Multi-view point cloud registration method and device based on adaptive symmetry and hybrid optimization

Abstract

The application provides a multi-view point cloud registration method and device based on adaptive symmetry and hybrid optimization, and relates to the technical field of computer vision. The application firstly acquires multi-view point cloud sequence data of a target scene, performs feature extraction and pair matching by using a YOHO network, calculates an initial relative transformation matrix, and constructs an overlap rate prediction matrix. Then, an adaptive redundant symmetric viewpoint graph is constructed based on a dynamic edge number upper and lower limit constraint and a logical union mechanism, single-direction edge interference is eliminated to form a bidirectional symmetric undirected topological graph. In a two-stage hybrid robust optimization part, a spectral graph synchronization algorithm is used for global solving to obtain a rough global pose, geometric consistency checking is performed to cut off residual outlier mismatched edges, and finally, a covariance decoupling model and a Huber robust kernel function are combined in Lie algebra space to perform nonlinear fine tuning with the rough pose as an initial value, so that sub-millimeter registration accuracy is realized. The application effectively solves the problems of initial value trap and insufficient accuracy of traditional algorithms.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, 3D point cloud processing and spatial registration technology, and more specifically, to a multi-view point cloud registration method and apparatus based on adaptive symmetry and hybrid optimization. Background Technology

[0002] With the rapid development of 3D laser scanning technology, autonomous driving, robot simultaneous localization and mapping (SLAM), and digital twin technology, the demand for large-scale 3D point cloud acquisition and panoramic reconstruction is increasing. In practical applications, due to the limited field of view of sensors, it is usually necessary to collect local point cloud data from multiple different perspectives. These local fragments are then stitched together into a unified global coordinate system using multi-view point cloud registration technology to construct a complete 3D scene model. This technology has broad application prospects in fields such as high-precision map construction for autonomous driving, complex building surveying, robot navigation, and digital twin model construction.

[0003] Traditional multi-view point cloud registration methods typically employ a two-stage pipeline architecture. The front end establishes relative pose relationships between adjacent or non-adjacent views through pairwise registration, while the back end eliminates drift errors accumulated during registration through pose graph optimization, thereby obtaining globally consistent pose estimates. In ideal low-noise, high-overlap scenarios, these methods achieve good registration results. However, in complex large-scene reconstruction tasks, traditional methods face severe challenges. In weak feature regions, highly symmetrical structures, or extreme scenarios with extremely low overlap, front-end feature matching is prone to generating numerous erroneous relative poses. These erroneous matches manifest as outlier mismatches or false positive loop closures in graph optimization. Traditional nonlinear graph optimization methods, when processing data with severe topological noise, are highly susceptible to local minima due to the high non-convexity of the objective function. Without a high-precision initial pose as prior guidance, the optimization process may even collapse, severely limiting the robustness and reliability of registration algorithms in complex scenarios.

[0004] To address the sensitivity of global optimization to initial values, researchers have recently introduced global synchronization methods based on spectral graph theory. These methods, through eigenvalue decomposition of the graph Laplacian matrix, can resist interference and directly output the global pose even with a certain proportion of noisy edges, without relying on precise initial estimates. However, existing spectral graph algorithms still have inherent limitations when applied to practical multi-viewpoint cloud registration. In the viewpoint topology graph construction stage, mainstream methods generally employ a rigid fixed Top-K edge strategy, selecting the top K edges with the highest overlap scores for each viewpoint node as candidate connections. This strategy is acceptable in scenarios with sufficient overlap, but in low-overlap scenarios, the Top-K strategy forcibly introduces a large number of ghost noise edges with virtually no overlap or extremely low overlap, severely compromising the reliability of the topology graph. Furthermore, the presence of unidirectional edges leads to a lack of symmetry in the graph structure, violating the mathematical requirement of spectral decomposition for the symmetry of the Laplacian matrix, and affecting the stability and accuracy of eigenvalue decomposition. Furthermore, the spectral theory is essentially a linear approximation method, lacking explicit constraints on the geometric properties of rigid body motion in three-dimensional space. Its optimization objective deviates from the physical meaning of the SE(3) Lie group space, resulting in a bottleneck in the final registration accuracy, making it difficult to meet the engineering application requirements at the sub-millimeter level.

[0005] In view of the above, this application is hereby submitted. Summary of the Invention

[0006] This invention aims to provide a multi-view point cloud registration method and apparatus based on adaptive symmetry and hybrid optimization, in order to solve the problems of existing technologies in scenarios with weak features, high symmetry and extremely low overlap rate. The front-end feature matching is prone to generating a large number of seriously mismatched outliers, which leads to severe topological noise interference in the back-end pose graph optimization, resulting in accumulated error amplification, graph structure breakage or even complete optimization collapse. This makes it difficult to meet the needs of high-precision spatial mapping and digital twin construction in large scenes.

[0007] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0008] A multi-view point cloud registration method based on adaptive symmetry and hybrid optimization includes: S1, acquire multi-view point cloud sequence data of the target scene; S2, use the pre-trained YOHO 3D local feature extraction network to perform feature extraction and pair matching on the multi-view point cloud sequence data, calculate the initial relative transformation matrix between each point cloud pair, and evaluate the effective inlier rate of the matching to obtain the overlap rate prediction matrix. S3, based on the dynamic edge number upper and lower limit constraints and the preset absolute overlap score threshold, the overlap rate prediction matrix is ​​adaptively filtered to obtain the set of unidirectional matching adjacent nodes. Then, a logical union symmetry mechanism is introduced to force the unidirectional matching graph to be converted into a bidirectional symmetric undirected topological graph. S4. Based on the undirected topological graph and the initial relative transformation matrix, a two-stage hybrid robust optimization is performed to solve the global absolute pose: In the first stage, the undirected topological graph is solved globally using the spectral graph synchronization algorithm. The eigenvector corresponding to the second smallest eigenvalue of the graph Laplacian matrix is ​​calculated using the power iteration method as the rotation component of the global pose, and the translation component is solved by combining translation constraints to output a coarse global absolute pose. In the second stage, geometric consistency checks and graph structure pruning are performed based on the coarse global absolute pose. Using the coarse global absolute pose as a priori initial value, nonlinear factor graph joint fine-tuning is performed in the Lie algebra space to output the optimized high-precision multi-view global absolute pose. S5, based on the high-precision multi-view global absolute pose, the multi-view point cloud sequence is stitched together into a globally consistent complete 3D point cloud.

[0009] The present invention also provides a multi-view point cloud registration device based on adaptive symmetry and hybrid optimization, comprising: The data acquisition unit is used to acquire multi-view point cloud sequence data of the target scene; The feature matching unit is used to extract features and perform pair matching on the multi-view point cloud sequence data using a pre-trained YOHO 3D local feature extraction network, calculate the initial relative transformation matrix between each point cloud pair, and evaluate the effective inlier rate of the match to obtain the overlap rate prediction matrix. An adaptive graph topology construction unit is used to adaptively filter the overlap rate prediction matrix based on dynamic edge number upper and lower limits and a preset absolute overlap score threshold to obtain a set of unidirectional matching adjacent nodes. Then, a logical union symmetry mechanism is introduced to force the unidirectional matching graph to be converted into a bidirectional symmetric undirected topology graph. A two-stage hybrid robust optimization unit is used to perform two-stage hybrid robust optimization based on the undirected topological graph and the initial relative transformation matrix to solve the global absolute pose: In the first stage, the undirected topological graph is solved globally using a spectral graph synchronization algorithm. The eigenvectors corresponding to the second smallest eigenvalues ​​of the graph Laplacian matrix are calculated using the power iteration method as the rotation components of the global pose, and the translation components are solved by combining translation constraints to output a coarse global absolute pose. In the second stage, geometric consistency checks and graph structure pruning are performed based on the coarse global absolute pose. Using the coarse global absolute pose as a priori initial value, nonlinear factor graph joint fine-tuning is performed in the Lie algebra space to output the optimized high-precision multi-view global absolute pose. The point cloud stitching unit is used to stitch together the multi-view point cloud sequence into a globally consistent complete 3D point cloud based on the high-precision multi-view global absolute pose.

