Construction method, system and equipment of high-precision map and medium

Abstract

The invention discloses a construction method, system and equipment of a high-precision map and a medium. The construction method comprises the following steps: S1, generating a bird's-eye view feature map based on a static map with a standard resolution and sensor data; s2, further extracting a static feature Fsd and a dynamic environment feature Ft based on the bird's-eye view feature map; s3, based on a cross attention mechanism, fusing the static feature Fsd and the dynamic feature Ft; and S4, decoding the fused features into a high-precision map through a decoder. According to the invention, when the environment structure changes, the road information of the current frame can be reasoned by combining the high-definition map information of the previous frame with the standard map information, and the depth feature extracted by the convolutional neural network and the original standard map information are combined. The system can more accurately perceive and understand the environment around the vehicle, so that the overall performance and decision-making efficiency of the automatic driving system are improved.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving positioning technology, and in particular to a method, system, device and medium for constructing a high-precision map. Background Art

[0002] Many autonomous vehicle systems rely on information encoded in high-precision maps, which are transformed into the vehicle reference coordinate system at runtime to assist in decision making. High-precision maps typically represent the topology, geometry, and properties of the road. Such systems require precise knowledge of the vehicle's position relative to the map.

[0003] The creation of high-precision maps usually requires a lot of resources to oversee the map encoding and annotation process. In addition, as the structure of the environment changes, the map data usually becomes outdated, which requires expensive maintenance procedures. In the period between the change of the environment structure and the map update cycle, the system that relies on high-precision maps may have security risks.

[0004] Addressing this challenge requires a conceptually different approach to the perception stack for autonomous vehicles. Instead of relying on highly detailed maps, human drivers directly combine real-time scene understanding with knowledge from traditional standard-resolution street maps. Mimicking this capability requires new perception techniques that combine raw sensor data with coarse standard-resolution maps to obtain representations similar to current high-precision maps. However, a strategy to use a priori standard-resolution maps to assist HD map construction is currently lacking. Summary of the invention

[0005] The present invention provides a method, system, device and medium for constructing a high-precision map to solve the above-mentioned problems.

[0006] In one aspect, the present invention provides a method for constructing a high-precision map, the method comprising the following steps: step S1: generating a bird's-eye view feature map based on a static map of standard resolution and sensor data; step S2: further extracting static features F based on the bird's-eye view feature map sd and dynamic environment characteristics F t ; Step S3: Based on the cross attention mechanism, the static feature F sd and dynamic characteristics F t Step S4: The fused features are decoded into a high-precision map through a decoder.

[0007] In one implementation of the present invention, the step S1 specifically includes: step S11: rasterizing the static map of standard resolution to obtain the topological structure of the road and the geometric information and attribute information of the lane; step S12: inputting the rasterized map into a pre-trained convolutional neural network for deep feature extraction to abstract high-level semantic features; step S13: based on the extracted high-level semantic features, converting them into the bird's-eye view feature map.

[0008] In one implementation of the present invention, the method further includes: calculating a transformation matrix of a vehicle coordinate system according to the displacement information of the vehicle; mapping the bird's-eye view feature at the previous moment to the coordinate system at the current moment, thereby predicting the first bird's-eye view feature F at the current moment. pre ; Generate a second bird's-eye view feature at the current moment based on the camera image and the lidar point cloud; pre The bird's-eye view feature F is obtained by fusing it with the second bird's-eye view feature sd .

[0009] In one implementation of the present invention, the method further comprises: accumulating a set of N consecutive point clouds in the vehicle coordinate system Each point cloud combines the measurement data of all lidar sensors within a specified time interval [t-1, t], and each time interval is also associated with N sets of images from each of the M cameras; each point cloud is colored with associated RGB values ​​by projecting the points into the image coordinate system; Each colored point cloud is projected into the current vehicle coordinate system; the accumulated colored point cloud Projection into the bird’s-eye view feature frame is performed by discretizing the space around the vehicle into grids in the XY plane and aggregating the intensity, height, and RGB color values ​​in each grid.

[0010] In one implementation of the present invention, the method further includes: for intensity features, using mean aggregation; for RGB features, selecting the median to remove outliers; for height features, using the minimum value to filter out moving objects and generate an elevation map of the ground surface.

