A point cloud registration method based on graph topology layering

Abstract

This invention discloses a point cloud registration method based on graph topology hierarchical structure, comprising: constructing a feature compatibility map based on geometric invariants according to the initial correspondence set between the source point cloud and the target point cloud to be registered, and performing feature aggregation and confidence prediction on each correspondence to obtain an original confidence score; selecting several discretely distributed seed points with high original confidence scores from the initial correspondence set based on a discretized seed sampling strategy using graph topology density; using the seed points as anchor points, performing a hierarchical cascaded search mechanism based on a dynamic resource pool to obtain a complete inner point cluster; and using a multidimensional spatial distribution perception scoring model to jointly evaluate each node in the complete inner point cluster and solve for the optimal rigid body transformation parameters to achieve registration of the source point cloud and the target point cloud. This method improves the computational efficiency, accuracy, and stability of point cloud registration under conditions of low overlap rate, large scale, and high outlier rate.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and 3D reconstruction technology, specifically relating to a point cloud registration method based on graph topology layering. Background Technology

[0002] With the rapid development of 3D sensing technologies such as LiDAR and RGB-D depth cameras, 3D point cloud data has become a core data form for perceiving the physical world in fields such as computer vision, robot navigation, autonomous driving, and digital twins. Point cloud registration, as a key foundational technology for processing 3D data, has the core task of unifying multiple local point cloud data acquired at different times and from different perspectives into the same global coordinate system by calculating rigid body transformation matrices, thereby achieving complete scene reconstruction or localization.

[0003] In practical applications, point cloud registration in various scenarios places different demands on algorithms. For example, in autonomous driving scenarios, onboard LiDAR continuously collects point cloud data of the surrounding environment. To construct high-precision 3D maps or perform real-time localization and map building, the system must accurately stitch together the point clouds of consecutive frames. However, when the vehicle moves quickly or makes large turns, the overlap area of ​​the point clouds between consecutive frames may be very limited, and outdoor scenes are filled with noise interference from dynamic objects, which places extremely high demands on the robustness of the registration algorithm. Furthermore, when digitally scanning large buildings, industrial parks, or cultural relics sites, the sheer size of the objects often necessitates scanning in blocks from hundreds of different angles. When stitching these blocks together, the large differences in shooting angles often result in an overlap rate of less than 30%, a low overlap problem, and the massive scale of data poses a severe challenge to the computational efficiency of the algorithm.

[0004] Currently, existing point cloud registration methods are mainly divided into two categories: traditional geometric methods and deep learning-based methods. Traditional geometric methods primarily rely on the geometric invariance constraints of rigid body transformations to select interior points, and their technical approaches can be broadly categorized into iterative sampling and globally deterministic methods. The most classic method is RANSAC (Random Sample Consensus) and its variants, which find interior points by iteratively sampling a minimum subset and performing hypothesis testing. For example, the RANSAC algorithm proposed by Fischler et al. (Fischler, Martin A, and Robert C. Bolles. "Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography." Communications of the ACM 24.6 (1981): 381-395.) uses the random sample consensus principle to eliminate outliers, but its computational complexity increases exponentially in scenarios with a high outlier rate. To overcome the uncertainty of RANSAC, researchers have proposed methods based on global optimization. Yang et al.'s TEASER++ (Yang, Heng, Jingnan Shi, and Luca Carlone. "TEASER: Fast and certifiable point cloud registration." IEEE Transactions on Robotics 37.2 (2020): 314-333.) achieves a certifiable optimal solution through truncated least squares. Chen et al.'s SC²-PCR (Chen, Zhi, et al. "SC2-PCR: A second order spatial compatibility for efficient and robust point cloud registration." CVPR. 2022.) utilizes second-order spatial consistency to construct a compatibility matrix and performs spectral clustering ranking. Zhang et al.'s MAC (Zhang, Xiyu, et al. "3D registration with maximal cliques." CVPR. 2023.) transforms the mismatch elimination problem into a maximal clique search problem, finding a consistent set by relaxing the maximal clique constraint.

[0005] With the development of deep learning technology, deep learning-based methods model mismatch elimination as a binary classification or sequence prediction problem, directly using the correspondence as input. For example, the DGR proposed by Choy et al. (Choy, Christopher, Wei Dong, and Vladlen Koltun. "Deep global registration." CVPR. 2020.) uses a six-dimensional convolutional network to predict the confidence of each correspondence and combines it with the weighted Procrustes algorithm to solve the pose. The PointDSC proposed by Bai et al. (Bai, Xuyang, et al. "PointDSC: Robust point cloud registration using deep spatial consistency." CVPR. 2021.) introduces a spatial consistency nonlocal module and combines neural networks with spectral matching algorithms to estimate the inlier probability, but the spatial consistency signal is prone to becoming blurred under high outlier rates. The VBReg method proposed by Jiang et al. (Jiang, Haobo, et al. "Robust outlier rejection for 3D registration with variational Bayes." CVPR. 2023.) utilizes a variational Bayesian inference framework to probabilistically evaluate interior points. The Hunter method proposed by Yao et al. (Yao, Runzhao, et al. "Hunter: Exploring high-order consistency for point cloud registration with severe outliers." IEEE TPAMI 45.12 (2023): 14760-14776.) introduces a hypergraph structure to capture high-order consistency information to address extreme outlier scenarios, but its hyperedge enumeration process limits the algorithm's scalability.

[0006] While existing traditional geometric methods and deep learning-based methods have made some progress in point cloud registration, they still generally have significant limitations when facing low overlap rates, high outlier rates, and large-scale complex scenes. Specifically, iterative sampling methods, represented by RANSAC, experience exponentially increasing computational complexity in high outlier scenarios. Global deterministic methods based on spectral matching or maximum clique search, although theoretically capable of finding optimal solutions, suffer from excessively high computational complexity, making them unsuitable for real-time performance in large-scale outdoor scenes. For deep learning-based methods, the commonly used one-time global search or flattened K-nearest neighbor retrieval strategies based on Euclidean distance often ignore the deep graph topology between corresponding points. This results in difficulties in effectively aggregating sparse and fragmented internal point clusters in low overlap scenarios and a failure to fully utilize confidence information for adaptive search. Furthermore, existing seed-based search frameworks lack awareness of graph cluster structure during sampling, which can easily lead to seed points being over-concentrated in the dominant region and missing key fragmented interior points. At the same time, relying solely on the number of interior points during the hypothesis evaluation phase cannot effectively distinguish between geometrically degenerate solutions and true solutions, and lacks joint consideration of the spatial distribution dispersion and coverage of interior points.