[0010] The present invention also provides a multi-view point cloud registration device based on adaptive symmetry and hybrid optimization, including a processor and a memory. The memory stores a computer program that can be executed by the processor to implement the multi-view point cloud registration method based on adaptive symmetry and hybrid optimization as described above.

[0011] The present invention also provides a computer-readable storage medium storing computer-readable instructions, which, when executed by a processor of the device on which the computer-readable storage medium resides, implement the multi-view point cloud registration method based on adaptive symmetry and hybrid optimization as described above.

[0012] In summary, compared with the prior art, the present invention has the following beneficial effects: This invention effectively solves problems such as extremely low overlap rate noise interference, traditional graph optimization easily falling into local optimum initial value traps, and insufficient global point cloud stitching accuracy when performing multi-view point cloud registration in complex large scenes, through the synergistic effect of adaptive redundant symmetric graph construction, synchronous global initial value solution of spectral graph, and hybrid robust fine-tuning in Lie algebra space.

[0013] This invention achieves noise suppression and accuracy improvement throughout the entire process of topology graph construction, global pose coarse localization, and nonlinear fine optimization. Compared with traditional pose graph optimization and pure spectrum graph synchronization algorithms, it has better registration accuracy, operational stability, and anti-interference ability under complex working conditions with low overlap and strong feature noise. It can be widely used in fields such as large-scene surveying and mapping, high-precision maps for autonomous driving, robot SLAM navigation, and digital twin modeling, and has extremely high engineering practical value.

[0014] This invention utilizes adaptive upper and lower bound constraints and a logical union mechanism to filter and force symmetry of viewpoint topology, extracting high-purity bidirectional dense closed-loop topology information with global connectivity from the scene, thereby cutting off false positive matching at the source and eliminating unidirectional edge interference.

[0015] This invention introduces a spectral graph initial value guidance mechanism into a two-stage optimization architecture. Through Laplacian eigenvalue decomposition and iterative reweighting, it achieves global suppression of topological out-of-topology noise and rapid solution of collision-free coarse pose, thereby significantly improving the noise resistance and convergence stability of global optimization and avoiding the dependence of traditional factor graph optimization on initial values.

[0016] In the high-precision fine-tuning stage of the factor graph, this invention combines coarse initial values ​​with rigorous geometric consistency checks and physical structure pruning to effectively remove residual mismatched outliers. It also utilizes the covariance decoupling mechanism and Huber robust error kernel function to complete nonlinear optimization in the Lie algebra space, generating high-quality global absolute pose with sub-millimeter precision and seamlessly stitched 3D large-scene point clouds. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of a multi-view point cloud registration method based on adaptive symmetry and hybrid optimization provided in Example 1.

[0019] Figure 2 The flowchart shows a multi-view point cloud registration method based on adaptive symmetry and hybrid optimization provided in Example 1.

[0020] Figure 3 This is a schematic diagram of the training and application architecture of the YOHO 3D local feature extraction network model provided in Example 1.

[0021] Figure 4 This is a schematic diagram of a multi-view point cloud registration device based on adaptive symmetry and hybrid optimization, provided in Embodiment 2.

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 described embodiments are only a part of the embodiments of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0024] Example 1 Embodiment 1 of the present invention provides a multi-view point cloud registration method based on adaptive symmetry and hybrid optimization, which can be implemented by a multi-view point cloud registration device based on adaptive symmetry and hybrid optimization (hereinafter referred to as multi-view point cloud registration device), specifically, executed by one or more processors within the multi-view point cloud registration device.

[0025] In this embodiment, the multi-view point cloud registration device can be an electronic device equipped with a processor. The processor carries a computer program for the multi-view point cloud registration method based on adaptive symmetry and hybrid optimization, and the computer program can be executed. Such devices include computers, smartphones, smart tablets, workstations, etc., without limitation.

[0026] This embodiment specifically addresses the common challenges in complex large-scale 3D reconstruction, such as extremely low overlap rate, high noise interference, and the tendency of traditional optimization algorithms to fall into initial value traps. By organically combining adaptive redundant symmetric mapping, a global solution mechanism for initial values ​​of the spectral graph, and a robust fine-tuning strategy using Lie algebra space, it achieves sub-millimeter-level high-precision global registration under extreme conditions.

[0027] like Figures 1-2 As shown, a multi-view point cloud registration method based on adaptive symmetry and hybrid optimization includes steps S1 to S5.

[0028] S1, acquire multi-view point cloud sequence data of the target scene.

[0029] During the data acquisition phase, this embodiment uses a terrestrial 3D laser scanner or a mobile mapping system equipped with a portable lidar to perform multi-view scanning data acquisition of the target scene.

[0030] The design of the scanning path fully considers the spatial distribution of the main three-dimensional geometric structure and local surface feature areas of the target scene. By planning different viewpoint positions, scanning angles and spatial overlap between adjacent scanning stations, it ensures that the acquired raw point cloud data has spatial continuity and can completely cover all parts of the complex large scene.

[0031] In practical applications, terrestrial 3D laser scanners are typically fixed on tripods for static data acquisition. The spatial overlap between adjacent scanning stations is generally set between 30% and 50% to ensure the reliability of subsequent feature matching and registration. Mobile mapping systems, on the other hand, use a mobile platform equipped with a lidar to move along a preset trajectory to achieve full coverage data acquisition of narrow spaces such as corridors and tunnels, as well as open areas such as large squares and factory workshops. The acquisition frequency of each frame of point cloud data is dynamically adjusted according to the platform's movement speed to ensure that there is sufficient overlap between adjacent frames for feature matching.

[0032] The raw point cloud data obtained can be processed by voxel downsampling to unify the point cloud data in space into a voxel grid of the same resolution, eliminating outliers introduced by sensor measurement noise and reducing the data size of subsequent calculations.

[0033] S2, using a pre-trained YOHO 3D local feature extraction network to perform feature extraction and pair matching on the multi-view point cloud sequence data, calculate the initial relative transformation matrix between each point cloud pair, and evaluate the effective inlier rate of the match to obtain the overlap rate prediction matrix.

[0034] The processed multi-view point cloud sequence data is input into the YOHO 3D local feature extraction network for feature extraction and pairwise matching.

[0035] Specifically, the point cloud of any two frames is first extracted using a pre-trained YOHO 3D local feature extraction network. and 3D local geometric feature descriptor; Then, based on the three-dimensional local geometric feature descriptor, a point-to-point correspondence set is established in the feature space through nearest neighbor search; Next, the random sampling consensus algorithm is used to filter the corresponding set of point pairs, and the initial relative transformation matrix that maximizes the number of points in the geometry is obtained; wherein, the initial relative transformation matrix is ​​composed of an initial relative rotation matrix and an initial relative translation vector; Finally, based on the initial relative transformation matrix, the number of effective in-point pairs that satisfy the preset in-point distance tolerance threshold is counted and the overlap rate score is calculated to construct the overlap rate prediction matrix.

[0036] The formula for calculating the elements in the overlap rate prediction matrix is ​​as follows: ; in, Point cloud arrive The predicted overlap score; For point-to-point corresponding set The total number of point pairs in the data; This is an indicator function that returns 1 if the condition within the parentheses is met, and 0 otherwise. and The corresponding points in the point cloud and The three-dimensional coordinate vector in the coordinate system, i.e., the first Point-to-point; This is the initial relative rotation matrix; This is the initial relative translation vector; The set tolerance threshold for interior point distance; Indicates the L2 norm distance; , They represent the first , Frame point cloud.

