Map construction method and device, equipment, storage medium and program product

Abstract

This application provides a map construction method, apparatus, device, storage medium, and program product. The method includes: acquiring multi-view camera images of a target location in a target area, and determining a bird's-eye view image corresponding to the multi-view camera images with the target location as the origin; inputting the bird's-eye view image into an image segmentation model, where, based on the learned correlation between image features and road elements, image segmentation is performed on the road elements in the bird's-eye view image to obtain segmentation masks corresponding to the road elements, and the pixel coordinates of the target location corresponding to the road elements are determined based on the segmentation masks; the world coordinates corresponding to the pixel coordinates are determined, and a road element-level map corresponding to the target area is constructed based on the world coordinates. This application allows for the low-cost and efficient construction of accurate road element-level maps corresponding to map data, thereby meeting the high-precision map requirements of autonomous driving.

Description

Technical Field

[0001] This application relates to the field of high-precision map construction technology, and in particular to a map construction method, apparatus, device, storage medium and program product. Background Technology

[0002] In the field of autonomous driving, high-precision maps, as a crucial source of prior environmental knowledge, need to provide centimeter-level absolute positions of road elements such as lane lines, traffic arrows, and pedestrian crossings to support precise vehicle positioning, reliable path planning, and in-depth environmental understanding. Currently, mainstream mapping solutions heavily rely on high-beam LiDAR to acquire dense 3D point clouds. While this provides accurate geometry, it suffers from inherent drawbacks such as high equipment costs, limited deployment scale, and difficulty in frequent updates, especially for dynamic or volatile elements like traffic signs and road markings, where maintenance costs are extremely high. Furthermore, LiDAR point clouds lack rich texture and color information, leading to difficulties in semantic understanding and hindering the automatic and accurate differentiation and labeling of various road elements in complex urban scenarios. Therefore, low-cost, crowdsourced vision-based map construction has become an important development direction.

[0003] In related technologies, road scene images are captured by a single camera and transformed into a bird's-eye view (BEV) image through inverse perspective mapping (IPM) to build maps. However, maps built in this way suffer from inaccurate map data and cannot meet the urgent need for high-precision maps in autonomous driving. Summary of the Invention

[0004] This application provides a map building method, apparatus, device, storage medium, and program product to build accurate road element-level maps of corresponding map data at low cost and high efficiency, thereby meeting the demand for high-precision maps for autonomous driving.

[0005] In a first aspect, embodiments of this application provide a map construction method, including:

[0006] Acquire multi-view camera images of the target location within the target area, and determine the corresponding bird's-eye view image with the target location as the origin.

[0007] The bird's-eye view image is input into the image segmentation model. In the image segmentation model, based on the learned correlation between image features and road elements, the road elements in the bird's-eye view image are segmented to obtain the segmentation mask corresponding to the road elements. The pixel coordinates of the target position corresponding to the road elements are determined based on the segmentation mask.

[0008] Determine the world coordinates corresponding to the pixel coordinates, and construct a road element-level map of the target area based on the world coordinates.

[0009] In one possible implementation, determining the bird's-eye view image corresponding to the multi-view camera image includes:

[0010] Determine the ground feature points corresponding to multi-view camera images;

[0011] Based on ground feature points, the relative poses of multiple vehicle-mounted cameras acquiring multi-view images are determined, and based on the relative poses, the inverse perspective transformation parameters are determined.

[0012] For the camera images corresponding to each viewpoint in the multi-view camera images, an inverse perspective transformation is performed on the camera images based on the inverse perspective transformation parameters to obtain the bird's-eye view images corresponding to the camera images.

[0013] Image fusion processing is performed on the bird's-eye view images corresponding to each viewpoint in the multi-view camera images to obtain the bird's-eye view images corresponding to the multi-view camera images.

[0014] In one possible implementation, determining the inverse perspective transformation parameters based on relative pose includes:

[0015] Determine the real-time image features corresponding to multi-view camera images;

[0016] Acquire the inertial measurement data corresponding to the vehicle, and correct the relative pose of the corresponding vehicle camera based on the inertial measurement data and real-time image features to obtain the corrected relative pose;

[0017] Based on the corrected relative pose, the parameters of the inverse perspective transformation are determined.

[0018] In one possible implementation, a road feature-level map of the target area is constructed based on world coordinates, including:

[0019] Obtain the world coordinates of road elements in multiple bird's-eye view images. The multiple bird's-eye view images are the bird's-eye view images corresponding to other locations in the target area as the origin.

[0020] Based on world coordinates, road elements that are close in location and of the same category are grouped and clustered to obtain the grouping and clustering results.

[0021] Based on the grouping and clustering results, a road feature-level map of the target area is constructed.

[0022] In one possible implementation, a road feature-level map of the target area is constructed based on the grouping clustering results, including:

[0023] In the grouping clustering results, the world coordinates of road elements corresponding to the maximum values ​​are retained based on the non-maximum suppression algorithm;

[0024] Based on the world coordinates of the road elements corresponding to the maximum values ​​of each group clustering result, a road element-level map of the target area is constructed.

[0025] In one possible implementation, determining the world coordinates corresponding to the pixel coordinates includes:

[0026] Based on the camera extrinsic parameters of the vehicle-mounted camera that acquires multi-view images, the pixel coordinates are transformed to the vehicle coordinate system to obtain the first coordinate data of the road elements in the vehicle coordinate system.

[0027] Based on vehicle trajectory data, the first coordinate data is transformed to the world coordinate system to obtain the world coordinates corresponding to the road elements.

[0028] Secondly, embodiments of this application provide a map building apparatus, comprising:

[0029] The acquisition module is used to acquire multi-view camera images of the target location in the target area and determine the bird's-eye view image corresponding to the multi-view camera images with the target location as the origin.

[0030] The recognition module is used to input the bird's-eye view image into the image segmentation model. In the image segmentation model, based on the learned correlation between image features and road elements, the road elements in the bird's-eye view image are segmented to obtain the segmentation mask corresponding to the road elements, and the pixel coordinates of the target position corresponding to the road elements are determined based on the segmentation mask.

