A two-stage high-precision three-dimensional point cloud semantic map construction method

Abstract

The application provides a two-stage high-precision three-dimensional point cloud semantic map construction method, steps are as follows: obtaining a first point cloud data set based on a three-dimensional point cloud map, a current vehicle pose and a preset radius; segmenting a current vehicle end image to obtain first visual semantic information by a segmentation model; performing first noise removal processing on the first point cloud data set and the first visual semantic information to obtain a sub-rough point cloud semantic data set; fusing at least one sub-rough point cloud semantic data set to obtain a rough point cloud semantic data set; performing second noise removal processing on the rough point cloud semantic data set based on second visual semantic information to obtain a sub-three-dimensional point cloud semantic map, and obtaining a three-dimensional point cloud semantic map according to at least one sub-three-dimensional point cloud semantic map. The method gradually improves the precision of the three-dimensional point cloud semantic map and the accuracy of the semantic information in a staged manner, breaks through the precision and efficiency problems that may exist in single-stage processing, and effectively improves the robustness and operability of the overall system.

Description

Technical Field

[0001] This invention relates to the field of point cloud processing technology, specifically to a two-stage high-precision three-dimensional point cloud semantic map construction method. Background Technology

[0002] In the field of autonomous driving, high-precision 3D point cloud maps are one of the core technologies for achieving accurate positioning and navigation. High-precision 3D point cloud maps provide detailed environmental information for autonomous driving systems, helping them perceive their surroundings in real time. However, 3D point cloud data itself only contains spatial coordinate information and cannot provide semantic information about the environment. To improve the map's usability and intelligence, it is often necessary to combine it with semantic information for further processing. This semantic information is usually extracted through image recognition, deep learning, or manual annotation, and then fused with 3D point cloud data to generate a 3D point cloud map with rich semantic information. A 3D point cloud map with semantic information not only provides accurate environmental understanding for systems such as autonomous driving and robot navigation, but also enhances the system's adaptability to complex situations.

[0003] In existing technologies, there are some problems and challenges in the construction of 3D point cloud maps and the integration of semantic information, especially in terms of processing accuracy, computational efficiency and adaptability to complex environments. Summary of the Invention

[0004] This invention addresses the problems existing in the prior art by providing a two-stage high-precision 3D point cloud semantic map construction method that gradually improves the accuracy of the semantic information of the 3D point cloud semantic map through a phased approach. This method overcomes the accuracy and efficiency problems that may exist in single-stage processing and effectively improves the robustness and operability of the overall system.

[0005] To achieve the above objectives, the technical solution adopted by this invention is as follows: A two-stage high-precision three-dimensional point cloud semantic map construction method, comprising the following steps:

[0006] Obtain the current vehicle pose of the autonomous vehicle;

[0007] Based on the 3D point cloud map, a first point cloud dataset is obtained according to the current vehicle pose and a preset radius;

[0008] The current vehicle-side image is segmented based on a deep learning image segmentation model to obtain first visual semantic information;

[0009] A sub-coarse point cloud semantic dataset is obtained based on the first point cloud dataset and the first visual semantic information;

[0010] When the current vehicle pose exceeds a preset range: fuse at least one of the sub-coarse point cloud semantic datasets to obtain a coarse point cloud semantic dataset; obtain second visual semantic information based on the first visual semantic information;

[0011] A sub-3D point cloud semantic map is obtained based on the second visual semantic information and the coarse point cloud semantic dataset.

[0012] When the current vehicle pose is outside the range of the 3D point cloud map: obtain a 3D point cloud semantic map based on at least one of the sub-3D point cloud semantic maps.

[0013] In some embodiments, the step of obtaining a sub-coarse point cloud semantic dataset based on the first point cloud dataset and the first visual semantic information includes:

[0014] Obtain first device parameters and second device parameters, wherein the first device parameters are parameters of the first device, the first device being a device used to acquire the original point cloud data, and the second device parameters are parameters of the second device, the second device being a device used to acquire the current vehicle-end image;

[0015] Based on the first visual semantic information, a second point cloud dataset is obtained by removing outliers from the first point cloud dataset according to the first device parameters and the second device parameters.

[0016] A first point cloud semantic dataset is obtained based on the first visual semantic information and the second point cloud dataset.

