Ship cabin shell multi-station cloud registration method based on loop constraint

Abstract

The application discloses a ship cabin shell multi-station cloud registration method based on a loop constraint, which comprises the following steps: extracting key points from a point cloud of a ship cabin shell by using an ISS algorithm; adopting a colored SHOT feature to describe the key points; completing preliminary pairwise coarse registration by a false matching elimination method; removing false matching in combination with a maximum consistency set constraint and a loop constraint to complete coarse registration; performing fine registration by using an iterative closest point algorithm; and completing point cloud pose optimization by using an LUM algorithm; and obtaining complete point cloud of the ship cabin shell. The method has the advantages that the possible false matching is checked by the loop constraint, error accumulation is eliminated, and the defects of long registration time, slow convergence and easy-to-error point matching in the prior art are overcome.

Description

Technical Field

[0001] This invention relates to the field of marine technology, and in particular to a multi-site cloud registration method for ship cabin shells based on lap-loop constraints. Background Technology

[0002] Point cloud registration is a widely used technique in fields such as 3D reconstruction, robot navigation, and stereo vision. Its main goal is to fuse multiple point cloud datasets from different perspectives into a complete model and make their geometry and features consistent. Therefore, improving the accuracy and speed of point cloud registration algorithms is a current research hotspot.

[0003] Shu Chengxun and He Yuntao from Beijing University of Aeronautics and Astronautics proposed a deep learning-based point cloud registration method. This method converts point clouds into depth images and uses a stacked convolutional autoencoder network to extract image features. The rotation parameters of the point cloud are then calculated using the image features through a registration network. Xie Tao, Dai Kun, Zhao Zhiwei, and others from Harbin Institute of Technology proposed a laser relocalization method for intelligent mobile robots based on dense point cloud registration. This method employs a hybrid local feature extraction and anchor-guided cross-attention mechanism to improve the accuracy of point cloud registration. Zhou Jian, Duan Cong, Xiao Jinsheng, and others from Wuhan University proposed a multi-view point cloud registration method based on high-precision maps. This method extracts superpoints from both high-precision map point clouds and multi-view point clouds, calculates their geometric information embedding, and then calculates the best matching superpoint pair to complete the point cloud data registration. Yao Jian, Wang Ying, and Li Li from Wuhan University proposed an end-to-end 3D point cloud registration method. The method acquires the point cloud features, image texture features, and mutual mapping relationship of two point clouds respectively. After resampling and calculation, resampled features and fused features are formed. Feature matching and transformation matrix estimation are then performed to register the point cloud data from the two perspectives into a single point cloud.

[0004] During shipbuilding, 3D laser scanning is used to measure the hull to check its construction accuracy, promptly identify and correct deviations, and ensure that the hull construction quality meets standards. Furthermore, the point cloud data acquired through scanning serves as an important reference for ship design and modification, helping designers better understand the hull shape, optimize designs, and improve ship performance. However, the hull measurement range is large, requiring multi-station scanning measurements to obtain complete data. The technical challenge of registering multi-station scanning point cloud data lies in the large data volume and poor overall consistency of the registration.

[0005] Currently, multi-station registration methods are mainly divided into sequential registration methods and joint registration methods. Sequential registration methods restore the scan registration order and create a fully connected graph, then extract the minimum spanning tree to determine the registration order from root to leaf. The drawback of this method is that pairwise registration is very time-consuming when processing large-scale point clouds. Joint registration methods use the results of all pairwise registrations to globally minimize the residuals, but they do not check for possible mismatches, which can lead to severe error accumulation and propagation.

[0006] Existing technologies primarily employ the Iterative Closest Point (ICP) algorithm, which requires two point clouds to have good initial positions, i.e., high overlap and similar data volume. However, in large-scale measurements such as ship hull scanning, there are often significant differences in the pose of point cloud data. Using the ICP algorithm then suffers from drawbacks such as long point cloud registration time, slow convergence, and susceptibility to point matching errors. Summary of the Invention

[0007] To overcome the aforementioned deficiencies in the existing technology, this invention provides a multi-site cloud registration method for ship cabin shells based on lap-loop constraints.

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

[0009] A multi-site cloud registration method for ship cabin shells based on lap-loop constraints includes the following steps:

[0010] Step 1: Use the ISS algorithm to extract key points from the point cloud of the ship's hull.

