A Point Cloud Registration Method and System Based on Multi-Source Semantic Information and Plane Normal Vectors

Abstract

This invention discloses a point cloud registration method and system based on multi-source semantic information and planar normal vectors. The point cloud registration method includes: acquiring first point cloud data from one site and second point cloud data from another site in a target area; performing semantic segmentation on the first and second point cloud data and classifying the segmented components; for components of a target category, extracting planar components from the target category; acquiring multi-dimensional feature data of the planar components; using the multi-dimensional feature data to match the planar components of the first and second point cloud data; calculating a rotation matrix using the matched components; and performing a global multi-plane consistency evaluation on the rotation matrix to obtain an initial registration matrix. This application achieves a balance between high registration success rate, high robustness, and high computational efficiency, thus providing a reliable initial pose for precise point cloud registration and multi-site fusion in complex urban scenarios.

Description

Technical Field

[0001] This invention relates to a point cloud registration method and system based on multi-source semantic information and planar normal vectors. Background Technology

[0002] Point cloud registration is a key foundational technology in fields such as 3D reconstruction, robot localization and mapping (SLAM), high-definition map production for autonomous driving, and digital twin construction.

[0003] The rapid development of technologies such as 3D laser scanning has made the acquisition of large-scale point clouds at the city level commonplace. The efficient, accurate and automatic registration of massive point clouds under different conditions has become the core bottleneck for the engineering application of related technologies.

[0004] Existing point cloud registration methods have shortcomings. The classic Iterative Closest Point (ICP) algorithm has high accuracy in the fine registration stage, but it is highly sensitive to the initial pose and is prone to getting trapped in local optima when there are large rotation and translation errors. The Normal Distributions Transform (NDT) algorithm has a slow convergence speed in sparse or open scenes and is sensitive to changes in point cloud density. Global registration methods such as Fast Global Registration (FGR) and 4-Points Congruent Sets (4PCS) rely on stable matching of local geometric features and are prone to failure in urban environments with missing textures or repetitive structures. At the same time, the computational cost of feature extraction and matching is large.

[0005] In recent years, point cloud registration methods based on deep learning (such as PointNetLK and Deep GlobalRegistration) have achieved certain accuracy improvements on specific datasets, but they generally suffer from problems such as slow registration speed, reliance on a large amount of labeled data during training, and sensitivity to sensor type and acquisition conditions, making it difficult to directly extend to complex and ever-changing real urban scenarios.

[0006] Existing point cloud registration techniques often fail to fully explore and integrate the aforementioned multi-source information. Plane-based registration methods (such as RANSAC-based plane matching or Plane ICP) are prone to mismatches and have limited anti-interference capabilities and stability. Summary of the Invention

[0007] The technical problem to be solved by this invention is to overcome the shortcomings of existing large-scale point cloud acquisition technologies at the city level, such as the tendency for mismatches to occur during the point cloud registration process, and the limited anti-interference ability and stability. This invention provides a point cloud registration method and system based on multi-source semantic information and planar normal vectors that achieves a high registration success rate, high robustness and high computational efficiency, thereby providing a reliable initial pose for precise point cloud registration and multi-site fusion in complex urban scenarios.

[0008] The present invention solves the above-mentioned technical problems through the following technical solution:

[0009] A point cloud registration method based on multi-source semantic information and plane normal vectors, characterized in that the point cloud registration method includes:

[0010] Acquire the first point cloud data of one site and the second point cloud data of another site in the target area;

[0011] Perform semantic segmentation on the first and second point cloud data and classify the segmented components;

[0012] For a component of a target category, extract the planar components from the target category;

[0013] Obtain multidimensional feature data of planar components;

[0014] The multidimensional feature data is used to perform planar component matching between the first point cloud data and the second point cloud data.

[0015] Calculate the rotation matrix using matching components;

[0016] The initial registration matrix is ​​obtained by performing a global multiplane consistency evaluation on the rotation matrix.

[0017] Preferably, the step of semantically segmenting the first point cloud data and the second point cloud data and classifying the segmented components includes:

[0018] A city-level semantic segmentation model was obtained by training using the PointNet++ network as the basic architecture.

