Rock mass point cloud registration method based on convex-concave features
By extracting the surface features of rocks using improved R-NRLC and DBSCAN algorithms, constructing a feature descriptor, and combining it with the RANSAC algorithm for rock mass point cloud registration, the problems of unstable feature extraction and insufficient registration accuracy in complex scenes of existing methods are solved, and high-precision and robust rock mass point cloud registration is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA HYDROELECTRIC ENGINEERING CONSULTING GROUP CHENGDU RESEARCH HYDROELECTRIC INVESTIGATION DESIGN AND INSTITUTE
- Filing Date
- 2026-01-29
- Publication Date
- 2026-05-08
AI Technical Summary
Existing rock mass point cloud registration methods suffer from unstable feature extraction, weak descriptor discrimination ability, and insufficient registration accuracy and robustness in complex scenarios, making them difficult to adapt to rock mass point cloud data with low overlap, high deflection angle, and uneven density.
An improved R-NRLC algorithm is used to identify concave and convex feature points. The DBSCAN clustering algorithm is used to filter out noise points, and a weighted descriptor for concave and convex features is constructed. Point pairs are screened by Euclidean distance and cosine similarity, and erroneous matches are eliminated by combining the RANSAC algorithm. The rigid transformation matrix is calculated to achieve registration.
Under conditions of low overlap, uneven density, and noise interference, high-precision and robust rock mass point cloud registration was achieved, improving matching accuracy and adaptability.
Smart Images
Figure CN121999021A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision and image processing, specifically to a method for registering point clouds of rock masses based on convex and concave features. Background Technology
[0002] In fields such as rock mass structure analysis, geological hazard monitoring, engineering surveying, and digital twin reconstruction, high-precision 3D point cloud data is a crucial foundation for accurate analysis and modeling. Compared to traditional 2D images, 3D point clouds can more intuitively reflect the spatial geometric features and structural morphology of the rock mass surface. In recent years, the development of lidar and 3D reconstruction technologies has made high-precision scanning of rock surfaces possible. However, due to limitations such as complex terrain obstruction and observation perspective, a single scan is insufficient to fully capture the 3D structure of the rock mass. Registration and fusion of point cloud data from multiple perspectives and batches are necessary to construct a complete digital model of the rock mass.
[0003] However, rock mass point cloud registration faces many technical bottlenecks: the rock mass surface has a complex concave-convex structure and scarce natural textures, and is easily affected by environmental noise and uneven point cloud density during the scanning process, making it difficult for traditional registration methods to extract stable and reliable features; geometric registration methods based on iterative nearest points rely on high overlap rate and good initial alignment, and are prone to getting trapped in local optima in low overlap rate and strong noise scenarios; registration methods based on conventional feature descriptors lack the specific characterization of the concave-convex structure of the rock mass, and are not robust enough in distinguishing similar geometric shapes; while deep learning-based registration methods rely on large-scale labeled datasets and have high hardware computing power requirements, making them difficult to adapt to the complex application scenarios in engineering sites.
[0004] In existing technologies, some methods attempt to optimize registration results through feature point extraction and descriptor construction, but they generally suffer from problems such as poor stability of feature points and insufficient sensitivity of descriptors to concave and convex structures. In the registration of rock mass point clouds with low overlap rate, high deflection angle and uneven density, it is difficult to balance registration accuracy and robustness.
[0005] Therefore, there is an urgent need for a highly reliable rock mass point cloud registration technology that can accurately capture the concave and convex features of rock masses, effectively suppress noise interference, and adapt to complex scenarios, so as to overcome the application limitations of existing methods. Summary of the Invention
[0006] The technical problem to be solved by this invention is to provide a method for registering rock mass point clouds based on convex and concave features, which solves the problems of unstable feature extraction, weak descriptor discrimination ability, and insufficient registration accuracy and robustness of existing rock mass point cloud registration methods in complex scenarios.