[0011] Step 2: Use colored SHOT features to describe the key points.

[0012] Step 3: Initial pairwise coarse registration is completed using the mismatch removal method; mismatches are removed by combining the maximum consistency set constraint and the loop closure constraint to complete the coarse registration.

[0013] Step 4: Use the iterative nearest point algorithm for fine registration and the LUM algorithm to optimize the point cloud pose; obtain the complete point cloud of the ship's outer shell.

[0014] Furthermore, in step 1, the step of extracting key points from the point cloud of the ship's hull using the ISS algorithm includes:

[0015] Assume the point cloud data has N points, and the coordinates of each point are p. i (x i y i , z i );

[0016] First, for each point p in the point cloud i Establish a local coordinate system, and specify the search radius R for each point p.i All kd-trees are constructed with a search radius R and weights w. ij for:

[0017]

[0018] Calculate point p i The covariance matrix and the eigenvalues ​​of the covariance matrix at each point Sort the eigenvalues ​​from smallest to largest:

[0019]

[0020] |λ i ED cov (p i )|=0

[0021] Finally, set thresholds ε1 and ε2, where ε1 and ε2 are no greater than 1. The key point is one that satisfies the following constraints:

[0022] and

[0023] Furthermore, step 2, which involves using colored SHOT features to describe the key points, includes:

[0024] First, a spherical coordinate system with radius R is established around each sampling point. The spherical coordinate system is divided into 2 parts along the radial direction, 2 parts along the latitude, and 8 parts along the longitude, which are divided into 32 spatial grid points.

[0025] Then, the shape and color information of each spatial grid point are statistically analyzed and a histogram is plotted; the cosine of the angle between the feature point normal vector and the neighboring point normal vector is used to represent the value of each cell in the histogram;

[0026] Then, the color information of the point cloud of the ship's outer shell is histogram encoded using the CIE color space combined with the L1 norm criterion, and a color histogram is constructed at the sampling point.

[0027] Finally, the histograms obtained from shape information and color information are fused and connected to obtain the final descriptor, realizing the description of the colored SHOT features of the point cloud of the ship's hull.

[0028] Furthermore, step 3 includes:

[0029] Constructing all scans into a fully connected graph G s (V,E);

[0030] In a fully connected graph G s In (V,E), the variable v i ∈V represents each scan, variable e iLet E represent the edge between two nodes, and define the transformation from S to T as T. s，t ;

[0031] For graph G s Each (V,E) in (V,E) i V j A coarse pair registration is achieved using a mismatch elimination method based on graph reliability metrics.

[0032] The initial pose of the point cloud of the ship's hull cannot be successfully obtained when the overlap between two scans is low or the scene content is too large, resulting in the failure of coarse registration of pairs.

[0033] By using closure constraints and prior knowledge to remove most uncertain edges, and then using maximum consistency set constraints to identify uncertain edges that cannot be reasoned out by closure constraints, mismatches can be detected and removed.

[0034] Furthermore, methods for closure constraints include:

[0035] In loop closure detection, depth-first search is used in graph G. s Extracting a unique loop structure from (V,E) Where l is the number of loops;

[0036] Loop Includes transformation T 1,2 ,T 2,3 ,T 3,4 ,…∈L i The result of the dot product should be an identity matrix, i.e., T. 1,2 ·T 2,3 ·T 3,4 …=E; During this process, the error of pairwise registration also accumulates. The error of dot product registration is defined as:

[0037]

[0038] Where, ΔR ε Represents rotational error, Δt ε Represents translation error;

[0039] R ε By transforming the Rodrigues rotation formula, we obtain the rotation error in axis-angle form.

[0040] Translation error is defined as t ε =||Δt ε ||2;

[0041] The rotation error and translation error are assigned the same weight, and at t max and θ maxBy truncating the loop, the rotation error and translation error are combined to obtain the overall combined error ψ of the closed loop:

[0042]

[0043] in, h is the number of edges contained in a loop;

[0044] t max and θ max These are the maximum translation and rotation errors predefined during pairwise registration. The validity of the closed loop is determined by the following formula:

[0045]

[0046] Validity means whether the registration error in the loop closure meets the predefined threshold. If the loop closure is valid, the transformation between scans is correctly estimated through pairwise registration; otherwise, there may be erroneous transformations in the loop closure.

