A multi-lrf point cloud registration method against noise and occlusion

Abstract

The application discloses an anti-noise and anti-shielding multiple LRF point cloud registration method, comprising the following steps: selecting key points from source point clouds and target point clouds respectively, and obtaining an initial corresponding point pair set according to the key points; a multiple local reference frame construction step; for each point pair in the initial corresponding point pair set, a plurality of local reference frames are constructed with the source point cloud key point p and the target point cloud key point q of each point pair as the center respectively; a multiple conversion hypothesis generation step; a conversion hypothesis verification and optimal solution screening step; an iterative optimization and registration result output step. The multiple LRF point cloud registration method has the advantages that the consistency and stability of the local reference frame in a complex scene containing a symmetrical structure or a regular geometric body can be significantly improved, and the registration algorithm can still output a high-precision registration result on strong interference and low-quality point cloud data.

Description

Technical Field

[0001] This invention relates to a point cloud registration technique, and more particularly to a noise- and occlusion-resistant multi-LRF point cloud registration method. Background Technology

[0002] Point cloud registration is a technique that aligns 3D point cloud data scanned from different angles or at different times to create a complete model. Its core is calculating rotation and translation parameters to ensure data overlap. Essentially, point cloud registration solves the rigid transformation relationship between two point clouds, using mathematical calculations to transform the source point cloud to match the target point cloud. As a fundamental problem in 3D computer vision and photogrammetry, point cloud registration aims to align multiple point cloud datasets located in different coordinate systems to the same reference coordinate system through spatial transformation. Point cloud registration has wide applications in 3D reconstruction, simultaneous localization and mapping (SMR), robot navigation, and reverse engineering. With the widespread use of 3D scanning equipment and the increasing complexity of application scenarios, point cloud data is often affected by noise, occlusion, uneven local point cloud density, and low overlap rates, placing higher demands on the robustness and accuracy of registration algorithms.

[0003] Currently, mainstream point cloud registration methods can be broadly categorized into two types: geometric constraint-based registration methods and local feature-based registration methods. Geometric constraint-based methods use distance and angle constraints to select a consistent set from initial corresponding point pairs to calculate the transformation matrix. However, when the correct matching rate in the initial corresponding point pairs is low, geometric constraints exhibit significant ambiguity, making it difficult to effectively eliminate incorrect matches. Local feature-based point cloud registration methods generate transformation hypotheses by constructing local reference frames or local reference axes, and have become a research hotspot due to their strong robustness to noise and occlusion.

[0004] Chinese patent application CN202510684882 discloses an iterative registration optimization algorithm based on angle clustering. It assesses the compatibility of corresponding point pairs by constructing a compatibility matrix and uses an angle clustering strategy to cluster rotation angles to generate conversion hypotheses. While this invention improves registration accuracy to some extent in low overlap scenarios, it requires selecting two corresponding point pairs simultaneously and filtering a third pair using the compatibility matrix to generate conversion hypotheses, making the calculation process complex. Furthermore, the accuracy of the compatibility matrix construction is affected when the outlier rate of corresponding point pairs is high, leading to decreased algorithm stability. Chinese patent application CN201510833425.5 discloses a method suitable for automated multi-view registration of multi-station terrestrial laser point cloud data. It achieves global registration through semantic feature point extraction and feature matching, uses triangular geometric constraints and semantic constraints for feature point matching, and eliminates erroneous matches through geometric consistency clustering. While this invention can effectively resist the effects of noise and occlusion, the extraction of its semantic feature points relies on a series of preprocessing operations such as data slicing, distance clustering, and geometric primitive fitting. It has high requirements for the quality of point cloud data, and the extraction of semantic features depends on prior knowledge, making it difficult to guarantee the repeatability of feature points in complex scenarios.

[0005] While the aforementioned existing technologies attempt to address the interference resistance problem in point cloud registration from different perspectives, their technical solutions all have limitations. Although registration methods based on local reference frames do not rely on prior knowledge and are computationally efficient, their core bottleneck lies in the fact that the quality of the constructed local reference frame directly determines the success or failure of the registration. Existing technologies typically construct the x-axis of the local reference frame by finding the salient point with the largest projection height along the z-axis, and then determining the x-axis direction by projecting the line connecting this point and the key point onto an orthogonal plane. However, in actual point cloud data, when the local neighborhood has specific geometric structures such as symmetrical planes, regular edges, or flat regions, multiple neighboring points along the z-axis may have the same or approximately the same projection extremum. Existing algorithms can only randomly select one of these points as the salient point, leading to ambiguity in the constructed x-axis direction. This random fluctuation in the x-axis direction causes inconsistencies in the local reference frames constructed for the same key point under different scans or viewpoints, significantly reducing the repeatability of the local reference frame. This, in turn, introduces uncertainty errors into the transformation hypothesis generation process, ultimately affecting the accuracy and stability of the registration.

[0006] In practical engineering applications, the aforementioned technical problems manifest as follows: when the object to be registered contains symmetrical structures such as the end face of a cylinder, regular planes such as building walls, or long straight edges such as roof ridges, the registration results output by existing algorithms often have angular deviations or fail to converge, requiring manual intervention or multiple repeated scans to obtain ideal registration results. With the deepening application of 3D laser scanning technology in fields such as automated measurement and robot navigation, symmetrical structures and regular geometries are prevalent in the working environment, and the aforementioned technical problems have become a key bottleneck restricting the widespread application of point cloud registration technology in practical engineering.

[0007] Therefore, there is an urgent need for a point cloud registration method that can maintain the stability of the local reference frame under complex geometries to address the aforementioned shortcomings of existing technologies. Summary of the Invention

[0008] To overcome the shortcomings of the existing technologies, this invention provides a noise- and occlusion-resistant multi-LRF point cloud registration method, which significantly improves the consistency and stability of the local reference frame in complex scenes containing symmetrical structures or regular geometries, and enables the registration algorithm to output high-precision registration results even on point cloud data with strong interference and low quality.

[0009] The present invention adopts the following technical solution to solve the technical problem.

[0010] A noise- and occlusion-resistant multi-LRF point cloud registration method, characterized by the following steps:

[0011] S1: Select key points from the source point cloud and the target point cloud respectively, and obtain an initial set of corresponding point pairs based on the key points;

[0012] S2: Multiple local reference frame construction steps; For each point pair in the initial set of corresponding point pairs, multiple local reference frames are constructed with the source point cloud keypoint p and the target point cloud keypoint q as the center for each point pair;

[0013] S3: Multiple transformation hypothesis generation step; Based on the multiple local reference frames obtained in step 2, obtain the rotation matrix R. n Translation vector t n This process yields N candidate transformation hypotheses; these N candidate transformation hypotheses are then output as the transformation hypothesis set generated for the corresponding point pair to the next step.