[0011] In one implementation of the present invention, the method further comprises: fusing the three obtained bird's-eye view feature maps using a convolutional neural network, and then fusing them with the bird's-eye view feature map F sd , the first bird's-eye view feature F pre Further fusion obtains the current bird's-eye view feature F t ; The current bird's-eye view feature F t .

[0012] In one implementation of the present invention, in step S3, the cross attention mechanism allows the model to dynamically adjust the focus area of ​​another feature map based on information from one feature map, where feature F t contains the dynamic environment information detected by the current sensor, feature F sd Provides static background information of road topology.

[0013] The present invention also provides a high-precision map construction system, the system comprising: a feature map generation module, used to generate a bird's-eye view feature map based on a static map of standard resolution and sensor data; a feature extraction module, used to further extract static features F based on the bird's-eye view feature map sd and dynamic environment characteristics F t ; Attention mechanism module, used to convert the static feature F sd and dynamic characteristics F t The decoding module is used to decode the fused features into a high-precision map through a decoder.

[0014] The present invention also provides a high-precision map construction device, the device comprising: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can complete: step S1: generating a bird's-eye view feature map based on a static map of standard resolution and sensor data; step S2: further extracting static features F based on the bird's-eye view feature map sd and dynamic environment characteristics F t ; Step S3: Based on the cross attention mechanism, the static feature F sd and dynamic characteristics F t Step S4: The fused features are decoded into a high-precision map through a decoder.

[0015] The present invention also provides a non-volatile computer storage medium for constructing a high-precision map, storing computer executable instructions, wherein the computer executable instructions are executed by a processor to implement: step S1: generating a bird's-eye view feature map based on a static map of standard resolution and sensor data; step S2: further extracting static features F based on the bird's-eye view feature map sd and dynamic environment characteristics F t ; Step S3: Based on the cross attention mechanism, the static feature F sd and dynamic characteristics F t Step S4: The fused features are decoded into a high-precision map through a decoder.

[0016] The present invention provides a method, system, device and medium for constructing a high-precision map, which has the following features:

[0017] Beneficial effects:

[0018] (1) By integrating temporal information, when the structure of the environment changes, the HD map information of the previous frame can be combined with the standard map information to infer the road information of the current frame to achieve real-time update of the HD map. At the same time, the model's perception of the dynamic environment and prediction accuracy are enhanced, especially in complex scenarios such as autonomous driving that require real-time decision-making;

[0019] (2) By combining the deep features extracted by the convolutional neural network with the original standard map information, the system can more accurately perceive and understand the environment around the vehicle, thereby improving the overall performance and decision-making efficiency of the autonomous driving system;

[0020] (3) The fusion method through the cross-attention mechanism not only improves the richness of feature expression, but also enhances the model's ability to understand complex scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0022] Figure 1 A flow chart of a method for constructing a high-precision map provided by an embodiment of the present invention;

[0023] Figure 2 An algorithm flow chart provided for an embodiment of the present invention;

[0024] Figure 3 The OSM city geographic information data map provided by the embodiment of the present invention;

[0025] Figure 4 A rasterized map provided by an embodiment of the present invention;

[0026] Figure 5 The experimental results and label comparison diagrams provided by the embodiments of the present invention; wherein (a), (b), (c), and (d) are label diagrams, and (e), (f), (g), and (h) are actual prediction diagrams;

[0027] Figure 6 A composition diagram of a high-precision map construction system provided by an embodiment of the present invention;

[0028] Figure 7 A schematic diagram of a high-precision map construction device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0030] The embodiments of the present invention provide a method, system, device and medium for constructing a high-precision map. The technical solution proposed in the embodiments of the present invention is described in detail below with reference to the accompanying drawings.

[0031] Figure 1 A flow chart of a method for constructing a high-precision map provided by an embodiment of the present invention. Figure 1 As shown, the method mainly includes the following steps:

[0032] Step S1: Generate a bird's-eye view feature map based on a static map of standard resolution and sensor data;

[0033] Step S2: further extracting static features F based on the bird's-eye view feature map sd and dynamic environment characteristics F t ;

[0034] Step S3: Based on the cross attention mechanism, the static feature F sd and dynamic characteristics F t to integrate;

[0035] Step S4: Decode the fused features into a high-precision map through a decoder.