[0019] Semantic segmentation is performed on the first point cloud data and the second point cloud data, and point-by-point semantic classification is performed on the segmented components;

[0020] A semantic topology graph is constructed based on the spatial adjacency relationship of semantic categories. The semantic topology graph is used to describe the spatial adjacency relationship, relative position relationship and scale characteristics between different semantic regions.

[0021] Preferably, the step of semantically segmenting the first point cloud data and the second point cloud data and classifying the segmented components includes:

[0022] The first point cloud data and the second point cloud data are preprocessed. The preprocessing includes one or more of the following: voxel mesh downsampling, normal vector estimation, reflection intensity normalization, RGB color normalization, and extraction of local curvature features.

[0023] Semantic segmentation is performed on the first and second point cloud data, and the segmented components are classified.

[0024] Preferably, the extraction of planar components from the target category includes:

[0025] Select the point with the smallest curvature among the components of the target category as the seed point;

[0026] The seed point is iteratively expanded to include neighboring points to obtain planar components of the target category based on the growth parameters preset for the target category and the growth algorithm. The growth parameters include semantic boundary data.

[0027] Preferably, the multidimensional feature data includes one or more of geometric features, intensity features, color features, and spatial distribution features. Geometric features include one or more of plane normal vectors, plane centroid positions, and plane areas. Intensity features include one or more of mean reflection intensity, standard deviation of intensity, and histogram of reflection intensity. Color features include one or more of mean color, standard deviation of color, and histogram of color in the RGB color space. Spatial distribution features include one or more of plane boundary complexity, point distribution density, and morphological consistency index. The component matching of the first point cloud data and the second point cloud data using the multidimensional feature data includes:

[0028] Obtain the similarity of the multidimensional features of the components in the first point cloud data and the second point cloud data, filter the matching plane according to the similarity threshold of each dimension, and add the matching plane to the candidate matching pool.

[0029] Preferably, the calculation of the rotation matrix using the matching component includes:

[0030] For a matching plane in the candidate matching pool, the rotation matrix is ​​calculated using the normal vector of the matching plane, and the rotation axis and rotation angle are determined using the angle between the normal vectors.

[0031] The translation vector is calculated by matching the spatial differences between the centroids of the plane.

[0032] Combining the rotation matrix with the translation vector yields the rotation matrix of the matching plane.

[0033] Preferably, the step of performing a global multiplane consistency evaluation on the rotation matrix to obtain the initial registration matrix includes:

[0034] For the target rotation matrix of the matching plane, apply the target rotation matrix to each candidate matching plane in the candidate matching pool;

[0035] Perform a global multiplane consistency evaluation on the rotated candidate matching planes;

[0036] The optimal rotation matrix from the evaluation results is used as the initial registration matrix.

[0037] Preferably, the global multiplane consistency evaluation of the rotated candidate matching plane includes:

[0038] The evaluation result of the rotation matrix is ​​obtained by calculating the weighted average of the cosine similarity of the normal vectors, the intensity similarity, the color similarity, and the area similarity.

[0039] Preferably, the point cloud registration method includes:

[0040] The first point cloud data and the second point cloud data are initially registered using the initial registration matrix.

[0041] The ICP algorithm is used to perform fine registration and iterative optimization on the first and second point cloud data after initial registration.

[0042] Output the final registration result of the first point cloud data and the second point cloud data.

[0043] The present invention also provides a point cloud registration system based on multi-source semantic information and plane normal vectors, characterized in that the point cloud registration system includes a lidar, and the point cloud registration system is used to implement the point cloud registration method based on multi-source semantic information and plane normal vectors as described above.

[0044] Based on common knowledge in the field, the above-mentioned preferred conditions can be combined arbitrarily to obtain various preferred embodiments of the present invention.

[0045] The positive and progressive effects of this invention are as follows:

[0046] This application achieves a balance between high registration success rate, high robustness, and high computational efficiency, thus providing a reliable initial pose for precise point cloud registration and multi-site fusion in complex urban scenarios.

[0047] Specifically, it significantly improves the success rate of automatic registration of point clouds in large-scale, textureless scenes, and promotes the fully automatic update process of high-precision maps for autonomous driving.