[0047] Add edges from valid loops to the valid edge set, and add edges from invalid loops to the uncertain edge set;

[0048] The effective edge set contains edges that are all correctly estimated transformations. These edges can be used as prior knowledge to remove uncertain edges from the uncertain edge set. This process is defined as knowledge sharing. After knowledge sharing, the following conditions will be met: Where VES represents the valid edge set and UES represents the uncertain edge set.

[0049] Furthermore, methods for maximum consistency set constraints include:

[0050] Let the corresponding point set be Where k is the number of corresponding points, and (pi,qj) is the set of corresponding points obtained through feature extraction and matching in two scans;

[0051] Pairwise coarse matching removes erroneous matches in (pi, qj) while simultaneously seeking the maximum set of correctly matched points. This set is defined as the maximum consistency set I, as shown in the following formula:

[0052]

[0053] By scanning the pairs of constraints mentioned above (V) i V j The optimal transformation R,t is obtained using the least squares method.

[0054] Furthermore, removing mismatches using a joint constraint of maximum consistency set constraint and loop closure constraint will improve graph G. s (V,E) is divided into several subgraphs. Subgraph S i(V,E) is treated as an independent graph for subsequent fine registration, and then fused together.

[0055] Furthermore, step 4, the step of performing fine registration using the iterative nearest point algorithm, includes:

[0056] The voxel grid downsampled point cloud from the original scan is used as input for fine registration. Minimum spanning tree is used to extract graphs without loops and with well-registered pairs. The MSAC algorithm is used to define the weight of each edge based on the matching score between two scans. Then, the pairwise iterative nearest point algorithm is used to register the edges to obtain a more accurate transformation. Each node can be matched to the root node along the edge of the minimum spanning tree to obtain a unified coordinate system.

[0057] Furthermore, step 4, which involves using the LUM algorithm to optimize the point cloud pose, includes:

[0058] The LUM algorithm is used to eliminate accumulated errors by assigning closed loops, and the graph is used to optimize the pose of the point cloud by utilizing all available matching results.

[0059] Furthermore, step 4, which involves using the LUM algorithm to optimize the point cloud pose, also includes:

[0060] In the process of optimizing point cloud pose using the LUM algorithm, the covariance matrix C of the six pose parameters is calculated. i,j To minimize the weighted error function W, optimization is equivalent to minimizing the Mahalanobis distance.

[0061]

[0062] Among them, D i,j A measure of relative transformation, which includes point-to-point correspondences by searching for the nearest point; This represents the initial value for the relative transformation, for each node X. i Indicates the scanning posture.

[0063] The beneficial effects of this invention are as follows: The advantage of this method is that it eliminates error accumulation by checking for possible erroneous matches through loop closure constraints, and overcomes the disadvantages of existing technologies such as long point cloud registration time, slow convergence, and easy point matching errors. Attached Figure Description

[0064] Figure 1 This is a point cloud registration framework diagram of the hull shell of the ship according to a preferred embodiment of the present invention.

[0065] Figure 2 This is a fully connected diagram of a preferred embodiment of the present invention.

[0066] Figure 3 This is a knowledge-sharing diagram representing a preferred embodiment of the present invention.

[0067] Figure 4 This is a diagram showing the removal of mismatches according to a preferred embodiment of the present invention.

[0068] Figure 5 This is a preferred embodiment of the MST extraction diagram based on MSAC scoring of the present invention.

[0069] Figure 6 The pose diagram of the LUM algorithm in a preferred embodiment of the present invention. Detailed Implementation

[0070] The present invention will be described more clearly and completely below with reference to a preferred embodiment and the accompanying drawings.

[0071] This invention proposes a multi-site cloud registration method based on loop closure constraints for ship cabin shells. Key points are extracted using ISS, and CSHOT descriptors are used to describe the features of the key points. Preliminary pairwise coarse registration is completed by a graph reliability metric-based mismatch removal method (GROR). Then, mismatches are removed by combining maximum consistency set constraints and loop closure constraints to complete the coarse registration. Finally, ICP is used for fine registration, and the LUM algorithm is used to complete the pose optimization.

[0072] like Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 As shown, a multi-site cloud registration method for ship cabin shell based on lap-loop constraints includes the following steps:

[0073] Step 1: Use the ISS algorithm to extract key points from the point cloud of the ship's hull.

[0074] Step 2: Use colored SHOT features to describe the key points.