[0017] The first outlier processing is performed on the first point cloud semantic dataset to obtain the sub-coarse point cloud semantic dataset.

[0018] In some embodiments, the step of performing the first outlier processing on the first point cloud semantic dataset to obtain the sub-coarse point cloud semantic dataset is as follows:

[0019] Based on the first point cloud semantic dataset, obtain the depth information and Z-axis distribution information of the point cloud data;

[0020] Based on the depth information and the Z-axis distribution information, the second point cloud semantic dataset is obtained by removing outliers from the first point cloud semantic dataset.

[0021] The second point cloud semantic dataset is clustered using a clustering algorithm, and the largest clustering region is extracted to obtain the sub-coarse point cloud semantic dataset.

[0022] In some embodiments, the step of obtaining the first point cloud semantic dataset based on the first visual semantic information and the second point cloud dataset is as follows:

[0023] At least one first key semantic category is obtained based on the first visual semantic information;

[0024] The first point cloud semantic dataset is obtained by labeling each point cloud data in the second point cloud dataset with the first key semantic category.

[0025] In some embodiments, the step of obtaining a sub-3D point cloud semantic map based on the second visual semantic information and the coarse point cloud semantic dataset is as follows:

[0026] A third point cloud semantic dataset is obtained by reducing the point cloud density in the coarse point cloud semantic dataset.

[0027] At least one second key semantic category is obtained based on the second visual semantic information;

[0028] Based on the at least one second key semantic category, a clustering algorithm is used to cluster the third point cloud semantic dataset and extract dense regions to obtain at least one first sub-point cloud semantic dataset.

[0029] Each first sub-point cloud semantic dataset is subjected to intensity filtering to obtain at least one second sub-point cloud semantic dataset.

[0030] At least one intermediate 3D point cloud semantic map is obtained by characterizing each second sub-point cloud semantic dataset using a line model.

[0031] The sub-3D point cloud semantic map is obtained by fusing the at least one intermediate 3D point cloud semantic map.

[0032] In some embodiments, the key semantic categories include curbs, lane lines, poles, and tree trunks.

[0033] In some embodiments, point cloud data corresponding to different first key semantic categories are labeled with different colors.

[0034] In some embodiments, the clustering algorithm is the DBSCAN clustering algorithm.

[0035] In some embodiments, when obtaining the first point cloud dataset based on the three-dimensional point cloud map and according to the current vehicle pose and preset radius, the three-dimensional point cloud map is constructed as a multi-dimensional spatial index structure.

[0036] In some embodiments, the timestamp of the current vehicle pose and the timestamp of the current vehicle-end image are consistent. Compared with the prior art, the present invention has the following advantages:

[0037] This disclosure combines point cloud data acquisition with visual semantic information segmentation to effectively construct a high-precision 3D point cloud semantic map with rich visual semantic information. Through two stages of coarse fusion and fine processing, it successfully achieves fine filtering and noise removal of point cloud data, improving the map's accuracy and expressive power. The resulting high-precision 3D point cloud semantic map not only possesses high-precision geometric information but also includes key semantic categories in the environment, enhancing the autonomous driving system's understanding and decision-making capabilities in complex environments and providing a solid foundation for high-precision positioning of subsequent autonomous vehicles.

[0038] This disclosure improves the accuracy of 3D point cloud semantic maps and semantic information through a phased approach, overcoming the accuracy and efficiency problems that may exist in single-stage processing.

[0039] This disclosure generates intermediate results during the two-stage processing, facilitating real-time monitoring and debugging, quickly locating and resolving potential problems, thereby ensuring stable system operation. This step-by-step optimization approach effectively improves the overall system robustness and operability. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the two-stage high-precision 3D point cloud semantic map construction method disclosed in this paper.

[0041] Figure 2 This is a flowchart illustrating a two-stage high-precision three-dimensional point cloud semantic map construction method for constructing a three-dimensional point cloud map into a multi-dimensional spatial index structure, as described in this embodiment of the disclosure.