[0031] A construction module is used to determine the world coordinates corresponding to pixel coordinates and construct a road element-level map of the target area based on the world coordinates.

[0032] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor; the memory stores computer execution instructions; the processor executes the computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0033] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0034] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0035] The map construction method, apparatus, device, storage medium, and program products provided in this application acquire multi-view camera images of target locations in a target area and generate bird's-eye view images corresponding to the multi-view camera images with the target location as the origin. This allows for the use of multi-view information to compensate for the limitations of a single viewpoint, comprehensively and meticulously capturing various details in the road scene. By determining the segmentation mask of road elements in the bird's-eye view image based on an image segmentation model and obtaining the pixel coordinates of the target locations corresponding to the road elements, and leveraging the powerful feature extraction and classification capabilities of artificial intelligence, various road elements in the bird's-eye view image can be accurately identified, effectively reducing the error and workload of manual annotation and improving recognition efficiency and accuracy. Based on the identified road elements and their corresponding target location pixel coordinates, the specific location of each road element in the camera image can be clearly defined, providing key information for subsequent coordinate transformation. Converting pixel coordinates to world coordinates transforms the relative position information in the corresponding camera image into the absolute position in the actual scene, making the positioning of road elements more realistic. Finally, a road element-level map of the target area is constructed based on world coordinates, improving the integrity and accuracy of map data. This results in the low-cost and efficient construction of a precise road element-level map corresponding to the map data, thereby meeting the demand for high-precision maps for autonomous driving. Attached Figure Description

[0036] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0037] Figure 1 A schematic diagram of the scene constructed for the map provided in this application;

[0038] Figure 2 Flowchart of the map construction method provided in this application Figure 1 ;

[0039] Figure 3 A bird's-eye view image annotation diagram illustrating the embodiments provided in this application;

[0040] Figure 4 A schematic diagram illustrating the transformation from multi-view camera images to bird's-eye view images provided in this application;

[0041] Figure 5 A schematic diagram of road elements obtained by fusing and deduplicating multiple frames of bird's-eye view images provided in this application;

[0042] Figure 6 A simplified flowchart illustrating the map construction method provided in this application;

[0043] Figure 7 A schematic diagram of the map building apparatus provided in this application;

[0044] Figure 8 A schematic diagram of the structure of the electronic device provided in this application.

[0045] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments.

[0046] The correspondence between the map reference numerals and the road element names is as follows:

[0047] 130001, solid white line (solid edge line);

[0048] 130002, Single white dashed line;

[0049] 130003, short and thick dashed line. Detailed Implementation

[0050] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0051] First, let me explain the terms used in this application:

[0052] Bird's-eye view images are created using specific techniques, such as inverse perspective transformation, to convert ordinary images with perspective effects, like photos of ground scenes taken from different angles, into a perspective effect of looking vertically downwards from directly above. This eliminates distortion problems caused by perspective distortion, such as objects appearing larger when closer and smaller when farther away. It allows objects in the image to be displayed in a way that more accurately reflects their actual spatial layout and proportions. It can intuitively and comprehensively present information such as the planar layout of objects or scenes and the relative positional relationships between different parts. Whether it is urban architecture, road traffic scenes, or interior spaces, bird's-eye view images can help people understand their overall structure and detailed distribution more clearly and accurately.

[0053] World coordinates is an absolute coordinate system defined within a grand and unified three-dimensional spatial framework. It does not depend on any local or specific object or observation perspective, but assigns a unique and fixed spatial location identifier to all objects in the entire virtual or real scene. Just as precise coordinates are set for each celestial body in the vast universe, with the help of world coordinates, information about objects in the scene obtained at different locations and at different times can be accurately correlated and integrated, thereby achieving precise location of objects in the complete scene, tracking of motion trajectories, and seamless fusion and collaborative analysis of multiple data sources.

[0054] Related technologies rely on a single camera to convert perspective images captured by the camera into bird's-eye views through inverse perspective transformation (IPM). This method typically processes images from only one or a few perspectives, transforming information such as road scenes in the camera images into a top-down view to present elements such as lane lines and vehicles.

[0055] The limited field of view of a single camera or a small number of cameras makes it difficult to fully cover the environment around the vehicle, resulting in significant blind spots. Furthermore, the cameras lack the ability to capture elements in areas such as the sides and rear of the vehicle, leading to incomplete bird's-eye view information. Therefore, maps constructed in this way suffer from inaccurate map data and fail to meet the urgent need for high-precision maps in autonomous driving.

[0056] To address the aforementioned issues, this application provides a map construction scheme that utilizes the collaboration of multi-view cameras to effectively integrate image information captured by different cameras, forming a complete and coherent view of the vehicle's surrounding environment and compensating for the limitations of a single camera's field of view. Simultaneously, it extracts pixel-level information of road elements through an image segmentation model, converts pixel coordinates into world coordinates, and finally constructs a map based on the world coordinates, thereby improving the completeness and accuracy of map data and meeting the requirements of autonomous driving for high-precision maps.

[0057] The map construction scheme provided in this application can be applied to the fields of map surveying and mapping and high-precision map construction, as well as any field that requires high-resolution map production. This application does not impose any restrictions on this application.

[0058] Please see Figure 1 , Figure 1 A schematic diagram of the scene constructed for the map provided in this application. For example... Figure 1 As shown, this scenario includes a terminal device 110 and a server 120. The terminal device 110 and the server 120 communicate via a wired or wireless network. The terminal device 110 can upload its own data to the server 120, and can also retrieve data from the server 120.

[0059] Among them, terminal device 110 may include, but is not limited to, mobile phones, tablets, laptops, computers, voice interaction devices, vehicle terminals, aircraft, remote driving terminals, etc.; server 120 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. This document does not restrict the specific form of terminal devices and servers.

[0060] It should be noted that, Figure 1 The number of terminal devices 110 and servers 120 is merely illustrative; any number of terminal devices 110 and servers 120 can be used as needed.