[0007] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:
[0008] In a first aspect, the present invention provides a method for registration of rock mass point clouds based on convex and concave features, comprising the following steps:
[0009] S1. Input the rock mass point cloud data to be registered, including the source point cloud and the target point cloud;
[0010] S2. Calculate the geometric center of the point cloud, construct the kd-tree, and calculate the normal vector of the point cloud;
[0011] S3. By analyzing the point cloud normal vectors, identify the characteristic points of rock mass undulations and remove noise points using a clustering algorithm;
[0012] S4. Calculate the centroid of the clustering results and construct a local coordinate system based on the centroid;
[0013] S5. Construct a weighted descriptor for concavity and convexity features based on the local coordinate system, and incorporate concavity and convexity information during the descriptor generation process;
[0014] S6. Calculate the Euclidean distance between the concave and convex feature points of the two point clouds to be matched, filter point pairs that meet the threshold, and construct an initial set of matching point pairs;
[0015] S7. For the initial set of matching point pairs, remove incorrect matching point pairs and calculate the rigid transformation matrix;
[0016] S8. Apply the rigid transformation matrix to the source point cloud to achieve registration of the point clouds of the two rock masses.
[0017] Furthermore, in step S3, the point cloud normal vectors are analyzed using the R-NRLC algorithm to identify the concave and convex feature points of the rock mass:
[0018] Calculate the local neighborhood of each point center of mass The centroid is the current point. The coordinates and their eigenvectors sum;
[0019] eigenvectors along the normal vector of the current point Decomposed into vertical components and horizontal components For each neighboring point in each neighborhood of a point The vector formed Decomposed along the vertical and horizontal directions respectively, yielding two components. and ;
[0020] vertical component and Addition, horizontal components and Adding them together yields a new eigenvector, defined as follows: ;
[0021] according to With normal vector dot product results To distinguish the regional attributes of feature points: if the dot product result is positive, the point is located in a convex region; if it is negative, it is located in a concave region.
[0022] in accordance with With normal vector The sign of the dot product value will affect the local neighborhood. The neighborhood points within the local area are divided into two groups. The proportions of the two groups of points to the total number of points in the local neighborhood are counted and denoted as a and b.
[0023] The weights of the two sets of points are calculated using the sigmoid function; the sigmoid function is expressed as:
[0024] ;
[0025] in, To control the sensitivity of the noise suppression effect;
[0026] definition , ,Will and The weights of the neighboring points within the corresponding group are used to calculate the weighted average of the current point. The sum of the eigenvectors of all neighboring points ;
[0027] Set direction threshold and length threshold ,in, Used to control the orientation filtering of point cloud convex and concave regions. Used to filter for controlling the degree of variation in uneven areas;
[0028] calculate With normal vector absolute value of the dot product ,Will Greater than The points are marked as points of interest; then feature vectors are selected from the points of interest. Less than The points are used as the final convex and concave feature points.
[0029] Furthermore, in step S3, the clustering algorithm used is the DBSCA algorithm.
[0030] Furthermore, in step S4, the centroid of the clustering results is calculated, and a local coordinate system is constructed based on the centroid, including:
[0031] Calculate the geometric center of each point cloud cluster obtained through clustering, and use this geometric center as the stable centroid;
[0032] Principal component analysis is performed on each point cloud cluster to calculate the covariance matrix of the cluster, and three eigenvectors are obtained through eigenvalue decomposition. , , , respectively, correspond to the directions of the largest eigenvalue, middle eigenvalue, and smallest eigenvalue of the covariance matrix;
[0033] With the stable centroid as the origin, the eigenvector , , Construct a three-dimensional orthogonal local coordinate system by using the Z-axis, Y-axis, and X-axis directions as the local coordinate system, respectively.