[0037] In this embodiment, the YOHO 3D local feature extraction network is a deep learning model designed by integrating a rotational group convolution mechanism on the basis of a 3D sparse feature extraction framework (such as lightweight PointNet / FCGF). Its core innovation lies in achieving end-to-end extraction of rotation-invariant features of 3D point clouds through the Group Convolution operation. During the offline training phase of the model, multi-view scanned 3D large-scene point cloud data is selected as input samples. Each set of point cloud data contains spatial visual information such as 3D spatial geometry, local surface undulation features, and point distribution density from different viewpoints.

[0038] like Figure 3 As shown, the training samples are first standardized and positive and negative matching point pairs are constructed based on the real relative pose labels. The positive sample pair is a pair of point clouds that have a real geometric correspondence in space, while the negative sample pair is a pair of point clouds from different regions that do not overlap. This sample design strategy helps the network learn the feature response to key geometrically overlapping regions.

[0039] The preprocessed training data matrix Input a pre-built YOHO 3D local feature extraction network model, where, Indicates the first Local point cloud patches of a training sample It represents the geometric information of point cloud blocks (such as coordinates, normal vectors, etc.). The label represents the rotation matrix of the point cloud block. This indicates the category label of the point cloud patch (used for supervised learning). This represents the total number of training samples. The network outputs the local features of all samples in the training set, forming a feature matrix. ,in, Indicates the first The training sample of the th training sample The feature descriptor matrix is ​​used for subsequent loss calculation, model parameter update, and training of the YOHO network.

[0040] The distance cost between positive and negative sample point pairs in the feature space is adaptively optimized through a metric learning mechanism. Specifically, contrastive loss or triplet loss functions can be used as training objectives to make the feature distance of matched point pairs approach zero and maximize the feature distance of non-matched point pairs. During training, the backpropagation algorithm is used to update the network weights, and after multiple iterations and convergence, a local feature extractor with strong generalization ability is obtained.

[0041] In the online application phase of the model, the trained and solidified YOHO network extracts 3D local geometric feature descriptors from any two frames of point clouds in the input multi-view point cloud sequence. For example... Figure 3 The test data matrix is ​​shown on the right. After inputting into the same YOHO network, the output is an inference feature descriptor matrix. Each feature descriptor is a high-dimensional rotation-invariant vector that encodes the local geometric structure information of the input point cloud. Indicates the first A test point cloud block, It represents the geometric information of point cloud blocks (such as coordinates, normal vectors, etc.). The rotation matrix labels of the test samples The category label representing the test sample (used for supervised learning). This represents the total number of test samples. Indicates the first The first test sample Dimensional feature descriptor.

[0042] For including The YOHO network processes the sequence of frame point clouds sequentially, combining all frames to generate... A feature descriptor matrix of dimension 1 is generated. Subsequently, a point pair correspondence set is established in the feature space using a nearest neighbor search algorithm. For feature matching between point cloud frames i and j, the feature descriptor output by the YOHO network is used for nearest neighbor matching. The matching threshold is set according to the Euclidean distance of the feature vectors, and matching pairs below the threshold constitute the initial candidate correspondence set.

[0043] After obtaining the initial correspondence set, a random sampling consensus algorithm is used to filter the correspondence set containing erroneous matches. Through an iterative process of random sampling and consensus verification, the initial relative transformation matrix that maximizes the number of geometric interior points is obtained. .

[0044] The core idea of ​​the random sampling consensus algorithm is to randomly select the smallest sample set from the initial corresponding set to solve the model parameters, then count the number of inliers that satisfy the current model, and repeat the iteration until a preset number of iterations is reached or a sufficient number of inliers are found.

[0045] In each iteration, three pairs of non-collinear corresponding points are randomly selected from the corresponding set. The rigid body transformation parameters are solved using singular value decomposition. Then, the reprojection error of all corresponding points under the current transformation is calculated. Corresponding points with errors less than a set threshold are marked as inliers. The number of inliers is counted, and the current optimal transformation matrix is ​​recorded. After sufficient iterations, the initial relative transformation matrix that maximizes the number of inliers is output. This matrix is ​​the initial relative pose estimate between the two frames of point clouds.

[0046] After obtaining the initial relative transformation matrix, the overlap rate score is calculated by counting the number of effective inlier pairs that meet the distance threshold, thereby constructing a global overlap rate prediction matrix.

[0047] The physical meaning of the global overlap rate prediction matrix formula is to count the number of interior point pairs whose reprojection error is less than a threshold under the initial relative transformation, divide by the total number of corresponding point pairs to obtain the overlap rate score, and construct a matrix of size after traversing all point cloud pairs. × The overlap rate prediction matrix is ​​used. Higher element values ​​in the overlap rate prediction matrix indicate a greater degree of spatial overlap between the corresponding two frame point clouds and better feature matching quality, providing a reliable observation data foundation for subsequent graph topology construction and optimization.

[0048] S3. Based on the dynamic edge number upper and lower limit constraints and the preset absolute overlap score threshold, the overlap rate prediction matrix is ​​adaptively filtered to obtain the set of unidirectional matching adjacent nodes. Then, a logical union symmetry mechanism is introduced to force the unidirectional matching graph to be converted into a bidirectional symmetric undirected topological graph.

[0049] Step S3 constructs a viewpoint topology graph based on the overlap rate prediction matrix. The design goal of this module is to solve the problem that the fixed Top-K edge strategy in traditional spectrograph algorithms introduces a large number of ghost noise edges in low overlap rate scenarios, while eliminating the graph topology asymmetry caused by unidirectional edges.

[0050] Specifically, firstly, a preset absolute overlap score threshold is used to perform global binarization filtering on the overlap rate prediction matrix to extract an initial candidate neighbor node set. The extraction rule is expressed as follows: ; in, For the first The frame point cloud corresponds to a preset absolute overlap threshold. The initial set of candidate adjacent node indices; For point clouds arrive Predicted overlap score; absolute overlap threshold Used for adaptively cutting off incorrectly matched edges with no overlap or very low overlap.

[0051] In practical applications, The threshold is typically set to 0.05, meaning that when the overlap score between two point clouds is less than 5%, it is determined that there is not enough geometrically overlapping area between them for feature matching, and the candidate edge is directly discarded to avoid introducing incorrect geometric constraints. This threshold setting takes into account the minimum overlap requirements of adjacent viewpoints in real scanning scenes, while leaving appropriate tolerance space to adapt to the differences in geometric complexity of different scenes.

[0052] To address the differences in local overlap density across different point cloud scenarios, the initial candidate adjacent node set is adaptively truncated using upper and lower limits of edge degree, resulting in a dynamically constrained set of unidirectionally matched adjacent nodes, expressed as: ; in, For the first The set of adjacent nodes corresponding to the frame point cloud after dynamic constraint and one-way matching; Represents the initial candidate set of adjacent nodes The total number of candidate nodes; and These are the preset minimum guaranteed number of connected edges per node and the maximum guaranteed number of connected edges per node; Indicates the first The overlap ratio score vector corresponding to the frame Among them, the top overlapping scores are selected. Each of the adjacent nodes.

[0053] In typical parameter settings, Setting it to 8 ensures that each node maintains a connection with at least 8 adjacent nodes, satisfies the global graph optimization requirement for dense cycles, and prevents the graph topology from being too sparse, which could lead to unstable solutions. Set it to 20 to prevent the computational burden and noise amplification effect caused by excessive redundant connections.