[0014] S4: Hypothesis verification and optimal solution selection steps; based on the rotation matrix R of each of the N candidate transformation hypotheses in the transformation hypothesis set. n and a translation vector t n This allows us to obtain the optimal transformation parameters and the correct matching point pairs for the current iteration.

[0015] S5: Iterative optimization and registration result output steps; determine whether the maximum number of iterations or the number of interior points has reached the threshold. If so, add the rotation matrix from the current optimal transformation parameters. Translation vector This will be output as the final registration result.

[0016] The present invention also features a noise- and occlusion-resistant multi-LRF point cloud registration method, which is characterized by:

[0017] Furthermore, step S1 includes a key point extraction and feature description step and an initial corresponding point pair set acquisition step.

[0018] Furthermore, the key point extraction and feature description process includes the following steps:

[0019] S111: Obtain the spatial coordinates of a 3D point;

[0020] S112: Select key points based on the spatial coordinates;

[0021] S113: Calculate the three-dimensional coordinates of the neighboring points of the key point;

[0022] S114: Determine the direction of the local reference axis;

[0023] S115: Divide the spatial sub-regions; using the local reference axis direction as the north pole direction of the spherical neighborhood and as the dividing benchmark, divide the spherical neighborhood along the spatial direction to form multiple spatial sub-regions;

[0024] S116: Obtain the local feature descriptor of the key point.

[0025] Furthermore, when determining the direction of the local reference axis, a sign redirection method is used to adjust the direction of the local reference axis to point to the main side of the distribution of neighboring points.

[0026] Furthermore, the process of obtaining the initial set of corresponding point pairs includes the following steps:

[0027] S121: Receive local feature descriptor;

[0028] S122: Calculate the feature distance between local feature descriptors;

[0029] S123: Select candidate corresponding points;

[0030] S124: Gather each key point in the source point cloud and its candidate corresponding points into an initial set of corresponding point pairs;

[0031] S125: Simplify the initial set of corresponding point pairs to obtain the simplified initial set of corresponding point pairs.

[0032] Furthermore, step S125 also includes a step of performing a geometric consistency check on the simplified initial set of corresponding points to obtain a further simplified set of initial sets of corresponding points.

[0033] Further, step S2 includes the following steps:

[0034] S21: Determine the final z-axis vector z(p) of the source point cloud keypoint p of the multiple local reference frame;

[0035] S22: According to the final The x-axis vector z(p) is used to construct the x-axis vector of the multiple local reference frames. Two strategies are employed to obtain the final x-axis vector x1(p) of the first strategy, the main x-axis direction x2(p) of the second strategy, and one or more candidate x-axis directions. k (p);

[0036] S23: Based on the final z-axis vector z(p) obtained in step 21, the final x-axis vector x1(p) obtained in step 22, the main x-axis direction x2(p), and K1 candidate x-axis directions x k (p), to obtain the y-axis vector y j (p);

[0037] S24: Combining the final z-axis vector z(p), the final x-axis vector x1(p), the main x-axis direction x2(p), and the K1 candidate x-axis directions x k (p) and y-axis vector y j (p), obtaining U types of local reference frames (LRFs) for the source point cloud keypoint p. u (p); U = K1+2, index u=1,2,...,U;

[0038] S25: Similarly, obtain V local reference frames (LRFs) for the keypoints q in the target point cloud. v (q); V = K2+2, index v=1,2,...,V.

[0039] Further, step S3 includes the following steps:

[0040] S31: Receive U types of local reference frames (LRFs) from the source point cloud keypoint p in step S24. u (p), receive V local reference frames (LRFs) of the target point cloud keypoints q from step S25. v (q); Set the number of matches N = min(U, V), extract the first N local reference frames of the source point cloud and the target point cloud respectively, and match them one by one;

[0041] S32: Calculate R according to the rotation matrix formula. n =(LRFn (q)) T ·LRF n (p) Calculate the rotation matrix R corresponding to the nth local reference frame. n , where n=1,2,...,N, T represents matrix transpose, and · represents matrix multiplication;

[0042] S33: Calculate t according to the translation vector formula. n = q - R n ·p calculates the translation vector corresponding to the nth local reference frame, where q and p are the three-dimensional coordinate vectors of the key points in the target point cloud and the key points in the source point cloud, respectively;

[0043] S34: will be determined by the rotation matrix R n Translation vector t n The constructed transformation parameter pair is used as the nth candidate transformation hypothesis. The above operation is performed on all N local reference frames to obtain N candidate transformation hypotheses. These N candidate transformation hypotheses are output as the transformation hypothesis set generated for the corresponding point pair to step S4.

[0044] Further, step S4 includes the following steps:

[0045] S41: Receive N candidate transformation hypotheses generated from the current corresponding point pair from step S3. Each candidate transformation hypothesis contains a rotation matrix R. n and a translation vector t n ;

[0046] S42: For each point in the transformed source point cloud, search for the point in the target point cloud with the closest Euclidean distance and record the closest distance value;

[0047] S43: The number of points in the source point cloud after statistical transformation whose distance to the nearest point in the target point cloud is less than a preset inlier distance threshold is used as the number of inlier points corresponding to the candidate transformation hypothesis.

[0048] S44: Compare the number of interior points corresponding to the N candidate transformation hypotheses, and determine the candidate transformation hypothesis with the largest number of interior points as the optimal transformation parameter for the current iteration;

[0049] S45: Record the optimal transformation parameters and their corresponding set of interior points, and output the corresponding point pairs in the set of interior points as the correct matching point pairs to step S5.

[0050] Further, step S5 includes the following steps:

[0051] Step S5 includes the following steps:

[0052] S51: Receive the optimal transformation parameters of the current iteration and its corresponding set of interior points from step S4, and use the set of interior points as the initial set of corresponding point pairs for the next iteration;

[0053] S52: Determine whether the current iteration count has reached the preset maximum iteration count, or whether the number of inliers corresponding to the current optimal transformation parameter has reached the preset inlier count threshold;

[0054] S53: Apply the final registration result to the source point cloud, according to the formula... Calculate the final coordinates of each point in the source point cloud after transformation, and complete the accurate registration between the source point cloud and the target point cloud.

