Lane-level map construction method and device, equipment and storage medium

Abstract

The invention provides a lane-level map construction method and device, equipment and a storage medium, and relates to the technical field of map navigation. According to the method, track data is used as a basis, through an automatic processing flow, lane-level map construction can be accurately carried out, and the accuracy, efficiency and real-time performance of lane-level map construction are greatly improved. By means of the image processing model, key areas such as intersections and forks in the map can be accurately recognized and extracted, manual intervention and errors are reduced, and the precision of the map is improved. Besides, various complex traffic scenes such as no-passing areas and complex intersections can be processed, map elements can be generated for different traffic scenes through sampling of track points and image processing, and adaptability and generalization of the construction method are improved.

Description

Technical Field

[0001] The present application relates to the field of map navigation technology, and in particular to a lane-level map construction method, apparatus, device, and storage medium. Background Art

[0002] Lane-level maps require detailed construction based on actual road information. Generally speaking, using positioning technology to record vehicle trajectory data is an effective method for cost-effective lane-level map construction and updates. However, the complex and ever-changing real-world road environments, coupled with diverse traffic regulations, present numerous challenges to lane-level map construction.

[0003] Specifically, roads often experience changes in road direction, intersection configuration, and impassable areas due to road construction, new road construction, and other factors. Lane-level maps must accurately reflect these real-time changes to ensure navigation accuracy and safety. Furthermore, the road network is dotted with various types of intersections, including level crossings, roundabouts, and T-junctions. These intersections exhibit significant differences in geometry and traffic regulations, placing higher demands on lane-level map construction. Furthermore, the inherent positioning accuracy of vehicle trajectory data, coupled with the potential for noise introduced by abnormal user driving behavior, further exacerbates the difficulty of constructing lane-level maps based on vehicle trajectory data. This results in poor accuracy when constructing lane-level maps based on vehicle trajectory data. Summary of the Invention

[0004] To address the above issues, this application provides a lane-level map construction method, which includes the following:

[0005] In a first aspect, the present application provides a lane-level map construction method, the method comprising: obtaining trajectory coordinate data, sampling the coordinate data as trajectory points;

[0006] Obtaining a trajectory grid image based on the trajectory points;

[0007] Inputting the trajectory grid image into a pre-trained image processing model to extract the target area;

[0008] Based on the extracted target area, a lane-level map is constructed.

[0009] In a second aspect, the present application provides a lane-level map construction device, which includes:

[0010] an acquisition unit, configured to acquire trajectory coordinate data and sample the coordinate data as trajectory points;

[0011] An image generating unit, configured to obtain a trajectory grid image based on the trajectory points;

[0012] an extraction unit, configured to input the trajectory grid image into a pre-trained image processing model to extract the target area;

[0013] The construction unit is used to construct a lane-level map based on the extracted target area.

[0014] In a third aspect, the present application provides a device comprising a memory and a processor, wherein the memory is used to store instructions or codes, and the processor is used to execute the instructions or codes so that the device executes the lane-level map construction method introduced in any implementation of the first aspect.

[0015] In a fourth aspect, the present application provides a computer-readable storage medium storing a code. When the code is executed, the device executing the code implements the lane-level map construction method described in any implementation of the first aspect.

[0016] The present application provides a lane-level map construction method. When executing the method, trajectory coordinate data is first obtained, and the coordinate data is sampled as trajectory points. By collecting real-time position information of the vehicle during driving, rich and dynamic traffic data can be obtained. This data has a wide range of sources and strong real-time performance, and can reflect the actual use of the road and changes in traffic flow. Sampling the vehicle's trajectory coordinate data into trajectory points can ensure the accuracy of the data. Then, based on the trajectory points, a trajectory raster image is obtained, which provides an intuitive and easy-to-process intermediate form for lane-level map construction. The trajectory raster image can clearly display the topological structure and spatial distribution of the road network, making complex trajectory data visual. In addition, the rasterization process can discretize the continuous trajectory data, which is convenient for subsequent image processing models to analyze and process.

[0017] The trajectory grid image is then input into a pre-trained image processing model to extract the target area. Finally, a lane-level map is constructed based on the extracted target area. Analyzing the trajectory grid image using a pre-trained image processing model can efficiently extract target areas, such as temporary diversion areas, intersection areas, crossroads, and other key areas. This deep learning-based image processing method has powerful feature recognition capabilities, can learn generalizable implicit features, and can adapt to different road scenarios and geometric forms. Compared with traditional methods, it can more accurately identify complex traffic scenarios, such as roundabouts and T-junctions, and can effectively handle noise and errors in trajectory data, improving the reliability and accuracy of the extraction results.