[0007] In summary, existing point cloud registration techniques suffer from problems such as low computational efficiency, difficulty in searching fragmented intrapoints, incomplete sampling coverage, and a single evaluation mechanism when facing complex scenarios with low overlap, large scale, and high outlier rate. Therefore, there is an urgent need for a robust point cloud registration method that can balance computational efficiency and registration accuracy under conditions of low overlap, large scale, and high outlier rate. Summary of the Invention

[0008] To address the aforementioned problems in the existing technology, this invention provides a point cloud registration method based on graph topology layering. The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, this invention proposes a point cloud registration method based on graph topology layering, comprising the following steps: Step 1: Construct a feature compatibility map based on geometric invariants according to the initial correspondence set between the source point cloud and the target point cloud to be registered, and combine a multi-layer graph neural network to perform feature aggregation and confidence prediction on each correspondence in the initial correspondence set to obtain the original confidence score of each correspondence as an interior point; Step 2: Based on the graph topology density discretization seed sampling strategy, select several discretely distributed correspondences with high original confidence scores from the initial correspondence set as seed points; Step 3: Using the seed point as the anchor point, a hierarchical cascaded search mechanism based on the dynamic resource pool is used to perform a hierarchical cascaded search on the feature compatibility map to obtain a complete cluster of interior points. Step 4: Use a multidimensional spatial distribution perception scoring model to jointly evaluate each node in the complete interior point cluster. After eliminating geometrically degenerate solutions, solve for the optimal rigid body transformation parameters to achieve registration between the source point cloud and the target point cloud.

[0009] The beneficial effects of this invention are: This invention proposes a point cloud registration method based on graph topology hierarchies. First, a feature compatibility graph based on geometric invariants is established. The topological connectivity of graph nodes bridges physical spatial fractures, effectively aggregating sparse intrapoint clusters distributed across different separated regions, significantly improving registration success rate under low overlap conditions. Second, a discretization seed sampling strategy based on graph topology density is proposed. By suppressing the connected neighborhoods of selected nodes in the graph, the seed points are forced to maintain structural diversity, ensuring coverage of small but crucial fragmented regions for registration, avoiding search failures and improving robustness. Third, a hierarchical cascaded search mechanism based on a dynamic resource pool is introduced, utilizing the graph topology for progressive exploration. This reduces computational complexity while maintaining search accuracy, thus meeting the real-time registration requirements of large-scale scenarios. Finally, a multi-dimensional scoring model is introduced to effectively eliminate geometric degradation, improving the reliability and accuracy of the final pose estimation. Through these operations, the computational efficiency, accuracy, and stability of point cloud registration under low overlap, large-scale, and high outlier conditions are significantly improved.

[0010] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0011] Figure 1 A flowchart illustrating a point cloud registration method based on graph topology layering provided in an embodiment of the present invention; Figure 2 A flowchart illustrating another point cloud registration method based on graph topology layering provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a hierarchical cascaded search based on a dynamic resource pool, provided in an embodiment of the present invention. Figure 4 This is a visual comparison of the registration results of the method of this invention and current mainstream advanced methods on an indoor scene dataset. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] This invention proposes a point cloud registration method based on graph topological layering. Firstly, addressing the issue that Euclidean distance-based neighborhood search strategies easily fail under field-of-view occlusion or feature sparsity, this method establishes a non-Euclidean space search strategy based on geometrically compatible graphs. This strategy leverages the topological connectivity of graph nodes to bridge physical spatial breaks, effectively aggregating sparse intra-point clusters distributed across different separated regions, significantly improving registration success rate under low overlap conditions. Secondly, addressing the lack of graph cluster structure awareness and the tendency to get trapped in local optima in dominant regions in existing sampling methods, this invention proposes a discretized seed sampling strategy based on graph topological density. By suppressing the connected neighbors of selected nodes in the graph, this strategy forces seed points to maintain structural diversity, ensuring coverage of small but crucial fragmented regions for registration. Thirdly, addressing the exponential increase in computational cost for existing globally deterministic methods when processing point clouds of tens of thousands or more, this method... To address the shortcomings of previous methods, a hierarchical cascaded search mechanism based on a dynamic resource pool is introduced. This mechanism utilizes graph topology for progressive exploration, reducing computational complexity while maintaining search accuracy, thus meeting the real-time registration requirements in large-scale scenarios. Finally, to address the issue that relying solely on the number of interior points for optimal assumptions can easily lead to mismatches (such as incorrectly aligned planes), a multi-dimensional scoring model incorporating spatial dispersion and geometric coverage is introduced. By jointly evaluating the number of interior points, spatial dispersion, and alignment accuracy, geometric degradation is effectively eliminated, improving the reliability and accuracy of the final pose estimation.

[0014] Please see the joint reference. Figure 1 and Figure 2 , Figure 1 This is a flowchart illustrating a point cloud registration method based on graph topology layering provided in an embodiment of the present invention. Figure 2 This is a flowchart illustrating another point cloud registration method based on graph topology layering provided by an embodiment of the present invention. The point cloud registration method based on graph topology layering provided by the present invention mainly includes the following steps: Step 1: Construct a feature compatibility map based on geometric invariants according to the initial correspondence set between the source point cloud and the target point cloud to be registered, and combine a multi-layer graph neural network to perform feature aggregation and confidence prediction on each correspondence in the initial correspondence set to obtain the original confidence score of each correspondence as an interior point.

[0015] Understandably, input data processing and problem definition are required before step 1.

[0016] First, the input source point cloud to be registered. and target point cloud Feature extraction and nearest neighbor matching are performed to obtain an initial set of correspondences containing noise.

[0017] Specifically, let the source point cloud to be registered be... The target point cloud is ,in and These represent the coordinates of a point in three-dimensional space. and These represent the number of points in the point cloud. The objective of this invention is to estimate an optimal rigid body transformation parameter, including the rotation matrix. Translation vector This makes the source cloud After transformation, it is compared with the target point cloud. Achieve spatial alignment within overlapping areas.

[0018] This embodiment uses the source point cloud to be registered. and target point cloud Feature extraction and nearest neighbor matching are performed to obtain an initial set of correspondences containing noise, denoted as: ; in, Represents the first in the initial set of correspondences. One correspondence, and For the first For the matching point, This represents the total number of corresponding relationships.

[0019] For any correspondence If a point satisfies the following spatial distance constraints, it is considered an interior point: ; in, The preset interior point distance threshold, This represents the Euclidean distance.

[0020] Then, the optimization objective for solving the rigid body transformation parameters is defined.