[0054] The dynamic truncation mechanism automatically adjusts the final number of connected edges based on the initial number of candidate adjacent nodes for each node. When the initial number of candidate edges for a node is less than [a certain value], [the mechanism will adjust the number of connected edges]. To prevent over-filtering, all initial candidate edges are retained. When the number of initial candidate edges for a node exceeds a certain threshold... Only retain the highest overlapping score. Edges are used to prevent excessive redundancy. This adaptive degree constraint strategy enables this embodiment to adapt to point cloud scenarios with different density distributions, ensuring the global connectivity of the topology graph while avoiding excessive accumulation of noisy edges.

[0055] After completing the dynamic constraints on node degree, the unidirectional matching graph is forcibly converted into a bidirectional symmetric undirected topology graph through a logical union mechanism.

[0056] Specifically, the set of adjacent nodes matched unidirectionally after dynamic constraints. Generate a directed initial mask matrix, whose elements are: ,when hour, ;when hour, ; Subsequently, the asymmetry of the topological graph is eliminated through a logical union operation, generating an undirected symmetric mask matrix. The formula is as follows: ; in, This is the undirected edge state mask after symmetry transformation. 1 indicates that there is an edge, and 0 indicates that there is no edge; Represents a logical OR operation.

[0057] This strategy guarantees that as long as there is any unidirectional candidate edge between point cloud frames i and j or j and i, a physical bidirectional connection is forcibly established, satisfying the strict mathematical requirement of the Laplacian matrix symmetry for subsequent spectral graph eigenvalue decomposition. In a practical physical sense, the logical union symmetry mechanism ensures that as long as there is sufficient unidirectional overlap between two perspectives, a geometric relationship is determined between the two perspectives, and a bidirectional constraint is established. This design effectively solves the problem of graph topological breakage caused by unidirectional matching failure in traditional methods.

[0058] After determining the global undirected topology, the overlap rate score and the geometric matching inlier rate are fused to calculate the weighted adjacency matrix of the undirected topology graph, which is used for weight allocation in subsequent robust optimization. Simultaneously, diagonal elements are set to zero to eliminate self-loop interference. The expression is: ; in, These are the edge weight coefficients of the weighted adjacency matrix; For point clouds , The geometric in-point rate observed during the feature matching phase; Point cloud arrive The predicted overlap score.

[0059] Edge weight coefficient It also reflects the degree of overlap and matching quality between the two perspectives: the higher the predicted overlap rate score, the more geometric structures that can be associated between the two perspectives, and the higher the geometric in-point rate, the more reliable the matching result. The constraint forces the diagonal elements to be set to zero to cut off the interference of self-closing loops on eigenvalue decomposition, so that the bidirectional dense closed loop with high overlap rate and high matching confidence plays a dominant guiding role in subsequent optimization.

[0060] Geometric interior point ratio For point clouds and The reliability quantification index for feature matching is calculated as the ratio of the number of matched intra-points to the total number of matched point pairs. It is used in conjunction with the overlap rate score for weighted calculation, jointly characterizing the matching quality between pairwise point clouds, so that the topological edge weights simultaneously take into account both the degree of overlap coverage and the geometric matching accuracy.

[0061] This results in an undirected topological graph with strictly symmetric edge weights and no topological direction.

[0062] S4. Based on the undirected topological graph and the initial relative transformation matrix, a two-stage hybrid robust optimization is performed to solve the global absolute pose: In the first stage, the undirected topological graph is solved globally using the spectral graph synchronization algorithm. The eigenvector corresponding to the second smallest eigenvalue of the graph Laplacian matrix is ​​calculated using the power iteration method as the rotation component of the global pose, and the translation component is solved by combining translation constraints to output a coarse global absolute pose. In the second stage, geometric consistency checks and graph structure pruning are performed based on the coarse global absolute pose. Using the coarse global absolute pose as the prior initial value, nonlinear factor graph joint fine-tuning is performed in the Lie algebra space to output the optimized high-precision multi-view global absolute pose.

[0063] This step performs global absolute pose calculation based on the constructed undirected topological graph and the initial relative transformation matrix. In the first stage, the spectral graph synchronization algorithm is used to globally solve the constructed topological graph, which can quickly obtain a rough global absolute pose without the need for preset initial values ​​for iteration.

[0064] In this embodiment, the core idea of ​​the spectral graph synchronization algorithm is to transform the global pose estimation problem into an eigenvalue decomposition problem of the graph Laplacian matrix. First, the weighted adjacency matrix is ​​normalized to a transition probability matrix. Then, the eigenvectors corresponding to the second smallest eigenvalues ​​of the graph Laplacian matrix are calculated using a power iteration method as the rotation components of the global pose. During iterative reweighting, the residuals of each edge are calculated based on the current pose estimation, and the edge weights in the adjacency matrix are adaptively adjusted to reduce the contribution of outliers, thereby suppressing the influence of topological noise. The spectral graph synchronization algorithm gradually improves the global consistency of pose estimation through 50 iterations, ultimately outputting a coarse global absolute pose without local minima traps.

[0065] Specifically, the weighted adjacency matrix of the undirected topological graph is first row-normalized to obtain the graph model transition probability matrix, which serves as the input for spectral graph synchronization. The expression is as follows: ; ; in, The row-normalized transition probability matrix serves as the input for the power-law iteration to solve for the eigenvectors. It is a weighted adjacency matrix; for The node degree matrix is ​​a diagonal matrix, with only diagonal elements being non-zero. express The inverse matrix; For the first The degree of a node, i.e., the degree of a node. The sum of the weights of all adjacent edges; The bidirectional edge weight coefficients of the graph topology are also... Element; This represents the total number of nodes in the topology graph. Then based on the node degree matrix with weighted adjacency matrix Construct the standard graph Laplacian matrix to transform the pose solution into an eigenvalue decomposition problem, expressed as: ; in, The graph is a Laplace matrix; Next, based on the transition probability matrix The power iteration method is used to process the Laplace matrix. Perform eigenvalue decomposition and extract eigenvalues. The eigenvector corresponding to the second smallest eigenvalue is used as the global coarse estimate of multi-view rotation, and the coarse rotation estimate of each view is output. Based on coarse rotation estimates from various perspectives, combined with coarse translation estimates derived from relative transformations, a coarse pose is obtained for the current estimation. Specifically, a set of linear constraint equations is established using pairwise relative translation observations obtained through matching. Under the condition of fixed coarse rotation, a global coarse translation estimate is obtained through least squares. The coarse rotation estimates and coarse translation estimates are then combined to obtain the current coarse pose estimate. This avoids translation drift caused by the lack of a global rotation reference in traditional standalone translation solutions, ensuring global consistency and eliminating local topological misalignment of the entire coarse pose set, serving as a reliable initial value for subsequent fine optimization.

[0066] The spatial residual of each topological edge is calculated using the currently estimated coarse pose (i.e., the predicted relative transformation corresponding to each topological edge is calculated, compared with the original observed relative transformation, and the spatial residual of the edge is calculated to quantify the magnitude of the matching error of a single edge).

[0067] The edge weights of the weighted adjacency matrix are adaptively adjusted based on the spatial residuals to reduce outlier edge weights and suppress topological noise. A larger residual (outlier mismatched edges) results in a lower weight. The lower the weight, the less interference from noisy edges on the global solution, thus completing the weighted adjacency matrix. Iterative reweighted update.

[0068] The edge weights of the weighted adjacency matrix and the estimated coarse pose are updated iteratively to obtain the coarse global absolute pose for each viewpoint.

[0069] In this embodiment, power iteration is a numerical iterative algorithm used to solve for the maximum eigenvalue of the modulus of a square matrix and its corresponding eigenvector; in this application, transformation adaptation is used to find the Laplace transform (...). The eigenvectors corresponding to the second smallest eigenvalues ​​(coarse rotation) are used to achieve global coarse rotation solutions without artificial initial values.