[0036] The algorithm flow chart of the present invention is as follows Figure 2As shown, in an embodiment of the present invention, OpenStreetMap (OSM) is used as the input of the Standard Definition (SD) map encoding module. As an open source map platform, OSM contains rich information about various geographical features, such as roads, traffic signs, building areas, intersections, lanes, bridges, etc. These data provide key environmental prior knowledge for the autonomous driving system. First, the present invention uses the SD map information obtained by pre-positioning to obtain the topological structure of the road and the geometry and attribute information of the lane by rasterizing it. This process converts the vector data in OSM into a rasterized form suitable for neural network processing, which contains key information such as road layout, lane lines, and related road network connectivity. The rasterized SD map not only retains the semantic information of the original map, but also provides structured input for subsequent convolution operations. Then, the present invention inputs the rasterized SD map into a pre-trained convolutional neural network (CNN) for deep feature extraction. The CNN network is designed to capture important patterns and structural information in the map, such as the geometry of the road, the number of lanes, the location of traffic signs, and the relationship with the surrounding environment. Through multi-layer convolution operations, CNN can gradually abstract high-level semantic features of SD maps, thereby forming a global understanding of the road network and its topological relationships. The extracted features are then transformed into the Bird's-Eye-View (BEV) space to ensure alignment with sensor data (such as lidar point clouds and camera images) to maintain the spatial consistency of the data. This transformation not only retains the semantic, location and relationship information in the original OSM data, but also enhances the applicability of the features in the actual physical space. Finally, the present invention obtains a feature map Fsd that retains the key features in the OSM environmental annotations. The feature map carries rich map prior information, including road structure, intersections, and driving directions, etc., which provides a basis for the feature fusion process of subsequent tasks (such as multi-source data fusion, path planning, and perception decision-making). By combining the deep features extracted by CNN and the original SD map information, the system can more accurately perceive and understand the environment around the vehicle, thereby improving the overall performance and decision-making efficiency of the autonomous driving system.

[0037] Furthermore, the vehicle-mounted camera and lidar are selected as the input of the sensor data encoding module. The module first converts the lidar point cloud and camera image to the BEV perspective. In order to generate BEV features, the present invention first accumulates a set of N continuous point clouds in the vehicle coordinate system. Each point cloud combines the measurements of all lidar sensors within a specified time interval [t-1, t]. Each time interval is also associated with N sets of images from each of the M cameras whose fields of view cover a portion of the point cloud P that may overlap. VThen, we associate an RGB value to each point by projecting the point into the image coordinate system. Next, we Each colored point cloud is projected into the current vehicle coordinate system. Although N-1 transformations are required per frame, each point cloud P V Coloring is only required once. Finally, the accumulated colored point cloud The images are projected into the BEV frame by discretizing the space around the vehicle into grids in the XY plane and aggregating the intensity, height, and RGB color values ​​in each grid. The present invention also selects specific aggregation functions for different feature types. For example, for intensity features, the present invention uses mean aggregation; for RGB features, in order to remove outliers, the present invention selects the median; and for height features, the present invention uses the minimum value to effectively filter out moving objects and generate an elevation map of the ground surface. The three BEV feature maps are then fused using a convolutional neural network (CNN), and the fused BEV features are then combined with the F obtained in 1. sd , and the current BEV feature F predicted based on the BEV feature of the previous frame pre Further integration obtains the current BEV feature F t , which integrates dynamic features and static features (F sd ) and time series characteristics (F pre ), reflecting the environmental dynamics at the current moment.