[0075] Step 3: Initial pairwise coarse registration is completed using the mismatch removal method; mismatches are removed by combining the maximum consistency set constraint and the loop closure constraint to complete the coarse registration.

[0076] Step 4: Use the iterative nearest point algorithm for fine registration and the LUM algorithm to optimize the point cloud pose; obtain the complete point cloud of the ship's outer shell.

[0077] The multi-site cloud registration method for ship cabin shell based on lap-loop constraints includes the following specific steps.

[0078] (1) ISS (Intrinsic Shape Signature) key point extraction

[0079] Step 1, which involves extracting key points from the point cloud of the ship's hull using the ISS algorithm, includes:

[0080] Assume the point cloud data has N points, and the coordinates of each point are p. i (x i y i , z i );

[0081] First, for each point p in the point cloud i Establish a local coordinate system, and specify the search radius R for each point p. i All kd-trees are constructed with a search radius R and weights w. ij for:

[0082]

[0083] Calculate point p i The covariance matrix and the eigenvalues ​​of the covariance matrix at each point Sort the eigenvalues ​​from smallest to largest:

[0084]

[0085] |λ i ED cov (p i )|=0

[0086] Finally, set thresholds ε1 and ε2, where ε1 and ε2 are no greater than 1. The key point is the one that satisfies the following constraints:

[0087] and

[0088] (2) Point cloud CSHOT feature description

[0089] SHOT (Signature of Histograms of Orientations) is a descriptor based on local features. For colored cabin point cloud data, we use colored SHOT features for feature description. "Colored SHOT features" is abbreviated as "CSHOT". The CSHOT feature description steps are as follows: First, establish a spherical coordinate system with radius R around each sampling point. Divide the spherical coordinate system into 2 parts radially, 2 parts latitudely, and 8 parts longitudely, resulting in 32 spatial grid points. Then, statistically analyze the shape and color information of each spatial grid point and create a histogram. The cosine of the angle between the feature point normal vector and the neighboring point normal vectors represents the value of each cell in the histogram.

[0090]

[0091] Among them, z k It is the Z-axis of the current spherical coordinate system, i.e., the normal vector of the feature point, nq This is the normal vector of all points contained within the current spatial grid. Then, using the Lab color mode defined by the CIE (Commission International Eclairage) organization, that is, employing the CIE color space combined with the L1 norm criterion, histogram encoding of the color information of the ship's outer shell point cloud is performed. A color histogram is constructed at sampling point p, as shown in the following formula:

[0092] γ=||Lab q -Lab p ||1

[0093] Among them, Lab q It is the Lab information of the nearest point q, Lab p This refers to the Lab information of sampling point p. Finally, the histogram obtained from the shape information and the histogram obtained from the color information are fused and concatenated to obtain the final descriptor, thus realizing the description of the CSHOT features of the ship's hull point cloud:

[0094] D(p)=D shape (p)∪D texture (p).

[0095] (3) Multi-station registration based on lap-loop constraints

[0096] This patent proposes a multi-station registration framework based on loop closure constraints. By constructing a graph of paired relationships, it achieves registration of the point cloud of the ship's outer shell. First, all scans are constructed into a fully connected graph. Then, GROR (Global Registration based on Oriented Relationships) is used to perform coarse pairwise registration on all scans in the graph, yielding the corresponding coarse relative transformations. To detect and remove mismatches, maximum consistency set constraints and loop closure constraints are used to achieve loop-based coarse matching from geometric and graph information, respectively. Finally, using the initial values ​​obtained from the coarse registration, Iterative Closest Point (ICP) fine registration is performed, and the point cloud pose is optimized using a graph-optimized Global Consistent Scan Matching (LUM) algorithm.

[0097] In a fully connected graph G s In (V,E), the variable v i ∈V represents each scan (node), variable e i Let E represent the edge between two nodes, and define the transformation from S to T as T. s，t Assuming there are 5 scan points, the generated fully connected graph is as follows: Figure 2 As shown.

[0098] For graph G s Each (V,E) in (V,E) i V jThe original coarse registration method used was Graph Reliability Metric-Based Mismatch Removal (GROR) to achieve pairwise coarse registration. However, the initial pose of the ship's hull point cloud could not be successfully acquired when the overlap between two scans was low or the scene content was too large, causing the pairwise coarse registration to fail. To address this, an improved method—a loop closure-based coarse matching strategy—is proposed. First, loop closure constraints (LC) use prior knowledge to remove most uncertain edges. Then, maximum consistency set constraints (MCS) are used to identify uncertain edges that cannot be reasoned about by loop closure constraints. This effectively detects and removes mismatches.