[0070] Although the rough global absolute pose has limited accuracy, it has good global consistency and no obvious topological misalignment, providing a reliable initial value for subsequent fine optimization.

[0071] The second stage performs geometric consistency checks based on a coarse global absolute pose and physically cuts off residual outlier mismatched edges that deviate significantly from the prediction to achieve deep pruning of the graph structure.

[0072] Specifically, firstly, based on the rough global absolute pose, the predicted relative rotation matrix and the predicted relative translation matrix between viewpoints are derived.

[0073] Then, rotation consistency assessment is performed: by calculating the difference between the predicted relative rotation matrix and the initial relative rotation matrix, and utilizing the properties of the rotation matrix trace, the matrix difference is converted into radian rotation angle residuals to characterize the rotation consistency deviation. The calculation formula is as follows: ； in, Point cloud and Rotational angle residuals between; The predicted relative rotation matrix is ​​derived based on the rough global absolute pose. This is the initial relative rotation matrix; This represents the trace operation of a matrix; The inverse cosine function is used to convert matrix differences into radian angle differences. This is the transpose symbol.

[0074] This formula utilizes the relationship between the product of the trace of a rotation matrix and its transpose to convert the matrix difference into an angular difference in radians, which physically represents the geodesic distance between two rotation matrices.

[0075] Next, a translation consistency assessment is performed: the translation consistency deviation is characterized by calculating the three-dimensional Euclidean distance between the predicted translation vector and the observed translation vector. The calculation formula is as follows: ; in, Point cloud and The translation distance residual between them; The predicted relative translation vector is derived based on the rough global absolute pose. This is the initial relative translation vector; Table L2 norm distance.

[0076] This formula measures the absolute deviation between translation vectors, reflecting the overall consistency of pose estimation in three-dimensional space.

[0077] Then perform physical cutoff of outliers: set an angle tolerance threshold. Distance tolerance threshold The rotation angle residual and the translation distance residual are jointly and double-checked to determine whether an edge is a pure interior edge, thus obtaining the set of pure interior edges, expressed as: ; in, Mark the pure interior point edge; when If the edge is determined to be a pure interior vertex edge, it is retained and the process proceeds to nonlinear factor graph optimization; otherwise, when Physically cut off the edge, without assigning it any robust core weights or optimizing resources.

[0078] In typical parameter settings, It can be set to 0.35 radians (approximately 20 degrees) to tolerate larger rotational errors introduced by the rough initial value. Set to 0.5 meters to tolerate scale errors in translation estimation.

[0079] This geometric consistency-based dual verification mechanism effectively removes extreme erroneous matching edges where predictions and observations are severely mismatched, preventing them from interfering with subsequent fine-tuning.

[0080] Subsequently, the coarse global absolute pose is used as a priori initial value input to the nonlinear factor graph optimizer. Combined with the covariance noise model of decoupled rotation and translation and the Huber robust error kernel function, the multi-view absolute pose is jointly nonlinearly fine-tuned in the Lie algebra space.

[0081] The pure interior edges retained after geometric consistency checks break the isotropic assumption of conventional pose graph optimization.

[0082] Specifically, based on the set of pure interior points and edges, the standard deviation of rotation measurement noise and the standard deviation of translation measurement noise are independently calculated for each edge, and a decoupled covariance model is constructed. The calculation formula is as follows: ; ; in, and These represent the standard deviations of rotational noise and translational noise for the corresponding connected edges, respectively. These are the edge weight coefficients of the weighted adjacency matrix; and These are the preset noise scale factors for rotating and translational foundations, which are 0.02 and 0.05 respectively in typical settings; To prevent tiny positive numbers with a denominator of zero.

[0083] The core innovation of this covariance model lies in modeling the measurement noise of rotation and translation separately, so that edges with high overlap and high interior point ratio have greater weight in rotation optimization, while the contribution of edges with different quality in translation optimization is also adjusted accordingly, avoiding the problem of different measurements being treated uniformly under the traditional isotropic assumption.

[0084] Then, the rough global absolute pose is used as the initial value for nonlinear optimization. A factor graph optimizer with a decoupled covariance model as the measurement noise model is constructed. The nonlinear Levenberg-Marquardt algorithm is used to minimize the global objective cost function constrained by the Huber robust error kernel function. The iteration is continued until convergence, and the high-precision multi-view global absolute pose of the point cloud at each viewpoint is output, which is the joint fine-tuning part of the Lie algebra space.

[0085] The expression for the global objective cost function is as follows: ; in, Represents the global objective cost function for factor graph optimization; For pure interior point edge marking, when Only when this edge is selected should it participate in the optimization accumulation; Represents the Huber robust error kernel function; For the reason and Constructed diagonal covariance information matrix; To map to Lie algebra Pose residual vector in space; This is the transpose symbol.

[0086] In this embodiment, the Huber robust error kernel function utilizes its characteristic of dynamically adjusting the penalty gradient to further suppress and weaken the interference of residual small mismatched edges during the fine-tuning process: when the residual is small, a quadratic penalty is used to accelerate convergence, and when the residual is large, a linear penalty is used to limit the influence of outliers.

[0087] The Levenberg-Marquardt algorithm adaptively adjusts the damping parameters, switching between gradient descent and the Gauss-Newton method to achieve fast convergence while maintaining convergence stability. The upper limit for optimization iterations can be set to 100 times or until the relative change in the objective function is less than a preset threshold, ensuring automatic termination of iteration after achieving sub-millimeter registration accuracy.

[0088] Diagonal covariance information matrix The measurement confidence level is used to characterize each topological connection edge. This invention employs a rotation and translation decoupling modeling approach, adaptively allocating noise weights based on the quality of the matched edges, allowing high-quality matched edges to dominate the optimization process, while reducing the constraint weight of low-quality edges.

[0089] Pose residual vector To map the deviation between the observed relative pose and the optimized predicted pose to the six-dimensional residual vector composed of the three-dimensional rotation residual and the three-dimensional translation residual obtained by the Lie algebra space of se(3), continuous, smooth, nonlinear iterative fine-tuning of the pose can be achieved by constructing residual constraints in the Lie algebra space, thus avoiding the optimization discontinuity problem caused by the orthogonal constraints of the rotation matrix.

[0090] se(3) is the Lie algebra corresponding to the special Euclidean group; se(3) is used to characterize the rigid body pose from any world coordinate system to the local coordinate system during a three-dimensional rigid body transformation (rotation + translation). All belong to the se(3) group. se(3) maps discrete matrix-form rigid body transformations to a 6-dimensional continuous vector space, which contains a 3-dimensional rotation Lie algebra se(3) (corresponding to rotation) + a 3-dimensional translation component (corresponding to translation).

[0091] In this embodiment, the pose residual is constructed in the se(3) Lie algebra space, which eliminates the optimization difficulties caused by the orthogonality constraint of the rotation matrix. It also achieves high-precision nonlinear fine-tuning by combining the decoupled covariance and Huber robust function.

[0092] S5, based on the high-precision multi-view global absolute pose, the multi-view point cloud sequence is stitched together into a globally consistent complete 3D point cloud.

[0093] After the nonlinear fine-tuning converges, the high-precision global absolute pose of the point clouds from each viewpoint is extracted, and the independent local view point clouds are transformed and stitched together into a unified global coordinate system.

[0094] Specifically, the expression for stitching together a multi-view point cloud sequence into a globally consistent complete 3D point cloud based on the high-precision multi-view global absolute pose is as follows: ; in, A three-dimensional coordinate vector representing a globally consistent complete three-dimensional point cloud; and The first The global rotation matrix and global translation vector for high-precision multi-view global absolute pose; Indicates the first The original 3D coordinate vector in the local camera coordinate system corresponding to the high-precision multi-view global absolute pose.

