, the boundaries of their point cloud data are uniformly sampled. The purpose of uniform sampling is to select points at fixed intervals on the boundary to ensure that the sampling points can better represent the shape of the boundary. The number of sampling points selected depends on the complexity of the boundary and the required vectorization accuracy. The multiple points obtained by sampling are connected in sequence to form a polygon, which is the vectorized representation of the regional element.

[0045] Furthermore, regional elements of the road sign category can be represented by rectangular frames because they are usually regular shapes. Rectangular frames can concisely describe their position and size. Specifically, the rotating calcaneal method can be used to obtain the minimum area circumscribed rectangle. By rotating a rectangular frame with the minimum area, the convex hull boundary of the point cloud is gradually "squeezed" until the minimum area rectangle that can completely enclose the convex hull is found. During each rotation, the size and position of the rectangle will be adjusted according to the current contact point. For irregularly shaped regional elements (such as sidewalks, diversion areas, etc.), polygons are used to represent them. Polygons can better adapt to areas of different shapes.

[0046] In some embodiments, the method comprises:

[0047] Step S60: obtaining a unit vectorized map according to the vectorized representations of all ground static elements in the target scene.

[0048] Specifically, the vectorized representations of all ground static elements in the target scene (including line elements and polygon elements) are integrated to create a unit vectorized map. The unit vectorized map is a local, detailed map that contains information about all important ground static elements in the current scene, such as lane lines, sidewalks, parking areas, etc.

[0049] Step S70: Incrementally merge the unit vectorized map with the global vectorized map. When the overlap length of any polyline in the unit vectorized map and any polyline in the global vectorized map is greater than a preset length threshold, the two polylines are merged. When the intersection-and-union ratio between any polygon in the unit vectorized map and any polygon in the global vectorized map is greater than a preset intersection-and-union ratio threshold, the two polygons are merged.

[0050] Specifically, when the spatial overlap length of any polyline in the unit vectorized map and any polyline in the global vectorized map is greater than a preset length threshold (for example, 3 meters), the two polylines are considered to represent the same line. In this case, the two polylines are merged into one polyline to maintain the accuracy and consistency of the map.

[0051] When the Intersection over Union (IoU) between any polygon in the unit vectorized map and any polygon in the global vectorized map exceeds a preset IoU threshold (e.g., 0.5), the two polygons are considered to represent the same area. IoU is the ratio of the area of the intersection of two polygons to the area of their union, and it is a measure of the degree of overlap between two polygons. If this condition is met, the two polygons are merged into a single polygon to ensure the accuracy of regional elements in the global map.

[0052] Through these merging conditions, the global vectorized map can be updated efficiently while maintaining the integrity and accuracy of the map. This incremental merging method helps reduce redundant information, improves the efficiency of map updates, and can adapt to various changes encountered by autonomous vehicles during driving.

[0053] In some embodiments, step S10 includes:

[0054] Step S101, obtaining GPS positioning data, IMU data, and LiDAR point cloud data for multiple vehicle journeys;

[0055] Specifically, GPS positioning data provides the vehicle's position information on the earth's surface, IMU data is the inertial measurement unit (IMU) providing the vehicle's motion state information, including acceleration, angular velocity, etc., and lidar point cloud data is the three-dimensional point cloud data generated by the lidar scanning the surrounding environment, which contains detailed geometric information of the vehicle's surrounding environment.

[0056] Step S102: filtering the dynamic element point cloud data in the laser radar point cloud data to obtain the laser radar point cloud data of the ground static elements.

[0057] Specifically, in order to reduce computing resource consumption and eliminate noise points with low reflective energy in the distance, this embodiment first filters the point cloud with a radius greater than 100m. Furthermore, this embodiment focuses on the annotation of static elements on the ground. Since the movement of traffic participants (such as vehicles, pedestrians, cyclists, etc.) may cause noise points in the point cloud map, the point cloud data of these dynamic obstacles need to be filtered before mapping. For example, by using RANSAC for ground fitting, only the results of the ground fitting of each frame of point cloud are retained, thereby avoiding the influence of dynamic obstacles.

[0058] Step S103, performing dedistortion processing on the lidar point cloud data of the ground static elements according to the IMU data;

[0059] Specifically, the point cloud data frame obtained from each round of scanning summarizes a series of scan results at different time points within the rotation cycle. However, when the vehicle itself is in motion, the observation posture of the lidar will change in each rotation cycle, which directly makes it difficult to achieve natural alignment between the scan segments within the same frame, thereby causing motion distortion. To address this challenge, this embodiment uses the data provided by the pre-integration operation performed by the IMU to accurately infer the dynamic behavior of the vehicle body, and based on this, accurately calculate the motion distortion parameters, and then apply targeted correction measures to complete the dedistortion processing of the lidar point cloud data, thereby eliminating the distortion problem caused by motion.