[0018] Finally, lane-level mapping is completed based on the extracted target areas. This process generates high-precision, high-resolution maps. Because the entire process is based on actual trajectory data, the constructed map truly reflects actual road usage and traffic flow characteristics.

[0019] In summary, the present invention achieves efficient and low-cost lane-level map construction through innovative data processing and modeling methods, while improving the accuracy and real-time performance of the map. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0021] Figure 1 A flowchart of a lane-level map construction method provided in an embodiment of the present application;

[0022] Figure 2 A schematic diagram of the structure of a target area provided in an embodiment of the present application;

[0023] Figure 3 A schematic diagram of image segmentation provided in an embodiment of the present application;

[0024] Figure 4 A schematic diagram of a process flow for constructing an image processing model provided in an embodiment of the present application;

[0025] Figure 5 A schematic structural diagram of a lane-level map construction device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0026] As described in the background technology of this application, when constructing lane-level maps, the massive amount of motor vehicle trajectory data recorded by positioning technologies such as satellite positioning, RTK (real-time kinematic measurement), and RTD (real-time differential positioning) can achieve low-cost construction and efficient updating of lane-level maps such as HD high-precision maps and HQ high-quality maps. However, in the process of processing trajectory data and constructing lane-level maps, the processing of specific areas of the road is particularly critical and complex. For example, the intersection area is an area where trajectories in different directions intersect, and its geometric form is diverse, including level crossings, roundabouts, T-junctions, etc., and the traffic rules and trajectory distribution are complex. In addition, the positioning accuracy errors in the trajectory data itself and the noise introduced by abnormal driving behavior of users further increase the difficulty of data processing. Therefore, if the scope of a specific area can be effectively identified, it will provide an important reference for subsequent road scene modeling, realize the automatic recognition and segmentation of trajectory scenes, and thus assist in the construction of lane-level maps based on trajectory data.

[0027] In order to solve the above technical problems, the embodiment of the present application provides a lane-level map construction method, which uses trajectory data as a basis and can accurately construct maps through an automated processing process, greatly improving the accuracy of lane-level map construction. With the help of an image processing model, key areas in the map, such as impassable roads and intersections, can be accurately identified and extracted, reducing manual intervention and errors and improving the accuracy of the map. In addition, the present invention can handle various complex traffic scenarios, such as construction sections and complex intersections. Through the sampling and image processing of trajectory points, map elements can be generated for different traffic scenarios, improving the adaptability and generalization of the construction method.

[0028] In order to make the purpose, technical solutions and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0029] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions.

[0030] Figure 1 A flow chart of a lane-level map construction method provided in an embodiment of the present application. Figure 1As shown, the lane-level map construction method provided in the embodiment of the present application may include:

[0031] S101 , obtaining trajectory coordinate data, sampling the trajectory coordinate data, and obtaining trajectory points.

[0032] Track coordinate data, such as the vehicle's driving track coordinate data, is the location information of the vehicle during travel recorded by satellite positioning such as GPS, or high-precision positioning technology such as RTK and RTD. It is presented as a series of coordinate points, usually in the form of longitude and latitude, reflecting the vehicle's movement path in time and space. In addition, the track coordinate data may also include track coordinate data such as cycling, walking or running. In the introduction of the embodiments of this application, the vehicle's driving track data is used as an example for explanation.

[0033] The vehicle trajectory coordinate data obtained in this application is trajectory data acquired within a preset time range in a preset area, serving as the original vehicle trajectory data. The original vehicle trajectory data is then uniformly sampled at equal time intervals, converting the sampled data into trajectory points. Specifically, trajectory points are extracted from the original vehicle trajectory data at fixed time intervals, such as every 1 second or every 2 seconds, for sampling. This ensures data uniformity and facilitates subsequent analysis.

[0034] In this way, by sampling the vehicle's real trajectory data as trajectory points and performing subsequent processing based on the trajectory points, it is helpful to distinguish the stop line at the intersection and the normal traffic area, so that the location of the target area can be identified more accurately.

[0035] S102: Obtain a trajectory grid image based on the trajectory points.

[0036] The trajectory grid image generated based on trajectory points is a technology for visualizing traffic data. It maps vehicle trajectory points onto a two-dimensional grid, or raster, and displays traffic flow, density, and distribution in the form of an image. This trajectory grid image can present traffic data in an intuitive visual manner, quickly helping traffic planners and researchers understand the spatial distribution and density changes of traffic flow. For example, by indicating the size of traffic flow through color depth or grayscale levels, high-traffic areas can be clearly identified. The following is a detailed description of the specific implementation process of obtaining a trajectory grid image based on trajectory points in an embodiment of the present application.