[0095] The final output after convergence The system generates high-precision global rotation matrices and global translation vectors. The transformed and fused multi-view 3D point clouds are output in standard formats (such as PLY, PCD, or LAS) and visualized. The visualization supports point cloud shading, voxel rendering, and multi-resolution layered display, facilitating intuitive evaluation and subsequent processing of the registration results.

[0096] In practical applications of high-precision map construction for autonomous driving, the method of this invention demonstrates significant technical advantages. When autonomous vehicles perform LiDAR scanning in urban scenes, the overlap rate of point clouds between adjacent scanned positions may drop below 20% due to factors such as building occlusion and narrow streets. Traditional methods are prone to generating a large number of erroneous matches under such extreme conditions, leading to pose estimation failure. This embodiment automatically filters candidate edges with low overlap rates through an adaptive redundant symmetric graph construction mechanism and ensures the global connectivity of the topology graph using logical union operations, enabling the construction of a reliable viewpoint association graph even under extremely low overlap rate conditions. The spectral graph synchronization algorithm quickly solves for a globally consistent coarse pose without the need for prior initial values, effectively avoiding the dependence on initial values ​​in traditional factor graph optimization. Geometric consistency checks and dual threshold verification mechanisms further eliminate residual erroneous matches, providing a clean set of interior points and edges for subsequent fine optimization. Finally, by decoupling the covariance model and Huber robust kernel function for nonlinear fine-tuning in the Lie algebra space, a high-precision global absolute pose with sub-millimeter accuracy is output, providing a reliable positioning benchmark for high-precision map construction for autonomous vehicles.

[0097] In the fields of large-scale mapping and digital twin construction, the embodiments of this invention also demonstrate excellent performance. When using mobile mapping systems to perform 3D modeling of large industrial parks, historical building complexes, or urban underground pipe networks, the scanning trajectory may span several kilometers and involve hundreds of scanning stations. Traditional methods are prone to error accumulation when processing such large-scale data, leading to significant misalignment in the final model. This embodiment employs a two-stage hybrid robust optimization architecture. In the spectral map stage, it achieves rapid coarse solution of the global pose, and in the factor map stage, it performs fine optimization using the coarse pose as the initial value, effectively suppressing the propagation of long-term accumulated errors. The Huber robust kernel function adaptively weights and attenuates residual small error points during fine-tuning, further improving the stability and accuracy of the registration results.

[0098] To verify the effectiveness of the method proposed in this invention, in a preferred embodiment, the publicly available 3DMatch three-dimensional point cloud dataset is selected as the research object, covering complex and extreme working conditions such as extremely low overlap rate, repetitive symmetrical geometry, lack of significant surface features, and sensor scanning noise.

[0099] The experimental parameters are set as follows: interior point distance tolerance threshold absolute overlap threshold Minimum number of connected edges Maximum number of connected edges The number of synchronous iterations of the spectrum is 50, and the rotation angle residual threshold is... Translation distance residual threshold The noise scale factor of the rotating foundation is Translational foundation noise scale factor The Huber kernel threshold parameter was set to 1.0, and the upper limit for fine-tuning the Levenberg-Marquardt algorithm was 100 iterations. The experimental comparison results are shown in Table 1.

[0100] Table 1. Performance comparison results of the method of this invention with other algorithms

[0101] Among them, 3D-RR(%) (3D Registration Recall, the overall 3D point cloud registration recall rate) is used to determine whether the poses of all frames in the entire scene have been successfully registered. The higher the value, the higher the overall registration success rate and the stronger the algorithm robustness.

[0102] 3DL-RR(%) (3D Local Registration Recall) focuses on the accuracy of local matching between point cloud pairs (more sensitive to translation / local consistency). A higher value indicates higher accuracy in local matching, with no misalignment or splicing defects.

[0103] EIGSE3 (Eigen-based SE(3) Synchronization, a pure spectral synchronization algorithm based on eigenvalue decomposition) is a classic early pose synchronization scheme. It solves the pose using only spectral eigenvalue decomposition and is easily affected by noise / low overlap. It is the baseline algorithm for comparison.

[0104] IRLS (Iteratively Reweighted Least Squares) is often used for pose graph optimization. It suppresses outliers by iteratively weighting, but it is highly dependent on the initial values ​​and is prone to getting trapped in local optima in low-overlap scenarios.

[0105] RotAvg (Rotation Averaging) only performs synchronous optimization on the rotation component before solving for the translation. It only focuses on rotation consistency and does not perform robust optimization on translation and the overall topology, making it prone to failure in complex scenarios.

[0106] LITS (Iterative Local-to-Global Transformation Synchronization) is an improved pose synchronization algorithm for noisy scenes. Compared with traditional algorithms, it has a certain noise resistance, but it does not adopt the "adaptive symmetric graphing + two-stage optimization" architecture.

[0107] HARA (A Hierarchical Approach for Robust Rotation Averaging) is designed for pose synchronization problems. It filters out outliers hierarchically through triplet support and then performs nonlinear optimization. It can effectively handle high-noise, low-overlap rotation synchronization problems and is one of the classic baseline algorithms in this field. However, it does not combine spectral graph initial value solution and decoupled covariance model.

[0108] SGHR (Spectral Graph with Hierarchical Robustness) is an improved baseline scheme in recent years. It is more robust than pure spectrograph algorithms, but it lacks geometric consistency pruning and Lie algebra fine-tuning steps.

[0109] The method of this invention combines adaptive symmetric graph construction, synchronous coarse solution of spectral graph, geometric pruning, and Lie algebra robust fine-tuning. Through two-stage hybrid optimization and multi-level noise suppression, it achieves optimal performance in low-overlap scenarios.

[0110] Quantitative experimental results show that the method of this invention significantly outperforms traditional pure spectrogram algorithms (such as EIGSE3) and conventional graph optimization models (such as IRLS and RotAvg) in both overall 3D registration recall (3D-RR) and local registration recall (3DL-RR), and also outperforms recently improved robust synchronization algorithms (such as SGHR and HARA). In complex real-world scenarios with low overlap and strong feature noise, the method achieves a substantial leap in 3D registration accuracy and system robustness, validating the effectiveness of the collaborative architecture of adaptive symmetric graph construction, spectrogram initial value guidance, and Lie algebra robust fine-tuning.

[0111] In summary, compared with the prior art, the present invention has the following beneficial effects: This invention first introduces an adaptive redundant symmetric graph construction strategy in the topology construction stage. By utilizing dynamic upper and lower limits on the number of edges and a logical union mechanism, false positive matching is eliminated at the source, unidirectional edge interference is forcibly eliminated, and a high-purity bidirectional dense closed loop is constructed. Subsequently, this invention innovatively proposes a two-stage hybrid optimization architecture of "robust factor graph fine-tuning guided by initial spectrogram values": the first stage cleverly utilizes the advantages of spectrogram theory, which "does not require initial values ​​and is resistant to global noise," to quickly solve for a collision-free initial global absolute pose; the second stage performs strict geometric consistency physical pruning on interior points and edges, and inputs the initial spectrogram value into the factor graph optimizer. Through an innovative "rotation and translation covariance decoupling" model and Huber robust error kernel function, extreme nonlinear fine-tuning is performed in the Lie algebra space.