[0034] Furthermore, in step S5, a weighted descriptor for concavity and convexity features is constructed based on the local coordinate system, including:
[0035] In the local coordinate system, several spherical sub-regions are divided with the target point as the center, and the neighboring points of the target point are projected onto the corresponding spherical sub-regions according to their spatial positions.
[0036] For each spherical sub-region, the directional distribution information is extracted based on the angle between its normal vector and the coordinate axis of the local coordinate system.
[0037] Define a concavity / convexity weighting factor, which is determined by the concavity / convexity index of a point. and distance weight It was jointly determined that, Indicates a convex point. Indicates a concave point;
[0038] In the process of interpolating directional information, the concavity and convexity weighting factor is incorporated into the directional distribution statistics, and the directional information in each spherical sub-region is weighted and accumulated to generate a weighted directional histogram of that sub-region.
[0039] The weighted orientation histograms of all spherical sub-regions are spliced together in a preset order to form a dot signature descriptor containing convex and concave feature information.
[0040] Furthermore, in step S6, the selection of point pairs that meet the threshold includes:
[0041] Calculate the Euclidean distance and cosine similarity between the weighted descriptors of the convex and concave feature points of the source point cloud and the weighted descriptors of the convex and concave feature points of the target point cloud;
[0042] Set a Euclidean distance threshold and a cosine similarity threshold, and filter point pairs that simultaneously satisfy the condition that the Euclidean distance is less than the Euclidean distance threshold and the cosine similarity is greater than the cosine similarity threshold;
[0043] The selected point pairs are sorted from high to low according to their cosine similarity, and point pairs that do not meet the geometric consistency are removed, thus forming the initial set of matching point pairs.
[0044] Furthermore, in step S7, for the initial set of matched point pairs, the RANSAC algorithm is used to remove erroneous matched point pairs, and the rigid transformation matrix is calculated, specifically including:
[0045] S71. Randomly select non-coplanar matching point pairs from the initial matching point pair set, and calculate the provisional rigid transformation matrix based on the SVD algorithm;
[0046] S72. Apply the provisional rigid transformation matrix to the feature points of the source point cloud and calculate the Euclidean distance error between the transformed source point and the corresponding target point;
[0047] S73. Mark the point pairs whose Euclidean distance error is less than the set distance error threshold as interior points, and count the number of interior points;
[0048] S74. Repeat steps S71-S73. After setting the number of iterations, select the provisional rigid transformation matrix obtained in the iteration with the largest number of interior points as the estimated optimal rigid transformation matrix.
[0049] Furthermore, in step S8, applying the rigid transformation matrix to the source point cloud includes:
[0050] ;
[0051] in, The coordinates of the registered source point cloud; The coordinates of the original source point cloud; These are the rotation parameters in the rigid transformation matrix; represents the translation parameter in the rigid transformation matrix.
[0052] Secondly, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method for registration of rock mass point clouds based on convex and concave features.
[0053] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for registration of rock mass point clouds based on convex and concave features.
[0054] The beneficial effects of this invention are:
[0055] This invention first utilizes an improved R-NRLC algorithm to extract concave and convex feature points on the surface of the rock mass, ensuring the stability of key points under noise and density heterogeneity conditions. Then, a local coordinate system is established centered on each concave and convex feature point, and its neighborhood directional point signature histogram descriptor is calculated. Concave and convexity weights are introduced during the interpolation voting process to enhance the sensitivity and stability of the descriptor to geometric differences. Next, the matching confidence is comprehensively evaluated based on the Euclidean distance and cosine similarity of the descriptor vectors, and initial corresponding point pairs are selected. Finally, the RANSAC algorithm is used to eliminate erroneous matches and estimate the rigid body transformation matrix. The calculated transformation is then applied to the source point cloud data to achieve high-precision registration. This invention can achieve robust rock mass point cloud registration under conditions of low overlap, density heterogeneity, and noise interference, exhibiting high matching accuracy and adaptability. Attached Figure Description
[0056] Figure 1 This is a flowchart of the method for registering rock mass point clouds based on convex and concave features in an embodiment of the present invention. Detailed Implementation
[0057] This invention aims to provide a method for registering rock mass point clouds based on convex and concave features, addressing the problems of unstable feature extraction, weak descriptor discriminative ability, and insufficient registration accuracy and robustness in complex scenes using existing rock mass point cloud registration methods. The core idea is to extract convex and concave feature points on the rock surface using an improved R-NRLC method, and construct feature descriptors based on Euclidean distance and convex / concave information to accurately characterize the geometric features of the rock surface. An optimized clustering mechanism is introduced to effectively suppress local noise interference and enhance the stability of feature matching. Furthermore, by calculating the edit distance between descriptors, corresponding point pairs are accurately found, and the registration matrix is calculated using congruent triangle transformation.