[0048] By eliminating the reliance on image features, this method solves the problem of traditional top-down view methods failing completely in complex scenarios such as multi-level interchanges and tunnels.

[0049] It has extremely high computing efficiency and can run in real time on vehicle computing platforms or robot embedded systems.

[0050] The method has a clear principle, few parameters, and is easy to implement in engineering and for secondary development. Attached Figure Description

[0051] Figure 1 This is a flowchart of the point cloud registration method in Embodiment 1 of the present invention.

[0052] Figure 2 This is a schematic diagram of the effect of point cloud data of the target area in Embodiment 1 of the present invention. Detailed Implementation

[0053] The present invention will be further illustrated by way of embodiments below, but the present invention is not limited to the scope of the embodiments described herein.

[0054] Example 1

[0055] This embodiment provides a point cloud registration system based on multi-source semantic information and plane normal vectors. The point cloud registration system includes a data acquisition module and a data processing module.

[0056] In this embodiment, the data acquisition module is a lidar.

[0057] The data processing module can be a high-performance processing chip in a LiDAR, or it can be a desktop computer, tablet computer, laptop computer, or a server with even stronger computing power.

[0058] The server may be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0059] In the specification and claims of this application, the terms “comprising” and “having”, and any variations thereof, are used to indicate a non-exclusive inclusion, for example, a process, method, product, or apparatus that includes a series of method steps or modules is not necessarily limited to those steps or modules that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0060] The data acquisition module is used to acquire the first point cloud data from one station and the second point cloud data from another station in the target area. (See below) Figure 2 In this embodiment, the target area is a typical urban scene.

[0061] The data processing module is used for:

[0062] Perform semantic segmentation on the first and second point cloud data and classify the segmented components;

[0063] For a component of a target category, extract the planar components from the target category;

[0064] Obtain multidimensional feature data of planar components;

[0065] The multidimensional feature data is used to perform planar component matching between the first point cloud data and the second point cloud data.

[0066] Calculate the rotation matrix using matching components;

[0067] The initial registration matrix is ​​obtained by performing a global multiplane consistency evaluation on the rotation matrix.

[0068] The data processing module is also used for:

[0069] A city-level semantic segmentation model was obtained by training using the PointNet++ network as the basic architecture.

[0070] Semantic segmentation is performed on the first point cloud data and the second point cloud data, and point-by-point semantic classification is performed on the segmented components;

[0071] A semantic topology graph is constructed based on the spatial adjacency relationship of semantic categories. The semantic topology graph is used to describe the spatial adjacency relationship, relative position relationship and scale characteristics between different semantic regions.

[0072] The data processing module is also used for:

[0073] The first point cloud data and the second point cloud data are preprocessed. The preprocessing includes one or more of the following: voxel mesh downsampling, normal vector estimation, reflection intensity normalization, RGB color normalization, and extraction of local curvature features.

[0074] Semantic segmentation is performed on the first and second point cloud data, and the segmented components are classified.

[0075] The data processing module is also used for:

[0076] Select the point with the smallest curvature among the components of the target category as the seed point;

[0077] The seed point is iteratively expanded to include neighboring points to obtain planar components of the target category based on the growth parameters preset for the target category and the growth algorithm. The growth parameters include semantic boundary data.

[0078] The aforementioned feature data includes one or more of geometric features, intensity features, color features, and spatial distribution features. Geometric features include one or more of plane normal vectors, plane centroid positions, and plane areas. Intensity features include one or more of mean reflection intensity, standard deviation of intensity, and histogram of reflection intensity. Color features include one or more of mean color, standard deviation of color, and histogram of color in the RGB color space. Spatial distribution features include one or more of plane boundary complexity, point distribution density, and morphological consistency indices. The data processing module is further used for:

[0079] Obtain the similarity of the multidimensional features of the components in the first point cloud data and the second point cloud data, filter the matching plane according to the similarity threshold of each dimension, and add the matching plane to the candidate matching pool.

[0080] The data processing module is also used for:

[0081] For a matching plane in the candidate matching pool, the rotation matrix is ​​calculated using the normal vector of the matching plane, and the rotation axis and rotation angle are determined using the angle between the normal vectors.