[0061] In an exemplary embodiment, the map construction method provided in this application can be executed by a terminal device 110. For example, the terminal device 110 can acquire multi-view camera images of a target location within a target area of ​​a pre-constructed map, and can determine a bird's-eye view image corresponding to the multi-view camera images with the target location as the origin. Then, the bird's-eye view image is input into a pre-trained image segmentation model. In the image segmentation model, based on the learned association between image features and road elements, image segmentation is performed on the road elements in the bird's-eye view image to obtain a segmentation mask corresponding to the road elements. Based on the segmentation mask, the pixel coordinates of the target location corresponding to the road elements are determined. Furthermore, a road element-level map of the target area can be constructed based on the world coordinates corresponding to the pixel coordinates and based on the world coordinates.

[0062] In another exemplary embodiment, server 120 may have similar functions to terminal device 110 to perform the image processing method provided in this embodiment. For example, server 120 may acquire multi-view camera images of a target location within a pre-built map's target area, and determine a bird's-eye view image corresponding to the multi-view camera images with the target location as the origin; then, the bird's-eye view image is input into a pre-trained image segmentation model to perform image segmentation on road elements in the bird's-eye view image, obtaining segmentation masks corresponding to the road elements, and determining the pixel coordinates of the target location corresponding to the road elements based on the segmentation masks; subsequently, a road element-level map of the target area can be constructed based on the world coordinates corresponding to the pixel coordinates and based on the world coordinates.

[0063] In another exemplary embodiment, the terminal device 110 and the server 120 may also jointly execute the map construction method provided in the embodiments of this application. For example, after acquiring a multi-view camera image, the server 120 determines the corresponding bird's-eye view image. Then, it determines the segmentation mask corresponding to the bird's-eye view image using an image segmentation model, and determines the road elements and their corresponding pixel coordinates in the bird's-eye view image based on the segmentation mask. The server then sends the road elements and their corresponding pixel coordinates in the bird's-eye view image to the terminal device 110, whereby the terminal device 110 determines the world coordinates corresponding to the pixel coordinates, constructs a road element-level map based on the world coordinates, and displays it in a designated display area.

[0064] In the above example, regarding acquiring multi-view camera images of a vehicle, it can be understood that in one possible implementation, multiple cameras, such as in-vehicle cameras, are deployed on the vehicle. These multiple cameras can then capture images from multiple different perspectives surrounding the vehicle, resulting in multi-view camera images. It should be noted that the deployment of multiple cameras must be based on the premise that all environmental information around the vehicle can be captured; that is, the acquired multi-view camera images must contain all available environmental information surrounding the vehicle.

[0065] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0066] Figure 2 Flowchart of the map construction method provided in this application Figure 1 .like Figure 2 As shown, the map construction method includes at least steps S201 to S203, which are described in detail below:

[0067] Step S201: Acquire multi-view camera images of the target location in the target area, and determine the bird's-eye view image corresponding to the multi-view camera images with the target location as the origin.

[0068] For example, please refer to Figure 3 During vehicle operation, multiple onboard cameras can capture multi-view camera images with the vehicle center as the origin. Then, based on the bird's-eye view image corresponding to the multi-view camera image with the vehicle center as the origin, the target location in this example is the vehicle center.

[0069] Specifically, multiple cameras can be strategically positioned on the vehicle to ensure comprehensive coverage of key areas surrounding the target object from different angles. For example, six cameras can be installed at the front, rear, left, and right of the vehicle to obtain all-around information about the road and surrounding environment. Furthermore, the intrinsic parameters (such as focal length and principal point coordinates) and extrinsic parameters (the camera's position and orientation in the vehicle coordinate system, including rotation matrices and translation vectors) of each camera must be accurately measured and recorded. After acquiring the original images captured by the multi-angle cameras—the "multi-view camera images" in this application—the images from the six cameras are processed separately. Inverse perspective transformation (IPM) is used to remap the pixels with perspective effects in the original six-view camera images to a new coordinate system according to geometric transformation rules, transforming the originally tilted and distorted perspective images into a bird's-eye view image viewed vertically downwards from directly above.

[0070] Step S202: Input the bird's-eye view image into the image segmentation model. In the image segmentation model, based on the learned correlation between image features and road elements, perform image segmentation on the road elements in the bird's-eye view image to obtain the segmentation mask corresponding to the road elements, and determine the pixel coordinates of the target position corresponding to the road elements based on the segmentation mask.

[0071] In this step, the acquired bird's-eye view image is input into a pre-trained image segmentation model. This model typically employs a deep learning network with encoding and decoding structures, such as DDRNet. During the encoding stage, multi-layer convolutional neural networks extract multi-scale feature information from the bird's-eye view image, based on the learned correlation between image features and road elements. In the decoding stage, operations such as upsampling or transposed convolution are used to restore the image features to the same size as the input image. Simultaneously, skip connections and other methods are used to fuse features from different levels to improve segmentation accuracy. Furthermore, during training, the model optimizes its parameters using a loss function (such as cross-entropy loss) to accurately distinguish pixel regions of different semantic categories. When the bird's-eye view image is input, the image segmentation model assigns a semantic category label to each pixel in the image, generating a segmentation mask of the same size as the input image. By parsing the pixel category information in the segmentation mask, the type of road element contained in the bird's-eye view image and its corresponding pixel location can be determined. This yields the recognition result of the bird's-eye view image, enabling automatic identification and localization of road elements in the image, and obtaining the pixel coordinates of the target position of the corresponding road element centered on the vehicle origin.