[0058] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0059] This embodiment provides a method for rock mass point cloud registration based on convex and concave features. See [link to relevant documentation]. Figure 1 It includes the following implementation process:
[0060] S1. Input the rock mass point cloud data to be registered.
[0061] In this step, the rock mass point cloud data to be registered is input, including the source point cloud and the target point cloud. The purpose is to clarify the input object for registration and provide basic data for subsequent processing.
[0062] Specifically, the two rock mass point cloud data obtained by input lidar scanning are defined as the source point cloud (point cloud to be transformed) and the target point cloud (reference point cloud), respectively. The point cloud data contains the three-dimensional coordinates (X, Y, Z) information of each point. The two point clouds have some overlapping areas, and the initial poses have translational or rotational differences. The point cloud data contains discrete noise points caused by scanning noise, environmental interference, etc.
[0063] S2. Calculate the normal vector of the point cloud;
[0064] In this step, the geometric center of the point cloud is calculated, a kd-tree is constructed, and the normal vectors of the point cloud are calculated. The calculation of the geometric center of the point cloud provides a reference for subsequent local neighborhood analysis and cluster centroid calculation; the construction of the kd-tree solves the problem of low efficiency in high-dimensional point cloud nearest neighbor search, and provides efficient data index support for subsequent normal vector calculation and feature extraction; the normal vector is an important parameter representing the geometric pose of the point cloud surface, and can provide a foundation for subsequent convex and concave feature recognition and local coordinate system construction.
[0065] Specifically, the geometric center of a point cloud is the mean of the coordinates of all points in the cloud. To calculate the point cloud normal vector, a kd-tree is constructed using the 3D coordinates of the point cloud, and k-nearest neighbor search is used to quickly locate the local neighborhood point set of each point. By performing principal component analysis on the local neighborhood point set of each point, the covariance matrix of the neighborhood point set is calculated. Then, by performing eigenvalue decomposition on the covariance matrix, the characteristic vector corresponding to the smallest eigenvalue is taken as the normal vector of that point.
[0066] S3. Identify convex and concave feature points
[0067] In this step, the characteristic points of the rock mass's unevenness are identified by analyzing the point cloud normal vectors, and noise points are filtered out using a clustering algorithm. The aim is to accurately identify the characteristic points of the rock mass's unevenness, suppress noise interference, and retain the set of valid feature points with geometric saliency.
[0068] In this embodiment, the improved R-NRLC algorithm is first used to analyze the concavity and convexity features of the point cloud surface, extracting candidate feature points with geometric saliency. Then, the DBSCAN clustering algorithm is used to suppress noise and group regions from these initial feature points, obtaining a stable and reliable set of convex and concave feature points, specifically including:
[0069] S31. Preliminary identification of concave and convex feature points based on the R-NRLC algorithm
[0070] Calculate the local neighborhood of each point center of mass The centroid is the current point. The coordinates and their eigenvectors sum;
[0071] eigenvectors along the normal vector of the current point Decomposed into vertical components and horizontal components For each neighboring point in each neighborhood of a point The vector formed Decomposed along the vertical and horizontal directions respectively, yielding two components. and ;
[0072] vertical component and Addition, horizontal components and Adding them together yields a new eigenvector, defined as follows: ;
[0073] according to With normal vector dot product results To distinguish the regional attributes of feature points: if the dot product result is positive, the point is located in a convex region; if it is negative, it is located in a concave region.