[0037] After the sampling in step S101 is completed, the trajectory points corresponding to the driving trajectory coordinate data in the preset area are obtained, and then the preset area is divided into several geographic coordinate grids, where each grid is a two-dimensional space unit, and its boundaries are usually defined by the latitude and longitude range. For example, an urban area can be divided into grids of 100 meters × 100 meters. In order to ensure the integrity of the trajectory data in the target area such as the intersection, there will be a certain overlap between adjacent grids. This can avoid the trajectory being truncated at the intersection due to grid division, thereby fully representing the behavior of the vehicle passing through the intersection. The sampled trajectory points are assigned to the corresponding grids, and each trajectory point is classified into one or more grids according to its coordinate position. When the trajectory point is located in the overlapping area, it is classified into multiple grids.

[0038] Due to the presence of three-dimensional traffic structures such as overpasses on actual roads, vehicle trajectories on upper and lower levels may overlap or become confused, thus affecting the accuracy of data analysis. To avoid this problem, the present embodiment processes trajectory data using an elevation layering method. Specifically, by performing elevation layering on the trajectories within the grids in the partitioning results, the trajectory point data of each layer is projected onto the ground and converted into image pixel coordinates.

[0039] Elevation stratification refers to the division of trajectory data into different layers based on the vehicle's elevation information. Specifically, vehicle trajectory data contains three-dimensional coordinates: longitude, latitude, and elevation. Elevation information can be obtained using on-board positioning equipment. Layers can then be divided by setting different elevation thresholds based on the structural characteristics of the overpass. For example, the upper and lower levels of an overpass can be defined as different elevation intervals. Assuming the elevation range of the upper level of the overpass is 10 to 20 meters, and the elevation range of the lower level is 0 to 5 meters, the trajectory data can be divided into two layers based on these thresholds. Each trajectory point is assigned to the corresponding layer based on its elevation value. In this way, the trajectory data of the upper and lower layers are separated, avoiding mutual interference.

[0040] After completing the elevation stratification, the trajectory data for each layer is further processed, projecting the trajectory points from three-dimensional space onto a two-dimensional ground plane. This strips away the elevation information, retaining only the horizontal trajectory distribution. The projected two-dimensional trajectory point coordinates are then converted to image pixel coordinates. Specifically, this is done by mapping the original geographic coordinates (latitude and longitude) to image pixel coordinates, allowing for the subsequent generation of a trajectory raster image.

[0041] By layering the elevation data, trajectory data for each layer is processed independently, preventing confusion between upper and lower layers. Because traffic flows on the upper and lower levels of the overpass are clearly separated, vehicle trajectories on the upper layer do not overlap with those on the lower layer. For example, expressway trajectories on the upper layer and ground road trajectories on the lower layer are processed separately, without interfering with each other. When generating trajectory raster images, trajectory data for each layer can be displayed independently, providing a clearer picture of traffic conditions.

[0042] The above describes the processing process after obtaining the trajectory points. Through the above processing, the original geographic coordinates of the trajectory points are mapped to the image pixel coordinates, which facilitates the rendering of the trajectory points into a trajectory raster image according to the density of the trajectory points on each layer of the image. In the embodiments of the present application, the process of rendering the trajectory points into the trajectory raster image includes multiple implementation methods. The following describes the process of rendering the trajectory points into the trajectory raster image in the present application through a specific embodiment:

[0043] In an implementation process of an embodiment of the present application, rendering the trajectory points into a trajectory grid image includes:

[0044] The area within the grid is divided into grids of preset resolution; the number of trajectory points in each grid unit is counted to generate a density distribution matrix; and a trajectory grid image is generated based on the density distribution matrix.

[0045] The grid area refers to the area where the trajectory points of each layer are located after the elevation layering process. For example, the upper and lower layers of an overpass are processed as independent grid areas. The preset resolution refers to the size of the grid, usually in meters, such as 10 meters × 10 meters. The choice of resolution depends on the accuracy requirements and computing resources of the analysis. The target area is then divided into multiple small grid cells based on the preset resolution, each of which is called a grid. The size and shape of the grid are usually regular, such as square or rectangular. For each grid cell, the number of trajectory points contained in it is counted. The number of trajectory points reflects the traffic flow or activity density in the area. The number of trajectory points in each grid cell is recorded to form a matrix. The rows and columns of the matrix correspond to the rows and columns of the grid, respectively, and the value of each element in the matrix represents the number of trajectory points in the corresponding grid cell. Then, based on the density distribution matrix, a trajectory grid image is generated. The final image intuitively shows the density distribution of trajectory points in the target area. The density distribution of trajectory points can clearly identify areas with high and low traffic flow, and further identify the target area. For example, in a grid, areas with dense trajectory points are drivable roads, while areas with fewer or no trajectory points can be considered intersections.