[0060] Step S104 , constructing a point cloud map of each vehicle journey segment based on the GPS positioning data, IMU data, and lidar point cloud data of ground static elements of each vehicle journey segment.

[0061] Specifically, the process of constructing the point cloud map for each vehicle journey in this embodiment is as follows:

[0062] Initialization phase: The IMU (inertial measurement unit) provides the vehicle's acceleration and angular velocity data. Through pre-integration operations, a preliminary estimate of the vehicle's posture (position and direction) in each segment of the journey can be made. For example, assuming the vehicle starts moving, the IMU records the acceleration and rotation of the vehicle from stationary to moving. Through this data, the approximate motion trajectory and direction of the vehicle can be calculated.

[0063] LiDAR data frame matching: Using LiDAR scan data, by comparing the point clouds between consecutive data frames, the vehicle's motion can be more accurately estimated. For example, while a vehicle is driving, the LiDAR collects point cloud data of the surrounding environment multiple times per second. By comparing the same features (such as roadside trees or buildings) in consecutive frames, the vehicle's position changes can be more accurately determined.

[0064] Point cloud downsampling: To improve processing efficiency, the point cloud data is downsampled to reduce the data volume. For example, the voxel downsampling threshold is 0.1m, which divides the point cloud data into voxels (three-dimensional pixels) of 0.1m*0.1m*0.1m. Only one point is retained in each voxel.

[0065] Incorporating GPS positioning data: To accurately correct and prevent positioning drift over time, GPS positioning data is used to provide absolute pose measurements. If a vehicle relies solely on IMU and lidar data for positioning over a period of time, cumulative errors may occur. By regularly receiving GPS signals, these errors can be corrected to ensure accurate positioning of the vehicle.

[0066] A loop closure detection mechanism is introduced: when a vehicle returns to a previously passed location, this situation will be detected and identified, thereby correcting the positioning error accumulated during long-term operation. For example, if a vehicle drives on a circular road and eventually returns to the starting point, the loop closure detection can identify that the vehicle has returned to the same location and correct any positioning error accumulated during the driving process by comparing the data of the starting and end points.

[0067] Step S105 , aggregating the point cloud maps of multiple vehicle journeys to obtain the ground static element point cloud map.

[0068] Specifically, multiple local point cloud map data are finally aggregated to form a complete ground static element point cloud map, which contains the ground static element information in all vehicle trips.

[0069] In some embodiments, step S105 includes:

[0070] Step_a1, obtain the vehicle posture at each time stamp in each vehicle journey based on the IMU data of each vehicle journey;

[0071] Specifically, step_a1 uses IMU data to calculate the vehicle posture at each timestamp, including information such as position, direction, and speed.

[0072] Step_a2: Obtain the relationship between two vehicle postures corresponding to adjacent timestamps in the same vehicle trip based on the vehicle posture at each timestamp in each vehicle trip.

[0073] Specifically, step_a2 determines the change in vehicle posture between consecutive timestamps within the same trip, that is, how the vehicle moves from one posture to another.

[0074] Step_a3: Match the lidar point cloud data of ground static elements of different vehicle journeys to obtain point cloud frame matching pairs between different vehicle journeys, and obtain the relationship between the two vehicle postures corresponding to the point cloud frame matching pairs between different vehicle journeys.

[0075] Specifically, by trying to find similar point cloud data frames in different trips, these frames may contain the same ground static elements. Scan-Context can be used as the point cloud scene descriptor of ICP (Iterative Closest Point) to match these point cloud frames. Based on the two postures corresponding to the point cloud frame matching pairs, the vehicle posture relationship between different trips is determined, that is, how to align the map of one trip with the map of another trip.

[0076] step_a4, the relationship between two vehicle postures corresponding to adjacent timestamps within the same vehicle trip and the relationship between two vehicle postures corresponding to point cloud frame matching between different vehicle trips, aggregating the point cloud maps of multiple vehicle trips to obtain the ground static element point cloud map.

[0077] Specifically, the information obtained in the previous steps will be used to aggregate the point cloud maps of multiple trips. The pose relationships of adjacent timestamps within the same trip and the pose relationships of point cloud frame matches between different trips will be combined to create a continuous and consistent point cloud map of ground static elements. That is, the point cloud data will be converted from the local coordinate system of each trip to a unified coordinate system based on the vehicle's pose information to ensure the consistency and accuracy of the entire map.