[0021] The core task of this invention is to start from an initial set containing a large number of outliers. In this process, the maximum subset of interior points that satisfies geometric consistency is efficiently selected. Then the optimization objective for solving the rigid body transformation parameters can be defined as follows: ; in, To be the optimal transformation parameters, This is an indicator function.

[0022] To solve the above optimization objective, this embodiment first uses the geometric invariants (length and angle) in rigid body transformation to construct a graph that can describe the inherent consistency of the correspondence, and then learns features through a graph neural network to predict the reliability of each correspondence.

[0023] Optionally, in this embodiment, step 1 mainly includes three core sub-steps: geometric consistency measurement and graph topology generation, feature aggregation based on relative position encoding, and confidence prediction, as detailed below: 11) Take each correspondence in the initial correspondence set as a node, calculate the relative distance difference and the angle difference between the normal vectors of any two nodes in the source point cloud and the target point cloud. If the preset distance tolerance threshold and angle tolerance threshold are satisfied at the same time, establish a connection edge between the two nodes to generate a feature compatibility map.

[0024] Specifically, for the initial set of correspondences Each correspondence Connect the points in the source point cloud and points in the target point cloud .

[0025] To construct a compatibility map This invention defines a graph node set. Correspondence set One-to-one correspondence, meaning each node represents a pair of corresponding relationships. For any two nodes and Calculate the relative distance difference between the source point cloud and the target point cloud respectively. Difference between the angles of the normal vector and the normal vector : ; ; in, and Representing points respectively and Surface normal vector at that location, Represents the Euclidean norm. A connecting edge is established in the graph only if both distance and angle constraints are satisfied. : ; in, and These are preset distance tolerance thresholds and angle tolerance thresholds, respectively. This connection rule ensures that edges in the graph exist only between correspondences with highly consistent geometric structures.

[0026] It should be noted that, to prevent excessive computation due to overly dense graphs, this embodiment can set a maximum node degree. Cut off.

[0027] 12) Initialize a high-dimensional feature vector for each node that integrates the local geometric descriptors of the source and target points, and introduce a relative position spatial encoding mechanism. Then, use a graph neural network to perform feature aggregation and update on the high-dimensional feature vector, so that the high-dimensional feature vector of each node integrates its own local similarity information and the global geometric structure consistency information of the neighborhood to obtain the final node features.

[0028] Specifically, in order to make the network invariant to rotation and translation, this invention does not directly use the absolute coordinates of points for feature learning, but introduces a relative position spatial encoding mechanism.

[0029] First, for each node Initialize high-dimensional feature vectors This feature is derived from the source point. and target point The 3D coordinates are stitched together and then mapped using a multilayer perceptron (MLP). This feature integrates the local geometric descriptors of the source and target points.

[0030] Subsequently, through Layered graphical neural networks perform feature updates. In the first... During the information transmission process of each layer, for each edge Extract the relative position vector from the source point cloud. And using spatial encoding functions Map it to directional features.

[0031] Optionally, as an implementation method, the graph neural network in this embodiment adopts a multi-layer equivariant message passing network structure, and uses a spherical harmonic function to encode the relative position vector. In each layer, the features of the neighboring nodes and the fusion information of the relative position encoding are aggregated for each node, and finally the feature aggregation and update are performed through residual connections.

[0032] For example, the multilayer graph neural network in this embodiment uses a multilayer equivariant message passing network to update node features. Unlike existing technologies, this embodiment designs a more lightweight yet efficient 3-layer network structure with the following specific parameters: Network depth: Set the message passing layer number L=3.

[0033] Feature Dimensions: The number of feature channels in each hidden layer is set to [128, 256, 256]. This means that the first layer outputs 128-dimensional features, and the second and third layers output 256-dimensional features.

[0034] Position encoding: Using spherical harmonic functions to represent the relative position vectors of corresponding relationships in the source point cloud. Encode the feature with an encoding order of 2 to ensure its equivariance to rigid body transformations.

[0035] Message aggregation: In each layer, a node aggregates the features of its neighbors with the fusion information of their relative position codes, and updates its own features through residual connections.

[0036] In this embodiment, the feature aggregation update formula is: ; In the formula, Represents a node In the The feature vector of the layer; Represents a node In the The feature vector of the layer; Represents a node The set of neighboring nodes in the feature compatibility graph G; This is a feature transformation function used to fuse the feature information of neighboring nodes with their relative geometric position information; Represents a node In the The feature vector of the layer; Indicates feature concatenation operation; This is a relative position encoding function; Source Point Cloud The point in the middle, For target point cloud The point in the middle.

[0037] Through the above aggregation mechanism, the feature vector of each node... It not only includes its own local similarity information, but also aggregates its global geometric structure consistency information in the graph neighborhood.

[0038] 13) Based on the final node features, using a multilayer perceptron and a sigmoid activation function, the confidence score of the corresponding node is output as the confidence score of the interior point.

[0039] Specifically, this embodiment uses a two-layer MLP structure to calculate the original confidence score, and uses the Sigmoid activation function to constrain the original confidence score within the [0,1] interval. The original confidence score is then expressed as: ; In the formula, Represents a node The original confidence score, This represents the Sigmoid activation function. This indicates the operation of a multilayer sensor. This represents the final node features obtained after L layers of feature aggregation.

[0040] Step 2: Based on the graph topology density discretization seed sampling strategy, select several discretely distributed correspondences with high original confidence scores from the initial correspondence set as seed points.

[0041] Specifically, after obtaining the original confidence scores for each correspondence in step 1, directly selecting the Top-K points with the highest scores as seed points can easily lead to excessive concentration of seed points in a dominant region with significant structural features, while ignoring fragmented intrapoint clusters separated due to occlusion or viewpoint changes. To address this, this step proposes a discretized seed sampling strategy based on graph topological density, introducing graph-theoretical distance as a suppression factor to force seed points to maintain a discrete distribution on the graph structure.

[0042] Optionally, in this embodiment, step 2 mainly includes three core sub-steps: graph topological distance definition and initialization, dynamic penalty scoring mechanism, and seed point iterative optimization, as detailed below: 21) Establish an initial candidate set based on the original confidence scores, and initialize the selected seed point set to an empty set; define the graph topological distance between any two nodes in the feature compatibility graph as the shortest path hop count in the feature compatibility graph, and set the graph topological distance between non-connected nodes to infinity.

[0043] Specifically, based on the set of confidence scores output in step 1 Establish an initial candidate set and initialize the selected seed point set to an empty set. To measure the structural correlation between nodes, we define any two nodes in the graph... and graph topological distance between For its compatibility map The shortest path hop count in the graph. For sparse graphs, Dijkstra's algorithm is used to quickly calculate this distance. If two nodes are not connected in the graph, their distance is set to infinity.

