A robust registration method for air-ground multi-source point clouds based on structure entropy sampling
By using a structural entropy sampling method and a robust optimization function, the noise and small-scale error problems existing in air-to-ground multi-source point cloud registration are solved, achieving high-precision and high-efficiency point cloud registration, which is suitable for air-to-ground fusion scenarios and improves the robustness and efficiency of the algorithm.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA RAILWAY DESIGN GRP CO LTD
- Filing Date
- 2025-09-19
- Publication Date
- 2026-07-10
AI Technical Summary
Existing air-ground multi-source point cloud registration methods struggle to achieve high-precision and high-efficiency registration in multi-source heterogeneous data environments, especially when there are noisy, incomplete, and small-scale error point clouds. They are also prone to getting trapped in local optima, leading to unstable registration results.
A structural entropy-based sampling method is adopted. By extracting the spatial coordinates, normal vectors, and curvature of the point cloud, standardization preprocessing is performed, including scale normalization and decentralization. Key points are screened, a sampling point set is constructed, and robust optimization functions and Anderson Acceleration optimization techniques are used to accelerate optimization. By iteratively solving for the optimal rigid transformation matrix, highly robust point cloud registration is achieved.
Despite the presence of minor initial errors, occlusion, and sparse sampling, high-precision and efficient registration between point clouds from different sources was achieved, improving the algorithm's generalization performance in air-to-ground fusion scenarios, as well as its robustness and efficiency.
Smart Images
Figure CN121280490B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of 3D point cloud registration technology, specifically relating to a robust registration method for multi-source point clouds based on structural entropy sampling. Background Technology
[0002] With the widespread deployment of various sensors such as UAV platforms and ground-based LiDAR, acquiring multi-source point cloud data from both air and ground has become increasingly convenient. However, due to differences in the viewing angles of different sensors, variations in ranging principles, and external environmental disturbances, the acquired point clouds often suffer from registration errors, noise, and outliers, severely impacting the accuracy of 3D reconstruction and measurement. Especially in sub-resolution level detail matching, traditional iterative nearest-point algorithms are prone to getting trapped in local optima, heavily reliant on initial values, and sensitive to outliers, leading to unstable registration results. Therefore, it is essential to introduce more robust and efficient data preprocessing and sampling strategies to improve the convergence and overall accuracy of registration algorithms.
[0003] In recent years, to improve the robustness and efficiency of point cloud registration, academia and industry have proposed a variety of improvement strategies. Some methods are based on region features and keypoint detection, achieving more reliable initial registration by extracting surface geometry or color descriptors; others are deep learning-driven feature learning methods, which automatically extract representations for adaptive registration through networks, but these methods have high requirements for training data and generalization performance; hybrid sampling strategies attempt to combine uniform sampling with importance sampling to balance structure preservation and efficiency.
[0004] However, the aforementioned methods often fail to achieve a balance between decentralized processing and adaptive sampling in multi-source heterogeneous data environments and on noisy, incomplete, small-scale error point clouds. This can even lead to the amplification of small errors or registration failure. Therefore, improvements are still needed in registration stability, global accuracy, and avoiding getting trapped in local optima when registering small error data. Summary of the Invention
[0005] The purpose of this invention is to provide a robust registration method for multi-source point clouds based on structural entropy sampling. Even under conditions of small initial errors, occlusion, sparse sampling, and other interferences, it can still achieve high-precision and high-efficiency registration between point clouds from different sources. It is especially suitable for scenarios such as air-to-ground fusion and small-error model registration.
[0006] Therefore, the present invention adopts the following technical solution:
[0007] A robust registration method for multi-source point clouds in the air and ground based on structural entropy sampling includes:
[0008] S1: Read the source point cloud data and target point cloud data respectively, and extract their spatial coordinates, normal vectors, and curvatures to obtain the coordinate set of the source point cloud. The coordinate set of the target point cloud ;
[0009] S2, standardizes the source point cloud and target point cloud obtained from S1, including scale normalization and decentralization, to eliminate scale bias and spatial translation differences of point clouds from different sources.