[0046] The trajectory raster images generated through the above process provide an intuitive visual representation of the spatial distribution of traffic flow, enabling traffic planners and researchers to quickly identify hotspots and low-flow areas, thereby better understanding target areas on the road. Furthermore, trajectory raster images provide an important basis for traffic planning. For example, by analyzing density distribution, the locations of intersections and special road sections can be determined, thereby improving the accuracy of map creation.

[0047] In an implementation process of an embodiment of the present application, rendering the trajectory points into a trajectory grid image includes:

[0048] The area within the grid is divided into grids of a preset resolution; the number of trajectory points in each grid cell is counted to generate a density distribution matrix; a histogram equalization algorithm is applied to the density distribution matrix to enhance the visual recognition of image details by expanding the dynamic distribution range of density values; and a trajectory grid image is generated based on the processed density distribution matrix.

[0049] In this implementation process, the process of rendering trajectory points into trajectory grid images not only includes basic grid division and density statistics, but also introduces a histogram equalization algorithm to optimize the visual effect of the image. Specifically, the size of the grid is preset, such as 0.1mX0.1m, 0.5mX0.5m, etc. Then the area within the grid is divided into grids of preset sizes, where each grid cell is of the same size for subsequent trajectory point statistics. Then the number of trajectory points in each grid cell is counted, where the number of trajectory points reflects the traffic activity density in the area. The statistical results are stored in the form of a matrix, where the rows and columns of the matrix correspond to the layout of the grid, and each element value in the matrix represents the number of trajectory points in the corresponding grid cell. After obtaining the density matrix, the histogram equalization algorithm is applied to the density distribution matrix, where histogram equalization is an image processing technology used to adjust the contrast of the image and enhance the details and visual recognition of the image by expanding the dynamic range of the pixel values ​​of the image. In a trajectory raster image, the density distribution matrix values ​​may be concentrated in certain intervals. For example, most grid cells may have a small number of trajectory points, while a few areas may have a large number of trajectory points. Using a histogram equalization algorithm, the density values ​​are redistributed to achieve a more uniform grayscale or color distribution in the image, thereby highlighting details in low-density areas while avoiding oversaturation in high-density areas. The resulting image intuitively displays the density distribution of trajectory points within the target area. Histogram equalization results in clearer image details, making them easier to observe and analyze.

[0050] In the process of rendering the trajectory points into a trajectory raster image, a histogram equalization algorithm is used to expand the dynamic range of density values, making details in low-density areas more prominent and avoiding oversaturation in high-density areas, thereby improving the overall visual recognition of the image. In other words, by introducing the histogram equalization algorithm, the embodiments of the present application not only achieve raster visualization of the trajectory points, but also significantly enhance the visual effect and analytical value of the image.

[0051] In an implementation process of an embodiment of the present application, rendering the trajectory points into a trajectory grid image includes:

[0052] The area within the grid is divided into grids of a preset resolution; the number of trajectory points in each grid cell is counted to generate a density distribution matrix; a three-channel color image is generated based on the density distribution matrix to obtain a trajectory grid image, in which different colors are used to thermally represent the density of trajectory points.

[0053] In the embodiment of the present application, the process of rendering the trajectory points into the trajectory grid image enhances the visualization effect by introducing a three-channel color image, namely an RGB color image. The specific implementation process is as follows:

[0054] The grid area is divided into a grid of a preset resolution. Each grid area is then divided into multiple small grid cells of a preset grid size, each of which is of uniform size. These cells are used for subsequent trajectory point statistics. The number of trajectory points within each grid cell is counted. The number of trajectory points reflects the density of traffic activity in the area. The statistical results are stored in a matrix format, with the rows and columns of the matrix corresponding to the grid layout, and each element in the matrix represents the number of trajectory points within the corresponding grid cell. A three-channel color image is then generated based on the density distribution matrix. The three-channel color image consists of three channels: red (R), green (G), and blue (B), with pixel values ​​in each channel ranging from 0 to 255. By adjusting the values ​​of the three channels, a rich color palette can be generated. Using different colors to represent trajectory point density can be achieved: cold tones such as blue or green represent low-density areas; warm tones such as yellow or red represent high-density areas; and intermediate tones such as cyan or orange represent intermediate-density areas. This color gradient visually demonstrates the changes in trajectory point density. The resulting three-channel color image clearly displays the density distribution of trajectory points. Through color changes, high-density areas such as stop lines at intersections, low-density areas such as side roads and side roads, and intermediate-density areas such as expressways can be clearly identified. The color heat map provides a more intuitive view of trajectory point density, facilitating the rapid identification of hotspots and low-traffic areas. Compared to grayscale images, three-channel color images provide richer visual information, making the contrast between low- and high-density areas more pronounced.