[0038] Furthermore, the transformation matrix of the vehicle coordinate system is calculated based on the displacement information of the vehicle, and the matrix describes the position change of the vehicle between consecutive moments. Using the transformation matrix, the present invention maps the BEV features of the previous moment to the coordinate system of the current moment, thereby predicting the BEV features of the current moment. This process is equivalent to encoding time information into the BEV features, so that the model can use historical information to speculate on future scenes, thereby improving the accuracy of the prediction. After predicting the BEV features of the current moment, the present invention further fuses them with the BEV features of the current moment generated based on camera images and lidar point clouds. Camera images provide rich texture and color information, while lidar point clouds provide accurate three-dimensional geometric structures. By fusing these multi-source data, the model can obtain a BEV feature representation with high detail and accuracy. In addition, the BEV feature F extracted from the SD map sd It also participates in the fusion as a priori information, providing the model with global clues such as terrain and road structure in a static environment. Finally, after this multi-level feature fusion, the present invention obtains the BEV feature F at the current moment. tThis fusion mechanism not only effectively integrates data from multiple sources and time steps, but also greatly enhances the model's perception of dynamic environments and prediction accuracy, especially in complex scenarios such as autonomous driving that require real-time decision-making.

[0039] Furthermore, in the decoder module, the present invention uses a cross-attention mechanism to fuse F t and F sd During decoding, the cross-attention mechanism allows the model to dynamically adjust the focus area of ​​another feature map based on information from one feature map. t contains the dynamic environment information detected by the current sensor, and F sd It provides static background information such as road topology. The cross-attention mechanism can learn to weigh this information at different times and spatial locations to generate more accurate and coherent fusion features. This fusion method not only improves the richness of feature expression, but also enhances the model's ability to understand complex scenes. In multi-task learning, the cross-attention mechanism can also ensure that the model can balance feature inputs from different sources when processing multiple targets, effectively reducing feature conflicts or information loss.

[0040] Finally, in the process of generating high-precision HD maps, the model will further process the decoded fused features. Specifically, the fused features output by the decoder will undergo a series of spatial and semantic parsing operations to integrate dynamic and static information into a unified BEV map. By finely segmenting and detecting roads, lane lines, signs, and other traffic elements, the model can generate HD maps that contain topological structures, road surface information, and lane boundaries. This information can be rendered through visualization tools, or used as the basic data of HD maps for further application in path planning, obstacle detection, and positioning modules for autonomous driving.

[0041] The following is a specific example of the present invention, which was experimented on an autonomous driving platform. The experimental platform uses a multi-sensor configuration, including a laser radar, a multi-view camera on a vehicle, and a standard resolution (SD) map obtained through OpenStreetMap (OSM). The system is designed to improve the perception of dynamic road scenes by autonomous vehicles, especially in complex urban traffic environments.

[0042] In this experiment, the present invention downloaded the geographic information data of a city center area from OpenStreetMap (OSM), such as Figure 3 The data includes road topology, lane lines, traffic signs, buildings and other information. The data is first rasterized to convert the vector road and lane information into a two-dimensional raster map suitable for neural network processing. The rasterized SD map contains the road structure, lane attributes and the location of key traffic facilities, such as Figure 4 shown.

[0043] In order to capture the key features in the map, the present invention inputs these rasterized maps into a pre-trained convolutional neural network (CNN) for feature extraction. CNN extracts high-level semantic features of road topology through multi-layer convolution operations, including global information such as lane geometry, road connections, and important intersections. These features are projected into the bird's-eye view (BEV) space to generate a feature map F sd , used for subsequent feature fusion.

[0044] In the experiment, the vehicle was equipped with a 16-line laser radar and a multi-view camera to perceive the environment around the vehicle. The sensor data was accumulated over multiple frames to generate a three-dimensional point cloud and multiple sets of camera images. First, the present invention registered the point cloud data of the laser radar with the camera image. Each point cloud was projected into the camera image and colored with RGB values ​​to generate point cloud data with texture information.

[0045] Then, these point clouds with color information are projected into the BEV coordinate system to form a BEV feature map generated by the lidar and the camera. For specific features such as intensity, RGB and height, the present invention uses mean, median and minimum aggregation respectively to reduce the influence of outliers and generate reliable BEV feature maps.

[0046] The 3D feature map is input into the convolutional neural network to further fuse different feature channels and generate the BEV feature F at the current moment. t .

[0047] In order to enhance the model's understanding of dynamic scenes, the present invention designs a temporal fusion mechanism. In this mechanism, the present invention calculates the transformation matrix between the ego vehicle coordinate system through the vehicle's odometer data, and uses this matrix to map the BEV features of the previous moment to the current moment. This step generates a predicted BEV feature F pre , which contains historical information.