[0042] Figure 3 This is a flowchart illustrating the coarse fusion stage in an embodiment of this disclosure;

[0043] Figure 4 This is a flowchart illustrating the detailed processing stage in an embodiment of this disclosure;

[0044] Figure 5 This refers to the 3D point cloud map constructed in the embodiments of this disclosure;

[0045] Figure 6 This is a 3D point cloud map formed from the first point cloud semantic dataset in this embodiment of the present disclosure;

[0046] Figure 7 This is a 3D point cloud map formed from a coarse point cloud semantic dataset in this embodiment of the disclosure;

[0047] Figure 8 This is a high-precision three-dimensional point cloud semantic map in the embodiments of this disclosure. Detailed Implementation

[0048] To address the challenge of constructing 3D point cloud semantic maps with high processing accuracy and computational efficiency in complex environments, current techniques for building 3D point cloud maps with semantic information can be broadly categorized into the following three methods:

[0049] 1. Methods for manually drawing semantic information

[0050] This method first acquires high-precision 3D point cloud data using sensors such as LiDAR, and then manually annotates this data semantically. The advantage of this method is that it ensures the accuracy of semantic information because the annotation is done manually, guaranteeing high precision. However, the disadvantages are also very obvious: manual annotation is labor-intensive, inefficient, and highly dependent on human experience and time. Furthermore, this method struggles to cope with large-scale, dynamically changing environments and cannot provide real-time updates, making it unsuitable for scenarios requiring rapid response in autonomous driving systems.

[0051] 2. A Deep Learning-Based Semantic Segmentation Method for 3D Point Clouds

[0052] This method utilizes deep learning networks to automatically identify semantic information in point clouds, reducing the workload of manual annotation. However, 3D point cloud data is typically sparse and irregular, making direct semantic segmentation extremely difficult. Currently, deep learning models exhibit low accuracy on 3D point clouds, especially when faced with noise, occlusion, and viewpoint variations, posing significant challenges to accuracy improvement. Furthermore, training deep learning models requires large amounts of high-quality labeled data and incurs substantial computational costs, making it difficult to meet the demands of real-time applications. This leads to a method based on the fusion of 3D point cloud maps and visual information.

[0053] 3. A method based on the fusion of 3D point cloud maps and visual semantic information

[0054] This method incorporates camera sensors to acquire visual semantic information in addition to the 3D point cloud data provided by LiDAR. A deep learning network combines camera and LiDAR data for semantic detection and segmentation, integrating the semantic information into the 3D point cloud map. This approach combines the advantages of visual information and LiDAR, providing richer and more accurate semantic annotations, particularly effective in complex environments. Compared to segmenting 3D point clouds separately, image segmentation algorithms are more mature, with simpler and more accurate annotation processes, and are easier to operate. However, achieving efficient fusion between the two sensors remains a technical challenge, impacting the quality of the final map construction.

[0055] Currently, the industry mainly uses methods 1 and 3. Based on method 3, this disclosure proposes an improved solution. To clearly illustrate the technical features of this solution, the implementation methods of this application will be described in detail below with reference to the accompanying drawings and embodiments. This will allow for a full understanding and implementation of how this application uses technical means to solve technical problems and achieve corresponding technical effects. The embodiments of this application and the various features within those embodiments can be combined with each other without conflict, and the resulting technical solutions are all within the protection scope of this application.

[0056] See Figure 1 This disclosure provides a two-stage high-precision 3D point cloud semantic map construction method, including the following steps:

[0057] Constructing a 3D point cloud map; the 3D point cloud map constructed in this embodiment is shown in [reference needed]. Figure 5 It is possible to collect raw data related to urban roads through autonomous vehicles, or to collect raw data related to urban roads through other methods and pre-build a 3D point cloud map. Specifically, by using the traditional Simultaneous Localization and Mapping (SLAM) algorithm, the raw data related to urban roads can be input into the SLAM algorithm to build a 3D point cloud map. The 3D point cloud map is usually a 3D point cloud map of the urban area where the vehicle is currently located. Each point cloud data in the 3D point cloud map carries global coordinate information in the UTM coordinate system.