[0055] S103: Input the trajectory grid image into a pre-trained image processing model to extract the target area.

[0056] The pre-trained image processing model is an image detection and segmentation deep neural network model, which is used to detect and / or segment the image to complete the extraction of the target area. In one implementation of the embodiment of the present application, the target area includes: one or more of intersections, road intersections, forks, road sections and impassable areas. Image detection refers to the recognition of specific objects or features in the image. In this application, it is the recognition of the target area, for example, the recognition of any one or more of intersections, road intersections, forks, road sections and impassable areas. Figure 2 As shown, Figure 2 A schematic diagram of the target area structure provided in an embodiment of the present application. This includes road section areas 1-8, intersection areas, and fork areas. An intersection refers to the intersection of multiple roads with different travel directions, including level crossings, T-junctions, and roundabouts. A fork refers to the intersection of multiple roads with the same travel direction, including ramp entrances and exits, overpass forks, and early right-turn forks. A road section refers to the road area excluding intersections and forks. The smallest granularity is an independent road section area containing only a single travel direction.

[0057] Image segmentation involves dividing an image into multiple arbitrarily shaped regions, each corresponding to a different object or background. For example, it can segment the geometric shapes of identified intersections, road junctions, forks, road sections, and impassable areas.

[0058] The above-mentioned detection model can adopt common target detection models such as YOLO, SSD, DETR or DINO, and the segmentation model can adopt segmentation models such as YOLOCAT, Mask RCNN, Mash DINO, or One Former, or a model that can realize detection and segmentation at the same time. The specific type of the model is not limited in the embodiment of the present application. Directly locate the track surface with a rectangular frame or use precise polygonal contours to locate, and select the appropriate detection or segmentation method according to the actual track point situation. For example, when the track surface is close to a rectangle or a regular shape, rectangular frame positioning can meet the needs; if the track surface is complex in shape, such as a road intersection, roundabout, etc., precise polygonal contour positioning can provide more accurate results. As Figure 3 As shown, Figure 3 A schematic diagram of image segmentation is provided in an embodiment of the present application. The rectangular frame positioning method is adopted in the figure, where the left side is the trajectory grid image, the right side is the corresponding identified intersection area, the solid line is the image segmentation area, and the dotted line is the image detection frame.

[0059] S104: Construct a lane-level map based on the extracted target area.

[0060] After extracting the target area from the trajectory raster image using a pre-trained image processing model, a map can be constructed based on the extracted target area. The process from target area extraction to lane-level map construction includes three-dimensional mapping, Z-value processing, vector data generation of the target area, and lane-level map construction. Through these steps, high-precision, dynamically updated maps can be generated, providing strong support for applications such as autonomous driving, intelligent transportation, and urban planning. This process includes mapping the extracted target area into three-dimensional space; performing Z-value post-processing on the mapped three-dimensional spatial data; generating vector data of the target area based on the processed data, and then constructing the map based on this vector data of the target area. The following describes this process in detail.

[0061] After extracting the target area from the trajectory raster image, such as a road section area, a fork area, or a junction area, these two-dimensional target areas are further mapped into three-dimensional space. The mapping process includes: converting the two-dimensional coordinates of the target area, such as longitude and latitude, into coordinates in three-dimensional space, including three dimensions: X, Y, and Z. Assigning a Z value, i.e., elevation information, to each target area. Elevation data can be obtained by directly using the original elevation information in the trajectory before Z value stratification, then fitting the two-dimensional shape to the corresponding original trajectory and assigning the value using the elevation of the original trajectory. Using terrain elevation data, such as a DEM digital elevation model, the two-dimensional shape and elevation information of the target area are combined to generate a three-dimensional model. In addition, the mapped three-dimensional spatial data needs to be post-processed with Z values ​​to ensure the accuracy and consistency of the data. Finally, vector data of the target area is generated based on the processed data. The vector data of the target area refers to the vector information describing the shape and direction of the target area. When the target area is an intersection, the vector data for the target area is intersection face vector data, which is vector information of the intersection's shape and is used to represent the intersection's geometric features and topological structure. Intersection face vector data enables the construction of detailed maps. Thus, through 3D mapping and Z-value processing, the vector data generated for the target area accurately describes the intersection's geometry and topological structure, providing the foundation for the construction of high-precision lane-level maps.