[0044] 22) Calculate the dynamic comprehensive score of each candidate node based on the dynamic penalty scoring mechanism.

[0045] Specifically, this invention employs an iterative greedy strategy for sampling. In the t-th iteration, instead of relying solely on the original confidence level, it calculates the confidence level for each candidate node. Dynamic composite score The score is calculated by subtracting the structural inhibition term from the seed point from the original confidence score. Therefore, the formula for calculating the dynamic composite score is: ; In the formula, For the front The dynamic comprehensive score of node k in each iteration. The original confidence score for node k; The preset penalty intensity coefficient is used to control the balance between dispersion and confidence. For the front The set of seed points already selected in the round; The decay kernel function based on graph distance is expressed as follows: ; in, For the spectral bandwidth parameter, In the feature compatibility map, nodes With the selected seed node The shortest path hop count between them These represent different nodes in the feature compatibility map.

[0046] 23) An iterative greedy strategy is adopted to iteratively update the dynamic comprehensive score. In each iteration, the node with the highest dynamic comprehensive score is selected and added to the seed point set. The iteration is repeated until the number of seed points reaches the preset threshold.

[0047] Specifically, in each iteration, the following operations are performed: A. Scoring Update: Calculate the dynamic composite score of all currently unselected nodes based on the formula above. For points that are closely connected to the selected seed points on the graph, i.e. Smaller nodes have larger suppression terms, leading to a significant decrease in the overall score; while for nodes located in different connected components or far apart, their suppression terms approach 0, preserving the original confidence advantage.

[0048] B. Extreme Value Selection: Select the node with the highest current overall score. Add to seed point set:

[0049] C. Termination Judgment: Repeat the "Score Update" and "Extreme Value Selection" steps until the number of seed points in the set reaches the preset threshold. .

[0050] Step 3: Using the seed point as the anchor point, a hierarchical cascaded search mechanism based on the dynamic resource pool is used to perform a hierarchical cascaded search on the feature compatibility map to obtain a complete cluster of interior points.

[0051] Specifically, this step uses the samples selected in step 2. Using high-confidence seed points as anchors, a hierarchical cascaded search is performed using the topological structure of the compatibility graph, aiming to aggregate complete clusters of interior points from the sparse graph. To address the computational redundancy problem under large-scale data, this invention specifically introduces a global dynamic resource pool mechanism.

[0052] Please see Figure 3 , Figure 3 This is a schematic diagram of a hierarchical cascaded search based on a dynamic resource pool, provided as an embodiment of the present invention. Optionally, in this embodiment, step 3 mainly includes four core sub-steps: global dynamic resource pool construction and management, first-level cascaded search, second-level cascaded search, and dynamic confidence propagation, as detailed below: 31) Construct a global dynamic resource pool that covers all corresponding relationships and set a global state mask vector; wherein, the initial state of the global state mask vector is 0. When the corresponding relationship is confirmed as an interior point, the corresponding mask value is set to 1, and in the future, only nodes with a mask value of 0 are searched and verified.

[0053] Specifically, before the search begins, a system is established that covers all corresponding relationships. global state mask vector The initial state is all 0s, indicating that all correspondences are in the "unallocated resource pool". When a correspondence is identified as an interior point and added to the candidate set during the search of any seed point, its corresponding Mask value is set to 1. In the subsequent expansion of seed points, only nodes with a Mask value of 0 are searched and verified. This mechanism ensures that each correspondence is evaluated only once, avoids repeated calculations of overlapping regions, and significantly reduces the time complexity of the algorithm.

[0054] 32) Construct the first cascade layer, including: retrieving unassigned neighbor nodes that are directly connected to the seed point in the feature compatibility graph; constructing a local induced subgraph by retrieving node feature similarity and calculating the centrality score; adding nodes with a centrality score higher than the dynamic threshold to the inlier set. .

[0055] Specifically, for each seed point, neighborhood aggregation based on spectral consistency... First, construct the first cascade layer. In the compatibility graph G, retrieve all points related to the seed point. There exists a set of neighbor nodes that are directly connected by an edge and are marked as "unassigned" in the resource pool. To remove spurious connections from the neighbor set, the node feature vectors generated in step 1 are used. Similarity retrieval is performed, and a local induced subgraph is constructed. The centrality score of each neighbor node is calculated by computed using the principal eigenvectors of the adjacency matrix, and nodes with scores above a dynamic threshold are retained. The node is added to the current hypothetical interior set. .

[0056] 33) Construct the second cascade layer, including: the set of interior points The nodes in the graph are used as the expansion source. The neighbor nodes of each node in the feature compatibility graph are searched, and nodes that meet the constraints are selected through geometric verification and added to the interior point set. .

[0057] Specifically, the quadratic extension based on geometric constraints uses the set of nodes identified as interior points in the first cascade layer. As a new source of expansion, its neighboring nodes in the graph (i.e., 2-Hop nodes relative to the seed point) are searched outwards to construct the second cascade layer. Since the spectral consistency signal may weaken as the graph topological distance increases, a direct geometric verification strategy is adopted in this layer. For candidate nodes... Calculate its relationship with Geometric compatibility of nodes in the middle; if constraints are satisfied, add them to the interior point set. .

[0058] 34) Construct a dynamic confidence propagation model, including: for the interior set The original confidence score of each node is corrected based on the corrected confidence score of its parent node. The results of each level are then merged to obtain the complete cluster of interior points corresponding to the seed point.

[0059] Specifically, in the aforementioned hierarchical expansion process, the admission determination of a node depends not only on its own original confidence level but also on the reliability of the nodes at the next higher level above its "parent node." This invention constructs a confidence inheritance and propagation model: Let node... In the Hierarchy ( ) is visited, and its parent node set is That is, the upper layer and Connected interior points.

[0060] The formula for correcting the original confidence of each node based on the corrected confidence of its parent node is as follows: ; In the formula, Represents a node In the Confidence score after hierarchical adjustment. Represents a node The original confidence score, Represents a node parent node In the Confidence score after hierarchical adjustment. This is a weighting balancing factor used to adjust the weights of its own features and structural priors. Represents a node The set of parent nodes.

[0061] Finally, by merging the results from each level, we obtain the complete set of interior point hypotheses corresponding to the seed point, which is the interior point cluster after hierarchical search, denoted as . .

[0062] Step 4: Use a multidimensional spatial distribution perception scoring model to jointly evaluate each node in the complete interior point cluster. After eliminating geometrically degenerate solutions, solve for the optimal rigid body transformation parameters to achieve registration between the source point cloud and the target point cloud.