[0055] Compared with existing technologies, the beneficial effects of this invention are reflected in:

[0056] This invention discloses a noise- and occlusion-resistant multi-LRF point cloud registration method, comprising the following steps:

[0057] S1: Key point extraction and feature description steps; Select key points from the source point cloud and the target point cloud respectively, obtain the position covariance matrix based on the spherical neighborhood of the key points and construct a normalized histogram, and use the normalized histogram as the feature description of the key points;

[0058] S2: Initial correspondence point pair set acquisition steps; establish a one-to-one correspondence between the key points of the source point cloud and the key points of the target point cloud, and gather the one-to-one correspondence between the key points of the source point cloud and the key points of the target point cloud into an initial correspondence point pair set;

[0059] S3: Multiple local reference frame construction steps; For each point pair in the initial set of corresponding point pairs, multiple local reference frames are constructed with the source point cloud keypoint p and the target point cloud keypoint q as the center for each point pair;

[0060] S4: Multiple transformation hypothesis generation steps;

[0061] S5: Transformation Hypothesis Verification and Optimal Solution Selection Steps;

[0062] S6: Iterative optimization and registration result output steps.

[0063] 1. In the construction of the local reference frame, this invention abandons the traditional single-frame construction mode and instead adopts a frame construction method based on multiple strategies. It not only constructs a highly robust z-axis through radially weighted covariance analysis, but also integrates weighted covariance analysis and multiple x-axis construction strategies based on extreme point projection. Specifically, addressing the ambiguity of the x-axis caused by randomly selecting a single direction due to multiple projection extreme points in the local geometry in existing technologies, this invention introduces a direction histogram-based weighted voting mechanism. This mechanism includes all points whose projection height is close to the extreme value in the candidate set, and determines one or more stable x-axis directions through weight accumulation and peak detection.

[0064] 2. This invention changes the registration logic of the prior art that relies on a single frame to generate a single transformation hypothesis, and uses the above-mentioned multiple local reference frames to synchronously generate multiple candidate transformation hypotheses for each initial corresponding point pair.

[0065] 3. This invention constructs a complete optimization mechanism that includes hypothesis generation, interior point quantity evaluation, and iterative feedback. It automatically selects the optimal hypothesis by using the objective criterion of maximizing the number of interior points, and uses the selected interior point set as the input for the next iteration to continuously improve the quality of the corresponding point pairs.

[0066] This invention overcomes the shortcomings of existing point cloud registration methods based on local reference frames, which suffer from poor repeatability and decreased registration accuracy and stability when dealing with complex geometric structures such as symmetrical planes, regular edges, or flat regions due to random fluctuations in the x-axis construction direction. Simultaneously, by employing a multi-frame fusion strategy, it addresses the technical bottleneck of a single local reference frame's inability to simultaneously accommodate multiple interference factors such as noise, occlusion, low overlap rates, and uneven local point cloud density. Through these improvements, the consistency and stability of the local reference frame are significantly enhanced in complex scenes containing symmetrical structures or regular geometries, thereby improving the reliability of the transformation hypothesis. Furthermore, through multiple hypothesis verification and iterative optimization mechanisms, the registration algorithm ultimately outputs high-precision registration results even on point cloud data with strong interference and low quality.

[0067] The noise- and occlusion-resistant multi-LRF point cloud registration method of the present invention has the advantages of significantly improving the consistency and stability of the local reference frame in complex scenes containing symmetrical structures or regular geometry, and enabling the registration algorithm to output high-precision registration results even on point cloud data with strong interference and low quality. Attached Figure Description

[0068] Figure 1 This is a flowchart of a multi-LRF point cloud registration method for noise and occlusion resistance according to the present invention.

[0069] Figure 2This is a flowchart of the multiple local reference frame construction process (S3) of a noise-resistant and occlusion-resistant multiple LRF point cloud registration method of the present invention.

[0070] Figure 3 This is a flowchart of the transformation hypothesis generation and verification process (S3-S4) of a noise-resistant and occlusion-resistant multi-LRF point cloud registration method of the present invention.

[0071] Figure 4 This is a flowchart of the iterative optimization and decision-making process (S6) of a noise-resistant and occlusion-resistant multi-LRF point cloud registration method according to the present invention.

[0072] Figure 5 This is a data flow and processing chain diagram of a noise-resistant and occlusion-resistant multi-LRF point cloud registration method according to the present invention.

[0073] Figure 6 This is a schematic diagram of the spherical neighborhood in S1 of a noise-resistant and occlusion-resistant multiple LRF point cloud registration method of the present invention.

[0074] Figure 7 This is a schematic diagram of the principal direction (z-axis) of the covariance of a noise-resistant and occlusion-resistant multi-LRF point cloud registration method according to the present invention (red dots represent key points, and circles represent point cloud distribution).

[0075] The present invention will be further described below through specific embodiments and in conjunction with the accompanying drawings. Detailed Implementation

[0076] See Figures 1 to 7 The present invention provides a noise- and occlusion-resistant multi-LRF point cloud registration method, comprising the following steps:

[0077] S1: Select key points from the source point cloud and the target point cloud respectively, and obtain an initial set of corresponding point pairs based on the key points;

[0078] S2: Multiple local reference frame construction steps; For each point pair in the initial set of corresponding point pairs, multiple local reference frames are constructed with the source point cloud keypoint p and the target point cloud keypoint q as the center for each point pair;

[0079] S3: Multiple transformation hypothesis generation step; Based on the multiple local reference frames obtained in step 2, obtain the rotation matrix R. n Translation vector t n This process yields N candidate transformation hypotheses; these N candidate transformation hypotheses are then output as the transformation hypothesis set generated for the corresponding point pair to the next step.

[0080] S4: Hypothesis verification and optimal solution selection steps; based on the rotation matrix R of each of the N candidate transformation hypotheses in the transformation hypothesis set.n and a translation vector t n This allows us to obtain the optimal transformation parameters and the correct matching point pairs for the current iteration.

[0081] S5: Iterative optimization and registration result output steps; determine whether the maximum number of iterations or the number of interior points has reached the threshold. If so, add the rotation matrix from the current optimal transformation parameters. Translation vector This will be output as the final registration result.

[0082] This invention provides a noise- and occlusion-resistant multi-LRF point cloud registration method, comprising the following steps: key point extraction and feature description, initial corresponding point pair acquisition, construction of multiple local reference frames, generation of multiple transformation hypotheses, verification of transformation hypotheses and selection of optimal solutions, and iterative optimization and registration result output. By constructing multiple local reference frames, multiple candidate transformation hypotheses are generated for each corresponding point pair, and the optimal hypothesis is selected based on maximizing the number of interior points. The interior points are iteratively fed back to the next processing step, and finally, a high-precision registration result is output. This method overcomes the shortcomings of a single local reference frame, which cannot take into account multiple interferences such as noise, occlusion, and low overlap rate, and significantly improves the adaptability and robustness of the registration algorithm in complex scenes.

[0083] In specific implementation, step S1 includes a key point extraction and feature description step and an initial corresponding point pair set acquisition step.