[0099] Maximum consistency set constraint, let the corresponding point set be Where k is the number of corresponding points, (p i ,q j The corresponding point set is obtained through feature extraction and matching in two scans. Pairwise coarse matching removal (p i ,q j In the case of mismatches, we seek the maximum number of correctly matched points, and define this set as the maximum consistency set I, as shown in the following formula:

[0100]

[0101] By scanning the pairs of constraints mentioned above (V) i V j The optimal transformation R,t is obtained using the least squares method.

[0102] The maximum consistency set constraint is only a geometric constraint for graph G. s Some information in (V,E) is not used. To make fuller use of the information in the graph and improve registration accuracy in certain extreme scenarios, we add loop closure constraints (LC). In loop closure detection, depth-first search (DFS) is used in graph G. s Extracting a unique loop structure from (V,E) Where l is the number of loops. Includes transformation T 1,2 ,T 2,3 ,T 3,4 ,…∈L i The result of the dot product should be an identity matrix, i.e., T. 1,2 ·T 2,3 ·T 3,4 …=E. During this process, the error in pairwise registration also accumulates. The error in dot product registration is defined as:

[0103]

[0104] Where, ΔR ε Represents rotational error, Δt ε Represents the translation error. Let R... εBy transforming the Rodrigues rotation formula, we obtain the rotation error in axis-angle form.

[0105] Similarly, the translation error is defined as t ε =||Δt ε 2. Assign equal weights to rotation error and translation error, and simultaneously at t max and θ max By truncating the loop, the rotation error and translation error are combined to obtain the overall combined error ψ of the closed loop:

[0106]

[0107] in, h is the number of edges contained in a cycle. t max and θ max These are the maximum translation and rotation errors predefined during pairwise registration. The validity of the closed loop is determined by the following formula:

[0108]

[0109] Validity means whether the registration error in the loop closure meets a predefined threshold. If the loop closure is valid, the transformation between scans is correctly estimated through pairwise registration; otherwise, there may be erroneous transformations in the loop closure. We put edges from valid loop closures into the valid edge set (VES) and edges from invalid loop closures into the uncertain edge set (UES). However, some common edges exist in different loop closures, and these common edges may be put into both VES and UES. The edges in VES are all correctly estimated transformations, and they can be used as prior knowledge to remove uncertain edges in UES. We define this process as knowledge sharing. After knowledge sharing, the following conditions will be met: Assuming the loop formed by scanning {2,3,4} and {3,4,5} is a valid loop, and the loop formed by scanning {2,4,5} is an invalid loop, the knowledge sharing process is as follows: Figure 3 As shown.

[0110] exist Figure 4 In the diagram, {e1,e2,e3} and {e4,e5,e6} are valid loops, and these edges are added to the VES (Version Equivalent Set), represented by green lines in the diagram. {e2,e6,e7} is an invalid loop, and its contained edges are added to the UES (Version Invalid Set), represented by orange lines in the diagram. Since {e2,e6} exists in both the UES and VES, they are removed from the UES. It can be deduced that if a loop contains only one uncertain edge, this edge must be a mismatch, and mismatched edges are added to the Invalid Edge Set (IES).

[0111] Using a combined constraint of maximum consistency set and loop closure to remove mismatches will transform graph G. s(V,E) is divided into several subgraphs. Subgraph S i (V,E) is treated as an independent graph for subsequent fine registration, and then fused together.

[0112] (4) Global precision matching

[0113] Loop-based coarse registration removes most mismatches and identifies usable transforms (VES). The graph composed of VES... Since these edges only contain relevant transformation information, the coordinate systems are not unified. Therefore, it is necessary to determine the order in which all scans enter a unified coordinate system.

[0114] We use the original scanned voxel grid downsampled point cloud as input for fine registration, and use MST (Minimum Spanning Tree) to extract the paired registered graph without loops. MST requires the weight of each edge to define the shortest path. Since G... v (V,E) contains only reliable matches, so the weight of each edge is defined using the MSAC (Maximum Likelihood Consensus) matching score between two scans. The weight ρ ε Defined as:

[0115]

[0116] Where, ε j It is the residual error, δ ρ It is the cutoff threshold.