[0058] The system acquires raw point cloud data and current vehicle-side images based on autonomous vehicles. The raw point cloud data records the current vehicle pose on a 3D point cloud map. The current vehicle pose is the vehicle's pose at the current moment, i.e., the vehicle's coordinates and attitude information in the UTM coordinate system at the current moment. The real-time vehicle pose is the current vehicle pose across consecutive frames, i.e., a collection of vehicle coordinates and attitude information in the UTM coordinate system across consecutive frames at the current moment. The current vehicle-side image is a collection of images acquired by a camera installed on the vehicle. In some embodiments, the timestamps of the raw point cloud data (i.e., the current vehicle pose) and the current vehicle-side image are consistent, facilitating subsequent data fusion (i.e., the fusion of point cloud data and visual semantic information) to ensure temporal consistency in subsequent data fusion.

[0059] Obtain the preset range and preset radius; divide the 3D point cloud map into at least one preset range;

[0060] The current vehicle pose of the autonomous vehicle is obtained on a 3D point cloud map based on the raw point cloud data.

[0061] Based on a 3D point cloud map, a first point cloud dataset is obtained according to the current vehicle pose and a preset radius. Local point cloud data corresponding to the vehicle coordinates in the current frame (i.e., the current moment) is quickly found in the 3D point cloud map. Taking the vehicle coordinates in the current frame as the center, a local area within a preset radius is calculated as the region of interest. The preset radius is usually 50 meters, but it can be adjusted according to needs. The point cloud data corresponding to this local area is extracted from the 3D point cloud map to obtain the first point cloud dataset. Each frame of the current vehicle pose corresponds to a first point cloud dataset, providing local information for subsequent processing.

[0062] See Figure 2 In some embodiments, when obtaining the first point cloud dataset based on a 3D point cloud map, the current vehicle pose, and a preset radius, the 3D point cloud map is constructed into a multi-dimensional spatial index structure, i.e., a multi-dimensional spatial index structure is obtained based on the 3D point cloud map. Constructing the 3D point cloud map into a multi-dimensional spatial index structure can accelerate the search and processing efficiency of point cloud data corresponding to the current vehicle pose and the region of interest in the 3D point cloud map. Preferably, the multi-dimensional spatial index structure is a KDTree.

[0063] The first visual semantic information is obtained by segmenting the current vehicle image based on a deep learning image segmentation model. The first visual semantic information is the visual semantic information obtained by segmenting from the current vehicle image. The deep learning image segmentation model is obtained through pre-training. Preferably, the deep learning image segmentation model is a semantic segmentation network based on a convolutional neural network (CNN) that is pre-trained. This first visual semantic information will serve as supplementary information for subsequent map processing, enhancing the semantic expression capability of the map and improving the autonomous driving system's understanding of the environment.

[0064] A sub-coarse point cloud semantic dataset is obtained based on the first point cloud dataset and the first visual semantic information; specifically, the first point cloud dataset and the first visual semantic information are coarsely fused and a first noise removal process is performed to obtain the sub-coarse point cloud semantic dataset; the purpose of the first noise removal process is to retain valuable information, that is, to initially retain point cloud data related to the first visual semantic information and to initially remove outliers and noise outside the region of interest.

[0065] See Figure 7 When the current vehicle pose exceeds the preset range: fuse at least one sub-coarse point cloud semantic dataset to obtain a coarse point cloud semantic dataset; obtain second visual semantic information based on first visual semantic information; obtain one first visual semantic information for each frame of the current vehicle image, obtain at least one first visual semantic information within the same preset range, and obtain the second visual semantic information by taking the union of at least one first visual semantic information.

[0066] The above describes the coarse fusion stage, which obtains a coarse point cloud semantic dataset through preliminary screening, and then proceeds to the fine processing stage. The goal of the fine processing stage is to further remove noise points, refine the effective information in the point cloud data, and generate the final high-precision 3D point cloud semantic map. The fine processing stage involves obtaining a sub-3D point cloud semantic map based on the second visual semantic information and the coarse point cloud semantic dataset. Specifically, the second visual semantic information is used to perform a second noise removal process on the coarse point cloud semantic dataset to obtain the sub-3D point cloud semantic map. See [link to relevant documentation]. Figure 8 The purpose of the second noise removal process is to further remove point cloud data that is irrelevant to the second visual semantic information, and to further remove outliers;

[0067] When the current vehicle pose exceeds the range of the 3D point cloud map: a 3D point cloud semantic map is obtained based on at least one sub-3D point cloud semantic map. Each preset range corresponds to a sub-3D point cloud semantic map. When the autonomous vehicle travels through the entire range of the 3D point cloud map, at least one sub-3D point cloud semantic map is combined to obtain a 3D point cloud semantic map.