[0084] In the process of key point extraction and feature description, key points are selected from the source point cloud and the target point cloud respectively. The spatial distribution and geometric features of the inner neighborhood points of the spherical neighborhood of the key point are encoded and combined to construct the local feature descriptor of the key point. This descriptor is used for feature matching in the subsequent process to find a correct and reliable correspondence between the source point cloud and the target point cloud.

[0085] In practice, the key point extraction and feature description process includes the following steps:

[0086] S111: Obtain the spatial coordinates of a 3D point;

[0087] Read the spatial coordinates of all 3D points in the source point cloud data file; read the spatial coordinates of all 3D points in the target point cloud data file.

[0088] S112: Select key points based on the spatial coordinates;

[0089] In specific implementation, key points can be selected using methods such as uniform sampling, ISS (Intrinsic Shape Signatures) algorithm, NARF (Normal Aligned Radial Feature) algorithm, or SIFT (Scale-Invariant Feature Transform). Key points are selected from the source point cloud and the target point cloud. Specifically, a uniform sampling method is used to select key points in both the source and target point clouds, ensuring that the minimum distance between any two key points is greater than a preset distance threshold. The preset distance threshold is twice the average distance mr between nearest neighbor pairs; the preset descriptor radius is 15 times the average distance mr between nearest neighbor pairs.

[0090] S113: Calculate the 3D coordinates of the neighboring points of the key point; for each selected key point, determine a spherical neighborhood centered on the key point with a preset descriptor radius as its length, and obtain the 3D coordinates of all neighboring points within this spherical neighborhood; such as Figure 7 A schematic diagram of the spherical neighborhood of point p and its neighboring points;

[0091] S114: Determine the direction of the local reference axis; calculate the position covariance matrix of all neighborhood points in the spherical neighborhood relative to the key point, perform eigenvalue decomposition on the position covariance matrix, take the eigenvector corresponding to the smallest eigenvalue as the initial direction of the local reference axis, and use the sign redirection method to adjust the direction of the local reference axis to point to the main side of the neighborhood point distribution.

[0092] S115: Divide the spatial sub-regions; using the local reference axis direction as the north pole direction of the spherical neighborhood and as the dividing benchmark, divide the spherical neighborhood along the spatial direction to form multiple spatial sub-regions;

[0093] S116: Obtain the local feature descriptor of the key point.

[0094] Based on the divided spatial sub-regions, the spatial distribution and geometric features of the neighboring points within the region are extracted and encoded, and then combined to construct the local feature descriptor of the key point and output to step S121.

[0095] In specific implementation, in step S114, when determining the direction of the local reference axis, a symbol redirection method is used to adjust the direction of the local reference axis to point to the main side of the distribution of neighboring points.

[0096] The symbol redirection method is as follows: calculate the average value of the vectors pointing from the key point to all neighboring points in its spherical neighborhood, multiply the initial direction of the local reference axis by this average value, and if the result of the multiplication is greater than or equal to zero, keep the initial direction unchanged; otherwise, invert it.

[0097] In practice, the process of obtaining the initial set of corresponding point pairs includes the following steps:

[0098] Establish a one-to-one correspondence between the key points of the source point cloud and the key points of the target point cloud, and gather the one-to-one correspondence between the key points of the source point cloud and the key points of the target point cloud into an initial set of corresponding point pairs.

[0099] S121: Receive local feature descriptors; Receive local feature descriptors of all key points in the source point cloud and local feature descriptors of all key points in the target point cloud from step S1;

[0100] S122: Calculate the feature distance between local feature descriptors; for each key point in the source point cloud, calculate the distance or difference (e.g., Euclidean distance) between the local feature descriptor of each key point in the source point cloud and the local feature descriptor of each key point in the target point cloud, and use the calculation result as a feature similarity measure between the two key points.

[0101] S123: Select candidate corresponding points; Sort all key points in the target point cloud in ascending order of feature similarity, select the key point in the target point cloud with the smallest feature similarity as the candidate corresponding point of the key point in the source point cloud, and record the corresponding point pair and its feature similarity value; Each key point in the source point cloud has a corresponding key point in the target point cloud (i.e., the candidate corresponding point of the key point in the source point cloud), thus establishing a one-to-one mapping relationship between the key points in the source point cloud and the key points in the target point cloud, i.e., forming a point pair; Each mapping pair (i.e., the initial corresponding point pair) has a corresponding feature similarity value.

[0102] S124: Collect each key point in the source point cloud and its candidate corresponding points into an initial set of corresponding point pairs; after performing the above-mentioned selection of candidate corresponding points on all key points in the source point cloud, an initial set of corresponding point pairs is obtained; the initial set of corresponding point pairs includes all key points in the source point cloud and their candidate corresponding points; the number of point pairs contained in the initial set of corresponding point pairs is equal to the number of key points in the source point cloud.

[0103] S125: Simplify the initial set of corresponding point pairs to obtain a simplified initial set of corresponding point pairs. Sort all candidate corresponding point pairs in the initial set of corresponding point pairs according to their feature similarity values ​​from smallest to largest, and select the top 200 corresponding point pairs with the smallest feature similarity to form the preliminary simplified initial set of corresponding point pairs.

[0104] In specific implementation, step S125 also includes a step of performing a geometric consistency check on the simplified initial set of corresponding points to obtain a further simplified set of initial sets of corresponding points.

[0105] In step S125, after selecting the first 200 corresponding point pairs with the smallest feature similarity to form an initial set of corresponding point pairs after preliminary simplification, a geometric consistency test is performed on each corresponding point pair in the set. Point pairs in the set whose Euclidean distance between the source point cloud key points and the target point cloud key points is greater than a preset geometric consistency threshold are removed. The remaining point pairs after the two simplifications are output as the final set of initial corresponding point pairs to step S211. The preset geometric consistency threshold is set to 5 times the average distance mr between the nearest neighbor point pairs.

[0106] In specific implementation, step S2 includes the following steps:

[0107] S21: Determine the final z-axis vector z(p) of the source point cloud keypoint p of the multiple local reference frame;

[0108] S211: Determine the local point set Q based on the initial set of corresponding point pairs;

[0109] The initial set of corresponding point pairs, simplified twice, is received in step S125. For each point pair in this initial set, a spherical neighborhood is determined centered on the source point cloud keypoint p and the target point cloud keypoint q. First, a spherical neighborhood with a preset radius is determined in the source point cloud centered on keypoint p, and all neighboring points (target point cloud keypoint q) within this spherical neighborhood are obtained to form a local point set Q. Similarly, the same neighborhood acquisition operation is performed in the target point cloud centered on target point cloud keypoint q. Therefore, in the specific operation, a local point set Q is obtained by determining a spherical neighborhood centered on keypoint p in the source point cloud; at the same time, another local point set is obtained by determining a spherical neighborhood centered on target point cloud keypoint q. For simplicity, the calculation steps below are all illustrated using the source point cloud keypoint p and its local point set Q as examples; the i-th neighboring point in the local point set Q is denoted as q. i .