[0048] Then, the predicted BEV feature F pre The BEV feature F generated by the sensor data t and the feature F from the SD map sd Multi-source fusion is performed. Camera images provide rich environmental details, LiDAR ensures the accuracy of 3D geometric information, and SD map features provide global road topology information.

[0049] By fusing these multi-source features, the model can accurately perceive the dynamic environment around the vehicle, especially in complex traffic intersections, lane changes, etc., the system exhibits extremely high stability.

[0050] In the decoder module, the present invention realizes the fusion of the current dynamic environment and static map information through the cross attention mechanism. In the decoding stage, the cross attention mechanism allows the model to dynamically adjust the feature map F t and F sd For example, when a vehicle enters an intersection, the model automatically pays more attention to the road topology information provided in the SD map through the attention mechanism, while on straight roads, it relies more on the dynamic environment information generated by the sensor.

[0051] This attention mechanism significantly enhances the expressiveness of features, enabling the model to balance multi-source input features, reduce feature conflicts, and ensure the integrity of information transfer. In experiments, this cross-attention mechanism enables the model to more efficiently perform multi-task processing in complex scenarios, such as road detection, lane detection, and obstacle perception. Figure 5 The comparison between the experimental prediction algorithm and the label graph of the present invention is shown. (a), (b), (c), and (d) are label graphs, and (e), (f), (g), and (h) are actual prediction graphs. The label graph is what should be generated, and the actual prediction graph is what is actually generated.

[0052] Through this specific embodiment, the present invention verifies the effectiveness of the feature fusion method using OSM maps as environmental prior knowledge, combined with sensor data and cross-attention mechanism in autonomous driving scenarios. This method not only improves the model's perception and understanding of complex scenes, but also enhances the overall performance of the system in multi-tasking.

[0053] The above is a method for constructing a high-precision map provided by an embodiment of the present invention. Based on the same inventive concept, an embodiment of the present invention also provides a system for constructing a high-precision map. Figure 6 A composition diagram of a high-precision map construction system provided by an embodiment of the present invention, such as Figure 6 As shown, the system mainly includes: a feature map generation module 601, which is used to generate a bird's-eye view feature map based on a static map of standard resolution and sensor data; a feature extraction module 602, which is used to further extract static features F based on the bird's-eye view feature map sd and dynamic environment characteristics F t ; Attention mechanism module 603, used to convert the static feature F sd and dynamic characteristics F t The fusion is performed; a decoding module 604 is used to decode the fused features into a high-precision map through a decoder.

[0054] The above is a high-precision map construction system provided by an embodiment of the present invention. Based on the same inventive concept, an embodiment of the present invention also provides a high-precision map construction device. Figure 7 A schematic diagram of a high-precision map construction device provided by an embodiment of the present invention, such as Figure 7 As shown, the device mainly includes: at least one processor 701; and a memory 702 connected to the at least one processor in communication; wherein the memory 702 stores instructions that can be executed by the at least one processor 701, and the instructions are executed by the at least one processor 701, so that the at least one processor 701 can complete: step S1: generating a bird's-eye view feature map based on a static map of standard resolution and sensor data; step S2: further extracting static features F based on the bird's-eye view feature map sd and dynamic environment characteristics F t ; Step S3: Based on the cross attention mechanism, the static feature F sd and dynamic characteristics F t Step S4: The fused features are decoded into a high-precision map through a decoder.

[0055] In addition, the embodiment of the present invention further provides a non-volatile computer storage medium for constructing a high-precision map, which stores computer executable instructions, and the computer executable instructions are executed by a processor to implement: step S1: generating a bird's-eye view feature map based on a static map of standard resolution and sensor data; step S2: further extracting static features F based on the bird's-eye view feature map sd and dynamic environment characteristics F t ; Step S3: Based on the cross attention mechanism, the static feature F sd and dynamic characteristics F t Step S4: The fused features are decoded into a high-precision map through a decoder.