[0078] In some embodiments, step S105 includes:

[0079] step_b1, takes the vehicle posture at each timestamp in each vehicle trip as a node;

[0080] Specifically, the vehicle posture at each timestamp in each vehicle trip is regarded as an independent node, and each node contains the vehicle's position, direction and other relevant information at that timestamp.

[0081] Step_b2: The relationship between two vehicle postures corresponding to adjacent timestamps within the same vehicle trip is taken as an edge within the same trip;

[0082] Specifically, within the same vehicle trip, the vehicle posture nodes at adjacent timestamps are connected through edges. These edges represent the changes in the vehicle posture from one timestamp to the next.

[0083] Step_b3: The relationship between the two vehicle postures corresponding to the point cloud frame matching pairs between different vehicle trips is used as the edge between different trips;

[0084] Specifically, matching pairs of point cloud frames between different vehicle trips are identified, and the vehicle poses in the matching pairs are connected by edges, which represent the association of vehicle poses between different trips.

[0085] step_b4, generating a posture graph based on the nodes, edges within the same trip, and edges between different trips;

[0086] Specifically, a pose graph is constructed using nodes and edges. A pose graph is a graph structure that represents the pose relationships of vehicles at different timestamps and different trips.

[0087] step_b5, optimizing the posture graph according to the constraints of the edges within the same trip and the edges between different trips;

[0088] Specifically, the optimization process considers edge constraints. Constraints for edges within the same trip and between different trips can be designed based on actual technical requirements. For example, edges within the same trip can include continuity constraints, as the vehicle postures at adjacent timestamps should be continuous, meaning the vehicle should not undergo sudden changes within a short period of time. For example, the vehicle's speed and acceleration should vary within a reasonable range, rather than suddenly increasing or decreasing from one timestamp to the next. If the vehicle's speed at timestamp t1 is v, then the speed at timestamp t2 (t2>t1) should be v plus a reasonable change determined by the acceleration a and the time period Δt, i.e., v+a*Δt. Constraints for edges between different trips can include spatial consistency constraints, meaning two matching postures should be spatially close enough to reflect different points in time at the same physical location. For example, a distance threshold can be set, and two postures are considered matched only when the Euclidean distance between them is less than this threshold.

[0089] Step_b6, aggregating the point cloud maps of multiple vehicle journeys according to the optimized posture graph to obtain the ground static element point cloud map;

[0090] Specifically, by applying optimized pose estimation, point cloud data from different journeys can be accurately fused into a unified coordinate system, thereby creating a complete and accurate point cloud map of static elements on the ground.

[0091] Another embodiment of the present invention provides an automatic labeling device for static ground elements, including a module for executing the method described in the above embodiment. The module can be implemented based on software, hardware, or a combination of software and hardware.

[0092] It should be noted that the device provided in this embodiment can be used to execute the method described in the above embodiment. Therefore, the content not described in detail in this embodiment can be obtained by referring to the content of the method of the above embodiment, so it will not be repeated here.

[0093] Another embodiment of the present invention provides an automatic labeling device for static elements on the ground, comprising:

[0094] Communication interface, used for communicating with other electronic devices;

[0095] a memory for storing computer program instructions;

[0096] The processor is configured to execute the computer program instructions to support the apparatus in implementing the method as described in the above embodiment.

[0097] In this embodiment, the memory mainly includes a program storage area and a data storage area, wherein the program storage area can store operating devices, at least one application required for a function, etc., and the data storage area can store related data, etc. In addition, the memory can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, and a flash card, etc., or the memory can also be other volatile solid-state memory devices.

[0098] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor, or the processor may be any conventional processor. The processor is the control center of the device and uses various interfaces and lines to connect the various parts of the device.

[0099] Another embodiment of the present invention provides a computer program product, including computer program instructions, wherein the computer program instructions instruct a computer device to perform operations corresponding to the method described in the above embodiment.

[0100] Specifically, the computer program product includes a series of computer program instructions. These computer program instructions are codes written in a computer program that define how to perform specific operations. These computer program instructions are designed to be loaded onto a computer device and instruct the device to perform specific operations, which are the steps in the method for automatically labeling static ground elements described in the above embodiment. In this way, the computer program product of this embodiment provides a complete software solution that can be run on various computer devices to implement the method for automatically labeling static ground elements described in the above embodiment.