[0110] S212: Calculate the i-th neighboring point q in the local point set Q. i radial weight w i ;

[0111] For the i-th neighboring point q i Calculate the i-th neighboring point q i The Euclidean distance ||pq between the center point p and the center point p i Then, calculate the i-th neighboring point q according to formula (1). i radial weight w i ;

[0112] w i =r1 -‖pq i ‖ (1);

[0113] In formula (1), r1 is the first radius; p represents the spatial coordinates of the key point p in the source point cloud (expressed as a vector); q i Let q represent the i-th neighborhood point of the target point cloud. i Spatial coordinates (expressed as vectors); ‖·‖ represents finding the magnitude of the vector.

[0114] S213: Construct the weighted covariance matrix Cov(Q);

[0115] The calculation process of the weighted covariance matrix Cov(Q) is shown in the following formula (2);

[0116] (2);

[0117] In formula (2), |Q| is the number of neighborhood points in the local point set Q; T represents the matrix transpose operation.

[0118] S214: Calculate the initial... Axial direction v min ;

[0119] The weighted covariance matrix Cov(Q) is decomposed into eigenvalues ​​to obtain three eigenvalues ​​and three corresponding eigenvectors. The eigenvector corresponding to the smallest eigenvalue is taken as the initial z-axis direction v. min ;

[0120] S215: Final determination Axis vector z(p).

[0121] Calculate the average value of the vector pointing from the center point p to all neighboring points in the local point set Q, and determine whether the direction of the initial z-axis vector z(p) needs to be re-signed according to formula (3); if the condition is met... Under this condition, the final z-axis vector z(p) = v min If the above conditions are not met, then the initial z-axis direction is reversed, i.e., z(p) = -v. min Then, the final z-axis vector z(p) is output to steps S22 and 23; see the following formula (3);

[0122] (3).

[0123] In formula (3), p represents the spatial coordinates of the source point cloud keypoint p (expressed as a vector); q i Let q represent the i-th neighborhood point of the target point cloud. i Spatial coordinates (expressed as vectors).

[0124] S22: According to the final The x-axis vector z(p) is used to construct the x-axis vector of the multiple local reference frames. Two strategies are employed to obtain the final x-axis vector x1(p) of the first strategy, the main x-axis direction x2(p) of the second strategy, and one or more candidate x-axis directions. k (p);

[0125] In practice, two different strategies are used to construct the x-axis vector of the multiple local reference frames.

[0126] The first strategy includes the following steps (the calculation process is similar to step S21 above):

[0127] 1. Receive the source point cloud key point p and its z-axis vector z(p) from step S21. (With the source point cloud key point p as the center) determine the spherical neighborhood of the second radius r2, and obtain all the neighborhood points in the spherical neighborhood to form a local point set Q; the second radius r2 is set to 15 times the average distance mr between the nearest neighbor pairs.

[0128] 2. For the i-th neighborhood point q in the local point set Q i Calculate the i-th neighboring point q i The Euclidean distance between the i-th point cloud keypoint p and the source point cloud keypoint p is calculated, and then the i-th neighboring point q is calculated according to formula (4). i radial weight w i ;

[0129] w i = r2 -‖pq i ‖ (4);

[0130] In formula (4), r2 is the second radius; p represents the spatial coordinates of the key point p in the source point cloud (expressed as a vector); q i Let q represent the i-th neighborhood point of the target point cloud. i Spatial coordinates (expressed as vectors); ‖·‖ represents finding the magnitude of the vector.

[0131] 3. Then, construct the weighted covariance matrix Cov(Q) according to the above formula (2);

[0132] 4. Perform eigenvalue decomposition on the weighted covariance matrix Cov(Q) to obtain three eigenvalues ​​and three corresponding eigenvectors. Take the eigenvector corresponding to the largest eigenvalue as the initial x-axis direction v. max ;

[0133] 5. Redirect the initial x-axis direction according to formula (5) to obtain the final x-axis vector x1(p) of the first strategy, ensuring that the x-axis points to the main direction of the distribution of neighborhood points, and output the x-axis vector x1(p) to step S23.

[0134] (5)

[0135] The initial x-axis direction is re-signed to obtain the final x-axis vector x1(p) of the first strategy; the specific judgment condition is as follows: Calculate all neighborhood points q in the local point set Q pointed to by the key point p. i The sum of vectors (i.e.) ), and the initial x-axis direction v max Perform a dot product operation with the sum of the vectors; if the dot product result is greater than or equal to zero, keep the initial direction unchanged, that is, the final x-axis vector x1(p) = v max ;

[0136] If the dot product is less than zero, then the initial x-axis direction is reversed, i.e., the final x-axis vector x1(p) = -v max This eliminates directional ambiguity, thereby ensuring that the final x-axis points in the main direction of the neighborhood point distribution, and outputs the x-axis vector x1(p) to step S23.

[0137] The second strategy includes the following steps:

[0138] 1. Receive the source point cloud key point p and its z-axis vector z(p) from step S21. (With the source point cloud key point p as the center) Determine the spherical neighborhood of the third radius r3 and obtain all the neighborhood points in the spherical neighborhood to form a local point set Q. The third radius r3 is set to 15 times the average distance mr between the nearest neighbor pairs.

[0139] 2. For the i-th neighborhood point q in the local point set Q i Calculate its projected height h i See formula (6) below;

[0140] (6)

[0141] In formula (6), "·" represents the dot product of vectors. Based on (6), all neighboring points q are traversed. i Calculate all neighborhood points q using formula (6) i Projection height h i .

[0142] 3. At all neighboring points q i Projection height h i Select projection height h i The maximum value h max Then, set ε as a preset threshold, and apply it to all conditions that satisfy |h. i -h max The neighborhood point q of | < ε i Add to the candidate salient point set. The tolerance ε is set to 0.05 times the average distance mr between nearest neighbor pairs;

[0143] 4. For each candidate point q in the candidate salient point set c Calculate candidate point q c The vector v relative to the keypoint p in the source point cloud c , where v c = h i -h max According to formula (7), the vector v c Projecting onto a plane orthogonal to the z-axis yields the projection vector u. c Then, for the projection vector u c Normalization is performed to obtain candidate points q c The unit projection vector.

[0144] (7);

[0145] In formula (7), "·" represents the dot product of vectors.