[0072] Optionally, in some feasible embodiments, the image segmentation model can be trained in the following manner. See also... Figure 3 , Figure 3This is a schematic diagram illustrating bird's-eye view image annotation as provided in the embodiments of this application. Specifically, key elements in the bird's-eye view image, such as lane markings, arrows, stop lines, and pedestrian crossings, can be annotated at the pixel level using manual or semi-automatic annotation tools to obtain a bird's-eye view image dataset with pixel labels. Road elements can be located at the pixel level first, and then annotated, such as... Figure 3 As shown, label 130001 represents a solid white line (edge ​​line) in the bird's-eye view image, label 130002 represents a single dashed white line, and label 130003 represents a short, thick dashed line. During the annotation process, the accuracy and consistency of category classification are strictly ensured to generate a segmentation mask, ultimately forming a BEV image dataset with pixel-level labels. This bird's-eye view image dataset serves as a crucial supervisory signal for subsequent training, and its quality is paramount. Specifically, in the image segmentation model training phase, the generated labeled bird's-eye view image dataset is used as the processing object. A semantic segmentation network (such as DDRNet) is selected for training. DDRNet, with its encoding and decoding structure, can extract multi-scale features of the image. During training, the model takes the labeled bird's-eye view image as input and outputs a segmentation mask with the same size as the input. Each pixel is accurately classified into corresponding categories such as lane lines, pedestrian arrows, and stop lines. The model gradually establishes a mapping relationship from image to semantic category by continuously learning the pixel-level features contained in the labeled data. Meanwhile, the model parameters are continuously optimized using the loss function. After a series of training iterations, a semantic segmentation model is finally obtained. This semantic segmentation model has the ability to identify features in new BEV images and can achieve high-precision segmentation results.

[0073] Step S203: Based on the world coordinates corresponding to the pixel coordinates, construct a road element-level map of the target area based on the world coordinates.

[0074] In this step, after knowing the pixel coordinates of the camera image corresponding to the road features, the world coordinates corresponding to the road features can be determined based on these pixel coordinates, which is the absolute position of the road pixels. Thus, a road feature-level map of the target area can be constructed based on the absolute position of the road pixels.

[0075] For example, the extrinsic parameters (rotation matrix R and translation vector T) of the vehicle-mounted camera are used to transform the pixel coordinates in the bird's-eye view image into coordinates in the vehicle coordinate system using a coordinate transformation formula. Specifically, the projection matrix is ​​used to map each pixel from the image coordinate system to the vehicle coordinate system, i.e., pixel coordinates, with the vehicle center as the origin. Subsequently, combined with vehicle trajectory data, such as real-time position, attitude, and velocity information provided by inertial sensors (IMU), the pixel coordinates in the vehicle coordinate system are further transformed into absolute positions in the world coordinate system.

[0076] In the embodiments provided in this application, by acquiring multi-view camera images of the target location in the target area and generating a bird's-eye view image with the target location as the origin, the multi-view information can compensate for the limitations of a single viewpoint, comprehensively and meticulously capturing various details in the road scene. Based on an image segmentation model, the segmentation mask of road elements in the bird's-eye view image is determined, and the pixel coordinates of the target location corresponding to the road elements are obtained. Leveraging the powerful feature extraction and classification capabilities of artificial intelligence, various road elements in the bird's-eye view image can be accurately identified, effectively reducing the error and workload of manual annotation and improving recognition efficiency and accuracy. Based on the identified road elements and their corresponding target location pixel coordinates, the specific location of each road element in the camera image can be clearly defined, providing key information for subsequent coordinate transformation. Converting pixel coordinates to world coordinates transforms the relative position information in the corresponding camera image into the absolute position in the actual scene, making the positioning of road elements more realistic. Finally, a road element-level map of the target area is constructed based on world coordinates, improving the integrity and accuracy of the map data. This allows for the low-cost and efficient construction of a precise road element-level map corresponding to the map data, thereby meeting the requirements of autonomous driving for high-precision maps.

[0077] Based on the above embodiments, in one exemplary embodiment provided in this application, the specific implementation process of determining the bird's-eye view image corresponding to the multi-view camera image may further include steps S301 to S304, which are described in detail below:

[0078] Step S301: Determine the ground feature points corresponding to the multi-view camera images;

[0079] Step S302: Based on ground feature points, determine the relative poses between multiple vehicle-mounted cameras that acquire multi-view images, and based on the relative poses, determine the inverse perspective transformation parameters.

[0080] Step S303: For the camera images corresponding to each viewpoint in the multi-view camera images, perform inverse perspective transformation on the camera images based on the inverse perspective transformation parameters to obtain the bird's-eye view image corresponding to the camera image.

[0081] Step S304: Perform image fusion processing on the bird's-eye view images corresponding to each viewpoint in the multi-view camera images to obtain the bird's-eye view images corresponding to the multi-view camera images.

[0082] For example, determining the ground feature points corresponding to multi-view camera images requires identifying stable and repeatable feature points such as lane lines, curbs, and ground markings in each frame using appropriate feature detection algorithms (such as SIFT and ORB). Then, feature matching algorithms (such as the FLANN matcher) are used to associate pairs of identical feature points across cameras. Subsequently, based on multi-view geometry principles (such as epipolar constraints and fundamental matrix calculation) combined with the camera intrinsic parameter matrix, the relative poses (including rotation matrices and translation vectors) between multiple cameras are calculated using the pixel coordinates of the ground feature point pairs. Using this relative pose relationship as a constraint, the target inverse perspective transformation parameters that map the images from each camera viewpoint to a unified bird's-eye view plane are further solved, ultimately establishing the pixel correspondence between each camera image and the global top-down coordinate system.

[0083] Optionally, based on the calculated relative poses between the cameras, combined with ground plane assumptions (such as the Z=0 plane) and parameters such as camera height and pitch angle, an inverse perspective transformation matrix can be derived for image transformation. This inverse perspective transformation matrix ensures that the transformed ground feature points remain horizontally aligned and scaled uniformly in the bird's-eye view. Then, the target inverse perspective transformation parameters obtained based on the inverse perspective transformation matrix are applied to perform coordinate mapping on the multi-view camera images, transforming each pixel from the original viewpoint to the bird's-eye viewpoint. Overlapping areas are then processed using image stitching or fusion algorithms (such as weighted averaging or multi-band fusion) to eliminate stitching gaps, ultimately generating a complete bird's-eye view image covering the environment surrounding the vehicle.