[0074] In this sub-step, the geometric relationship of the local neighborhood is transformed into the directional relationship between the feature vector and the normal vector through vector decomposition and accumulation, thereby realizing the distinction of convex and concave region attributes.
[0075] S32. Noise Suppression and Point of Interest Filtering
[0076] in accordance with With normal vector The sign of the dot product value will affect the local neighborhood. The neighborhood points within the local area are divided into two groups. The proportions of the two groups of points to the total number of points in the local neighborhood are counted and denoted as a and b.
[0077] The weights of the two sets of points are calculated using the sigmoid function; the sigmoid function is expressed as:
[0078] ;
[0079] in, To control the sensitivity of the noise suppression effect;
[0080] definition , ,Will and The weights of the neighboring points within the corresponding group are used to calculate the weighted average of the current point. The sum of the eigenvectors of all neighboring points ;
[0081] Set direction threshold and length threshold ,in, Used to control the orientation filtering of point cloud convex and concave regions. Used to filter for controlling the degree of variation in uneven areas;
[0082] calculate With normal vector absolute value of the dot product ,Will Greater than The points are marked as points of interest; then feature vectors are selected from the points of interest. Less than The points are used as the final convex and concave feature points.
[0083] This sub-step assigns low weights to a few outlier groups in the neighborhood using the sigmoid function, which reduces their interference with feature vector calculation. At the same time, it uses regional attributes and dual thresholds to ensure that the points of interest not only conform to the directional characteristics of the concave and convex regions, but also have a reasonable degree of undulation, thus eliminating invalid points in the boundary regions and outlier points with excessive undulation.
[0084] S33. DBSCAN clustering algorithm to filter out noise points
[0085] Set the DBSCAN algorithm parameters: neighborhood radius and minimum number of points threshold.
[0086] Iterate through all points of interest and count the number of neighboring points within the neighborhood radius of each point: if the number is greater than or equal to the minimum number of points threshold, then mark it as a core point.
[0087] Point cloud clusters are formed by the core point and its density-reachable points. Isolated points (noise points) that do not constitute point cloud clusters are filtered out, and high-density point cloud clusters containing effective concavity and convexity information are retained.
[0088] This sub-step utilizes the density clustering properties of DBSCAN to accurately preserve continuous clusters of concave and convex structural points, further enhancing the stability of feature points and providing a reliable point set foundation for subsequent local coordinate system construction.
[0089] S4. Construct a local coordinate system
[0090] In this step, the centroid is calculated based on the aforementioned clustering results, and a local coordinate system is constructed based on the centroid. The purpose is to transform the geometric features of each point into relative relationships under local coordinates by constructing a local coordinate system based on the principal direction of the point cluster, thereby eliminating the influence of global rotation on feature description and achieving unified expression and alignment of features.
[0091] Specifically, in this embodiment, for each high-density point cloud cluster, the geometric center is calculated and used as the stable centroid.
[0092] For each point cluster, perform PCA analysis to calculate the covariance matrix of the point cluster, and obtain three eigenvectors through eigenvalue decomposition. (Maximum eigenvalue) (Middle eigenvalue) (Minimum eigenvalue).
[0093] With the stable center of mass as the origin, , , A three-dimensional orthogonal local coordinate system is constructed by using the Z-axis, Y-axis, and X-axis as the directions of the local coordinate system.
[0094] S5. Construct a concave / convex feature descriptor based on the local coordinate system.