[0146] 5. Draw a unit circle with the key point p as the origin. In the plane orthogonal to the final z-axis vector z(p), divide the 360-degree azimuth range into A directional intervals. Each interval covers an angle range of 360 / A degrees, and construct a directional histogram containing A counting units. The number of directional intervals A is set to 36.

[0147] For each candidate point q c Candidate point q is determined based on the azimuth angle of its unit projection vector. c The weight value w is accumulated within the specified direction interval. c Calculate according to formula (8);

[0148] (8);

[0149] In formula (8), h c For the candidate point q c The projected height, σ h This is a preset height difference attenuation coefficient. The preset height difference attenuation coefficient σ h Set to 0.1 times the first radius r1. ‖·‖ represents finding the magnitude of the vector.

[0150] 6. Perform Gaussian smoothing on the orientation histogram, search for all local peaks in the smoothed orientation histogram, determine the center direction of the interval containing the highest peak as the main x-axis direction x2(p), and determine the center directions of the intervals containing all k secondary peaks that are not less than 80% of the main peak as candidate x-axis directions x2(p). k (p), which generates one or more candidate x-axis directions x by the second strategy k(p) Output to step S23, where the number of candidates in the x-axis direction is k; the Gaussian smoothing process uses a Gaussian kernel with a standard deviation of 1 to convolve the histogram; the secondary peak value that is not less than 80% of the main peak value refers to the cumulative weight value of the interval where the secondary peak value is located being greater than or equal to 0.8 times the cumulative weight value of the interval of the highest peak value.

[0151] S23: Based on the final z-axis vector z(p) obtained in step 21, the final x-axis vector x1(p) obtained in step 22, the main x-axis direction x2(p), and K1 candidate x-axis directions x k (p) (where K1 is the number of candidate x-axis in the source point cloud), to obtain the y-axis vector y j (p);

[0152] Receive the z-axis vector z(p) from step S21, and receive the x-axis vector x1(p) generated by the first strategy and one or more x-axis vectors x2(p) and x from the second strategy from step S32. k (p), where k=1, 2, ... K1; for each received x-axis vector, generate the corresponding y-axis vector y using the vector cross product operation according to the right-hand rule. j (p), see formula (9) below;

[0153] (9);

[0154] In formula (9), × represents the cross product of three-dimensional vectors, x j (p) represents any of the above-mentioned x-axis vectors received; where j = 1, 2, ... (2 + K1); and the candidate x-axis direction x k The number of (p) is K1.

[0155] S24: Combining the above vectors, obtain U types of local reference frames (LRFs) for the source point cloud keypoint p. u (p) Since it includes 1 final x-axis, 1 primary x-axis and K1 candidate x-axis, the total number of local reference frames is U = K1 + 2, and the index u = 1, 2, ..., U;

[0156] The triplet consisting of z(p), x1(p) and its corresponding y1(p) is used as the first type of local reference frame LRF1(p);

[0157] The triplet consisting of z(p), x2(p) and its corresponding y2(p) is used as the second local reference frame LRF2(p);

[0158] This will be composed of z(p), each x k (p) and its corresponding y k(p) forms a triplet as the second + K1 type of local reference frame (LRF). u (p); where k = 1, 2, ... K1;

[0159] For the case where multiple x-axis vectors are generated by the second strategy, each x-axis vector x k (p) Each of these corresponds to the generation of an independent local reference frame (LRF). k (p), where k=1, 2, ... K1; these local reference frames are used in subsequent step S3 to generate independent candidate transformation hypotheses, so that the optimal local reference frame is automatically selected by comparing the number of interior points in subsequent verification step S4, thus eliminating the influence of x-axis ambiguity.

[0160] S25: Similarly, obtain V local reference frames (LRFs) for the keypoints q in the target point cloud. v (q) Since it contains 1 final x-axis, 1 main x-axis and K2 candidate x-axis (where K2 is the number of candidate x-axis in the target point cloud), the total number of local reference frames is V = K2 + 2, and the index v = 1, 2, ..., V.

[0161] Perform the above local reference frame construction operation on the source point cloud keypoint p and the target point cloud keypoint q in the initial corresponding point pair to obtain multiple local reference frames (LRFs) of the source point cloud keypoints. u (p) and multiple local reference frames (LRFs) for key points in the target point cloud. v (q), where u and v both represent the index numbers of the local reference frames. These local reference frames are combined and output to step S3.

[0162] In specific implementation, step S3 includes the following steps:

[0163] S31: Receive U types of local reference frames (LRFs) from the source point cloud keypoint p in step S24. u (p), receive V local reference frames (LRFs) of the target point cloud keypoints q from step S25. v (q); Set the number of matches N = min(U, V), extract the first N local reference frames of the source point cloud and the target point cloud respectively, and match them one by one;

[0164] Where u = 1, 2, ..., U, U = K1 + 2; v = 1, 2, ..., V, V = K2 + 2. U and V represent the total number of local reference frames of the two types; for each index u or v, the LRF is... u (p) and LRF v (q) is represented as a 3×3 matrix, where each column of the matrix corresponds to the direction vector of the x-axis, y-axis and z-axis, respectively;

[0165] S32: Calculate R according to the rotation matrix formula. n =(LRF n (q)) T ·LRF n (p) Calculate the rotation matrix R corresponding to the nth local reference frame. n , where n=1,2,...,N, T represents matrix transpose, and · represents matrix multiplication;

[0166] For each corresponding point pair, the rotation matrix R among the N candidate transformation hypotheses is calculated. n All are orthogonal matrices, satisfying R n T R n =I, where I is the identity matrix; n=1, 2, …N, where N=min(U, V), that is, N takes the smaller value between U and V.

[0167] S33: Calculate t according to the translation vector formula. n = q - R n ·p calculates the translation vector corresponding to the nth local reference frame, where q and p are the three-dimensional coordinate vectors of the key points in the target point cloud and the key points in the source point cloud, respectively;

[0168] The translation vector t n The calculation results ensure that the source point cloud key point p is strictly aligned with the target point cloud key point q in spatial position after transformation.

[0169] S34: will be determined by the rotation matrix R n Translation vector t n The constructed transformation parameter pair is used as the nth candidate transformation hypothesis. The above operation is performed on all N local reference frames to obtain N candidate transformation hypotheses. These N candidate transformation hypotheses are output as the transformation hypothesis set generated for the corresponding point pair to step S4.