[0084] Optional, please refer to Figure 4 , Figure 4 This application provides a schematic diagram of the transformation from multi-view camera images to bird's-eye view images, such as... Figure 4 As shown, six cameras are strategically positioned on the vehicle, ensuring they comprehensively cover the key scene area surrounding the target object from different angles. Six original images from different perspectives are then acquired. Simultaneously, the intrinsic parameters (such as focal length and principal point coordinates) and extrinsic parameters (the camera's position and orientation in the vehicle coordinate system, including rotation matrices and translation vectors) of these six cameras are measured and recorded. Further, after acquiring the original images from these six perspectives, the Inverse Perspective Transform (IPM) algorithm is applied to each image. This IPM algorithm remaps the pixels with perspective effects in the six original images to a new coordinate system according to geometric transformation rules, converting the originally tilted and distorted perspective images into a bird's-eye view image viewed vertically downwards from directly above.

[0085] It should be noted that in this embodiment, the inverse perspective transformation parameters need to be pre-calibrated using a calibration board or Simultaneous Localization and Mapping (SLAM) to ensure the ground plane assumption holds true. Specifically, if the inverse perspective transformation parameter calibration is achieved using a calibration board, the camera needs to acquire multi-view images of the calibration board at different positions and attitudes. Then, a feature point detection algorithm is used to extract the coordinates of the corner points or center point of the calibration board in each view image. Based on the center point coordinates and the known 3D dimensions of the calibration board, the camera's intrinsic and extrinsic parameter matrices are determined. Then, using the camera's intrinsic and extrinsic parameter verification combined with the ground plane assumption (such as the Z=0 plane), the homography matrix calculated is used as the inverse perspective transformation parameter. If the inverse perspective transformation parameter calibration is achieved using SLAM, the camera pose can be tracked in real time using visual or LiDAR data to construct an environmental map. Then, ground regions can be identified from the environmental map using methods such as feature point matching, keyframe optimization, and ground segmentation. By combining vehicle motion trajectory and vehicle inertial measurement data to correct camera pose drift, the inverse perspective transformation matrix calculated based on the corrected camera pose and the ground plane parameters (such as the Z=0 plane) corresponding to the ground area is used as the inverse perspective transformation parameter. This ensures that the ground plane assumption holds in dynamic scenes and improves the accuracy and robustness of the inverse perspective transformation.

[0086] In the embodiments provided in this application, high-precision pose estimation and inverse perspective transformation of multiple cameras are achieved by extracting ground feature points, thereby generating a seamlessly fused bird's-eye view image, effectively eliminating blind spots and maintaining scene geometric consistency.

[0087] Based on the above embodiments, in one exemplary embodiment provided in this application, the specific implementation process of determining the inverse perspective transformation parameters based on relative pose may further include steps S401 to S403, which are described in detail below:

[0088] Step S401: Determine the real-time image features corresponding to the multi-view camera images;

[0089] Step S402: Obtain the inertial measurement data corresponding to the vehicle, and correct the relative pose of the corresponding vehicle camera based on the inertial measurement data and real-time image features to obtain the corrected relative pose.

[0090] Step S403: Determine the inverse perspective transformation parameters based on the corrected relative pose.

[0091] For example, to ensure the accuracy of road feature-level maps constructed in dynamic environments, the relative pose of vehicle-mounted cameras can be corrected. Specifically, visually distinctive feature points can be extracted from multi-view camera images as real-time image features. Simultaneously, IMU data output by the vehicle's inertial measurement unit (IMU) can be acquired, including high-frequency angular velocity and linear acceleration data. Then, by pre-integrating the IMU data, the relative pose change estimate of the corresponding vehicle-mounted camera in a short time is obtained. The relative pose of the corresponding vehicle-mounted camera is then corrected based on the relative pose change estimate to obtain the corrected relative pose. Based on the corrected accurate relative pose relationship, the projection geometry relationship from each camera's viewpoint to the unified bird's-eye view plane is calculated, ultimately obtaining accurate inverse perspective transformation parameters that can adapt to changes in vehicle motion state.

[0092] Optionally, in some feasible embodiments, obtaining the relative pose change estimate of the corresponding vehicle-mounted camera in a short time by pre-integrating the IMU data can specifically be as follows: the motion prior provided by the IMU pre-integration is tightly coupled and optimized with the multi-view image features; a scale-consistent local map is established through visual-inertial joint initialization; and the state variables of each camera frame, including position, attitude, velocity and sensor deviation, are continuously optimized using a sliding window-based nonlinear optimization method, thereby correcting the relative pose error between cameras caused by vehicle vibration, road bumps or loose camera installation in real time.

[0093] In the embodiments provided in this application, dynamic correction of multi-camera pose is achieved by fusing real-time image features and inertial measurement data, which significantly improves the accuracy and robustness of inverse perspective transformation, thereby generating a more accurate and stable global bird's-eye view.

[0094] Based on the above embodiments, in one exemplary embodiment provided in this application, the specific implementation process of constructing a road element-level map of a target area based on world coordinates may further include steps S501 to S503, which are described in detail below:

[0095] Step S501: Obtain the world coordinates of road elements in multiple bird's-eye view images. The multiple bird's-eye view images are the bird's-eye view images corresponding to other locations in the target area as the origin.

[0096] Step S502: Based on world coordinates, group and cluster road elements that are close in location and of the same category to obtain the grouping and clustering results;

[0097] Step S503: Based on the grouping and clustering results, construct a road feature-level map of the target area.

[0098] For example, following the above embodiments, since a single frame of a BEV image may be affected by factors such as lighting, occlusion, and sensor noise, resulting in inaccurate road feature identification or missing information, multiple vehicle-mounted cameras (front, rear, left, and right) can be used to work collaboratively to obtain multi-view camera images. These multi-view camera images are then uniformly processed and converted into BEV images. Furthermore, in this embodiment, a bird's-eye view image generated from multiple vehicle-mounted camera images is generated at regular intervals (e.g., 9m). Each bird's-eye view image actually covers an area of ​​20m × 20m, thus obtaining multiple consecutive frames of bird's-eye view images.