[0068] The two-stage high-precision 3D point cloud semantic map construction method disclosed herein improves the accuracy of the map and the visual semantic information step by step, overcoming the accuracy and efficiency problems that may exist in single-stage processing. This method generates intermediate results during the two-stage processing, facilitating real-time monitoring and debugging, quickly locating and resolving potential problems, thereby ensuring stable system operation. This step-by-step optimization approach effectively improves the robustness and operability of the overall system.

[0069] See Figure 3 In some embodiments, the step of coarsely fusing the first point cloud dataset and the first visual semantic information and performing a first noise removal process to obtain a sub-coarse point cloud semantic dataset includes:

[0070] The system acquires first device parameters and second device parameters. The first device parameters are the parameters of the first device, which is a device used to acquire raw point cloud data. The first device is a high-precision sensor, including but not limited to one or more of LiDAR, IMU, or RTK. The first device is used to collect raw point cloud data. The raw point cloud data is used to construct a 3D point cloud map using a traditional SLAM algorithm, and the current vehicle pose is recorded. The second device parameters are the parameters of the second device, which is a device used to acquire the current vehicle-side image. Typically, the second device includes but is not limited to one or more of a camera, vehicle-mounted camera, 3D laser scanner, or sensor.

[0071] Based on the first visual semantic information, the second point cloud dataset is obtained by removing outliers from the first point cloud dataset according to the first device parameters and the second device parameters; the first outlier is a point that is not related to the first visual semantic information.

[0072] See Figure 6 The first point cloud semantic dataset is obtained based on the first visual semantic information and the second point cloud dataset; the first visual semantic information corresponding to each point cloud data in the second point cloud dataset is fused to obtain the first point cloud semantic dataset. After the fusion is completed, within a preset range, each second point cloud dataset corresponds to a first point cloud semantic dataset.

[0073] The first outlier processing is performed on the first point cloud semantic dataset to obtain the sub-coarse point cloud semantic dataset. The first outlier processing is to initially remove outliers and noise outside the region of interest.

[0074] In some embodiments, the step of performing a first outlier processing on the first point cloud semantic dataset to obtain a sub-coarse point cloud semantic dataset is as follows:

[0075] Depth information and Z-axis distribution information of point cloud data are obtained based on the first point cloud semantic dataset;

[0076] The second point cloud semantic dataset is obtained by removing outliers from the first point cloud semantic dataset based on depth information and Z-axis distribution information. By statistically analyzing the depth information and Z-axis distribution information (Z-axis distribution range) of the point cloud data in the first point cloud semantic dataset, the depth difference of the point cloud data in the first point cloud semantic dataset is obtained based on the depth information. The depth difference and Z-axis distribution range are used to filter the point cloud data in the first point cloud semantic dataset, removing outliers and point cloud data that are not in the region of interest. That is, the second outliers are points that are not in the region of interest, thus achieving the removal of the second outliers to obtain the second point cloud semantic dataset.

[0077] A clustering algorithm is used to cluster the second point cloud semantic dataset and extract the largest clustering region to obtain the sub-coarse point cloud semantic dataset. After extracting the largest clustering region, sparsely distributed point cloud data, corresponding first visual semantic information, and noise points are removed to ensure that the point cloud data retained in the sub-coarse point cloud semantic dataset is representative point cloud data. Preferably, the clustering algorithm is the DBSCAN clustering algorithm.

[0078] In some embodiments, the step of obtaining at least one first point cloud semantic dataset based on first visual semantic information and second point cloud dataset is as follows:

[0079] At least one first key semantic category is obtained based on first visual semantic information; in some embodiments, the first key semantic category includes one or more of curbs, lane lines, poles, or tree trunks; currently, most 3D point cloud semantic maps mainly rely on a limited number of key semantic categories such as lane lines, landmarks, and poles; these key semantic categories perform well in scenarios with clear visual semantic information (e.g., elevated roads), but often limit vehicle traffic capacity on complex urban and rural roads where poles are scarce or lane lines are unclear. To address this limitation, this disclosure introduces more types of visual semantic information, such as tree trunks and curbs, in the pre-setting stage. By integrating multiple key semantic categories, this disclosure significantly expands the applicable traffic area for vehicles, improves the adaptability and versatility of the map, especially in complex environments;

[0080] The first point cloud semantic dataset is obtained by labeling each point cloud data in the second point cloud dataset with the first key semantic category. The rich and diverse visual semantic information is fused with the point cloud data to expand the vehicle's passage area.