[0056] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0057] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0058] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0059] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0060] The various embodiments of the present invention 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 embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0061] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0062] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for constructing a high-precision map, characterized in that: The method comprises the following steps: Step S1: Generate a bird's-eye view feature map based on a static map of standard resolution and sensor data; Step S2: further extracting static features F based on the bird's-eye view feature map sd and dynamic environment characteristics F t ; Step S3: Based on the cross attention mechanism, the static feature F sd and dynamic characteristics F t to integrate; Step S4: Decode the fused features into a high-precision map through a decoder.

2. A method for constructing a high-precision map according to claim 1, characterized in that: In the step S1, it specifically includes: Step S11: rasterizing the static map of standard resolution to obtain the topological structure of the road and the geometric information and attribute information of the lanes; Step S12: inputting the rasterized map into a pre-trained convolutional neural network for deep feature extraction to abstract high-level semantic features; Step S13: based on the extracted high-level semantic features, converting into the bird's-eye view feature map.

3. The method for constructing a high-precision map according to claim 1, characterized in that: The method further comprises: Calculate the transformation matrix of the vehicle coordinate system according to the displacement information of the vehicle; Map the bird's-eye view feature of the previous moment to the coordinate system of the current moment to predict the first bird's-eye view feature F at the current moment pre ; Generate a second bird's-eye view feature at the current moment based on the camera image and the lidar point cloud; The first bird's-eye view feature F pre The bird's-eye view feature F is obtained by fusing it with the second bird's-eye view feature sd .

4. The method for constructing a high-precision map according to claim 1, characterized in that: The method further comprises: Accumulate a set of N consecutive point clouds in the vehicle coordinate system Each point cloud combines the measurements of all lidar sensors within a specified time interval [t-1, t], and each time interval is also associated with N sets of images from each of the M cameras; By projecting the points into the image coordinate system, each point cloud is colored with associated RGB values; Will Each colored point cloud is projected into the current vehicle coordinate system; The accumulated colored point cloud Projection into the bird’s-eye view feature frame is performed by discretizing the space around the vehicle into grids in the XY plane and aggregating the intensity, height, and RGB color values ​​in each grid.

5. A method for constructing a high-precision map according to claim 4, characterized in that: The method further comprises: For intensity features, mean aggregation was used; For RGB features, the median is selected to remove outliers; For height features, the minimum value is used to filter out moving objects and generate an elevation map of the ground surface.

6. A method for constructing a high-precision map according to claim 5, characterized in that: The method further comprises: The three bird’s-eye view feature maps are fused using a convolutional neural network and then combined with the bird’s-eye view feature F sd , the first bird's-eye view feature F pre Further fusion obtains the current bird's-eye view feature F t ; The current bird's-eye view feature F t .

7. The method for constructing a high-precision map according to claim 1, characterized in that: In step S3, the cross attention mechanism allows the model to dynamically adjust the focus area of ​​another feature map based on information from one feature map. The feature F t contains the dynamic environment information detected by the current sensor, feature F sd Provides static background information of road topology.

8. A high-precision map construction system, characterized in that: The system comprises: A feature map generation module is used to generate a bird's-eye view feature map based on a static map of standard resolution and sensor data; A feature extraction module is used to further extract static features F based on the bird's-eye view feature map. sd and dynamic environment characteristics F t ; The attention mechanism module is used to convert the static features F sd and dynamic characteristics F t to integrate; The decoding module is used to decode the fused features into a high-precision map through a decoder.

9. A high-precision map construction device, characterized in that: The device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to complete: Step S1: Generate a bird's-eye view feature map based on a static map of standard resolution and sensor data; Step S2: further extracting static features F based on the bird's-eye view feature map sd and dynamic environment characteristics F t ; Step S3: Based on the cross attention mechanism, the static feature F sd and dynamic characteristics F t to integrate; Step S4: Decode the fused features into a high-precision map through a decoder.

10. A non-volatile computer storage medium for constructing a high-precision map, storing computer executable instructions, characterized in that: The computer executable instructions are executed by a processor to implement: Step S1: Generate a bird's-eye view feature map based on a static map of standard resolution and sensor data; Step S2: further extracting static features F based on the bird's-eye view feature map sd and dynamic environment characteristics F t ; Step S3: Based on the cross attention mechanism, the static feature F sd and dynamic characteristics F t to integrate; Step S4: Decode the fused features into a high-precision map through a decoder.

Images