[0063] 41) Based on the multidimensional spatial distribution perception scoring model, the multidimensional spatial quality comprehensive score of each subset of interior points in the complete interior point cluster is obtained by jointly evaluating the three dimensions of interior point quantity, spatial dispersion and alignment accuracy.

[0064] Specifically, existing transformation solution evaluation methods typically rely solely on the number of interior points. When faced with textures containing numerous repetitive patterns or symmetrical structures, these methods easily fall into degenerate solutions where interior points are dense but the overall structure suffers from geometric misalignment. To address this, this invention jointly evaluates three dimensions: the scale of interior point count, spatial dispersion, and alignment accuracy, constructing a multidimensional spatial distribution-aware scoring model to calculate the comprehensive multidimensional spatial quality score for each candidate rigid body transformation matrix. The calculation formula is: ; ; In the formula, Let be the total number of interior points in the interior point set, representing the consensus scale of the registration. Spatial distribution dispersion is used to measure the dispersion of points within the source cloud around its centroid. Spatial divergence degree, To register residuals, the distribution scale parameter The exponential decay term is designed to strongly penalize geometric degenerate solutions with extremely concentrated spatial distributions. The average alignment residual is calculated as follows: Contains minimal constants The accuracy term is used to reward high-quality hypotheses with smaller alignment errors and tighter fits.

[0065] 42) For each subset of interior points in the complete interior point cluster, perform weighted singular value decomposition based on the corrected confidence score of each interior point, and solve in parallel the candidate rigid body transformation matrix corresponding to each subset of interior points; and select the rigid body transformation matrix corresponding to the subset of interior points with the highest comprehensive quality score in multidimensional space as the optimal solution.

[0066] Specifically, for all seed interior point clusters, the corrected confidence score obtained from dynamic propagation in step 3 is used as the weight corresponding to the interior point. Through weighted singular value decomposition, a candidate rigid body transformation matrix corresponding to the interior point cluster is solved in parallel, and the above comprehensive score is applied. The calculation involves sorting all candidate rigid body transformation solutions generated in parallel in descending order of their scores, and selecting the rigid body transformation matrix corresponding to the subset of interior points with the highest comprehensive quality score in multidimensional space as the current optimal solution.

[0067] Furthermore, in order to further eliminate the interference of residual external points at the global edge, the current optimal solution is fine-tuned by performing iterative weighted singular value decomposition on the global internal point set to obtain the final transformed solution.

[0068] ; ; In the formula, Gaussian weights For the correspondences with larger alignment residuals, a soft weighting is applied. After the iteration converges, the comprehensive score of the fine-tuned assumptions is re-evaluated, and the set of matrices with the highest comprehensive score (i.e., the one with a reasonable spatial distribution and the smallest alignment error) is selected. As the final optimal rigid body transformation parameters.

[0069] Thus, through the above operations, the optimal rotation matrix and translation vector have been obtained, which can be used to achieve point cloud registration.

[0070] It is understood that the feature aggregation network architecture with geometric invariants proposed in this invention requires model training before use. The specific training configuration is as follows: Optimizer: The Adam optimizer is used, with momentum parameters set to β1=0.9 and β2=0.999.

[0071] Learning rate strategy: The initial learning rate is set to 2.5 × 10. 5 It is adjusted in conjunction with a cosine learning rate scheduler, and linear warm-up is used for the first 10 epochs.

[0072] Loss function: A binary cross-entropy loss function with positive sample weights is used. The positive sample weights vary depending on the dataset. Different settings: indoor scenes are set to 2.0, and outdoor scenes are set to 1.5.

[0073] Regularization: The weight decay coefficient is set to 1×10. -6 .

[0074] In addition, other parameter settings for network training can be found in Table 1 below.

[0075] Table 1 Training Parameter Settings

[0076] This invention proposes a graph-based topological hierarchical point cloud registration method. First, it establishes a feature compatibility graph based on geometric invariants, utilizing the topological connectivity of graph nodes to bridge physical spatial fractures and effectively aggregate sparse intrapoint clusters distributed across different separated regions, significantly improving registration success rate under low overlap conditions. Second, it proposes a discretized seed sampling strategy based on graph topological density. By suppressing the connected neighborhoods of selected nodes in the graph, it forces seed points to maintain structural diversity, ensuring coverage of small but crucial fragmented regions for registration, avoiding search failures and improving robustness. Third, it introduces a hierarchical cascaded search mechanism based on a dynamic resource pool, using the graph topology for progressive exploration. This reduces computational complexity while maintaining search accuracy, thus meeting the real-time registration requirements of large-scale scenarios. Finally, it introduces a multi-dimensional scoring model to effectively eliminate geometric degradation, improving the reliability and accuracy of the final pose estimation. Through these operations, the computational efficiency, accuracy, and stability of point cloud registration under low overlap, large-scale, and high outlier conditions are significantly improved.

[0077] To verify the effectiveness and superiority of the point cloud registration method based on graph topology layering proposed in this invention, point cloud registration experiments were conducted on real scene datasets in indoor standard scenes, indoor low-overlap scenes, and outdoor large-scale scenes, and compared with existing mainstream registration methods. At the same time, ablation comparison experiments were conducted on the 3DMatch standard scene and 3DLoMatch low-overlap scene datasets.

[0078] Experiment 1: Point Cloud Registration for Indoor Standard Scene 1.1 Experimental conditions: This experiment uses the internationally recognized indoor 3D reconstruction benchmark dataset 3DMatch for validation. This dataset contains RGB-D scan data from eight different scenes, totaling 1623 pairs of test point cloud fragments. The experiment aims to verify the advantages of this invention in registration accuracy and computational efficiency compared to existing mainstream technologies in standard indoor scenes, especially when facing partial overlap and sensor noise.

[0079] 1.2 Experimental Content and Result Analysis: To verify the beneficial effects of the method of the present invention, it was compared with current mainstream traditional geometric methods (such as RANSAC, SC²-PCR, MAC) and deep learning methods (such as PointDSC, VBReg) in the same hardware environment.

[0080] Evaluation Indicator Explanation: Registration Recall Rate (RR): The proportion of samples that were successfully registered (success criteria: rotation error <15° and translation error <30cm).

[0081] Rotation Error (RE): The average rotation deviation of successfully registered samples.

[0082] Translation Error (TE): The average translation deviation of successfully registered samples.

[0083] Average Time: The average running time for processing a pair of point clouds.