[0010] S3: Perform structural entropy sampling on the standardized source point cloud and target point cloud obtained in S2, filter out key points, and construct a set of sampling points for registration;
[0011] S4 employs a robust point-to-point registration method to calculate the optimal rigid transformation matrix between the source point cloud and the target point cloud obtained by the structural entropy sampling in S3, thereby achieving highly robust point cloud registration.
[0012] S5, multiply the registration result obtained in S4 by the original point cloud scaling factor again. And add the geometric center obtained by S2 This maintains the point cloud structure consistent with the original coordinate system, completing the spatial restoration of the registration result; the translation vector in the rigid transformation matrix obtained in S4 Restored to the original scale: Construct the rigid transformation matrix in the original coordinates: ,in, It is a rotation matrix;
[0013] S6, the final output includes the transformation matrix. The point cloud is registered and spatially restored; the transformation matrix can be used for subsequent attitude analysis or map stitching, and the point cloud data is saved as a standard format file for easy system integration.
[0014] The normalization and decentralization operations described in step S2 above include:
[0015] S21, calculate the input source point cloud and target point cloud respectively. Find the maximum and minimum values of the boundary in three directions, and calculate their maximum norm as a scaling factor. :
[0016]
[0017]
[0018] in, , , These are the scaling factor, maximum value, and minimum value of the source point cloud, respectively; , , These are the scaling factor, maximum boundary value, and minimum boundary value of the target point cloud, respectively.
[0019] S22, scales the input source and target point cloud coordinates according to their corresponding scaling factors, so that all coordinates fall within the unit space range:
[0020]
[0021]
[0022] in, This indicates that the source point cloud is scaled according to the scaling factor. The scaled set of coordinates; This indicates that the target point cloud is scaled according to the scaling factor. The scaled set of coordinates;
[0023] S23, calculate the geometric center position of all points in the scaled source point cloud. Geometric center position of all points in the target point cloud :
[0024]
[0025]
[0026] in, The total number of points in the source point cloud. The total number of points in the target point cloud. The first point in the source cloud The position coordinates of each point For the target point cloud The position coordinates of each point;
[0027] S24, mean removal processing: using all the geometric center positions obtained in S23, the source point cloud and the target point cloud are translated as a whole, so that their geometric centers are moved to the origin:
[0028]
[0029]
[0030] After the above processing, the source point cloud and the target point cloud are transformed to a uniform scale and their centers of gravity are aligned, which helps to improve the numerical stability of subsequent sampling and registration.
[0031] The structural entropy sampling described in step S3 above includes the following steps:
[0032] S31, for each point, construct a structure containing its... The local neighborhood of the nearest neighbor. It can be set to 12~64;
[0033] S32, construct the covariance matrix based on the local neighborhood of each point, perform eigenvalue decomposition, and define the curvature of the point as the ratio of the minimum eigenvalue to the sum of eigenvalues:
[0034]
[0035] in, are eigenvalues, and It is the smallest eigenvalue;
[0036] S33, for each point Its neighboring points Calculate the following three types of difference information respectively:
[0037] (1) Difference in the angle between normal vectors: used to measure the degree of change in the surface orientation of two points:
[0038]
[0039] in , Let these be the normal vectors of the two points;
[0040] (2) Curvature difference squared term: reflects the degree of difference between the local curvature structures of two points:
[0041]
[0042] (3) Spatial distance term: Euclidean distance between two points, which can be used to assess the density of the distribution between the two points:
[0043]
[0044] in Represents the coordinates of a point;
[0045] S34, Using the above three types of difference information as structural entropy values, construct the structural entropy scoring function for this point. , representing the structural complexity between this point and its neighboring points:
[0046]
[0047] The higher the score, the richer and more representative the point is in terms of normal, curvature, and space compared to its neighboring points;
[0048] S35, tally all ratings mean with standard deviation Points that meet the following conditions are selected as key structural points:
[0049]
[0050] The proposed sampling method based on structural entropy preserves morphological representativeness while eliminating redundancy and outliers.
