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

Abstract

The invention provides a two-stage high-precision three-dimensional point cloud semantic map construction method, and the method comprises the steps: obtaining a first point cloud data set based on a three-dimensional point cloud map according to a current vehicle pose and a preset radius; the segmentation model segments the current vehicle end image to obtain first visual semantic information; coarsely fusing the first point cloud data set and the first visual semantic information, and performing first noise elimination processing to obtain a sub coarse point cloud semantic data set; fusing the at least one sub coarse point cloud semantic data set to obtain a coarse point cloud semantic data set; and performing second noise elimination processing on the coarse point cloud semantic data set based on the 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. According to the method, the precision of the three-dimensional point cloud semantic map and the accuracy of semantic information are gradually improved in a staged mode, the problems of precision and efficiency possibly existing in single-stage processing are solved, and the robustness and operability of the whole system are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of point cloud processing technology, and in particular to a two-stage high-precision three-dimensional point cloud semantic map construction method. Background Art

[0002] In the field of autonomous driving, high-precision 3D point cloud maps are one of the core technologies for achieving precise positioning and navigation. High-precision 3D point cloud maps can provide autonomous driving systems with detailed environmental information, 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 practicality and intelligence of maps, further processing incorporating semantic information is often necessary. This semantic information is typically extracted through image recognition, deep learning, or manual annotation, and fused with 3D point cloud data to generate 3D point cloud maps rich in semantic information. 3D point cloud maps with semantic information not only provide accurate environmental understanding for systems such as autonomous driving and robotic navigation, but also enhance the systems' adaptability to complex situations.

[0003] In the existing technology, there are some problems and challenges in the way of constructing three-dimensional point cloud maps and integrating semantic information, especially in terms of processing accuracy, computing efficiency and adaptability to complex environments. Summary of the Invention

[0004] In response to the problems existing in the prior art, the present invention provides a two-stage high-precision three-dimensional point cloud semantic map construction method, which gradually improves the precision of the three-dimensional point cloud semantic map and the accuracy of the semantic information in a phased manner. It breaks through 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 present invention adopts the following technical solution: a two-stage high-precision three-dimensional point cloud semantic map construction method, comprising the following steps: Get the current vehicle pose of the autonomous vehicle; Obtaining a first point cloud dataset based on the three-dimensional point cloud map, according to the current vehicle posture and a preset radius; Segmenting the current vehicle-side image based on a deep learning image segmentation model to obtain first visual semantic information; Obtaining a sub-coarse point cloud semantic dataset according to the first point cloud dataset and the first visual semantic information; When the current vehicle posture exceeds a preset range: fusing at least one of the sub-coarse point cloud semantic data sets to obtain a coarse point cloud semantic data set; obtaining second visual semantic information based on the first visual semantic information; Obtaining a sub-3D point cloud semantic map based on the second visual semantic information and the coarse point cloud semantic dataset; When the current vehicle posture exceeds the range of the three-dimensional point cloud map: obtaining a three-dimensional point cloud semantic map according to at least one of the sub-three-dimensional point cloud semantic maps.

[0006] In some embodiments, the step of obtaining a sub-coarse point cloud semantic dataset according to the first point cloud dataset and the first visual semantic information includes: Obtaining first device parameters and second device parameters, where the first device parameters are parameters of a first device used to obtain the original point cloud data, and the second device parameters are parameters of a second device used to obtain the current vehicle-side image; Based on the first visual semantic information, performing first outlier removal on the first point cloud dataset according to the first device parameters and the second device parameters to obtain a second point cloud dataset; Obtain a first point cloud semantic dataset according to the first visual semantic information and the second point cloud dataset; A first outlier processing is performed on the first point cloud semantic dataset to obtain the sub-coarse point cloud semantic dataset.

[0007] 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: Acquire depth information and Z-axis distribution information of point cloud data based on the first point cloud semantic dataset; Performing a second outlier removal on the first point cloud semantic dataset based on the depth information and the Z-axis distribution information to obtain a second point cloud semantic dataset; A clustering algorithm is used to perform clustering processing on the second point cloud semantic dataset and the largest cluster area is extracted to obtain the sub-coarse point cloud semantic dataset.