[0084] The experimental comparison results are shown in Table 2 below: Table 2 Comparison of Indoor Dataset Registration Results Methods

[0085] The experimental results, as shown in the table above, demonstrate that the method of this invention outperforms existing technologies in all key indicators. The registration recall rate of this invention reaches 88.56%, an improvement of approximately 4.6 percentage points compared to the current best-performing global search methods (MAC and SC²-PCR). Simultaneously, the rotation error is reduced to 1.67°, the lowest among all compared methods, proving the extremely high accuracy of the hierarchical cascaded search strategy in eliminating mismatches. While maintaining the highest accuracy, the average processing time of this invention is only 0.10 seconds, comparable to the lightweight PointDSC method. Compared to the traditional RANSAC method (2.86 seconds) and the MAC method (5.54 seconds), this invention achieves a speed improvement of 20 to 50 times, perfectly meeting the requirements of real-time processing. This invention successfully resolves the contradiction between the traditional method's "high accuracy but slow speed" and the early deep learning methods' "fast speed but limited accuracy," achieving the optimal balance between accuracy and efficiency.

[0086] also, Figure 4The paper also compares the registration performance of the method of this invention with that of current mainstream advanced methods (SC²-PCR, MAC, VBReg) on ​​typical difficult samples of the indoor 3DMatch dataset. Figures (a) to (c) show the registration performance of the existing SC²-PCR, MAC, and VBReg methods, respectively. It can be seen that the existing methods exhibit poor stability when facing scenes with low overlap or high geometric symmetry. Specifically, the SC²-PCR method in Figure (a) and the MAC method in Figure (b), as global methods based on spectral matching, are prone to making incorrect geometric assumptions when interior points are extremely sparse or unevenly distributed. As shown by the red dashed circles in the first and second rows, the point clouds show obvious misalignment and separation, failing to achieve correct alignment. The VBReg method in Figure (c), as a deep learning-based method, is prone to getting trapped in local optima when handling large-scale rigid body transformations. As shown by the red circles in the third row, although the point clouds roughly overlap, there are significant rotational deviations in the detailed structures. In contrast, Figure (d) shows the registration results of the present invention, in which the source point cloud (cyan) and the target point cloud (yellow) are closely and accurately aligned in all the samples shown.

[0087] Experiment 2: Point Cloud Registration for Indoor Low-Overlap Scenes 2.1 Experimental Conditions This embodiment uses the 3DLoMatch dataset for validation. This dataset is a subset of 3DMatch and is specifically designed to test the algorithm's performance under low overlap conditions. It contains image pairs with an overlap rate of only 10% to 30%, and exhibits significant viewpoint variations and geometric occlusion. 2.2 Experimental Content and Result Analysis: To address the characteristics of sparse and fragmented interior points in low overlap scenarios, this embodiment adaptively adjusts the algorithm parameters as follows: Feature extraction: FCGF deep learning feature descriptor is used to enhance feature recognition in low-overlap regions.

[0088] Seed sampling strategy: In order to cover possible small overlapping areas, the seed point sampling ratio is increased to p=0.15.

[0089] Layered search thresholds: In order to recall weak interior point signals in a lot of noise, the admission thresholds of each layer are appropriately relaxed. The threshold for Layer 1 is set to τ1=0.4 and the threshold for Layer 2 is set to τ2=0.2.

[0090] Candidate hypothesis optimization: Increase the number of candidate transformations for final optimization to K=10 to prevent the optimal solution from being missed in the early stages.

[0091] Table 3 below shows a performance comparison between the method of this invention and current mainstream methods on the 3DLoMatch dataset. The data shows that the method of this invention still maintains a high registration success rate even with extremely low overlap.

[0092] Table 3 Comparison Results of Registration Effects in Low-Overlap Indoor Scenes

[0093] Experimental results show that, under the harsh condition that the RANSAC algorithm only achieves a success rate of 46.38%, the method of this invention improves the registration recall rate to 61.59%, which is about 1.7 percentage points higher than the second-place (MAC) and 5.5 percentage points higher than PointDSC. This demonstrates the unique advantages of the "graph topology hierarchical expansion" strategy in traversing non-overlapping regions and aggregating and separating interior point clusters. The rotation error (3.23°) and translation error (9.75cm) of this invention are the lowest among all compared methods, indicating that in successfully registered samples, this invention can discover more and more accurate interior points, thereby calculating more accurate pose transformations. Experimental results verify that the "discretized seed sampling" and "dynamic resource pool search" mechanisms proposed in this invention can effectively address the problem of "island-like" interior point distribution in low-overlapping scenarios, avoiding registration failures caused by concentrated sampling in traditional methods.

[0094] Experiment 3: Point Cloud Registration for Large-Scale Outdoor Scenes 3.1 Experimental Conditions This embodiment uses the authoritative benchmark KITTI Odometry dataset in the field of autonomous driving for validation. This dataset is collected by vehicle-mounted LiDAR and includes various complex environments such as urban roads, highways, and rural roads. Compared with indoor scenes, outdoor scenes have a huge amount of point cloud data (the number of key points in a single frame is usually more than 5,000 to 20,000), sparse distribution, and contain a large number of ground planar features, which places extremely high demands on the computational efficiency and degradation resistance of the algorithm.

[0095] 3.2 Experimental Content and Result Analysis In this embodiment, sequences 08-10 from the KITTI dataset are selected as the test objects. First, voxel downsampling is performed on the radar point cloud. Interior point mining is then performed using the dynamic resource pool hierarchical search strategy of this invention. Considering the characteristics of outdoor scenes, this invention utilizes graph topology to quickly remove a large number of ground outliers and employs a spatial distribution scoring mechanism to preferentially select non-degenerate solutions with three-dimensional constraints.

[0096] Table 4 below shows the comparison results of the registration effect of the method of the present invention and the current mainstream methods in large-scale outdoor scenes.

[0097] Table 4 Comparison of Registration Results for Large-Scale Outdoor Scenes

[0098] Experimental results show that when processing large-scale data containing thousands to tens of thousands of corresponding points, the average processing time of the method of this invention is only 0.12 seconds. Compared with the MAC method based on maximum clique search (3.29 seconds), the speed is improved by 27 times; compared with the traditional RANSAC method (5.43 seconds), the speed is improved by 45 times. Even compared with the similar deep learning method PointDSC (0.28 seconds), the speed is more than twice as fast. The registration recall rate of this invention is as high as 99.82%, achieving almost perfect registration for all test samples, which is superior to all comparison methods. This shows that the scoring mechanism of this invention can effectively identify the common "ground plane degradation" problem in outdoor scenes, ensuring that the output pose has the correct 6 degrees of freedom constraints.