[0117] MST extraction based on MSAC scores, such as Figure 5 As shown.

[0118] Then, pairwise ICP registration is used on the edges to obtain a more accurate transformation, and each node can be matched to the root node along the edges of the MST, obtaining a unified coordinate system. However, since the MST is a graph without loops, errors accumulate along the edges from the root node to the leaf node. To solve this problem, the Least Squares Update Method (LUM) algorithm is used to further optimize G. v The pose in (V,E) is used. LUM(6DOF) is employed to eliminate accumulated errors through assigned loop closures, and graphs are used to optimize the pose of the point cloud by utilizing all available matching results. Figure 6 This is the pose graph of the LUM.

[0119] Figure 6 In the middle, each edge D i,j A metric representing a relative transformation, which includes point-to-point correspondences by searching for the nearest point. This represents the initial value for the relative transformation, for each node X. i This represents the scanning pose. During the optimization process, the covariance matrix C of the six pose parameters is calculated. i,j To minimize the weighted error function W, optimization is equivalent to minimizing the Mahalanobis distance.

[0120]

[0121] This invention belongs to the field of multi-site cloud data registration technology. It proposes a method for multi-site cloud data registration based on loop closure constraints. The key points and protection points of this invention are to detect and remove mismatches by using maximum consistency set constraints and loop closure constraints, to implement a coarse point cloud matching method using geometric and graph information, and to optimize the point cloud pose using a graph-optimized Global Consistent Scan Matching (LUM) algorithm.

[0122] The advantage of this method is that it eliminates error accumulation by checking for possible mismatches through loop closure constraints, and overcomes the disadvantages of existing technologies such as long point cloud registration time, slow convergence, and easy point matching errors.

[0123] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these are merely illustrative examples, and the scope of protection of the present invention is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of the present invention, but all such changes and modifications fall within the scope of protection of the present invention.

Claims

1. A multi-site cloud registration method for ship cabin shells based on lap-loop constraints, characterized in that, It includes the following steps: Step 1: Use the ISS algorithm to extract key points from the point cloud of the ship's hull. Step 2: Use colored SHOT features to describe the key points. Step 3: Initial pairwise coarse registration is completed using the mismatch removal method; mismatches are removed by combining the maximum consistency set constraint and the loop closure constraint to complete the coarse registration. Step 4: Use the iterative nearest point algorithm for fine registration, and use the LUM algorithm to optimize the point cloud pose. Obtain the complete point cloud of the ship's outer shell; Step 3 includes: Construct all scans into a fully connected graph ; In a fully connected graph In, variables The variable represents the node scanned in each iteration. Represents the edge between two nodes. arrive The transformation process is defined as ; For the graph Each of them A coarse pair registration is achieved using a mismatch elimination method based on graph reliability metrics. The initial pose of the hull point cloud could not be successfully obtained due to the low overlap between the two scans, resulting in the failure of the coarse registration of the pairs. By using closure constraints and prior knowledge to remove uncertain edges, and then using maximum consistency set constraints to identify uncertain edges that cannot be reasoned by closure constraints, false matches can be detected and removed. Methods for closure constraints include: In loop closure detection, depth-first search is used in the graph. Extracting unique loop structures ,in It is the number of loops; Loop Contains conversion The result of the dot product should be an identity matrix, i.e. During this process, the error of pairwise registration also accumulates. The error of dot product registration is defined as: in, Represents rotational error. Represents translation error; Will By transforming the Rodrigues rotation formula, we obtain the rotation error in axis-angle form. ; The overall translational deviation is defined as the matrix 2-norm. ; The rotation error and translation error are assigned the same weight, and at the same time... and The rotation error and translation error are combined by truncation to obtain the overall combined error of the closed loop. : ； in, , , It is the number of edges contained in a loop; and These are the maximum translation and rotation errors predefined during pairwise registration. The validity of the closed loop is determined by the following formula: ； Validity determination means judging whether the registration error in the loop closure meets the predefined threshold. If the loop closure is valid, the transformation between scans is correctly estimated through pairwise registration. Otherwise, there may be erroneous transformations in the loop closure. Add edges from valid loops to the valid edge set, and add edges from invalid loops to the uncertain edge set; The effective edge set contains edges that are all correctly estimated transformations. These edges serve as prior knowledge to remove uncertain edges from the uncertain edge set. This process is defined as knowledge sharing. After knowledge sharing, the following conditions will be met: Where VES represents the valid edge set and UES represents the uncertain edge set.