[0062] The above describes a lane-level map construction method provided by the present application. In this method, the vehicle's driving trajectory is converted into a trajectory raster image, and then the trajectory raster image is input into a pre-trained image processing model to recognize the target area through the pre-trained image processing model, and then the map is constructed based on the recognition result. In this way, the recognition result of the target area is more accurate, thereby improving the accuracy of the lane-level map construction result. The following is an introduction to the construction process of the image processing model in the present application in conjunction with another embodiment. Figure 4 As shown, Figure 4 A schematic diagram of a process flow for constructing an image processing model provided in an embodiment of the present application, wherein the construction process includes:

[0063] S401: Obtain vector data of the target area in a geographic coordinate system and trajectory sample data.

[0064] The vector data of the target area in the geographic coordinate system refers to the vector data used to accurately represent the position, shape and attributes of the target area in the geographic coordinate system, wherein the target area can be any of an intersection, an intersection in a road, a fork, a road section and an impassable area. The functions of the model finally obtained by training with different training data are different. For example, by training the model with the vector data of the intersection in the geographic coordinate system, a model for detecting and segmenting the intersection can be obtained. The vector data of the target area can be obtained through manual three-dimensional annotation or from the high-precision map data that has been produced.

[0065] In the embodiments of the present application, the trajectory sample data includes trajectory data of behaviors such as walking, cycling, and driving. In the embodiments of the present application, the vehicle's driving trajectory data is used as an example for explanation. The vehicle's driving trajectory sample data is the trajectory data of the vehicle when it is driving at an intersection, recording the vehicle's driving path, speed, and other information. This data is usually obtained through GPS or other positioning technologies.

[0066] The vector data of the target area in the geographic coordinate system and the vehicle's driving trajectory sample data are the basis for model training, providing intersection geometry information and vehicle behavior information in real scenes.

[0067] S402: Divide the vector data of the target area in the geographic coordinate system and the trajectory sample data into multiple sub-areas in the plane, and perform layering in the vertical direction.

[0068] The intersection surface true value vector data and driving trajectory sample data in the geographic coordinate system are divided into multiple sub-areas to facilitate subsequent processing by breaking down complex scenes into smaller units. In addition, the data needs to be layered in the vertical direction. For example, the upper and lower layers of the overpass are processed separately to avoid interference between trajectory data at different heights. In this way, the efficiency and accuracy of data processing can be improved by dividing the sub-areas and layering the processing, while avoiding data confusion in complex scenes. The specific division process is the same as the division process in step S102 above and will not be repeated here.

[0069] S403: Project the divided vector data of the target area in the geographic coordinate system onto a two-dimensional plane to generate two-dimensional pixel coordinates of the target area.

[0070] The target area's geographic coordinates (latitude and longitude) are converted to pixel coordinates to project the target area's vector data in 3D space onto a 2D plane. This involves projecting the target area's true vector data onto a 2D plane to generate the target area's 2D pixel coordinates. This conversion of 3D data into 2D pixel coordinates facilitates matching and processing with 2D image data.

[0071] S404 , generating a trajectory grid image based on the divided trajectory sample data projection.

[0072] The segmented vehicle trajectory sample data is projected to generate a trajectory grid image. This is a two-dimensional grid image, where each grid cell represents the density or other characteristics of a trajectory point. The generated trajectory grid image can be directly used to train image processing models, providing a visual representation of vehicle trajectories.

[0073] S405: Generate annotated true value training data based on the two-dimensional pixel coordinates of the target area and the trajectory grid image.

[0074] The 2D pixel coordinates of the target area are combined with the trajectory raster image to annotate the trajectory raster image. For example, the annotated data can indicate which areas in the trajectory raster image are intersections and which are road sections. This produces an annotated trajectory raster image, which serves as annotated ground-truth training data. This annotated training data is the foundation for deep learning model training, guiding the model in learning how to identify and segment target areas from input images.

[0075] To generate annotated ground-truth training data, the following operations are performed based on the 2D pixel coordinates of the target area and the trajectory raster image:

[0076] Using data reflow or manual annotation, the location and geometry of target areas, such as intersections, within massive amounts of trajectory raster map data are marked. The data reflow method utilizes the target area vector data from the already constructed high-precision map and projects this vector data onto the corresponding trajectory raster image coordinates. Manual annotation involves using annotation software to manually draw the approximate location of the intersection.