[0099] Table 5 below shows the input stress test results for different magnitudes of input stress compared to the current mainstream methods.

[0100] Table 5. Input pressure test results corresponding to different magnitudes of quantity.

[0101] Further stress tests, as shown in the table above, revealed that when the number of input points increased from 5,000 to 20,000, the computation time of existing technologies (such as SC²-PCR) increased exponentially (approximately 13.5 times), while the computation time of this invention, thanks to the pruning effect of the "dynamic resource pool," only increased linearly (only 3 times). At an extreme scale of 20,000 points, its running speed surpassed all comparative methods, demonstrating its core advantage in large-scale data processing.

[0102] Experiment 4: Ablation Experiment of Core Module To verify the specific contribution of the four core technical steps proposed in this invention (feature compatibility map construction, graph topology discrete sampling, hierarchical cascaded search, and multidimensional space scoring) to the final registration effect, this embodiment conducted ablation comparison experiments on the 3DMatch standard scene and the 3DLoMatch low overlap scene dataset, and the results are shown in Table 6.

[0103] Table 6 Comparison of ablation experiment results for core modules

[0104] Based on the above comparative data, the following conclusions can be drawn: The criticality of the multi-dimensional spatial scoring mechanism (comparative experiment G vs A): When the spatial distribution dimension is removed and scoring is based solely on the number of interior points, the registration success rate in low-overlap scenes drops sharply from 61.59% to 56.15% (a decrease of 5.44%). This indicates that under limited field of view, a large number of erroneous transformation assumptions (such as planar degradation) will produce an artificially high number of interior points, which must be effectively suppressed by the spatial dispersion index proposed in this invention. After replacing the graph topology sampling of this invention with traditional spatial NMS, the performance on 3DLoMatch decreased by 4.55%. This directly proves that when the interior point distribution is fragmented, simply dispersing seed points in physical space is insufficient; graph topology distance must be used to force seed points to cover different connected components to ensure the comprehensiveness of the search. Compared with simple voting or two-stage search, the three-layer cascaded search strategy of this invention significantly improves the recall capability of interior points while maintaining high accuracy, especially on 3DLoMatch, improving by approximately 2.6% to 4.3%. This verifies that the "dynamic threshold propagation" mechanism can effectively utilize the strong signals in the core region to drive the weak signals in the edge region, achieving in-depth mining of inlier clusters. Although the impact on the final recall rate is small, the inlier precision on 3DMatch decreased from 87.21% to 79.06% after removing the relative position encoding. This indicates that the feature compatibility map constructed in this invention can provide stronger geometric constraints and significantly reduce noise interference in the feature matching stage.

[0105] In summary, the point cloud registration method based on graph topology hierarchies proposed in this invention significantly improves registration stability in low overlap and fragmented scenarios compared to existing technologies. Existing technologies often fail in low overlap scenarios due to sparse feature distribution and fragmented, isolated internal point clusters. This invention employs a discretized seed sampling strategy based on graph topology density to force seed points to cover different connected components in the geometric compatibility graph. Utilizing a hierarchical cascaded search mechanism, it can start from reliable local seeds, traverse non-physically connected regions, and recall occluded and segmented edge internal points. On the publicly available indoor low overlap benchmark dataset 3DLoMatch, using FCGF features, the registration recall rate of this invention reaches 61.59%. This represents a 15.2 percentage point improvement compared to the traditional RANSAC method (46.38%) and a 1.74 percentage point improvement compared to the current state-of-the-art global search method MAC (59.85%). This demonstrates that the present invention can effectively recover fragmented correspondences even under extreme geometric occlusion conditions, exhibiting strong robustness. It significantly reduces computational complexity in large-scale scenarios. For large-scale scenarios such as outdoor autonomous driving, existing globally deterministic methods typically require constructing fully connected matrices or performing maximum clique searches, resulting in high time complexity and difficulty meeting real-time requirements. The present invention introduces a search mechanism based on a dynamic resource pool, avoiding redundant calculations through a global state mask, and achieving dynamic pruning of the search space, reducing computational complexity to near linear incremental increases. On the KITTI autonomous driving dataset, the average processing time for a pair of point clouds using the present invention is only 0.12 seconds, a 27-fold speedup compared to the MAC method (3.29 seconds) and a 45-fold speedup compared to RANSAC (5.43 seconds). In stress tests, when the number of input corresponding points increases from 5,000 to 20,000, existing techniques, such as SC²-PCR, increase in processing time by approximately 13.5 times, showing an exponential trend, while the present invention only increases by 3 times, showing a linear trend. With a massive input of 20,000 points, this invention outperformed all comparable methods with a processing speed of 0.24 seconds, demonstrating its core advantage in large-scale data processing. It effectively solves the mismatch problem caused by high symmetry and geometric degradation, achieving high registration accuracy. Existing methods typically only select the transformation matrix based on the number of interior points, which easily leads to geometric degradation when dealing with highly symmetric or texture-repetitive areas such as the ground and walls. The multi-dimensional spatial distribution perception scoring model proposed in this invention introduces spatial dispersion in addition to quantitative indicators, effectively suppressing the erroneous assumption of spatial distribution degradation. On the 3DMatch standard dataset, the rotation error of this invention is reduced to 1.67° and the translation error to 6.24 cm, both the lowest values ​​among all comparable methods.On the KITTI dataset, the registration recall rate of this invention reached 99.82%, achieving near-perfect registration for all test samples. This demonstrates that the scoring mechanism ensures the output pose has correct six-degree-of-freedom constraints, achieving sub-voxel level accuracy. It also reduces the dependence on the quality of the feature extractor. By constructing an SE(3) isovariant feature compatibility map, this invention enhances feature discriminability at the geometric level. Even with poor quality initial feature descriptors, this invention can still mine correct geometric patterns through graph topology.

[0106] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A point cloud registration method based on graph topological layering, characterized in that, Includes the following steps: Step 1: Construct a feature compatibility map based on geometric invariants according to the initial correspondence set between the source point cloud and the target point cloud to be registered, and combine a multi-layer graph neural network to perform feature aggregation and confidence prediction on each correspondence in the initial correspondence set to obtain the original confidence score of each correspondence as an inlier. Step 2: Based on the graph topology density discretization seed sampling strategy, select several discretely distributed correspondences with high original confidence scores from the initial correspondence set as seed points; Step 3: Using the seed point as the anchor point, perform a hierarchical cascaded search on the feature compatibility map using a hierarchical cascaded search mechanism based on a dynamic resource pool to obtain a complete cluster of interior points; Step 4: Use a multidimensional spatial distribution perception scoring model to jointly evaluate each node in the complete internal point cluster. After eliminating geometrically degenerate solutions, solve for the optimal rigid body transformation parameters to achieve registration between the source point cloud and the target point cloud.