[0082] The translation vector is calculated by matching the spatial differences between the centroids of the plane.

[0083] Combining the rotation matrix with the translation vector yields the rotation matrix of the matching plane.

[0084] The data processing module is also used for:

[0085] For the target rotation matrix of the matching plane, apply the target rotation matrix to each candidate matching plane in the candidate matching pool;

[0086] Perform a global multiplane consistency evaluation on the rotated candidate matching planes;

[0087] The optimal rotation matrix from the evaluation results is used as the initial registration matrix.

[0088] The data processing module is also used for:

[0089] The evaluation result of the rotation matrix is ​​obtained by calculating the weighted average of the cosine similarity of the normal vectors, the intensity similarity, the color similarity, and the area similarity.

[0090] The data processing module is also used for:

[0091] The first point cloud data and the second point cloud data are initially registered using the initial registration matrix.

[0092] The ICP algorithm is used to perform fine registration and iterative optimization on the first and second point cloud data after initial registration.

[0093] Output the final registration result of the first point cloud data and the second point cloud data.

[0094] In typical urban scenarios, 3D point clouds naturally contain rich semantic information and stable geometric structural features. For example, large-scale planar structures such as building facades, ground, and rooftops have highly stable normal vector distributions; objects of different materials (such as asphalt pavements, concrete sidewalks, metal railings, and glass curtain walls) exhibit significant differences in laser reflection intensity; urban elements (buildings, roads, vegetation, vehicles, etc.) have clear semantic constraints in spatial distribution and topological relationships; and RGB color information can further distinguish subtle differences between similar targets.

[0095] Existing point cloud registration techniques often fail to fully exploit and integrate the aforementioned multi-source information. Plane-based registration methods (such as RANSAC-based plane matching or Plane ICP) typically utilize only a single or a small number of maximum planes for registration, lacking global multi-plane consistency constraints. In complex scenarios with a large number of similar planar structures, they are prone to mismatches, exhibiting limited anti-interference capabilities and stability.

[0096] This embodiment proposes an initial pose estimation framework of "semantic perception - planar normal vector constraint - multi-feature weighted fusion", and then combines it with ICP fine registration to achieve final registration. It can fully exploit the semantic structure information and multimodal features of urban point clouds, significantly improving the success rate and robustness of registration.

[0097] Furthermore, by integrating city-level semantic analysis and multi-source registration features based on PointNet++, a registration foundation driven by both semantics and features is constructed. This not only enables refined semantic parsing of city point clouds, but also provides more comprehensive and reliable support for semantic perception registration through the synergistic empowerment of semantics and multiple features.

[0098] This embodiment takes the normal vector direction constraint as the core basis and uses a planar multi-dimensional feature representation scheme that weights and fuses reflection intensity, RGB color and spatial geometric information, breaking through the limitation of traditional methods that rely on only a single geometric information.

[0099] This embodiment proposes a global multi-plane consistency evaluation mechanism, which scores the overall consistency between all planes in the transformed source point cloud (first point cloud data) and the target point cloud (second point cloud data). Through a weighted collaborative mode of multi-source information, the discriminativeness and stability of plane matching are significantly improved, avoiding ambiguity and errors in single-plane matching.

[0100] Specifically, the operation method using the point cloud registration system is as follows:

[0101] First point cloud data and second point cloud data from another site

[0102] This embodiment uses urban scene point clouds acquired by adjacent station-based 3D laser scanning as an example to illustrate the point cloud registration method based on multi-source semantic information and planar normal vectors described in this invention. The point cloud size of each station is approximately 40 million points, and the overlap rate of point clouds between adjacent stations is approximately 30%.

[0103] Based on a pre-trained PointNet++ city-level semantic segmentation model, semantic segmentation is performed on source and target point clouds, classifying the point clouds into semantic categories such as building facades, ground, rooftops, vegetation, and vehicles. Semantic topological relationships are constructed based on the semantic segmentation results to record the spatial adjacency and relative positional relationships between different semantic regions, providing semantic constraints for subsequent planar matching.