[0112] This invention, through a deep integration of adaptive graph topology and hybrid optimization, completely eliminates accumulated errors in large-scene stitching, successfully overcoming the "initial value trap" that traditional nonlinear graph optimization is prone to collapse. It also compensates for the insufficient accuracy of simple spectral graph algorithms, enabling the model to achieve sub-millimeter-level high-precision registration even under extremely low overlap and high outlier noise conditions. Compared to existing technologies, this invention not only outputs globally absolute poses with extremely strong noise resistance but also generates seamlessly stitched, high-quality 3D large-scene point cloud models, providing reliable core technical support for fields such as high-precision map construction for autonomous driving, complex building surveying, and spatial positioning of intelligent robots.

[0113] Example 2 like Figure 4 As shown, the second embodiment of the present invention also provides a multi-view point cloud registration device based on adaptive symmetry and hybrid optimization, comprising: The data acquisition unit is used to acquire multi-view point cloud sequence data of the target scene; The feature matching unit is used to extract features and perform pair matching on the multi-view point cloud sequence data using a pre-trained YOHO 3D local feature extraction network, calculate the initial relative transformation matrix between each point cloud pair, and evaluate the effective inlier rate of the match to obtain the overlap rate prediction matrix. An adaptive graph topology construction unit is used to adaptively filter the overlap rate prediction matrix based on dynamic edge number upper and lower limits and a preset absolute overlap score threshold to obtain a set of unidirectional matching adjacent nodes. Then, a logical union symmetry mechanism is introduced to force the unidirectional matching graph to be converted into a bidirectional symmetric undirected topology graph. A two-stage hybrid robust optimization unit is used to perform two-stage hybrid robust optimization based on the undirected topological graph and the initial relative transformation matrix to solve the global absolute pose: In the first stage, the undirected topological graph is solved globally using a spectral graph synchronization algorithm. The eigenvectors corresponding to the second smallest eigenvalues ​​of the graph Laplacian matrix are calculated using the power iteration method as the rotation components of the global pose, and the translation components are solved by combining translation constraints to output a coarse global absolute pose. In the second stage, geometric consistency checks and graph structure pruning are performed based on the coarse global absolute pose. Using the coarse global absolute pose as a priori initial value, nonlinear factor graph joint fine-tuning is performed in the Lie algebra space to output the optimized high-precision multi-view global absolute pose. The point cloud stitching unit is used to stitch together the multi-view point cloud sequence into a globally consistent complete 3D point cloud based on the high-precision multi-view global absolute pose.

[0114] Example 3 The third embodiment of the present invention also provides a multi-view point cloud registration device based on adaptive symmetry and hybrid optimization, which includes a memory and a processor. The memory stores a computer program, which can be executed by the processor to realize the multi-view point cloud registration method based on adaptive symmetry and hybrid optimization as described above.

[0115] Example 4 The fourth embodiment of the present invention also provides a computer-readable storage medium storing computer-readable instructions. When the computer-readable instructions are executed by the processor of the device where the computer-readable storage medium is located, they implement the multi-view point cloud registration method based on adaptive symmetry and hybrid optimization as described above.

[0116] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-view point cloud registration method based on adaptive symmetry and hybrid optimization, characterized in that, include: S1, acquire multi-view point cloud sequence data of the target scene; S2, use the pre-trained YOHO 3D local feature extraction network to perform feature extraction and pair matching on the multi-view point cloud sequence data, calculate the initial relative transformation matrix between each point cloud pair, and evaluate the effective inlier rate of the matching to obtain the overlap rate prediction matrix. S3, based on the dynamic edge number upper and lower limit constraints and the preset absolute overlap score threshold, the overlap rate prediction matrix is ​​adaptively filtered to obtain the set of unidirectional matching adjacent nodes. Then, a logical union symmetry mechanism is introduced to force the unidirectional matching graph to be converted into a bidirectional symmetric undirected topological graph. The process of adaptively filtering the overlap rate prediction matrix based on dynamic edge number upper and lower limits and a preset absolute overlap score threshold includes: Using a preset absolute overlap score threshold, the overlap rate prediction matrix is ​​globally binarized and filtered to extract an initial candidate neighbor node set, represented as: ； in, For the first The frame point cloud corresponds to a preset absolute overlap threshold. The initial set of candidate adjacent node indices; For point clouds arrive The predicted overlap score; To address the differences in local overlap density across different point cloud scenarios, the initial candidate adjacent node set is adaptively truncated using upper and lower limits of edge degree, resulting in a dynamically constrained set of unidirectionally matched adjacent nodes, expressed as: ； in, For the first The set of adjacent nodes corresponding to the frame point cloud after dynamic constraint and one-way matching; This represents the total number of candidate nodes in the initial candidate adjacent node set; and These are the preset minimum guaranteed number of connected edges per node and the maximum guaranteed number of connected edges per node; Indicates the first Overlap rate score vector corresponding to frame point cloud Among them, the top overlapping scores are selected. Each of the adjacent nodes; S4. Based on the undirected topological graph and the initial relative transformation matrix, a two-stage hybrid robust optimization is performed to solve the global absolute pose: In the first stage, the undirected topological graph is solved globally using the spectral graph synchronization algorithm. The eigenvector corresponding to the second smallest eigenvalue of the graph Laplacian matrix is ​​calculated using the power iteration method as the rotation component of the global pose, and the translation component is solved by combining translation constraints to output a coarse global absolute pose. In the second stage, geometric consistency checks and graph structure pruning are performed based on the coarse global absolute pose. Using the coarse global absolute pose as a priori initial value, nonlinear factor graph joint fine-tuning is performed in the Lie algebra space to output the optimized high-precision multi-view global absolute pose. S5, based on the high-precision multi-view global absolute pose, the multi-view point cloud sequence is stitched together into a globally consistent complete 3D point cloud.

2. The multi-view point cloud registration method based on adaptive symmetry and hybrid optimization according to claim 1, characterized in that... The process of feature extraction and pairwise matching using the pre-trained YOHO 3D local feature extraction network includes: Extract point clouds from any two frames using a pre-trained YOHO 3D local feature extraction network. and 3D local geometric feature descriptor; Based on the three-dimensional local geometric feature descriptor, a point-to-point correspondence set is established in the feature space through nearest neighbor search; The point pair correspondence set is filtered using a random sampling consensus algorithm to obtain the initial relative transformation matrix that maximizes the number of points in the geometry; wherein, the initial relative transformation matrix is ​​composed of an initial relative rotation matrix and an initial relative translation vector; Based on the initial relative transformation matrix, the number of valid inlier pairs that satisfy the preset inlier distance tolerance threshold is counted and the overlap rate score is calculated to construct the overlap rate prediction matrix.

3. The multi-view point cloud registration method based on adaptive symmetry and hybrid optimization according to claim 2, characterized in that... The formula for calculating the elements in the overlap rate prediction matrix is ​​as follows: ； in, Point cloud arrive The predicted overlap score; For point-to-point corresponding set The total number of point pairs in the data; This is an indicator function that returns 1 if the condition within the parentheses is met, and 0 otherwise. and The corresponding points in the point cloud and The three-dimensional coordinate vector in the coordinate system, i.e., the first Point-to-point; This is the initial relative rotation matrix; This is the initial relative translation vector; The set tolerance threshold for interior point distance; Indicates the L2 norm distance; , They represent the first , Frame point cloud.

4. The multi-view point cloud registration method based on adaptive symmetry and hybrid optimization according to claim 2, characterized in that... The process of introducing a logical union symmetry mechanism to forcibly convert a one-way matching graph into a two-way symmetric undirected topological graph is as follows: Based on the set of adjacent nodes matched unidirectionally after dynamic constraints Generate a directed initial mask matrix, whose elements are: ,when hour, ;when hour, ; Subsequently, the asymmetry of the topological graph is eliminated through a logical union operation, generating an undirected symmetric mask matrix. The formula is as follows: ； in, This is the undirected edge state mask after symmetry transformation. 1 indicates that there is an edge, and 0 indicates that there is no edge; Represents a logical OR operation; By fusing the overlap rate score and the geometric matching inlier rate, a weighted adjacency matrix of the undirected topological graph is calculated for weight allocation in subsequent robust optimization. Simultaneously, diagonal elements are set to zero to eliminate self-loop interference. The expression is as follows: ； in, These are the edge weight coefficients of the weighted adjacency matrix; For point clouds , The geometric in-point rate observed during the feature matching phase; Point cloud arrive The predicted overlap score; This results in an undirected topological graph with strictly symmetric edge weights and no topological direction.