2. The point cloud registration method based on graph topology layering according to claim 1, characterized in that, Before step 1, the following is also included: Source cloud for registration and target point cloud Feature extraction and nearest neighbor matching are performed to obtain an initial set of correspondences containing noise. , denoted as: ； in, Represents the first in the initial set of correspondences. One correspondence, and For the first For the matching point, This represents the total number of corresponding relationships; The optimization objective for solving the rigid body transformation parameters is defined as follows: ； In the formula, This represents the optimal rotation matrix obtained by solving the problem. This represents the optimal translation vector obtained by solving the problem. Represents the rotation matrix. Represents the translation vector. For indicator functions, Represents Euclidean distance. This is the preset interior point distance threshold.

3. The point cloud registration method based on graph topology layering according to claim 1, characterized in that, Step 1 specifically includes: 11) Take each correspondence in the initial correspondence set as a node, calculate the relative distance difference and the angle difference between the normal vectors of any two nodes in the source point cloud and the target point cloud. If the preset distance tolerance threshold and angle tolerance threshold are satisfied at the same time, then establish a connection edge between the two nodes to generate a feature compatibility map. 12) Initialize a high-dimensional feature vector for each node that integrates the local geometric descriptors of the source point and the target point, and introduce a relative position spatial encoding mechanism. Then, perform feature aggregation and update on the high-dimensional feature vector through a graph neural network, so that the high-dimensional feature vector of each node integrates its own local similarity information and the global geometric structure consistency information of the neighborhood to obtain the final node features. 13) Based on the final node features, using a multilayer perceptron and a sigmoid activation function, output the confidence score of the node's correspondence as an interior point.

4. The point cloud registration method based on graph topology layering according to claim 3, characterized in that, In step 12), the graph neural network adopts a multi-layer equivariant message passing network structure and uses spherical harmonic functions to encode the relative position vector. In each layer, the features of each node are aggregated with the fusion information of the relative position encoding of its neighboring nodes. Finally, feature aggregation and updating are performed through residual connections. The update formula is as follows: ； In the formula, Represents a node In the The feature vector of the layer; Represents a node In the The feature vector of the layer; Represents a node The set of neighboring nodes in the feature compatibility graph G; It is the characteristic transformation function; Represents a node In the The feature vector of the layer; Indicates feature concatenation operation; This is a relative position encoding function; Source Point Cloud The point in the middle, For target point cloud The point in the middle.

5. The point cloud registration method based on graph topology layering according to claim 3, characterized in that, In step 13), a two-layer MLP structure is used to calculate the original confidence score, and the Sigmoid activation function is used to constrain the original confidence score within the interval [0,1]. The original confidence score is then expressed as: ； In the formula, Represents a node The original confidence score, This represents the Sigmoid activation function. This indicates the operation of a multilayer sensor. Indicates the process L The final node features obtained after layer feature aggregation.

6. The point cloud registration method based on graph topology layering according to claim 1, characterized in that, Step 2 specifically includes: 21) Establish an initial candidate set based on the original confidence scores, and initialize the selected seed point set to an empty set; define the graph topological distance between any two nodes in the feature compatibility graph as the shortest path hop count in the feature compatibility graph, and set the graph topological distance between non-connected nodes to infinity; 22) Calculate the dynamic comprehensive score for each candidate node based on the dynamic penalty scoring mechanism; 23) An iterative greedy strategy is adopted to iteratively update the dynamic comprehensive score. In each iteration, the node with the highest dynamic comprehensive score is selected and added to the seed point set. The iteration is repeated until the number of seed points reaches a preset threshold.

7. The point cloud registration method based on graph topology layering according to claim 6, characterized in that, In step 22), the formula for calculating the dynamic comprehensive score is: ； In the formula, For the front The dynamic comprehensive score of node k in each iteration. The original confidence score for node k; This is the preset penalty intensity coefficient; For the front The set of seed points already selected in the round; The decay kernel function based on graph distance is expressed as follows: ； in, For the spectral bandwidth parameter, For nodes in the feature compatibility map With the selected seed node The shortest path hop count between them These represent different nodes in the feature compatibility map.

8. The point cloud registration method based on graph topology layering according to claim 1, characterized in that, Step 3 includes: 31) Construct a global dynamic resource pool covering all corresponding relationships and set a global state mask vector; wherein, the initial state of the global state mask vector is 0, and when the corresponding relationship is confirmed as an interior point, the corresponding mask value is set to 1, and in the future, only nodes with a mask value of 0 are searched and verified. 32) Constructing the first cascade layer includes: retrieving unassigned neighbor nodes directly connected to the seed point in the feature compatibility graph; constructing a local induced subgraph through node feature similarity retrieval and calculating the centrality score; and adding nodes with a centrality score higher than a dynamic threshold to the inlier set. ; 33) Constructing the second cascade layer, including: using the set of interior points... The nodes in the graph serve as the expansion source. The neighbor nodes of each node in the feature compatibility graph are searched, and nodes that meet the constraints are selected through geometric verification and added to the interior point set. ; 34) Construct a dynamic confidence propagation model, including: for the aforementioned interior set The nodes in the process are modified based on the modified confidence scores of their parent nodes. The original confidence scores of each node are then modified, and the results from each level are merged to obtain the complete cluster of interior points corresponding to the seed point.

9. A point cloud registration method based on graph topology layering according to claim 8, characterized in that, In step 34), the formula for correcting the original confidence score of each node is as follows: ； In the formula, Represents a node In the Confidence score after hierarchical adjustment. Represents a node The original confidence score, Represents a node parent node In the Confidence score after hierarchical adjustment. As a weighting balancing factor, Represents a node The set of parent nodes.

10. A point cloud registration method based on graph topology layering according to claim 9, characterized in that, Step 4 includes: 41) Based on the multidimensional spatial distribution perception scoring model, each subset of interior points in the complete interior point cluster is scored by jointly evaluating three dimensions: the number of interior points, spatial dispersion, and alignment accuracy, so as to obtain the comprehensive multidimensional spatial quality score of each subset of interior points. 42) For each subset of interior points in the complete interior point cluster, perform weighted singular value decomposition based on the corrected confidence score of each interior point, and solve in parallel the candidate rigid body transformation matrix corresponding to each subset of interior points; and select the rigid body transformation matrix corresponding to the subset of interior points with the highest comprehensive quality score in multidimensional space as the optimal solution.

Images