[0051] In step S3 above, it is recommended that K The value is 32. The value is 1.0.
[0052] The robust point-to-point registration method described in step S4 above includes:
[0053] A robust optimization function is introduced to define the residual cost between point pairs to enhance the robustness of registration to outliers, as shown in the following formula:
[0054]
[0055] in, To determine the Euclidean distance between registration points. Adjustable scaling parameter; robust optimization function value The penalty coefficient is the number of points paired. The larger the penalty coefficient, the smaller the proportion of the point pair in the subsequent calculation of the error function, and vice versa.
[0056] By combining the Anderson Acceleration optimization method, the registration point pairs are updated iteratively, with the Euclidean distance of the registration point pairs representing the error. The optimal rigid transformation matrix is then solved iteratively. The goal is to minimize the cumulative error of the error function;
[0057] During the iteration process, the rigid transformation matrix from the source point cloud to the target point cloud is continuously solved and updated. :
[0058] ,
[0059] The iterative process is as follows: Apply the current rigid transformation matrix A rigid transformation is performed on the source point cloud to generate intermediate registration results. New registration point pairs are established between the transformed source and target point clouds. The robust optimization function value is recalculated to confirm the proportion of the error in the error function for the new registration point pair. The Euclidean distance between the current registration point pairs is calculated; if the error function value satisfies the minimization condition, the iteration terminates; otherwise, the next iteration begins. The rigid transformation matrix is updated using Anderson Acceleration technology, which integrates historical iteration information with the current estimate. This accelerates convergence; finally, the source point cloud after the last transformation is output as the registration result, and the optimal rigid transformation matrix is retained.
[0060] Compared with the prior art, the present invention has the following beneficial effects:
[0061] 1. This invention integrates normal vector, curvature and spatial distribution information through a sampling mechanism driven by structural entropy, and uses the structural entropy function to characterize the geometric uncertainty of points, and selects key structural points for registration, thereby enhancing the adaptability to scenes with occlusion and density inconsistencies.
[0062] 2. This invention unifies the scale and center of point clouds from different sources through mean removal and normalization preprocessing, making subsequent registration more stable and accurate, and improving the generalization performance of the algorithm in heterogeneous point cloud scenarios such as air-ground fusion.
[0063] 3. This invention improves the registration robustness at outliers by using a robust optimization function and introduces the Anderson acceleration method to speed up the optimization registration efficiency, thus balancing accuracy and convergence speed.
[0064] 4. This invention does not rely on semantic tags, markers or external matching priors, and can complete accurate registration using only the geometric properties of the point cloud itself, which has low dependency characteristics and good versatility.
[0065] 5. The algorithm of this invention can be directly embedded into existing point cloud mapping systems, SLAM systems or registration platforms as a pre-registration module or structural point extraction module, which is easy to deploy and highly adaptable to engineering. Attached Figure Description
[0066] Figure 1 This is a flowchart of a robust registration method for multi-source point clouds in air and ground according to an embodiment of the present invention;
[0067] Figure 2 This is a schematic diagram of uniform sampling (50,000 points) of air-to-ground fused point cloud data in an embodiment of the present invention.
[0068] Figure 3 This is a schematic diagram (31028 points) of the structural entropy sampling of air-to-ground fused point cloud data in an embodiment of the present invention.
[0069] Figure 4 This is a schematic diagram of the robust optimization function used in the embodiments of the present invention. =1.0);
[0070] Figure 5 and Figure 6 These are schematic diagrams showing the effect of the first group of air-to-ground fused point cloud data before and after registration using the method of the present invention in this embodiment of the invention.