[0008] In some embodiments, the step of obtaining the first point cloud semantic dataset according to the first visual semantic information and the second point cloud dataset is: Acquire at least one first key semantic category according to the first visual semantic information; Each point cloud data in the second point cloud data set is labeled with the first key semantic category to obtain the first point cloud semantic data set.

[0009] 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: Reducing the point cloud density in the coarse point cloud semantic dataset to obtain a third point cloud semantic dataset; Acquire at least one second key semantic category according to the second visual semantic information; Based on the at least one second key semantic category, clustering the third point cloud semantic dataset using a clustering algorithm and extracting dense areas to obtain at least one first sub-point cloud semantic dataset; Performing intensity screening on each of the first sub-point cloud semantic datasets to obtain at least one second sub-point cloud semantic dataset; Using a line model, characterize each of the second sub-point cloud semantic data sets to obtain at least one intermediate three-dimensional point cloud semantic map; The at least one intermediate three-dimensional point cloud semantic map is fused to obtain the sub-three-dimensional point cloud semantic map.

[0010] In some embodiments, the key semantic categories include curb, lane marking, pole, and tree trunk.

[0011] In some embodiments, point cloud data corresponding to different first key semantic categories are marked with different colors.

[0012] In some embodiments, the clustering algorithm is a DBSCAN clustering algorithm.

[0013] In some embodiments, when a first point cloud dataset is obtained based on the three-dimensional point cloud map and according to the current vehicle posture and a preset radius, the three-dimensional point cloud map is constructed as a multi-dimensional spatial index structure.

[0014] In some embodiments, the current vehicle posture and the timestamp of the current vehicle-side image are consistent. Compared with the prior art, the present invention has the following advantages: This paper combines point cloud data acquisition with visual semantic information segmentation to effectively construct a high-precision three-dimensional point cloud semantic map with rich visual semantic information. Through the two stages of coarse fusion and fine processing, it successfully achieves fine screening and noise removal of point cloud data, improving the map's accuracy and expressiveness. The resulting high-precision three-dimensional point cloud semantic map not only contains highly accurate geometric information but also includes key semantic categories in the environment. This enhances the autonomous driving system's understanding and decision-making capabilities in complex environments, providing a solid foundation for the high-precision positioning of subsequent autonomous vehicles.

[0015] The present disclosure gradually improves the precision of the three-dimensional point cloud semantic map and the accuracy of the semantic information in a phased manner, breaking through the precision and efficiency problems that may exist in single-stage processing.

[0016] This disclosure generates intermediate results during the two-stage process, facilitating real-time monitoring and debugging, quickly locating and resolving potential issues, and ensuring stable system operation. This gradual optimization approach effectively improves the robustness and operability of the overall system. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Schematic diagram of the process of constructing a two-stage high-precision three-dimensional point cloud semantic map disclosed in the present invention; Figure 2 A schematic flow chart of a two-stage high-precision 3D point cloud semantic map construction method for constructing a 3D point cloud map into a multi-dimensional spatial index structure in an embodiment of the present disclosure; Figure 3 Schematic diagram of the process of the rough fusion stage in the embodiment of the present disclosure; Figure 4 Schematic diagram of the process of the detailed processing stage in the embodiment of the present disclosure; Figure 5 A three-dimensional point cloud map constructed in an embodiment of the present disclosure; Figure 6 A three-dimensional point cloud map formed by the first point cloud semantic data set in an embodiment of the present disclosure; Figure 7 A three-dimensional point cloud map formed by a coarse point cloud semantic data set in an embodiment of the present disclosure; Figure 8 It is a high-precision three-dimensional point cloud semantic map in the embodiment of the present disclosure. DETAILED DESCRIPTION

[0018] To address the problem of building 3D point cloud semantic maps with processing accuracy and computational efficiency to adapt to complex environments, the current technologies for building 3D point cloud maps with semantic information are mainly divided into the following three categories: 1. Methods for Manually Drawing Semantic Information In this approach, high-precision three-dimensional point cloud data is first acquired using sensors such as LiDAR, and then semantically annotated manually. This method's advantage lies in its ability to ensure the accuracy of semantic information, as the annotation is performed manually, ensuring high precision. However, its disadvantages are also significant: manual annotation is labor-intensive, inefficient, and extremely dependent on human experience and time. Furthermore, this method struggles with large-scale, dynamically changing environments and lacks the ability to provide real-time updates, making it unsuitable for scenarios requiring rapid response in autonomous driving systems.