2. The multi-site cloud registration method for ship cabin shell based on lap-loop constraints as described in claim 1, characterized in that, Step 1, which involves extracting key points from the point cloud of the ship's hull using the ISS algorithm, includes: Assuming there are a total of point cloud data There are points, and the coordinates of each point are... ; First, for each point in the point cloud Establish a local coordinate system and specify the search radius for each point. Each point All based on search radius Construct a kd-tree with weights for: Calculation points The covariance matrix and the eigenvalues ​​of the covariance matrix at each point The eigenvalues ​​are sorted from smallest to largest: Finally, set the threshold. and threshold and The key point is that the value is no greater than 1 and the following constraints are met: 。 3. The multi-site cloud registration method for ship cabin shell based on lap-loop constraints as described in claim 1, characterized in that, Step 2, which uses colored SHOT features to describe the key points, includes: First, a spherical coordinate system with radius R is established around each sampling point. The spherical coordinate system is divided into 2 parts along the radial direction, 2 parts along the latitude, and 8 parts along the longitude, which are divided into 32 spatial grid points. Then, the shape and color information of each spatial grid point are statistically analyzed and a histogram is plotted; the cosine of the angle between the feature point normal vector and the neighboring point normal vector is used to represent the value of each cell in the histogram; Then, the color information of the point cloud of the ship's outer shell is histogram encoded using the CIE color space combined with the L1 norm criterion, and a color histogram is constructed at the sampling point. Finally, the histograms obtained from shape information and color information are fused and connected to obtain the final descriptor, realizing the description of the colored SHOT features of the point cloud of the ship's hull.

4. The multi-site cloud registration method for ship cabin hull based on lap-loop constraints as described in claim 1, characterized in that, Methods for maximum consistency set constraints include: Let the corresponding point set be ,in is the number of corresponding points, and (pi,qj) is the set of corresponding points obtained through feature extraction and matching in two scans; Pairwise coarse matching removes erroneous matches in (pi, qj) while simultaneously seeking the maximum set of correctly matched points. This set is defined as the maximum consistency set I, as shown in the following formula: ； The above constraints are used to scan pairs. The optimal transformation is obtained using the least squares method. , .

5. The multi-site cloud registration method for ship cabin shell based on lap-loop constraints as described in claim 4, characterized in that, Using a combined constraint of maximum consistency set constraint and loop closure constraint to remove mismatches will improve the graph. Divide into several subgraphs , subgraph The images are then treated as independent images for further fine registration before being merged.

6. The multi-site cloud registration method for ship cabin shell based on lap-loop constraints as described in claim 4, characterized in that, Step 4, the step of performing fine registration using the iterative nearest point algorithm, includes: The voxel grid downsampled point cloud from the original scan is used as input for fine registration. Minimum spanning tree is used to extract graphs without loops and with well-registered pairs. The MSAC algorithm is used to define the weight of each edge based on the matching score between two scans. Then, the pairwise iterative nearest point algorithm is used to register the edges to obtain a more accurate transformation. Each node can be matched to the root node along the edge of the minimum spanning tree to obtain a unified coordinate system.

7. The multi-site cloud registration method for ship cabin shell based on lap-loop constraints as described in claim 6, characterized in that, Step 4, which involves using the LUM algorithm to optimize the point cloud pose, includes: The LUM algorithm is used to eliminate accumulated errors by assigning closed loops, and the graph is used to optimize the pose of the point cloud by utilizing all available matching results.

8. The multi-site cloud registration method for ship cabin shell based on lap-loop constraints as described in claim 6, characterized in that, Step 4, which involves using the LUM algorithm to optimize the point cloud pose, also includes: In the process of optimizing point cloud pose using the LUM algorithm, the covariance matrix of the six pose parameters is calculated. Make the weighted error function Minimize, optimization is equivalent to minimizing Mahalanobis distance; in, A measure of relative transformation, which includes point-to-point correspondences by searching for the nearest point; Represents the initial values ​​for the relative transformation, for each node. Indicates the scanning posture.

Images