[0170] In specific implementation, step S4 includes the following steps:

[0171] S41: Receive N candidate transformation hypotheses generated from the current corresponding point pair from step S3. Each candidate transformation hypothesis contains a rotation matrix R. n and a translation vector t n ;

[0172] For each candidate transformation hypothesis, apply the candidate transformation hypothesis to all points in the source point cloud, according to the formula p′= R. n ·p +t n Calculate the transformed coordinates of each point in the source point cloud;

[0173] S42: For each point in the transformed source point cloud, search for the point in the target point cloud with the closest Euclidean distance and record the closest distance value;

[0174] S43: The number of points in the source point cloud after statistical transformation whose distance to the nearest point in the target point cloud is less than a preset inlier distance threshold is used as the number of inlier points corresponding to the candidate transformation hypothesis.

[0175] The preset interior point distance threshold is set to twice the average distance mr between nearest neighbor pairs. For each candidate transformation hypothesis, a spatial index structure is used to accelerate the nearest neighbor search when counting the number of interior points; the spatial index structure is a kd-tree, and a kd-tree is pre-built on the target point cloud to quickly query the nearest neighbor of each transformed point.

[0176] S44: Compare the number of interior points corresponding to the N candidate transformation hypotheses, and determine the candidate transformation hypothesis with the largest number of interior points as the optimal transformation parameter for the current iteration;

[0177] S45: Record the optimal transformation parameters and their corresponding set of interior points, and output the corresponding point pairs in the set of interior points as the correct matching point pairs to step S5.

[0178] In specific implementation, step S5 includes the following steps:

[0179] Step S5 includes the following steps:

[0180] S51: Receive the optimal transformation parameters of the current iteration and its corresponding set of interior points from step S4, and use the set of interior points as the initial set of corresponding point pairs for the next iteration;

[0181] S52: Determine whether the current iteration count has reached the preset maximum iteration count, or whether the number of inliers corresponding to the current optimal transformation parameters has reached the preset inlier count threshold; the preset maximum iteration count is set to 100 times, and the preset inlier count threshold is set to 30% of the total number of points in the source point cloud;

[0182] If the maximum number of iterations is not reached and the number of inliers does not reach the threshold, return to step S2 and reconstruct the multiple local reference frame and generate transformation hypotheses using the current set of inliers. When returning to step S2, the set of inliers used is used as the new initial set of corresponding point pairs. The number of corresponding point pairs in this set is usually much smaller than the number of initial corresponding point pairs, thereby reducing the computational load of subsequent iterations and improving the registration accuracy.

[0183] If the maximum number of iterations or the number of interior points has been reached, then the rotation matrix in the current optimal transformation parameters will be adjusted. Translation vector Output as the final registration result;

[0184] S53: Apply the final registration result to the source point cloud, according to the formula... Calculate the final coordinates of each point in the source point cloud after transformation, and complete the accurate registration between the source point cloud and the target point cloud.

[0185] This invention has the following technical features:

[0186] 1. This invention proposes a noise- and occlusion-resistant point cloud registration method using multiple local reference frames, comprising six steps: key point extraction and feature description, initial corresponding point pair acquisition, construction of multiple local reference frames, generation of multiple transformation hypotheses, verification of transformation hypotheses and selection of optimal solutions, and iterative optimization and registration result output. By constructing multiple local reference frames, multiple candidate transformation hypotheses are generated for each corresponding point pair, and the optimal hypothesis is selected based on maximizing the number of interior points. The interior points are iteratively fed back to the next processing step, and finally, a high-precision registration result is output. This method overcomes the shortcomings of a single local reference frame in taking into account multiple interferences such as noise, occlusion, and low overlap rate, and significantly improves the adaptability and robustness of the registration algorithm in complex scenes.

[0187] 2. This invention proposes a z-axis construction method based on a radially weighted covariance matrix. Within a first-radius neighborhood centered on a keypoint, neighborhood points are assigned radial weights that decrease with increasing distance to construct a weighted covariance matrix. The eigenvector corresponding to its minimum eigenvalue is used as the initial z-axis direction. Then, a sign reversal method based on the point mean is used to eliminate directional ambiguity, resulting in a stable z-axis. This method effectively reduces the interference of distant points on z-axis calculation, enhances the local reference frame's resistance to noise, occlusion, and incomplete edges, and ensures the repeatability of the z-axis under different viewpoints.

[0188] 3. This invention proposes an x-axis construction method based on weighted voting using a directional histogram. When multiple candidate points exist for the projection extremum along the z-axis, all points with projection heights close to the extremum are included in the candidate salient point set. The orthogonal projection vectors of each candidate point are calculated, a directional histogram is constructed, and weights considering the proximity of projection heights and distances are accumulated. After smoothing the histogram, all peak intervals are extracted. The center direction of the highest peak interval is taken as the main x-axis direction, and the center direction of the secondary peak interval, which is not less than 80% of the main peak, is taken as the candidate x-axis direction. This method can avoid the x-axis ambiguity caused by randomly selecting a single extremum point under complex geometric structures such as symmetrical planes, regular edges, or flat regions. It significantly improves the stability and repeatability of the local reference frame and lays the foundation for the reliability of subsequent transformation hypotheses.

[0189] 4. This invention proposes a registration optimization mechanism for generating and iteratively verifying multiple transformation hypotheses. It utilizes multiple constructed local reference frames to simultaneously generate multiple candidate rotation matrices and translation vectors for each initial corresponding point pair, forming multiple transformation hypotheses. During the verification phase, each hypothesis is applied to transform the source point cloud to the target point cloud coordinate system. The number of point pairs satisfying the interior point condition is counted, and the hypothesis with the largest number of interior points is used as the optimal transformation parameter for the current iteration. The set of interior points corresponding to this hypothesis is used as the initial corresponding point pair for the next iteration. Through repeated iterations until a preset stopping condition is met, the optimal transformation parameter is finally output. This method effectively integrates the advantages of multiple local reference frames, automatically filters the optimal hypothesis through the objective criterion of maximizing the number of interior points, and continuously refines the corresponding point pairs using iterative feedback, significantly improving the accuracy and robustness of the registration algorithm on low-quality, highly interfering point cloud data.