[0077] S406: Clean the labeled true value training data.

[0078] Data cleaning involves removing noise, outliers, and duplicate data from training data. For example, this involves removing incorrectly labeled images or incomplete trajectory data. Cleaned data is more accurate and helps improve model training results.

[0079] S407: Use the cleaned training data to train an image detection and segmentation deep neural network model.

[0080] When using cleaned training data to train a deep neural network model for image detection and segmentation, you first need to select an appropriate deep learning model. Then, use the cleaned, labeled training data to train the model, adjusting the model parameters to minimize the loss function and improve the model's recognition and segmentation accuracy. The trained model can automatically identify and segment target areas such as intersections from the input image, supporting the subsequent construction of lane-level maps.

[0081] The above are some specific implementations of a lane-level map construction method provided in the embodiment of the present application. Based on this, the present application also provides a corresponding device. The device provided in the embodiment of the present application will be introduced from the perspective of functional modularization.

[0082] Figure 5 A schematic diagram of the structure of a lane-level map construction device provided in an embodiment of the present application. Figure 5 As shown, the lane-level map construction device 500 provided in the embodiment of the present application includes:

[0083] An acquisition unit 510 is configured to acquire trajectory coordinate data and sample the trajectory coordinate data to obtain trajectory points;

[0084] An image generating unit 520 is configured to obtain a trajectory grid image based on the trajectory points;

[0085] An extraction unit 530 is used to input the trajectory grid image into a pre-trained image processing model to extract the target area;

[0086] The construction unit 540 is used to construct a lane-level map according to the extracted target area.

[0087] In one implementation of the embodiment of the present application, the image generating unit obtaining the trajectory grid image based on the trajectory points includes:

[0088] Dividing the trajectory points into a geographic coordinate grid;

[0089] The trajectories in the grid of the division result are layered by elevation, the trajectory point data of each layer are projected on the ground, and converted into image pixel coordinates;

[0090] The trajectory points are rendered into a trajectory grid image according to the density of the trajectory points on each layer of the image.

[0091] In one implementation of the embodiment of the present application, rendering the trajectory points into a trajectory grid image includes:

[0092] Divide each grid area into grids of preset resolution;

[0093] Count the number of trajectory points in each grid cell and generate a density distribution matrix;

[0094] A trajectory grid image is generated based on the density distribution matrix.

[0095] In one implementation of the embodiment of the present application, rendering the trajectory points into a trajectory grid image includes:

[0096] Divide the area within the grid into grids of preset resolution;

[0097] Count the number of trajectory points in each grid cell and generate a density distribution matrix;

[0098] Applying a histogram equalization algorithm to the density distribution matrix to enhance the visual recognition of image details by expanding the dynamic distribution range of density values;

[0099] Generate a trajectory raster image based on the processed density distribution matrix.

[0100] In one implementation of the embodiment of the present application, rendering the trajectory points into a trajectory grid image includes:

[0101] Divide the area within the grid into grids of preset resolution;

[0102] Count the number of trajectory points in each grid cell and generate a density distribution matrix;

[0103] A three-channel color image is generated based on the density distribution matrix to obtain a trajectory grid image, in which different colors are used to thermally represent the density of trajectory points.

[0104] In one implementation of the embodiment of the present application, the construction unit is specifically configured to map the extracted target area into a three-dimensional space;

[0105] Perform Z value post-processing on the mapped three-dimensional spatial data;

[0106] Vector data of the target area is generated according to the processed data, and a map is constructed based on the vector data of the target area.

[0107] In one implementation of the embodiment of the present application, the target area includes:

[0108] One or more of an intersection, a road junction, a fork, a road section, and an impassable area.

[0109] In one implementation of the embodiment of the present application, the training process of the image processing model includes:

[0110] Obtain vector data of the target area in the geographic coordinate system and trajectory sample data;

[0111] Dividing the vector data of the target area in the geographic coordinate system and the trajectory sample data into a plurality of sub-areas in a plane and layering them in a vertical direction;

[0112] Projecting the divided vector data of the target area in the geographic coordinate system onto a two-dimensional plane to generate two-dimensional pixel coordinates of the target area;

[0113] Generate a trajectory raster image based on the projection of the divided trajectory sample data;

[0114] Generate annotated ground-truth training data based on the two-dimensional pixel coordinates of the target area and the trajectory raster image;

[0115] Cleaning the annotated true value training data;

[0116] Use the cleaned training data to train a deep neural network model for image detection and segmentation.