[0095] In this step, a weighted descriptor for concavity and convexity features is constructed based on a local coordinate system, incorporating concavity and convexity information during the descriptor generation process. The core function of the descriptor is to transform three-dimensional geometric features into quantifiable vectors for similarity calculation. This step uses sub-region division, direction statistics, and concavity / convexity weighting to ensure that the descriptor contains both local spatial structure information and concavity / convexity attribute differences, thereby enhancing the ability to distinguish complex rock surfaces.
[0096] Specifically, in this embodiment, the process of constructing the convex-concave feature descriptor based on the local coordinate system is as follows:
[0097] S51. In the local coordinate system, construct a spherical neighborhood with the target point as the center, and divide the neighborhood into several spherical sub-regions. Project the neighborhood points of the target point onto the corresponding spherical sub-regions according to their spatial positions.
[0098] S52. For each spherical sub-region, extract the direction distribution information based on the angle between its normal vector and the coordinate axis of the local coordinate system.
[0099] S53. Define a concavity / convexity weighting factor, wherein the concavity / convexity weighting factor is determined by the concavity / convexity index of a point. and distance weight It was jointly determined that, Indicates a convex point. Indicates a concave point;
[0100] S54. In the process of interpolating directional information, the concavity and convexity weighting factor is incorporated into the directional distribution statistics, and the directional information in each spherical sub-region is weighted and accumulated to generate a weighted directional histogram of the sub-region.
[0101] S55. The weighted orientation histograms of all spherical sub-regions are spliced together in a preset order to form a dot signature descriptor containing convex and concave feature information.
[0102] S6. Filter out initial matching point pairs
[0103] In this step, the Euclidean distance between the concave and convex feature points of the two point clouds to be matched is calculated, and point pairs that meet the threshold are selected to construct an initial set of matched point pairs. The purpose is to measure the similarity of descriptors from two different dimensions, namely Euclidean distance and cosine similarity, so as to select initial matched point pairs and provide a high-quality candidate point pair set for subsequent robust estimation.
[0104] Specifically, this embodiment calculates the Euclidean distance and cosine similarity between the convex and concave feature descriptors of the source point cloud and the target point cloud, and sets a threshold for Euclidean distance and a threshold for cosine similarity. Point pairs that simultaneously satisfy the condition that the Euclidean distance is less than the Euclidean distance threshold and the cosine similarity is greater than the cosine similarity threshold are then selected. The selected point pairs are sorted from high to low according to their cosine similarity, and point pairs that do not satisfy geometric consistency are removed, thus forming an initial set of matched point pairs.
[0105] S7. Eliminate incorrect matching pairs and calculate the rigid transformation matrix.
[0106] In this step, for the initial set of matched point pairs, erroneous matched point pairs are removed, and a rigid transformation matrix is calculated. Since a certain proportion of erroneous matches still exist in the initial set of matched point pairs, directly using them to calculate the transformation matrix would lead to registration bias. Therefore, the RANSAC algorithm is used, employing a mechanism of random sampling, verification, and iteration. It estimates the matrix using a small number of correctly matched points (interior points), and then verifies the matrix's validity by checking the number of interior points, effectively eliminating the interference of erroneous matched point pairs. Finally, the matrix is optimized using the interior point set to further improve the accuracy of the transformation matrix, providing core parameters for the final registration.
[0107] Specifically, in this embodiment, the process of eliminating incorrect matching point pairs and calculating the rigid transformation matrix is as follows:
[0108] S71. Randomly select non-coplanar matching point pairs from the initial matching point pair set, and calculate the provisional rigid transformation matrix based on the SVD algorithm;
[0109] S72. Apply the provisional rigid transformation matrix to the feature points of the source point cloud and calculate the Euclidean distance error between the transformed source point and the corresponding target point;
[0110] S73. Mark the point pairs whose Euclidean distance error is less than the set distance error threshold as interior points, and count the number of interior points;
[0111] S74. Repeat steps S71-S73. After setting the number of iterations, select the provisional rigid transformation matrix obtained in the iteration with the largest number of interior points as the estimated optimal rigid transformation matrix.