[0104] The point cloud is preprocessed, and the source and target point clouds are downsampled using a voxel grid with a voxel size of 0.2m. The point cloud normal vector and curvature are estimated using a neighborhood search with a radius of 0.5m.

[0105] Semantic-aware plane segmentation is performed. Within each semantic category, a region-growing plane segmentation algorithm is executed. The normal vector smoothing threshold is set to 15°, and the minimum number of plane points is set to 1500 (approximately 3 m²). Only large-scale plane structures with an area greater than 20 m² are retained.

[0106] In this embodiment, typically 15 to 50 candidate planes, including the ground, building facades on both sides of the road, and isolation structures, can be extracted from each site cloud.

[0107] The multidimensional features of the planes are calculated by filtering the extracted planes and removing those with an absolute value greater than 0.7 for the Z-axis component of the normal vector. The plane normal vectors are then normalized in the XY plane to remove the influence of the Z-axis component. Simultaneously, the multidimensional features of each plane are calculated, including: plane normal vector, plane centroid position, plane area, average reflection intensity, and color statistical features.

[0108] A candidate plane matching pool with semantic constraints is constructed, matching only planes belonging to the same semantic category in the source and target point clouds. The similarity between planes is calculated based on their normal vectors, reflection intensity, color features, and geometric shape features, with the following thresholds: normal vector angle threshold set to 12°, relative area difference threshold set to 30%, average reflection intensity difference threshold set to 15%, and RGB mean difference threshold set to 20% for color features. When a pair of planes simultaneously satisfies both the semantic consistency constraint and the aforementioned multi-dimensional feature similarity thresholds, the pair is added to the candidate plane matching pool.

[0109] The candidate rigid body transformation is generated quickly. For each pair of matching planes in the candidate plane matching pool, the rotation matrix is ​​calculated based on the normal vector of the matching plane, and the translation vector is calculated by aligning the centroids of the matching planes, thereby generating the corresponding candidate rigid body transformation.

[0110] A global multi-plane consistency evaluation is performed. For each candidate rigid body transformation, it is applied to all candidate planes in the source point cloud A, and the overall consistency between the transformed set of planes and the set of planes in the target point cloud B is evaluated. For each pair of matching planes, the cosine similarity of the normal vector, the similarity of reflection intensity, the similarity of color, and the similarity of plane area are calculated. A weighted average method is used to calculate the global consistency score, with weights set to 0.5 for normal vector similarity, 0.2 for area similarity, 0.2 for reflection intensity similarity, and 0.1 for color similarity. When the number of plane pairs satisfying the similarity condition is not less than 5 pairs, the candidate rigid body transformation is considered a valid transformation.

[0111] Candidate transformation sorting and output: All valid candidate rigid body transformations are sorted according to the global consistency score. The candidate rigid body transformation with the highest score is selected as the initial registration pose of the source point cloud and the target point cloud. The ICP algorithm is used for fine registration iterative optimization, and the final registration result is output.

[0112] See Figure 1 Utilizing the aforementioned point cloud registration system based on multi-source semantic information and plane normal vectors, this embodiment also provides a point cloud registration method, including:

[0113] Step 100: Obtain the first point cloud data of one station and the second point cloud data of another station in the target area;

[0114] Step 101: Perform semantic segmentation on the first point cloud data and the second point cloud data, and classify the segmented components;

[0115] Step 102: For a component of a target category, extract the planar components from the target category;

[0116] Step 103: Obtain multidimensional feature data of planar components;

[0117] Step 104: Use the multidimensional feature data to perform planar component matching between the first point cloud data and the second point cloud data;

[0118] Step 105: Calculate the rotation matrix using the matching components;

[0119] Step 106: Perform a global multiplane consistency evaluation on the rotation matrix to obtain the initial registration matrix.

[0120] Step 101 specifically includes:

[0121] A city-level semantic segmentation model was obtained by training using the PointNet++ network as the basic architecture.

[0122] Semantic segmentation is performed on the first point cloud data and the second point cloud data, and point-by-point semantic classification is performed on the segmented components;

[0123] A semantic topology graph is constructed based on the spatial adjacency relationship of semantic categories. The semantic topology graph is used to describe the spatial adjacency relationship, relative positional relationship and scale features between different semantic regions.