[0081] In the first stage, the coarse fusion stage, a coarse data processing and fusion strategy is employed to quickly generate a preliminary 3D point cloud map and semantic information. The main goal of this stage is to remove irrelevant point cloud data, retaining only the regions of interest containing primary visual semantic information, although some noise may still exist. In this way, a basic 3D point cloud semantic map can be constructed in a relatively short time, enabling rapid response to environmental changes.

[0082] In the second stage, the fine-tuning stage, refined algorithms are used to further optimize the accuracy of the 3D point cloud semantic map, remove noise, and cluster the remaining data using second-vision semantic information. This allows for the formation of independent 3D point cloud semantic elements, and their geometric structures are characterized, ensuring high accuracy and structural integrity while reducing storage pressure, making downstream tasks more efficient and user-friendly.

[0083] See Figure 4 In some embodiments, the step of performing a second noise removal process on the coarse point cloud semantic dataset based on the second visual semantic information to obtain a sub-3D point cloud semantic map is as follows:

[0084] A third point cloud semantic dataset is obtained by reducing the point cloud density in the coarse point cloud semantic dataset; preferably, voxel filtering is used to reduce the point cloud density.

[0085] At least one second key semantic category is obtained based on the second visual semantic information; similarly, the second key semantic category includes one or more of the following: curb, lane line, pole, or tree trunk.

[0086] Based on at least one second key semantic category, a clustering algorithm is used to cluster the third point cloud semantic dataset and extract dense regions to obtain at least one first sub-point cloud semantic dataset; a clustering algorithm is used to cluster the point cloud data of each second key semantic category, extract dense regions and remove scattered noise points to obtain at least one first sub-point cloud semantic dataset, the number of first sub-point cloud semantic datasets corresponding to the number of key semantic categories; preferably, the clustering algorithm is the DBSCAN clustering algorithm;

[0087] Each first sub-point cloud semantic dataset is subjected to intensity filtering to obtain at least one second sub-point cloud semantic dataset. During the intensity filtering process, outlier points whose intensity is inconsistent with that of the majority of points are removed to ensure that valid point cloud data is retained in the second sub-point cloud semantic dataset.

[0088] The linear model is used to characterize each second sub-point cloud semantic dataset to obtain at least one sub-3D point cloud semantic map; linear structure information is extracted during the characterization process to generate an intermediate 3D point cloud semantic map.

[0089] At least one intermediate 3D point cloud semantic map is fused to obtain a sub-3D point cloud semantic map.

[0090] In some embodiments, point cloud data corresponding to different first key semantic categories are labeled with different colors, which can more intuitively distinguish semantic objects corresponding to different key semantic categories and facilitate subsequent analysis applications. This step usually requires calling a code interpreter tool and writing Python code to implement the color labeling of point cloud data.

[0091] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention do not depart from the essence and scope of the technical solution of the present invention.

Claims

1. A two-stage high-precision 3D point cloud semantic map construction method, characterized in that: Includes the following steps: Obtain the current vehicle pose of the autonomous vehicle; Based on a 3D point cloud map, a first point cloud dataset is obtained according to the current vehicle pose and a preset radius; The first visual semantic information is obtained by segmenting the current vehicle-side image based on a deep learning image segmentation model; A sub-coarse point cloud semantic dataset is obtained based on the first point cloud dataset and the first visual semantic information; the first point cloud dataset and the first visual semantic information are coarsely fused and a first noise removal process is performed to obtain the sub-coarse point cloud semantic dataset; the first noise removal process is to retain point cloud data related to the first visual semantic information and remove outliers and noise outside the region of interest. When the current vehicle pose exceeds a preset range: fuse at least one of the sub-coarse point cloud semantic datasets to obtain a coarse point cloud semantic dataset; obtain second visual semantic information based on the first visual semantic information; Within the same preset range, at least one first visual semantic information is obtained, and the union of at least one first visual semantic information is used to obtain the second visual semantic information. A sub-3D point cloud semantic map is obtained based on the second visual semantic information and the coarse point cloud semantic dataset. When the current vehicle pose is outside the range of the 3D point cloud map: obtain a 3D point cloud semantic map based on at least one of the sub-3D point cloud semantic maps.