[0112] S8. Registration is achieved using a rigid transformation matrix.
[0113] In this step, the rigid transformation matrix is applied to the source point cloud to achieve registration of the two rock mass point clouds. The purpose is to use the rigid transformation matrix to unify the coordinate system of the source point cloud to the coordinate system of the target point cloud, eliminating initial pose differences and ensuring accurate matching of the overlapping areas of the two point clouds to meet subsequent application requirements.
[0114] Specifically, in this embodiment, the rigid transformation matrix obtained above is applied to all points in the source point cloud, that is, a transformation is performed on each point in the source point cloud:
[0115] ;
[0116] in, The coordinates of the registered source point cloud; The coordinates of the original source point cloud; These are the rotation parameters in the rigid transformation matrix; represents the translation parameter in the rigid transformation matrix.
[0117] Although embodiments of the present invention have been described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present invention, and all such changes and alterations shall not depart from the protection scope of the present invention.
Claims
1. A method for registering point clouds of rock masses based on convex and concave features, characterized in that, Includes the following steps: S1. Input the rock mass point cloud data to be registered, including the source point cloud and the target point cloud; S2. Calculate the geometric center of the point cloud, construct the kd-tree, and calculate the normal vector of the point cloud; S3. By analyzing the point cloud normal vectors, identify the characteristic points of rock mass undulations and remove noise points using a clustering algorithm; S4. Calculate the centroid of the clustering results and construct a local coordinate system based on the centroid; S5. Construct a weighted descriptor for concavity and convexity features based on the local coordinate system, and incorporate concavity and convexity information during the descriptor generation process; S6. Calculate the Euclidean distance between the concave and convex feature points of the two point clouds to be matched, filter point pairs that meet the threshold, and construct an initial set of matching point pairs; S7. For the initial set of matching point pairs, remove incorrect matching point pairs and calculate the rigid transformation matrix; S8. Apply the rigid transformation matrix to the source point cloud to achieve registration of the point clouds of the two rock masses.
2. The method for registering rock mass point clouds based on convex and concave features as described in claim 1, characterized in that, In step S3, the point cloud normal vectors are analyzed using the R-NRLC algorithm to identify the concave and convex feature points of the rock mass: Calculate the local neighborhood of each point center of mass The centroid is the current point. The coordinates and their eigenvectors sum; eigenvectors along the normal vector of the current point Decomposed into vertical components and horizontal components ; For each neighboring point in each neighborhood of a point The vector formed Decomposed along the vertical and horizontal directions respectively, yielding two components. and ; vertical component and Addition, horizontal components and Adding them together yields a new eigenvector, defined as follows: ; according to With normal vector dot product results To distinguish the regional attributes of feature points: if the dot product result is positive, the point is located in a convex region; if it is negative, it is located in a concave region. in accordance with With normal vector The sign of the dot product value will affect the local neighborhood. The neighborhood points within the local area are divided into two groups. The proportions of the two groups of points to the total number of points in the local neighborhood are counted and denoted as a and b. The weights of the two sets of points are calculated using the sigmoid function; the sigmoid function is expressed as: ； in, To control the sensitivity of the noise suppression effect; definition , ,Will and The weights of the neighboring points within the corresponding group are used to calculate the weighted average of the current point. The sum of the eigenvectors of all neighboring points ; Set direction threshold and length threshold ,in, Used to control the orientation filtering of point cloud convex and concave regions. Used to filter for controlling the degree of variation in uneven areas; calculate With normal vector absolute value of the dot product ,Will Greater than The points are marked as points of interest; then feature vectors are selected from the points of interest. Less than The points are used as the final convex and concave feature points.
3. The method for registering rock mass point clouds based on convex and concave features as described in claim 1, characterized in that, In step S3, the clustering algorithm used is the DBSCA algorithm.