[0117] The embodiments of the present application also provide corresponding devices and computer storage media for implementing the solutions provided by the embodiments of the present application.

[0118] The device includes a memory and a processor, the memory is used to store instructions or codes, and the processor is used to execute the instructions or codes so that the device executes the method described in any embodiment of the present application.

[0119] The computer storage medium stores code, and when the code is executed, the device executing the code implements the method described in any embodiment of the present application.

[0120] Through the description of the above implementation methods, it can be known that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment methods can be implemented by means of software plus a general hardware platform. Based on this understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a storage medium, such as a read-only memory (ROM) / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network communication device such as a router) to execute the methods described in each embodiment of the present application or certain parts of the embodiments.

[0121] It is understandable that in the specific implementation of this application, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved, when the above embodiments of this application are applied to specific products or technologies, need to obtain user permission or consent, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of relevant countries and regions.

[0122] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0123] It should also be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and apparatus embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments. The device and apparatus embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components indicated as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0124] The above is merely one specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A lane-level map construction method, characterized in that: The method comprises: Acquiring trajectory coordinate data, sampling the trajectory coordinate data, and obtaining trajectory points; Obtaining a trajectory grid image based on the trajectory points; Inputting the trajectory grid image into a pre-trained image processing model to extract the target area; A lane-level map is constructed based on the extracted target area.

2. The method according to claim 1, characterized in that The obtaining of a trajectory grid image based on the trajectory points comprises: Dividing the trajectory points into a geographic coordinate grid; The trajectories in the grid of the division result are layered by elevation, the trajectory point data of each layer are projected on the ground, and converted into image pixel coordinates; The trajectory points are rendered into a trajectory grid image according to the density of the trajectory points on each layer of the image.

3. The method according to claim 2, characterized in that Rendering the trajectory points into a trajectory raster image comprises: Divide each grid area into grids of preset resolution; Count the number of trajectory points in each grid cell and generate a density distribution matrix; A trajectory grid image is generated based on the density distribution matrix.

4. The method according to claim 2, characterized in that Rendering the trajectory points into a trajectory raster image comprises: Divide the area within the grid into grids of preset resolution; Count the number of trajectory points in each grid cell and generate a density distribution matrix; Applying a histogram equalization algorithm to the density distribution matrix to enhance the visual recognition of image details by expanding the dynamic distribution range of density values; Generate a trajectory raster image based on the processed density distribution matrix.

5. The method according to claim 2, characterized in that Rendering the trajectory points into a trajectory raster image comprises: Divide the area within the grid into grids of preset resolution; Count the number of trajectory points in each grid cell and generate a density distribution matrix; A three-channel color image is generated based on the density distribution matrix to obtain a trajectory grid image, in which different colors are used to thermally represent the density of trajectory points.

6. The method according to claim 1, characterized in that The lane-level map construction based on the extracted target area includes: Mapping the extracted target area to three-dimensional space; Perform Z value post-processing on the mapped three-dimensional spatial data; Vector data of the target area is generated according to the processed data, and a map is constructed based on the vector data of the target area.

7. The method according to claim 1, characterized in that The target areas include: One or more of an intersection, a road junction, a fork, a road section, and an impassable area.

8. The method according to claim 1, characterized in that The training process of the image processing model includes: Obtain vector data of the target area in the geographic coordinate system and trajectory sample data; Dividing the vector data of the target area in the geographic coordinate system and the trajectory sample data into a plurality of sub-areas in a plane and layering them in a vertical direction; Projecting the divided vector data of the target area in the geographic coordinate system onto a two-dimensional plane to generate two-dimensional pixel coordinates of the target area; Generate a trajectory raster image based on the projection of the divided trajectory sample data; Generate annotated ground-truth training data based on the two-dimensional pixel coordinates of the target area and the trajectory raster image; Cleaning the annotated true value training data; Use the cleaned training data to train a deep neural network model for image detection and segmentation.

9. A lane-level map construction device, characterized in that: The device includes: an acquisition unit, configured to acquire trajectory coordinate data, sample the coordinate data, and obtain trajectory points; An image generating unit, configured to obtain a trajectory grid image based on the trajectory points; an extraction unit, configured to input the trajectory grid image into a pre-trained image processing model to extract a target area; The construction unit is used to construct a lane-level map based on the extracted target area.

10. A computing device, characterized in that The computing device includes: a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the method according to any one of claims 1 to 8 when executing the computer program.

Images