[0019] 2. 3D Point Cloud Semantic Segmentation Method Based on Deep Learning This method uses a deep learning network to automatically identify semantic information in point clouds, reducing the workload of manual annotation. However, 3D point cloud data is usually sparse and irregular, which makes it very difficult to directly perform semantic segmentation on it. At present, the performance accuracy of deep learning models on 3D point clouds is relatively low, especially when faced with problems such as noise, occlusion and perspective changes in point cloud data, which poses a great challenge to improving accuracy. In addition, deep learning model training requires a large amount of high-quality annotated data and has a high computational overhead, making it difficult to meet the needs of real-time applications. A method based on the fusion of 3D point cloud maps and visual information.

[0020] 3. Method based on fusion of 3D point cloud map and visual semantic information In this approach, in addition to the 3D point cloud data provided by LiDAR, camera sensors are also introduced to obtain visual semantic information. Through a deep learning network, semantic detection and segmentation are performed by combining camera and LiDAR data, and the semantic information is integrated into the 3D point cloud map. This method combines the advantages of visual information and LiDAR to provide richer and more accurate semantic annotation, especially in complex environments. Compared to segmenting 3D point clouds alone, image segmentation algorithms are more mature, the annotation process is simple and highly accurate, and the operation is also simpler. However, how to achieve efficient fusion between the two sensors remains a technical challenge, affecting the quality of the final map construction.

[0021] Currently, the industry mainly adopts Method 1 and Method 3. Based on Method 3, the present disclosure proposes an improved solution. In order to clearly illustrate the technical features of this solution, the following will describe in detail the implementation method of this application in combination with the accompanying drawings and examples, so that the application can fully understand how to use technical means to solve technical problems and achieve the corresponding technical effects and implement them accordingly. The embodiments of this application and the various features in the embodiments can be combined with each other without conflict, and the technical solutions formed are all within the scope of protection of this application.