[0101] While various embodiments of the present invention have been described above, the above descriptions are intended to be illustrative, non-exhaustive, and not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or technological improvements in the marketplace, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for automatically labeling static elements on the ground, characterized in that: The method comprises: Obtain a point cloud map of static ground elements of the target scene; Segmenting the ground static element point cloud map to obtain a plurality of point cloud map annotation units; Projecting the multiple point cloud map annotation units from a bird's-eye view perspective to obtain multiple point cloud map bird's-eye view images; the color of each pixel in the point cloud map bird's-eye view image is determined according to its corresponding point cloud intensity; Based on a preset image segmentation model, ground static elements in the multiple point cloud map bird's-eye view images are segmented to obtain multiple segmentation masks; For each segmentation mask, point cloud data of each ground static element is obtained according to its corresponding point cloud map annotation unit, and a vectorized representation of each ground static element is generated according to the point cloud data of each ground static element.

2. The method according to claim 1, characterized in that Generating a vectorized representation of each ground static element according to the point cloud data of each ground static element includes: For ground static elements of line element type, curve fitting is performed based on their point cloud data to obtain a cubic curve equation, a starting point and an end point. According to the cubic curve equation, the starting point and the end point, the point cloud data is downsampled to obtain multiple downsampling points, and the multiple downsampling points are connected in sequence to obtain a broken line for representing the ground static element.

3. The method according to claim 2, characterized in that Generating a vectorized representation of each ground static element according to the point cloud data of each ground static element includes: For a ground static element of the area element type, the boundary of its point cloud data is uniformly sampled to obtain a plurality of sampling points, and the plurality of sampling points are sequentially connected to obtain a polygon for representing the ground static element.

4. The method according to claim 3, characterized in that The method comprises: Obtaining a unit vectorized map according to vectorized representations of all ground static elements in the target scene; The unit vectorized map is incrementally merged with the global vectorized map; when the overlapping length of any polyline in the unit vectorized map and any polyline in the global vectorized map is greater than a preset length threshold, the two polylines are merged; when the intersection-and-union ratio between any polygon in the unit vectorized map and any polygon in the global vectorized map is greater than a preset intersection-and-union ratio threshold, the two polygons are merged.

5. The method according to claim 1, wherein The step of obtaining a ground static element point cloud map of a target scene includes: Obtain GPS positioning data, IMU data, and LiDAR point cloud data for multiple vehicle journeys; Filtering the dynamic element point cloud data in the laser radar point cloud data to obtain the laser radar point cloud data of the ground static elements; Dedistorting the laser radar point cloud data of the ground static elements according to the IMU data; Construct a point cloud map of each vehicle journey based on the GPS positioning data, IMU data, and lidar point cloud data of static elements on the ground. The point cloud maps of multiple vehicle journeys are aggregated to obtain the ground static element point cloud map.

6. The method according to claim 5, characterized in that The step of aggregating the point cloud maps of the plurality of vehicle journeys to obtain the ground static element point cloud map includes: The vehicle posture at each time stamp in each vehicle journey is obtained based on the IMU data of each vehicle journey; According to the vehicle posture at each time stamp in each vehicle trip, the relationship between two vehicle postures corresponding to adjacent time stamps in the same vehicle trip is obtained; Matching the lidar point cloud data of ground static elements of different vehicle journeys to obtain point cloud frame matching pairs between different vehicle journeys, and obtaining the relationship between the two vehicle postures corresponding to the point cloud frame matching pairs between different vehicle journeys; The relationship between two vehicle postures corresponding to adjacent timestamps within the same vehicle trip and the relationship between two vehicle postures corresponding to point cloud frame matching pairs between different vehicle trips are obtained by aggregating the point cloud maps of multiple vehicle trips to obtain the ground static element point cloud map.

7. The method according to claim 6, characterized in that The step of aggregating the point cloud maps of the plurality of vehicle journeys to obtain the ground static element point cloud map includes: The vehicle posture at each time stamp in each vehicle trip is taken as a node; The relationship between two vehicle postures corresponding to adjacent timestamps within the same vehicle trip is regarded as an edge within the same trip; The relationship between the two vehicle postures corresponding to the point cloud frame matching pairs between different vehicle trips is used as the edge between different trips; generating a posture graph based on the nodes, edges within the same trip, and edges between different trips; Optimizing the posture graph according to the constraints of the edges within the same trip and the edges between different trips; The point cloud maps of multiple vehicle journeys are aggregated according to the optimized posture graph to obtain the ground static element point cloud map.

8. A device for automatically labeling static elements on the ground, characterized in that: The method comprises a module for executing the method according to any one of claims 1 to 7.

9. A device for automatically labeling static elements on the ground, characterized in that: include: Communication interface, used for communicating with other electronic devices; a memory for storing computer program instructions; A processor, configured to execute the computer program instructions to enable the apparatus to implement the method according to any one of claims 1 to 7.

10. A computer program product, characterized in that The method comprises computer program instructions, wherein the computer program instructions instruct a computer device to perform operations corresponding to the method according to any one of claims 1 to 7.

Images