2. The two-stage high-precision three-dimensional point cloud semantic map construction method according to claim 1, characterized in that: The steps for obtaining a sub-coarse point cloud semantic dataset based on the first point cloud dataset and the first visual semantic information include: Obtain first device parameters and second device parameters, wherein the first device parameters are parameters of the first device, the first device being a device used to acquire raw point cloud data, and the second device parameters are parameters of the second device, the second device being a device used to acquire the current vehicle-end image; Based on the first visual semantic information, a second point cloud dataset is obtained by removing outliers from the first point cloud dataset according to the first device parameters and the second device parameters. A first point cloud semantic dataset is obtained based on the first visual semantic information and the second point cloud dataset. The first outlier processing is performed on the first point cloud semantic dataset to obtain the sub-coarse point cloud semantic dataset.

3. The two-stage high-precision three-dimensional point cloud semantic map construction method according to claim 2, characterized in that: The steps for obtaining the sub-coarse point cloud semantic dataset by performing the first outlier processing on the first point cloud semantic dataset are as follows: Based on the first point cloud semantic dataset, obtain the depth information and Z-axis distribution information of the point cloud data; Based on the depth information and the Z-axis distribution information, the second point cloud semantic dataset is obtained by removing outliers from the first point cloud semantic dataset. The second point cloud semantic dataset is clustered using a clustering algorithm, and the largest clustering region is extracted to obtain the sub-coarse point cloud semantic dataset.

4. The two-stage high-precision three-dimensional point cloud semantic map construction method according to claim 2, characterized in that: The steps for obtaining the first point cloud semantic dataset based on the first visual semantic information and the second point cloud dataset are as follows: At least one first key semantic category is obtained based on the first visual semantic information; The first point cloud semantic dataset is obtained by labeling each point cloud data in the second point cloud dataset with the first key semantic category.

5. The two-stage high-precision three-dimensional point cloud semantic map construction method according to claim 4, characterized in that: The steps for obtaining a sub-3D point cloud semantic map based on the second visual semantic information and the coarse point cloud semantic dataset are as follows: A third point cloud semantic dataset is obtained by reducing the point cloud density in the coarse point cloud semantic dataset. At least one second key semantic category is obtained based on the second visual semantic information; Based on the at least one second key semantic category, a clustering algorithm is used to cluster the third point cloud semantic dataset and extract dense regions to obtain at least one first sub-point cloud semantic dataset. Each first sub-point cloud semantic dataset is subjected to intensity filtering to obtain at least one second sub-point cloud semantic dataset. At least one intermediate 3D point cloud semantic map is obtained by characterizing each second sub-point cloud semantic dataset using a line model. The sub-3D point cloud semantic map is obtained by fusing the at least one intermediate 3D point cloud semantic map.

6. The two-stage high-precision three-dimensional point cloud semantic map construction method according to claim 4, characterized in that: The key semantic categories include curbs, lane lines, poles, and tree trunks.

7. The two-stage high-precision three-dimensional point cloud semantic map construction method according to claim 4, characterized in that: The point cloud data corresponding to different first key semantic categories are labeled with different colors.

8. The two-stage high-precision three-dimensional point cloud semantic map construction method according to claim 3 or 5, characterized in that: The clustering algorithm is the DBSCAN clustering algorithm.

9. The two-stage high-precision three-dimensional point cloud semantic map construction method according to claim 1, characterized in that: When obtaining the first point cloud dataset based on the 3D point cloud map and the current vehicle pose and preset radius, the 3D point cloud map is constructed as a multi-dimensional spatial index structure.

10. The two-stage high-precision three-dimensional point cloud semantic map construction method according to claim 1, characterized in that: The timestamps of the current vehicle pose and the current vehicle image are consistent.

Images