[0022] See also Figure 1 The present disclosure provides a two-stage method for constructing a high-precision three-dimensional point cloud semantic map, comprising the following steps: Construct a three-dimensional point cloud map; the three-dimensional point cloud map constructed in this embodiment is shown in FIG. Figure 5 ; Raw data related to urban roads can be collected by autonomous vehicles on urban roads, or other methods can be used to collect raw data related to urban roads and pre-construct a three-dimensional point cloud map. Specifically, a traditional simultaneous localization and mapping (SLAM) algorithm is used to input the raw data related to urban roads into the SLAM algorithm to construct a three-dimensional point cloud map. The three-dimensional point cloud map is usually a three-dimensional point cloud map of the urban area where the vehicle is currently located. Each point cloud data in the three-dimensional point cloud map carries global coordinate information in the UTM coordinate system; The point cloud raw data and the current vehicle-side image are obtained based on the autonomous driving vehicle; the current vehicle posture of the autonomous driving vehicle on the three-dimensional point cloud map can be recorded based on the point cloud raw data. The current vehicle posture is the posture of the vehicle at the current moment, that is, the vehicle coordinate information and vehicle posture information in the UTM coordinate system at the current moment. The real-time vehicle posture is the current vehicle posture of continuous frames, that is, the real-time vehicle posture is a collection of the vehicle coordinate information and vehicle posture information at the current moment in the UTM coordinate system of continuous frames; the current vehicle-side image is a collection of vehicle-side images at the current moment, and the current vehicle-side image is acquired by using a camera installed on the vehicle; in some embodiments, the point cloud raw data, that is, the current vehicle posture and the current vehicle-side image have the same timestamp, which facilitates subsequent data fusion, that is, the fusion of point cloud data and visual semantic information, to ensure the temporal consistency of subsequent data fusion; Obtaining a preset range and a preset radius; dividing the three-dimensional point cloud map into at least one preset range; Obtain the current vehicle posture of the autonomous driving vehicle on the three-dimensional point cloud map based on the original point cloud data; Based on the 3D point cloud map, the first point cloud dataset is obtained according to the current vehicle posture and a preset radius. The local point cloud data corresponding to the vehicle coordinate information of the current frame, i.e., the current moment, is quickly found in the 3D point cloud map. With the vehicle coordinates of the current frame as the center of the circle, a local area within a preset radius is calculated as the area of interest. The preset radius is usually 50 meters, which can be adjusted according to needs. The point cloud data corresponding to the local area is extracted from the 3D point cloud map to obtain the first point cloud dataset. Each frame of the current vehicle posture corresponds to a first point cloud dataset, providing local information for subsequent processing. See also Figure 2 In some embodiments, when a first point cloud dataset is obtained based on a three-dimensional point cloud map and according to a current vehicle posture and a preset radius, the three-dimensional point cloud map is constructed into a multidimensional spatial index structure, that is, a multidimensional spatial index structure is obtained based on the three-dimensional point cloud map; the three-dimensional point cloud map is constructed into a multidimensional spatial index structure, and the multidimensional spatial index structure can accelerate the search and processing efficiency of point cloud data corresponding to the current vehicle posture and the area of interest in the three-dimensional point cloud map; preferably, the multidimensional spatial index structure is a KDTree; Segmenting the current vehicle-side image based on a deep learning image segmentation model to obtain first visual semantic information; the first visual semantic information is visual semantic information segmented from the current vehicle-side image, and the deep learning image segmentation model is obtained through pre-training. Preferably, the deep learning image segmentation model is a pre-trained semantic segmentation network based on a convolutional neural network (CNN). 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; 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 roughly 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 preliminarily retain point cloud data related to the first visual semantic information and preliminarily remove outliers and noise not in the area of interest. See also Figure 7 When the current vehicle posture exceeds a preset range: fuse at least one sub-coarse point cloud semantic data set to obtain a coarse point cloud semantic data set; obtain second visual semantic information based on the first visual semantic information; obtain a first visual semantic information for each frame of the current vehicle-side 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 the at least one first visual semantic information; The above is the coarse fusion stage. In the coarse fusion stage, a coarse point cloud semantic dataset is obtained through preliminary screening, and then the fine processing stage is entered. The goal of the fine processing stage is to further remove noise points, finely extract effective information from the point cloud data, and generate the final high-precision 3D point cloud semantic map. The fine processing stage is to obtain a sub-3D point cloud semantic map based on the second visual semantic information and the coarse point cloud semantic dataset. Specifically, based on the second visual semantic information, the coarse point cloud semantic dataset is subjected to a second noise removal process to obtain a sub-3D point cloud semantic map. 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 further remove outliers; When the current vehicle posture exceeds the range of the three-dimensional point cloud map: a three-dimensional point cloud semantic map is obtained based on at least one sub-three-dimensional point cloud semantic map, each preset range corresponds to a sub-three-dimensional point cloud semantic map, and when the autonomous driving vehicle travels through the range of the entire three-dimensional point cloud map, the at least one sub-three-dimensional point cloud semantic map is integrated to obtain a three-dimensional point cloud semantic map.

[0023] The two-stage, high-precision 3D point cloud semantic map construction method provided by this disclosure gradually improves the map's precision and the accuracy of its visual semantic information in a phased approach, overcoming the accuracy and efficiency issues that may exist with single-stage processing. This method generates intermediate results during the two-stage process, facilitating real-time monitoring and debugging, quickly locating and resolving potential issues, and ensuring stable system operation. This gradual optimization approach effectively improves the robustness and operability of the overall system.

[0024] See also 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: Obtaining first device parameters and second device parameters, where the first device parameters are parameters of the first device, which is a device used to obtain raw point cloud data. The first device is a high-precision sensor, including but not limited to one or more of a laser radar (LiDAR), an inertial measurement unit (IMU), or RTK. The first device is used to collect raw point cloud data, and the raw point cloud data is used to construct a three-dimensional point cloud map using a traditional SLAM algorithm, and the current vehicle posture is recorded. The second device parameters are parameters of the second device, which is a device used to obtain the current vehicle-side image. Typically, the second device includes but is not limited to one or more of a camera, a vehicle-mounted camera, a three-dimensional laser scanner, or a sensor; Based on the first visual semantic information, removing first outliers from the first point cloud dataset according to first device parameters and second device parameters to obtain a second point cloud dataset; the first outliers are points that are irrelevant to the first visual semantic information; See also Figure 6 , obtaining a first point cloud semantic dataset based on the first visual semantic information and the second point cloud dataset; fusing the first visual semantic information corresponding to each point cloud data in the second point cloud dataset to obtain the first point cloud semantic dataset, and after the fusion is completed, within a preset range, each second point cloud dataset corresponds to a first point cloud semantic dataset; A first outlier processing is performed on the first point cloud semantic dataset to obtain a sub-coarse point cloud semantic dataset, where the first outlier processing is to preliminarily remove outliers and noise that is not in the area of interest.