[0099] Optional, please refer to Figure 5 , Figure 5 This is a schematic diagram of road features obtained by fusing and deduplicating multiple frames of bird's-eye view images provided in this application. Figure 5 As shown, specifically, multiple frames of bird's-eye view images collected from different locations within the target area can be acquired. Road elements (such as lane lines, stop lines, and arrows) identified by an image segmentation model in each frame are then unified to the world coordinate system through coordinate transformation, forming a multi-source observation point set corresponding to the spatiotemporal distribution of the target area. Based on the world coordinates and semantic labels of road elements at the target location within the target area, a distance- and density-based clustering algorithm is used to group the observation points. Grouping includes grouping observation points with a 3D spatial distance less than a preset threshold and the same semantic category into the same cluster, considering them to represent the same road element instance within the target area, thus obtaining multiple world coordinates for each road element. Subsequently, model fitting and optimization are performed on the multiple world coordinates. For example, spline curve fitting is used for the multiple world coordinates of lane lines to extract their continuous geometry, and template matching is used for the multiple world coordinates of traffic arrows to determine their type and orientation. Weighted fusion is then performed by combining the world coordinates of each observation point with the confidence levels corresponding to the semantic labels to improve the accuracy of road element location. Finally, the fitted vector geometry is associated with its semantic attributes to construct a structured road feature-level map covering the target area, containing accurate geometry and rich semantic information.

[0100] In the embodiments provided in this application, the world coordinate clustering analysis of road elements in multi-frame bird's-eye view images is used to achieve accurate association and spatial integration of road elements across different perspectives, thereby constructing a high-precision road element-level map covering the entire target area.

[0101] Based on the above embodiments, in one exemplary embodiment provided in this application, the specific implementation process of constructing a road element-level map of the target area based on the grouping clustering results may further include steps S601 and S602, which are described in detail below:

[0102] Step S601: In the grouping clustering results, based on the non-maximum suppression algorithm, retain the world coordinates of the road elements corresponding to the maxima;

[0103] Step S602: Based on the world coordinates of the road elements corresponding to the maximum values ​​of each group clustering result, construct a road element-level map of the target area.

[0104] Specifically, the nonmaximum suppression algorithm can retain the road element with the highest score and remove other redundant elements based on the confidence score of the road element.

[0105] For example, following the embodiments described above, in order to generate a stable and accurate feature-level map from continuously sampled bird's-eye view image perception results, a spatial sliding window sampling method can be used. Positions on the vehicle trajectory are selected as sampling points at fixed intervals (e.g., 9 meters), and a bird's-eye view image corresponding to each sampling point is obtained. Since a single frame of bird's-eye view image covers a limited area (e.g., 20 meters × 20 meters), the bird's-eye view images of adjacent sampling points overlap spatially, causing the same real road feature to be repeatedly perceived and mapped to the world coordinate system by multiple consecutive frames of bird's-eye view images.

[0106] After obtaining multiple observation coordinates of the same feature within the overlapping area, a non-maximum suppression algorithm can be used for processing. Specifically, the multiple observation locations of the same feature are first grouped according to their spatial proximity and semantic proximity. Then, within each group, the confidence scores attached to each observation result are compared. Only the observation result with the highest confidence score in the group is retained as the final location estimate of the road feature, while other repeated observations with lower confidence at the same location are suppressed.

[0107] In this embodiment, the non-maximum suppression algorithm effectively eliminates the positioning jitter and duplicate labeling problems caused by sensor noise, viewpoint changes, or instantaneous fluctuations in the perception model, ensuring the uniqueness and optimality of map element locations. Finally, based on all the optimal observed world coordinates and their semantic information selected through non-maximum suppression, a spatially consistent, non-redundant, and accuracy-optimized target area road element-level map is constructed.

[0108] Based on the above embodiments, in one exemplary embodiment provided in this application, the specific implementation process of determining the world coordinates corresponding to the pixel coordinates may further include steps S701 and S702, which are described in detail below:

[0109] Step S701: Based on the camera extrinsic parameters of the vehicle-mounted camera that acquires multi-view images, the pixel coordinates are transformed to the vehicle coordinate system to obtain the first coordinate data of the road elements in the vehicle coordinate system.

[0110] Step S702: Based on the vehicle trajectory data, the first coordinate data is converted to the world coordinate system to obtain the world coordinates corresponding to the road elements.

[0111] For example, the rotation matrix in the camera's extrinsic parameters can be used. Translation vector Geometric transformations are performed on the pixel coordinates. By adjusting the positional relationship between the camera's viewpoint and the vehicle's center, the pixels are mapped from the image plane to a vehicle coordinate system with the vehicle's center as the origin. The resulting coordinate data reflects the spatial position of road elements relative to the vehicle. Subsequently, combined with vehicle trajectory data (such as attitude angles provided by the IMU, position information and speed data provided by GPS), the coordinates in the vehicle coordinate system are transformed to the world coordinate system using a vehicle motion model, ultimately obtaining the absolute position of the road elements in the world coordinate system.

[0112] Optionally, in some feasible embodiments, the pixel coordinates can be determined using camera intrinsic parameters. Converted to normalized camera coordinates, specifically represented as follows:

[0113]

[0114] in, Normalized camera coordinates, This is the camera intrinsic parameter matrix.

[0115] Then, through the rotation matrix in the camera extrinsic parameters Translation vector Perform coordinate transformation to obtain the first coordinate data in vehicle coordinate system. Specifically, it is expressed as:

[0116]

[0117] Subsequently, this is combined with vehicle trajectory data (such as attitude angles provided by the vehicle's IMU, such as roll angle). Pitch angle Heading angle Location information provided by GPS The vehicle motion model transforms the coordinates in the vehicle coordinate system to the global world coordinate system. Specifically, this involves calculating the global rotation matrix based on the IMU's attitude angles. Represented as:

[0118]

[0119] Determine the global translation vector by combining GPS location information In the formula

[0120]

[0121] Furthermore, coordinate transformation is completed synchronously with vehicle trajectory data and image acquisition time, ensuring the accuracy and consistency of coordinate transformation.

[0122] In the embodiments provided in this application, the pixel coordinates of road elements are transformed to the vehicle coordinate system by using camera extrinsic parameters, and then further transformed to the world coordinate system by combining vehicle trajectory data. This enables accurate positioning of road elements from image space to real-world space, providing a reliable location benchmark for constructing a high-precision global map.