[0124] The first point cloud data and the second point cloud data are preprocessed. The preprocessing includes one or more of the following: voxel mesh downsampling, normal vector estimation, reflection intensity normalization, RGB color normalization, and extraction of local curvature features.

[0125] Semantic segmentation is performed on the first and second point cloud data, and the segmented components are classified.

[0126] Step 102 includes:

[0127] Select the point with the smallest curvature among the components of the target category as the seed point;

[0128] The seed point is iteratively expanded to include neighboring points to obtain planar components of the target category based on the growth parameters preset for the target category and the growth algorithm. The growth parameters include semantic boundary data.

[0129] The multiple feature data include one or more of geometric features, intensity features, color features, and spatial distribution features. Geometric features include one or more of plane normal vector, plane centroid position, and plane area. Intensity features include one or more of mean reflection intensity, standard deviation of intensity, and histogram of reflection intensity. Color features include one or more of mean color, standard deviation of color, and histogram of color in the RGB color space. Spatial distribution features include one or more of plane boundary complexity, point distribution density, and morphological consistency index. Step 104 includes:

[0130] Obtain the similarity of the multidimensional features of the components in the first point cloud data and the second point cloud data, filter the matching plane according to the similarity threshold of each dimension, and add the matching plane to the candidate matching pool.

[0131] Step 105 includes:

[0132] For a matching plane in the candidate matching pool, the rotation matrix is ​​calculated using the normal vector of the matching plane, and the rotation axis and rotation angle are determined using the angle between the normal vectors.

[0133] The translation vector is calculated by matching the spatial differences between the centroids of the plane.

[0134] Combining the rotation matrix with the translation vector yields the rotation matrix of the matching plane.

[0135] Step 106 includes:

[0136] For the target rotation matrix of the matching plane, apply the target rotation matrix to each candidate matching plane in the candidate matching pool;

[0137] A global multi-plane consistency evaluation is performed on the rotated candidate matching planes. The evaluation result of the rotation matrix is ​​obtained by calculating the weighted average of the normal vector cosine similarity, intensity similarity, color similarity, and area similarity.

[0138] The optimal rotation matrix from the evaluation results is used as the initial registration matrix.

[0139] Step 106 is followed by:

[0140] The first point cloud data and the second point cloud data are initially registered using the initial registration matrix.

[0141] The ICP algorithm is used to perform fine registration and iterative optimization on the first and second point cloud data after initial registration.

[0142] Output the final registration result of the first point cloud data and the second point cloud data.

[0143] This embodiment achieves a balance between high registration success rate, high robustness, and high computational efficiency, thus providing a reliable initial pose for precise point cloud registration and multi-site fusion in complex urban scenarios.

[0144] Specifically, it significantly improves the success rate of automatic registration of point clouds in large-scale, textureless scenes, and promotes the fully automatic update process of high-precision maps for autonomous driving.

[0145] By eliminating the reliance on image features, this method solves the problem of traditional top-down view methods failing completely in complex scenarios such as multi-level interchanges and tunnels.

[0146] It has extremely high computing efficiency and can run in real time on vehicle computing platforms or robot embedded systems.

[0147] The method has a clear principle, few parameters, and is easy to implement in engineering and for secondary development.

[0148] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these are merely illustrative examples, and the scope of protection of the present invention is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of the present invention, but all such changes and modifications fall within the scope of protection of the present invention.