[0025] In some embodiments, the step of performing first outlier processing on the first point cloud semantic dataset to obtain a sub-coarse point cloud semantic dataset is: Acquire depth information and Z-axis distribution information of the point cloud data based on the first point cloud semantic dataset; Based on the depth information and Z-axis distribution information, the first point cloud semantic dataset is respectively subjected to second outlier elimination to obtain the second point cloud semantic dataset; 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 according to the depth information, and the depth difference and Z-axis distribution range are used to filter the point cloud data in the first point cloud semantic dataset to remove outliers and point cloud data that are not in the area of interest in the first point cloud semantic dataset, that is, the second outlier points are points that are not in the area of interest, thereby achieving the second outlier elimination to obtain the second point cloud semantic dataset.

[0026] A clustering algorithm is used to cluster the second point cloud semantic dataset and extract the largest cluster area to obtain a sub-coarse point cloud semantic dataset. After extracting the largest cluster area, sparsely distributed point cloud data and the corresponding first visual semantic information and noise points are eliminated 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.

[0027] In some embodiments, the step of obtaining at least one first point cloud semantic dataset according to the first visual semantic information and the second point cloud dataset is: At least one first key semantic category is obtained based on the 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 three-dimensional point cloud semantic maps rely mainly on limited key semantic categories such as lane lines, landmarks, poles, etc.; these key semantic categories perform well in scenes with clear visual semantic information (such as elevated roads), but often limit the vehicle's ability to pass on complex urban and rural roads where poles are scarce or lane lines are fuzzy. In order to address this limitation, the present disclosure introduces more types of visual semantic information in the preset stage, such as tree trunks, curbs, etc. By integrating multiple key semantic categories, the present disclosure significantly expands the applicable passage area of vehicles and improves the adaptability and versatility of the map, especially in complex environments; Each point cloud data in the second point cloud data set is labeled with the first key semantic category to obtain the first point cloud semantic data set. Rich and diverse vision-related visual semantic information is integrated with the point cloud data to broaden the vehicle's travel area.

[0028] In the first stage, coarse fusion, a rough data processing and fusion strategy is used 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 and retain only the regions of interest that contain the primary visual semantic information, although some noise may still be present. This approach allows for the construction of a basic 3D point cloud semantic map in a relatively short time, enabling rapid response to environmental changes.

[0029] In the second stage, detailed processing, a refinement algorithm is used to further optimize the accuracy of the 3D point cloud semantic map, remove noise, and cluster the remaining data using secondary visual semantic information. This creates independent 3D point cloud semantic elements and characterizes their geometric structure, ensuring high accuracy and structural integrity while reducing storage requirements and making downstream tasks more efficient and easy to use.

[0030] See also Figure 4In 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: Reducing the point cloud density in the coarse point cloud semantic dataset to obtain a third point cloud semantic dataset; preferably, using voxel filtering to reduce the density of the point cloud; Acquire at least one second key semantic category based on the second visual semantic information; similarly, the second key semantic category includes one or more of a curb, a lane line, a pole, or a tree trunk; Based on at least one second key semantic category, a clustering algorithm is used to perform clustering processing on the third point cloud semantic dataset and extract dense areas to obtain at least one first sub-point cloud semantic dataset; a clustering algorithm is used to perform clustering processing on the point cloud data of each second key semantic category, extract dense areas and remove scattered noise points to obtain at least one first sub-point cloud semantic dataset, and the number of the first sub-point cloud semantic datasets corresponds to the number of key semantic categories; preferably, the clustering algorithm is a DBSCAN clustering algorithm; Performing intensity screening on each of the first sub-point cloud semantic datasets to obtain at least one second sub-point cloud semantic dataset; during the intensity screening process, outliers whose intensities are inconsistent with those of the majority of points are removed to ensure that valid point cloud data is retained in the second sub-point cloud semantic dataset; Using the line model, each second sub-point cloud semantic data set is characterized to obtain at least one sub-3D point cloud semantic map; during the characterization process, linear structure information is extracted to generate an intermediate 3D point cloud semantic map; At least one intermediate three-dimensional point cloud semantic map is fused to obtain a sub-three-dimensional point cloud semantic map.