[0123] Please see Figure 6 , Figure 6 A simplified flowchart illustrating the map construction method provided in this application. Figure 6 As shown, multi-view vehicle-mounted camera images can be acquired first. Then, an inverse perspective algorithm is used to convert these images into bird's-eye view images. Road elements in these bird's-eye view images can be labeled, and a model training dataset can be built based on the labeled bird's-eye view images. An image segmentation model is trained using this dataset. This model is then used to process the real-time bird's-eye view images to extract the pixel coordinates corresponding to road elements. Combined with the vehicle trajectory from the vehicle's external camera, the pixel coordinates are converted into world coordinates. Finally, the world coordinates of the same road element in each of the acquired multi-frame bird's-eye view images are deduplicated and fused to generate a vectorized map at the road element level. For detailed implementation processes, please refer to the descriptions in the aforementioned embodiments; they will not be repeated here.

[0124] In summary, the map construction method provided in this application has at least the following beneficial effects:

[0125] First, low cost and high precision.

[0126] Improved technical approach: Use images from six onboard cameras (front, rear, left, and right) combined with IPM to convert them into bird's-eye view images, instead of relying on traditional high-cost equipment such as LiDAR to obtain high-precision map data.

[0127] The resulting technical benefits are as follows: professional equipment such as LiDAR is expensive, while cameras are relatively inexpensive. High-precision map elements with absolute coordinates can be generated using only cameras and extrinsic parameters, avoiding the expenses associated with using high-cost equipment, thus achieving low-cost and high-precision map construction.

[0128] Second, multi-view fusion reduces blind spots.

[0129] Improved technical methods: Six cameras (front, rear, left, and right) are used on the vehicle to work together, and the acquired multi-view camera images are uniformly processed and converted into bird's-eye view images.

[0130] The resulting technical effects: Multiple cameras shooting from different angles can capture more comprehensive scene information around the vehicle. By fusing these multi-view images, the generated bird's-eye view can cover the field of vision in all directions, avoiding blind spots caused by the limited field of vision of a single or few cameras and reducing blind spots around the vehicle.

[0131] Third, improve the accuracy and completeness of map data.

[0132] By selecting a frame of bird's-eye view image generated from 6 cameras at regular intervals, the perception results of multiple consecutive frames of bird's-eye view images are fused, and the road elements at the same location are processed using a non-maximum suppression algorithm to obtain the optimal solution, thereby improving the accuracy and completeness of map data and solving problems such as incomplete information and inaccurate feature recognition that may exist in a single frame of bird's-eye view image.

[0133] The resulting technical effects are as follows: A single-frame bird's-eye view may be affected by factors such as lighting, occlusion, and sensor noise, leading to inaccurate feature identification or missing information. However, continuous multi-frame bird's-eye view images contain information from different points in time. Fusion of these frames can complement and corroborate each other, using non-maximum suppression algorithms to find the optimal solution for features at the same location, thereby improving the accuracy of map data. For example, if each bird's-eye view image actually covers an area of ​​20 meters × 20 meters, and a frame is selected every 9 meters, the fusion of multiple frames can cover a wider area and compensate for the incomplete information in some areas that may exist in a single frame, thus enhancing the completeness of the map data. By fusing the perception results of continuous multi-frame bird's-eye view images, the information from multiple frames can be integrated, reducing errors and noise that may exist in a single frame, making the map data more accurate. Continuous multi-frame bird's-eye view images cover a larger area; fusion can integrate the information from these areas, making the map data more complete and reducing information blind spots.

[0134] Fourth, automated closed-loop systems improve efficiency.

[0135] Improved technical methods: Achieved a fully automated closed loop from image acquisition, BEV conversion, feature annotation, model training to map data generation.

[0136] The resulting technical benefits are as follows: The entire process forms an end-to-end automated system, with each link closely connected and working collaboratively. Cameras continuously acquire images, which, after a series of processing steps, automatically generate high-precision map elements without much human intervention. This allows for timely reflection of the latest changes in the road environment, enabling real-time updates of map data. Simultaneously, it reduces the time and workload of manual operations, supports real-time updates of high-precision map data, and improves the efficiency of map construction.

[0137] Figure 7 A schematic diagram of the map building apparatus provided in this application. Figure 7 As shown, the map building apparatus 70 provided in this embodiment includes:

[0138] The acquisition module 710 is used to acquire multi-view camera images of the target location in the target area and determine the bird's-eye view image corresponding to the multi-view camera images with the target location as the origin.

[0139] The recognition module 720 is used to input the bird's-eye view image into the image segmentation model. In the image segmentation model, based on the learned correlation between image features and road elements, the road elements in the bird's-eye view image are segmented to obtain the segmentation mask corresponding to the road elements, and the pixel coordinates of the target position corresponding to the road elements are determined based on the segmentation mask.

[0140] Module 730 is used to determine the world coordinates corresponding to pixel coordinates and to construct a road feature-level map of the target area based on the world coordinates.

[0141] In one possible implementation, when determining the bird's-eye view image corresponding to the multi-view camera image, the acquisition module 710 is specifically used to: determine the ground feature points corresponding to the multi-view camera image.

[0142] Based on ground feature points, the relative poses of multiple vehicle-mounted cameras acquiring multi-view images are determined, and based on the relative poses, the inverse perspective transformation parameters are determined.

[0143] For the camera images corresponding to each viewpoint in the multi-view camera images, an inverse perspective transformation is performed on the camera images based on the inverse perspective transformation parameters to obtain the bird's-eye view images corresponding to the camera images.

[0144] Image fusion processing is performed on the bird's-eye view images corresponding to each viewpoint in the multi-view camera images to obtain the bird's-eye view images corresponding to the multi-view camera images.

[0145] In one possible implementation, when determining the inverse perspective transformation parameters based on the relative pose, the acquisition module 710 is specifically used for:

[0146] Determine the real-time image features corresponding to multi-view camera images;

[0147] Acquire the inertial measurement data corresponding to the vehicle, and correct the relative pose of the corresponding vehicle camera based on the inertial measurement data and real-time image features to obtain the corrected relative pose;

[0148] Based on the corrected relative pose, the parameters of the inverse perspective transformation are determined.