4. The method for registering rock mass point clouds based on convex and concave features as described in claim 1, characterized in that, In step S4, the centroid of the clustering results is calculated, and a local coordinate system is constructed based on the centroid, including: Calculate the geometric center of each point cloud cluster obtained through clustering, and use this geometric center as the stable centroid; Principal component analysis is performed on each point cloud cluster to calculate the covariance matrix of the cluster, and three eigenvectors are obtained through eigenvalue decomposition. , , , respectively, correspond to the directions of the largest eigenvalue, middle eigenvalue, and smallest eigenvalue of the covariance matrix; With the stable centroid as the origin, the eigenvector , , Construct a three-dimensional orthogonal local coordinate system by using the Z-axis, Y-axis, and X-axis directions as the local coordinate system, respectively.
5. The method for registering rock mass point clouds based on convex and concave features as described in claim 1, characterized in that, In step S5, a weighted descriptor for concavity and convexity features is constructed based on the local coordinate system, including: In the local coordinate system, several spherical sub-regions are divided with the target point as the center, and the neighboring points of the target point are projected onto the corresponding spherical sub-regions according to their spatial positions. For each spherical sub-region, the directional distribution information is extracted based on the angle between its normal vector and the coordinate axis of the local coordinate system. Define a concavity / convexity weighting factor, which is determined by the concavity / convexity index of a point. and distance weight It was jointly determined that, Indicates a convex point. Indicates a concave point; In the process of interpolating directional information, the concavity and convexity weighting factor is incorporated into the directional distribution statistics, and the directional information in each spherical sub-region is weighted and accumulated to generate a weighted directional histogram of that sub-region. The weighted orientation histograms of all spherical sub-regions are spliced together in a preset order to form a dot signature descriptor containing convex and concave feature information.
6. The method for registering rock mass point clouds based on convex and concave features as described in claim 5, characterized in that, In step S6, point pairs that meet the threshold are selected, including: Calculate the Euclidean distance and cosine similarity between the weighted descriptors of the convex and concave feature points of the source point cloud and the weighted descriptors of the convex and concave feature points of the target point cloud; Set a Euclidean distance threshold and a cosine similarity threshold, and filter point pairs that simultaneously satisfy the condition that the Euclidean distance is less than the Euclidean distance threshold and the cosine similarity is greater than the cosine similarity threshold; The selected point pairs are sorted from high to low according to their cosine similarity, and point pairs that do not meet the geometric consistency are removed, thus forming the initial set of matching point pairs.
7. The method for registering rock mass point clouds based on convex and concave features as described in claim 6, characterized in that, In step S7, for the initial set of matched point pairs, the RANSAC algorithm is used to remove erroneous matched point pairs, and the rigid transformation matrix is calculated, specifically including: S71. Randomly select non-coplanar matching point pairs from the initial matching point pair set, and calculate the provisional rigid transformation matrix based on the SVD algorithm; S72. Apply the provisional rigid transformation matrix to the feature points of the source point cloud and calculate the Euclidean distance error between the transformed source point and the corresponding target point; S73. Mark the point pairs whose Euclidean distance error is less than the set distance error threshold as interior points, and count the number of interior points; S74. Repeat steps S71-S73. After setting the number of iterations, select the provisional rigid transformation matrix obtained in the iteration with the largest number of interior points as the estimated optimal rigid transformation matrix.
8. The method for registering rock mass point clouds based on convex and concave features as described in claim 7, characterized in that, In step S8, the rigid transformation matrix is applied to the source point cloud, including: ； in, The coordinates of the registered source point cloud; The coordinates of the original source point cloud; These are the rotation parameters in the rigid transformation matrix; represents the translation parameter in the rigid transformation matrix.
9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method for rock mass point cloud registration based on convex and concave features as described in any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for rock mass point cloud registration based on convex and concave features as described in any one of claims 1 to 8.