Claims

1. A point cloud registration method based on multi-source semantic information and plane normal vector, characterized in that, The point cloud registration method includes: Acquire the first point cloud data of one site and the second point cloud data of another site in the target area; Perform semantic segmentation on the first and second point cloud data and classify the segmented components; For a component of a target category, extract the planar components from the target category; Obtain multidimensional feature data of planar components. The multidimensional feature data includes multiple features such as geometric features, intensity features, color features, and spatial distribution features. Geometric features include plane normal vectors. The multidimensional feature data is used to perform planar component matching between the first point cloud data and the second point cloud data. Calculate the rotation matrix using matching components; The initial registration matrix is ​​obtained by performing a global multiplane consistency evaluation on the rotation matrix. 2.The method of claim 1, wherein, The semantic segmentation of the first point cloud data and the second point cloud data, and the classification of the segmented components, include: A city-level semantic segmentation model was obtained by training using the PointNet++ network as the basic architecture. Semantic segmentation is performed on the first point cloud data and the second point cloud data, and point-by-point semantic classification is performed on the segmented components; A semantic topology graph is constructed based on the spatial adjacency relationship of semantic categories. The semantic topology graph is used to describe the spatial adjacency relationship, relative position relationship and scale characteristics between different semantic regions. 3.The method of claim 1, wherein, The semantic segmentation of the first point cloud data and the second point cloud data, and the classification of the segmented components, include: The first point cloud data and the second point cloud data are preprocessed. The preprocessing includes one or more of the following: voxel mesh downsampling, normal vector estimation, reflection intensity normalization, RGB color normalization, and extraction of local curvature features. Semantic segmentation is performed on the first and second point cloud data, and the segmented components are classified.

4. The point cloud registration method based on multi-source semantic information and plane normal vectors as described in claim 1, characterized in that, The extraction of planar components from the target category includes: Select the point with the smallest curvature among the components of the target category as the seed point; The seed point is iteratively expanded to include neighboring points to obtain planar components of the target category based on the growth parameters preset for the target category and the growth algorithm. The growth parameters include semantic boundary data.

5. The point cloud registration method based on multi-source semantic information and plane normal vectors as described in claim 1, characterized in that, The multidimensional feature data includes multiple features such as geometric features, intensity features, color features, and spatial distribution features. Geometric features include one or more of the following: plane normal vector, plane centroid position, and plane area. Intensity features include one or more of the following: mean reflection intensity, intensity standard deviation, and reflection intensity histogram. Color features include one or more of the following: mean color, standard deviation, and color histogram in the RGB color space. Spatial distribution features include one or more of the following: plane boundary complexity, point distribution density, and morphological consistency index. The process of matching planar components of the first point cloud data and the second point cloud data using the multidimensional feature data includes: Obtain the similarity of the multidimensional features of the components in the first point cloud data and the second point cloud data, filter the matching plane according to the similarity threshold of each dimension, and add the matching plane to the candidate matching pool.

6. The point cloud registration method based on multi-source semantic information and plane normal vectors as described in claim 5, characterized in that, The calculation of the rotation matrix using matching components includes: For a matching plane in the candidate matching pool, the rotation matrix is ​​calculated using the normal vector of the matching plane, and the rotation axis and rotation angle are determined using the angle between the normal vectors. The translation vector is calculated by matching the spatial differences between the centroids of the plane. Combining the rotation matrix with the translation vector yields the rotation matrix of the matching plane.

7. The point cloud registration method based on multi-source semantic information and plane normal vectors as described in claim 6, characterized in that, The step of obtaining the initial registration matrix by performing a global multiplane consistency evaluation on the rotation matrix includes: For the target rotation matrix of the matching plane, apply the target rotation matrix to each candidate matching plane in the candidate matching pool; Perform a global multiplane consistency evaluation on the rotated candidate matching planes; The optimal rotation matrix from the evaluation results is used as the initial registration matrix.

8. The point cloud registration method based on multi-source semantic information and plane normal vectors as described in claim 7, characterized in that, The global multiplane consistency evaluation of the rotated candidate matching plane includes: The evaluation result of the rotation matrix is ​​obtained by calculating the weighted average of the cosine similarity of the normal vectors, the intensity similarity, the color similarity, and the area similarity.

9. The point cloud registration method based on multi-source semantic information and plane normal vectors as described in claim 8, characterized in that, The point cloud registration method includes: The first point cloud data and the second point cloud data are initially registered using the initial registration matrix. The ICP algorithm is used to perform fine registration and iterative optimization on the first and second point cloud data after initial registration. Output the final registration result of the first point cloud data and the second point cloud data.

10. A point cloud registration system based on multi-source semantic information and planar normal vectors, characterized in that, The point cloud registration system includes a lidar, and the point cloud registration system is used to implement the point cloud registration method based on multi-source semantic information and plane normal vectors as described in any one of claims 1 to 9.

Images