5. The multi-view point cloud registration method based on adaptive symmetry and hybrid optimization according to claim 4, characterized in that... The specific process for outputting the coarse global absolute pose is as follows: Row normalization is performed on the weighted adjacency matrix of the undirected topological graph to obtain the graph model transition probability matrix, which serves as the input for spectral graph synchronization. The expression is as follows: ； ； in, This is the row-normalized transition probability matrix; It is a weighted adjacency matrix; for The node degree matrix is ​​a diagonal matrix, with only diagonal elements being non-zero; For the first The degree of the node corresponding to the frame point cloud, i.e., the node The sum of the weights of all adjacent edges; The weight coefficients for the bidirectional edges of the graph topology; This represents the total number of nodes in the topology graph. Based on the node degree matrix with weighted adjacency matrix Construct the standard graph Laplacian matrix to transform the pose solution into an eigenvalue decomposition problem, expressed as: ； in, The graph is a Laplace matrix; Based on the transition probability matrix, the Laplace matrix is ​​processed using the power iteration method. Perform eigenvalue decomposition and extract eigenvalues. The eigenvector corresponding to the second smallest eigenvalue is used as the global coarse estimate of multi-view rotation, and the coarse rotation estimate of each view is output. Based on the coarse rotation estimates from each viewpoint, combined with the coarse translation estimates calculated by the relative transformation, the current estimated coarse pose is obtained; Calculate the spatial residual for each topological edge using the currently estimated coarse pose; The edge weights of the weighted adjacency matrix are adaptively adjusted based on the spatial residual to reduce outlier edge weights and suppress topological noise. The edge weights of the weighted adjacency matrix and the estimated coarse pose are updated iteratively to obtain the coarse global absolute pose for each viewpoint.

6. The multi-view point cloud registration method based on adaptive symmetry and hybrid optimization according to claim 5, characterized in that... The process of performing geometric consistency checks and graph structure pruning based on the coarse global absolute pose is as follows: First, based on the rough global absolute pose, the predicted relative rotation matrix and the predicted relative translation matrix between viewpoints are derived; By utilizing the properties of the rotation matrix trace, the matrix difference is converted into a radian rotation angle residual to characterize the rotation consistency deviation. The calculation formula is as follows: ； in, Point cloud and Rotational angle residuals between; The predicted relative rotation matrix is ​​derived based on the rough global absolute pose. This is the initial relative rotation matrix; This represents the trace operation of a matrix; It is an inverse cosine function used to convert matrix differences into radian angle differences; It is the transpose symbol; Then, the three-dimensional Euclidean distance between the predicted translation vector and the observed translation vector is calculated to characterize the translation consistency deviation. The calculation formula is as follows: ； in, Point cloud and The translation distance residual between them; The predicted relative translation vector is derived based on the rough global absolute pose. This is the initial relative translation vector; Table L2 norm distance; Set angle tolerance threshold Distance tolerance threshold The rotation angle residual and the translation distance residual are jointly and double-checked to determine whether an edge is a pure interior edge, thus obtaining the set of pure interior edges, expressed as: ； in, Mark the pure interior point edge; when If the edge is a pure interior vertex edge, it is retained; otherwise, if the edge is a pure interior vertex edge, it is retained. Physically cut off the edge, without assigning it any robust core weights or optimizing resources.

7. The multi-view point cloud registration method based on adaptive symmetry and hybrid optimization according to claim 6, characterized in that... The process of joint fine-tuning of nonlinear factor graphs in the Lie algebra space, using the rough global absolute pose as the prior initial value, is as follows: Based on the set of pure interior points and edges, the standard deviations of rotation measurement noise and translation measurement noise are independently calculated for each edge, and a decoupled covariance model is constructed. The calculation formula is as follows: ； ； in, and These represent the standard deviations of rotational noise and translational noise for the corresponding connected edges, respectively. These are the edge weight coefficients of the weighted adjacency matrix; and These are the preset noise scale coefficients for rotating and translational foundations, respectively. To prevent the division of tiny positive numbers with a denominator of zero; The rough global absolute pose is used as the initial value for nonlinear optimization. A factor graph optimizer with a decoupled covariance model as the measurement noise model is constructed. The nonlinear Levenberg-Marquardt algorithm is used to minimize the global objective cost function constrained by the Huber robust error kernel function. The iteration is continued until convergence, and the high-precision multi-view global absolute pose of the point cloud at each viewpoint is output. The expression for the global objective cost function is as follows: ； in, Represents the global objective cost function for factor graph optimization; For pure interior point edge marking, when Only when this edge is selected should it participate in the optimization accumulation; Represents the Huber robust error kernel function; For the reason and Constructed diagonal covariance information matrix; To map to Lie algebra Pose residual vector in space; This is the transpose symbol.

8. The multi-view point cloud registration method based on adaptive symmetry and hybrid optimization according to claim 1, characterized in that... The expression for stitching together a multi-view point cloud sequence into a globally consistent complete 3D point cloud based on the high-precision multi-view global absolute pose is as follows: ； in, A three-dimensional coordinate vector representing a globally consistent complete three-dimensional point cloud; and The first The global rotation matrix and global translation vector for high-precision multi-view global absolute pose; Indicates the first The original 3D coordinate vector in the local camera coordinate system corresponding to the high-precision multi-view global absolute pose.

9. A multi-view point cloud registration device based on adaptive symmetry and hybrid optimization, used to implement the multi-view point cloud registration method based on adaptive symmetry and hybrid optimization as described in any one of claims 1-8, characterized in that, include: The data acquisition unit is used to acquire multi-view point cloud sequence data of the target scene; The feature matching unit is used to extract features and perform pair matching on the multi-view point cloud sequence data using a pre-trained YOHO 3D local feature extraction network, calculate the initial relative transformation matrix between each point cloud pair, and evaluate the effective inlier rate of the match to obtain the overlap rate prediction matrix. An adaptive graph topology construction unit is used to adaptively filter the overlap rate prediction matrix based on dynamic edge number upper and lower limits and a preset absolute overlap score threshold to obtain a set of unidirectional matching adjacent nodes. Then, a logical union symmetry mechanism is introduced to force the unidirectional matching graph to be converted into a bidirectional symmetric undirected topology graph. A two-stage hybrid robust optimization unit is used to perform two-stage hybrid robust optimization based on the undirected topological graph and the initial relative transformation matrix to solve the global absolute pose: In the first stage, the undirected topological graph is solved globally using the spectral graph synchronization algorithm, and the eigenvector corresponding to the second smallest eigenvalue of the graph Laplacian matrix is ​​calculated by the power iteration method as the rotation component of the global pose, and the translation component is solved by combining translation constraints to output a coarse global absolute pose; In the second stage, geometric consistency checks and graph structure pruning are performed based on the coarse global absolute pose. Using the coarse global absolute pose as a priori initial value, nonlinear factor graph joint fine-tuning is performed in the Lie algebra space to output the optimized high-precision multi-view global absolute pose. The point cloud stitching unit is used to stitch together the multi-view point cloud sequence into a globally consistent complete 3D point cloud based on the high-precision multi-view global absolute pose.

Images