[0031] In some embodiments, point cloud data corresponding to different first key semantic categories are marked with different colors, which can more intuitively distinguish the 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 color labeling of point cloud data.

[0032] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions of the technical solution of the present invention by ordinary technicians in this field do not deviate 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 by: The following steps are involved: Get the current vehicle pose of the autonomous vehicle; Obtaining a first point cloud dataset based on the three-dimensional point cloud map and according to the current vehicle posture and a preset radius; Segment the current vehicle-side image based on a deep learning image segmentation model to obtain first visual semantic information; Obtaining a sub-coarse point cloud semantic dataset according to the first point cloud dataset and the first visual semantic information; When the current vehicle posture exceeds a preset range: fusing at least one of the sub-coarse point cloud semantic data sets to obtain a coarse point cloud semantic data set; obtaining second visual semantic information based on the first visual semantic information; Obtaining a sub-3D point cloud semantic map based on the second visual semantic information and the coarse point cloud semantic dataset; When the current vehicle posture exceeds the range of the three-dimensional point cloud map: obtaining a three-dimensional point cloud semantic map according to at least one of the sub-three-dimensional point cloud semantic maps.

2. The two-stage high-precision 3D point cloud semantic map construction method according to claim 1, characterized in that: The step of obtaining a sub-coarse point cloud semantic dataset according to the first point cloud dataset and the first visual semantic information includes: Obtaining first device parameters and second device parameters, where the first device parameters are parameters of a first device used to obtain the original point cloud data, and the second device parameters are parameters of a second device used to obtain the current vehicle-side image; Based on the first visual semantic information, performing first outlier removal on the first point cloud dataset according to the first device parameters and the second device parameters to obtain a second point cloud dataset; Obtain a first point cloud semantic dataset according to the first visual semantic information and the second point cloud dataset; A 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 3D point cloud semantic map construction method according to claim 2, characterized in that: 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: Acquire depth information and Z-axis distribution information of point cloud data based on the first point cloud semantic dataset; Performing a second outlier removal on the first point cloud semantic dataset based on the depth information and the Z-axis distribution information to obtain a second point cloud semantic dataset; A clustering algorithm is used to perform clustering processing on the second point cloud semantic dataset and the largest cluster area is extracted to obtain the sub-coarse point cloud semantic dataset.

4. The two-stage high-precision 3D point cloud semantic map construction method according to claim 2, characterized in that: The step of obtaining the first point cloud semantic dataset according to the first visual semantic information and the second point cloud dataset is: Acquire at least one first key semantic category according to the first visual semantic information; Each point cloud data in the second point cloud data set is labeled with the first key semantic category to obtain the first point cloud semantic data set.

5. The two-stage high-precision 3D point cloud semantic map construction method according to claim 4, characterized in that: 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: Reducing the point cloud density in the coarse point cloud semantic dataset to obtain a third point cloud semantic dataset; Acquire at least one second key semantic category according to the second visual semantic information; Based on the at least one second key semantic category, clustering the third point cloud semantic dataset using a clustering algorithm and extracting dense areas to obtain at least one first sub-point cloud semantic dataset; Performing intensity screening on each of the first sub-point cloud semantic datasets to obtain at least one second sub-point cloud semantic dataset; Using a line model, characterize each of the second sub-point cloud semantic data sets to obtain at least one intermediate three-dimensional point cloud semantic map; The at least one intermediate three-dimensional point cloud semantic map is fused to obtain the sub-three-dimensional point cloud semantic map.

6. The two-stage high-precision 3D point cloud semantic map construction method according to claim 4, characterized in that: The key semantic categories include curb, lane marking, pole, and tree trunk.

7. The two-stage high-precision 3D 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 marked with different colors.

8. The two-stage high-precision 3D 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 3D point cloud semantic map construction method according to claim 1, characterized in that: When a first point cloud data set is obtained based on the three-dimensional point cloud map and according to the current vehicle posture and a preset radius, the three-dimensional point cloud map is constructed into a multi-dimensional spatial index structure.

10. The two-stage high-precision 3D point cloud semantic map construction method according to claim 1, characterized in that: The current vehicle posture and the timestamp of the current vehicle-side image are consistent.

Images