[0190] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0191] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A noise- and occlusion-resistant multi-LRF point cloud registration method, characterized in that, Includes the following steps: S1: Select key points from the source point cloud and the target point cloud respectively, and obtain an initial set of corresponding point pairs based on the key points; Step S1 includes the step of obtaining the initial set of corresponding point pairs; The process of obtaining the initial set of corresponding point pairs includes the following steps: receiving local feature descriptors; calculating the feature distance between local feature descriptors; selecting candidate corresponding points; gathering each key point in the source point cloud and its candidate corresponding points into an initial set of corresponding point pairs; simplifying the initial set of corresponding point pairs to obtain a simplified initial set of corresponding point pairs; S2: Multiple local reference frame construction steps; For each point pair in the initial set of corresponding point pairs, multiple local reference frames are constructed with the source point cloud keypoint p and the target point cloud keypoint q as the center for each point pair; S3: Multiple transformation hypothesis generation step; Based on the multiple local reference frames obtained in step 2, obtain the rotation matrix R. n Translation vector t n This process yields N candidate transformation hypotheses; these N candidate transformation hypotheses are then output as the transformation hypothesis set generated for the corresponding point pair to the next step. S4: Hypothesis verification and optimal solution selection steps; based on the rotation matrix R of each of the N candidate transformation hypotheses in the transformation hypothesis set. n and a translation vector t n This allows us to obtain the optimal transformation parameters and the correct matching point pairs for the current iteration. S5: Iterative optimization and registration result output steps; determine whether the maximum number of iterations or the number of interior points has reached the threshold. If so, add the rotation matrix R in the current optimal transformation parameters. * Translation vector t * This will be output as the final registration result.

2. The noise- and occlusion-resistant multi-LRF point cloud registration method according to claim 1, characterized in that, Step S1 includes key point extraction and feature description steps.

3. The noise- and occlusion-resistant multi-LRF point cloud registration method according to claim 2, characterized in that, The key point extraction and feature description process includes the following steps: S111: Obtain the spatial coordinates of a 3D point; S112: Select key points based on the spatial coordinates; S113: Calculate the 3D coordinates of the neighboring points of the key point; for each selected key point, determine a spherical neighborhood centered on the key point with a preset descriptor radius as the length, and obtain the 3D coordinates of all neighboring points within the spherical neighborhood; S114: Determine the direction of the local reference axis; S115: Divide the spatial sub-regions; using the local reference axis direction as the north pole direction of the spherical neighborhood and as the dividing benchmark, divide the spherical neighborhood along the spatial direction to form multiple spatial sub-regions; S116: Obtain the local feature descriptor of the key point.

4. The noise- and occlusion-resistant multi-LRF point cloud registration method according to claim 3, characterized in that, In step S114, when determining the direction of the local reference axis, a sign redirection method is used to adjust the direction of the local reference axis to point to the main side of the distribution of neighboring points.

5. The noise- and occlusion-resistant multi-LRF point cloud registration method according to claim 1, characterized in that, The process of simplifying the initial set of corresponding points to obtain a simplified initial set of corresponding points also includes a step of performing a geometric consistency check on the simplified initial set of corresponding points to obtain a further simplified initial set of corresponding points.

6. The noise- and occlusion-resistant multi-LRF point cloud registration method according to claim 1, characterized in that, Step S2 includes the following steps: S21: Determine the final z-axis vector z(p) of the source point cloud keypoint p of the multiple local reference frame; S22: According to the final The x-axis vector z(p) is used to construct the x-axis vector of the multiple local reference frames. Two strategies are employed to obtain the final x-axis vector x1(p) of the first strategy, the main x-axis direction x2(p) of the second strategy, and one or more candidate x-axis directions. k (p); S23: Based on the final z-axis vector z(p) obtained in step 21, the final x-axis vector x1(p) obtained in step 22, the main x-axis direction x2(p), and K1 candidate x-axis directions x k (p), to obtain the y-axis vector y j (p); S24: Combining the final z-axis vector z(p), the final x-axis vector x1(p), the main x-axis direction x2(p), and the K1 candidate x-axis directions x k (p) and y-axis vector y j (p), obtaining U types of local reference frames (LRFs) for the source point cloud keypoint p. u (p); U = K1+ 2, index u=1,2,...,U; S25: Similarly, obtain V local reference frames (LRFs) for the keypoints q in the target point cloud. v (q); V = K2+2, index v=1,2,...,V.

7. The noise- and occlusion-resistant multi-LRF point cloud registration method according to claim 6, characterized in that, Step S3 includes the following steps: S31: Receive U types of local reference frames (LRFs) from the source point cloud keypoint p in step S24. u (p), receive V local reference frames (LRFs) of the target point cloud keypoints q from step S25. v (q); Set the number of matches N = min(U, V), extract the first N local reference frames of the source point cloud and the target point cloud respectively, and match them one by one; S32: Calculate R according to the rotation matrix formula. n =(LRF n (q)) T ·LRF n (p) Calculate the rotation matrix R corresponding to the nth local reference frame. n , where n=1,2,...,N, T represents matrix transpose, and · represents matrix multiplication; S33: Calculate t according to the translation vector formula. n = q - R n ·p calculates the translation vector corresponding to the nth local reference frame, where q and p are the three-dimensional coordinate vectors of the key points in the target point cloud and the key points in the source point cloud, respectively; S34: will be determined by the rotation matrix R n Translation vector t n The transformation parameters constitute the nth candidate transformation hypothesis, and the above "will be performed on all N local reference frames" is applied. n Translation vector t n The operation of "using the constructed transformation parameter pair as the nth candidate transformation hypothesis" yields N candidate transformation hypotheses. These N candidate transformation hypotheses are then output as the transformation hypothesis set generated for the corresponding point pair to step S4.

8. The noise- and occlusion-resistant multi-LRF point cloud registration method according to claim 1, characterized in that, Step S4 includes the following steps: S41: Receive N candidate transformation hypotheses generated from the current corresponding point pair from step S3. Each candidate transformation hypothesis contains a rotation matrix R. n and a translation vector t n ; S42: For each point in the transformed source point cloud, search for the point in the target point cloud with the closest Euclidean distance and record the closest distance value; S43: The number of points in the source point cloud after statistical transformation whose distance to the nearest point in the target point cloud is less than a preset inlier distance threshold is used as the number of inlier points corresponding to the candidate transformation hypothesis. S44: Compare the number of interior points corresponding to the N candidate transformation hypotheses, and determine the candidate transformation hypothesis with the largest number of interior points as the optimal transformation parameter for the current iteration; S45: Record the optimal transformation parameters and their corresponding set of interior points, and output the corresponding point pairs in the set of interior points as the correct matching point pairs to step S5.

9. The noise- and occlusion-resistant multi-LRF point cloud registration method according to claim 8, characterized in that, Step S5 includes the following steps: Step S5 includes the following steps: S51: Receive the optimal transformation parameters of the current iteration and its corresponding set of interior points from step S4, and use the set of interior points as the initial set of corresponding point pairs for the next iteration; S52: Determine whether the current iteration count has reached the preset maximum iteration count, or whether the number of inliers corresponding to the current optimal transformation parameter has reached the preset inlier count threshold; S53: Apply the final registration result to the source point cloud, according to formula p final = R * ·p + t * Calculate the final coordinates of each point in the source point cloud after transformation, and complete the accurate registration between the source point cloud and the target point cloud.

Images