[0149] In one possible implementation, when constructing a road feature-level map of a target area based on world coordinates, the construction module 730 is specifically used for:

[0150] Obtain the world coordinates of road elements in multiple bird's-eye view images. The multiple bird's-eye view images are the bird's-eye view images corresponding to other locations in the target area as the origin.

[0151] Based on world coordinates, road elements that are close in location and of the same category are grouped and clustered to obtain the grouping and clustering results.

[0152] Based on the grouping and clustering results, a road feature-level map of the target area is constructed.

[0153] In one possible implementation, when constructing a road feature-level map of the target area based on the grouping clustering results, the construction module 730 is specifically used for:

[0154] In the grouping clustering results, the world coordinates of road elements corresponding to the maximum values ​​are retained based on the non-maximum suppression algorithm;

[0155] Based on the world coordinates of the road elements corresponding to the maximum values ​​of each group clustering result, a road element-level map of the target area is constructed.

[0156] In one possible implementation, the construction module 730, when used to determine the world coordinates corresponding to the pixel coordinates, specifically performs the following:

[0157] Based on the camera extrinsic parameters of the vehicle-mounted camera that acquires multi-view images, the pixel coordinates are transformed to the vehicle coordinate system to obtain the first coordinate data of the road elements in the vehicle coordinate system.

[0158] Based on vehicle trajectory data, the first coordinate data is transformed to the world coordinate system to obtain the world coordinates corresponding to the road elements.

[0159] The map building apparatus provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0160] Figure 8 A schematic diagram of the structure of the electronic device provided in this application. Figure 8 As shown, the electronic device 80 provided in this embodiment includes at least one processor 810 and a memory 820. Optionally, the device 80 further includes a communication component 830. The processor 810, memory 820, and communication component 830 are connected via a bus 840.

[0161] In a specific implementation, at least one processor 810 executes computer execution instructions stored in memory 820, causing at least one processor 810 to perform the above-described method.

[0162] The specific implementation process of processor 810 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0163] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0164] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0165] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0166] This application also provides a computer program product, including a computer program that, when executed by a processor or the like, implements the above-described method.

[0167] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor or the like, implement the above-described method.

[0168] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0169] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0170] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0171] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0172] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0173] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0174] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0175] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A map construction method, characterized in that, include: Acquire multi-view camera images of the target location in the target area, and determine the bird's-eye view image corresponding to the multi-view camera images with the target location as the origin; The bird's-eye view image is input into an image segmentation model. In the image segmentation model, based on the learned correlation between image features and road elements, the road elements in the bird's-eye view image are segmented to obtain the segmentation mask corresponding to the road elements. The pixel coordinates of the target position corresponding to the road elements are determined based on the segmentation mask. Determine the world coordinates corresponding to the pixel coordinates, and construct a road element-level map of the target area based on the world coordinates.

2. The method according to claim 1, characterized in that, Determining the bird's-eye view image corresponding to the multi-view camera image includes: Determine the ground feature points corresponding to the multi-view camera images; Based on the ground feature points, the relative poses between multiple vehicle-mounted cameras that acquire the multi-view images are determined, and based on the relative poses, the inverse perspective transformation parameters are determined. For each camera image corresponding to a viewpoint in the multi-view camera image, an inverse perspective transformation is performed on the camera image based on the inverse perspective transformation parameters to obtain the bird's-eye view image corresponding to the camera image; Image fusion processing is performed on the bird's-eye view images corresponding to each viewpoint in the multi-view camera images to obtain the bird's-eye view images corresponding to the multi-view camera images.

3. The method according to claim 2, characterized in that, The determination of inverse perspective transformation parameters based on the relative pose includes: Determine the real-time image features corresponding to the multi-view camera images; Acquire the inertial measurement data corresponding to the vehicle, and correct the relative pose of the corresponding vehicle camera based on the inertial measurement data and the real-time image features to obtain the corrected relative pose; Based on the corrected relative pose, the inverse perspective transformation parameters are determined.

4. The method according to any one of claims 1 to 3, characterized in that, The construction of a road feature-level map of the target area based on the world coordinates includes: Obtain the world coordinates of road elements in multiple bird's-eye view images, wherein the multiple bird's-eye view images are bird's-eye view images corresponding to other locations in the target area as the origin; Based on the world coordinates, road elements that are close in location and of the same category are grouped and clustered to obtain the grouping and clustering results. Based on the grouping and clustering results, a road feature-level map of the target area is constructed.

5. The method according to claim 4, characterized in that, The step of constructing a road feature-level map of the target area based on the grouping and clustering results includes: In the grouping clustering results, the world coordinates of road elements corresponding to the maximum values ​​are retained based on the non-maximum suppression algorithm; Based on the world coordinates of the road elements corresponding to the maximum values ​​of each group clustering result, a road element-level map of the target area is constructed.

6. The method according to any one of claims 1 to 3, characterized in that, Determining the world coordinates corresponding to the pixel coordinates includes: Based on the camera extrinsic parameters of the vehicle-mounted camera that acquires the multi-view images, the pixel coordinates are transformed to the vehicle coordinate system to obtain the first coordinate data of the road elements in the vehicle coordinate system. Based on the vehicle trajectory data, the first coordinate data is converted to the world coordinate system to obtain the world coordinates corresponding to the road elements.

7. A map building device, characterized in that, The device includes: The acquisition module is used to acquire multi-view camera images of the target location in the target area, and determine the bird's-eye view image corresponding to the multi-view camera images with the target location as the origin; The recognition module is used to input the bird's-eye view image into the image segmentation model. In the image segmentation model, based on the learned correlation between image features and road elements, the road elements in the bird's-eye view image are segmented to obtain the segmentation mask corresponding to the road elements, and the pixel coordinates of the target position corresponding to the road elements are determined based on the segmentation mask. A construction module is used to determine the world coordinates corresponding to the pixel coordinates and construct a road element-level map of the target area based on the world coordinates.

8. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed, are used to implement the method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed, implements the method of any one of claims 1 to 6.

Images