[0071] Figure 7 and Figure 8 This is a schematic diagram showing the effect of the second set of air-to-ground fused point cloud data before and after registration using the method of the present invention in an embodiment of the present invention. Detailed Implementation
[0072] The method of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.
[0073] See Figure 1 The present invention provides a robust registration method for multi-source point clouds based on structural entropy sampling, comprising the following steps:
[0074] S1, point cloud data loading and information extraction.
[0075] The source and target point cloud data are read separately through a file interface, and all point cloud data are unified into a single file. The matrix representation facilitates subsequent processing. Spatial coordinates, normal vectors, curvature, and optional color channel information are extracted from the point cloud data to obtain the coordinate set of the source point cloud. The coordinate set of the target point cloud .
[0076] S2, Normalization and Mean Removal Preprocessing.
[0077] Because point cloud data from different acquisition sources differ in scale units, coordinate references, and measurement ranges, direct registration may result in significant errors. Therefore, this invention standardizes the source and target point clouds obtained in S1 through normalization and mean removal operations. The specific steps are as follows:
[0078] S21, Determine the scaling factor and standardize the scale.
[0079] Calculate the source point cloud and target point cloud obtained by S1 respectively. Find the maximum and minimum values of the boundary in three directions, and calculate their maximum norm as a scaling factor. :
[0080]
[0081]
[0082] in, , , These are the scaling factor, maximum value, and minimum value of the source point cloud, respectively; , , These are the scaling factor, maximum boundary value, and minimum boundary value of the target point cloud, respectively.
[0083] S22, scale the source and target point cloud coordinates obtained from S1 according to their corresponding scaling factors, so that all coordinates fall within the unit space range:
[0084]
[0085]
[0086] in, This indicates that the source point cloud is scaled according to the scaling factor. The scaled set of coordinates; This indicates that the target point cloud is scaled according to the scaling factor. The scaled set of coordinates.
[0087] S23, calculate the geometric center position of all points in the scaled source point cloud. Geometric center position of all points in the target point cloud :
[0088]
[0089]
[0090] in, The total number of points in the source point cloud. The total number of points in the target point cloud. The first point in the source cloud The position coordinates of each point For the target point cloud The position coordinates of each point;
[0091] S24, mean removal processing: using all the geometric center positions obtained in S23, the source point cloud and the target point cloud are translated as a whole, so that their geometric centers are moved to the origin:
[0092]
[0093]
[0094] After the above processing, the source point cloud and the target point cloud are transformed to a uniform scale and their centers of gravity are aligned, which helps to improve the numerical stability of subsequent sampling and registration.
[0095] S3, Keypoint Sampling Based on Structural Entropy. To extract keypoints with strong geometric representation and high information content from the point cloud, a scoring strategy based on structural entropy is used to quantitatively evaluate each point in the standardized source and target point clouds obtained in S2, and keypoints with strong representation and stable structure are selected to construct a sampling point set for registration. This includes the following steps:
[0096] S31, for each point, construct a structure containing its... The local neighborhood of the nearest neighbor. It can be set to 12~64, recommended. K The value is 32.
[0097] S32, construct the covariance matrix based on the local neighborhood of each point, perform eigenvalue decomposition, and define the curvature of the point as the ratio of the minimum eigenvalue to the sum of eigenvalues:
[0098]
[0099] in, are eigenvalues, and It is the smallest eigenvalue.
[0100] S33, for each point Its neighboring points Calculate the following three types of difference information respectively:
[0101] (1) Difference in the angle between normal vectors: used to measure the degree of change in the surface orientation of two points:
[0102]
[0103] in , Let be the normal vectors of the two points.
[0104] (2) Curvature difference squared term: reflects the degree of difference between the local curvature structures of two points:
[0105]
[0106] (3) Spatial distance term: Euclidean distance between two points, which can be used to assess the density of the distribution between the two points:
[0107]
[0108] in Represents the coordinates of a point.
[0109] S34, Using the above three types of difference information as structural entropy values, construct the structural entropy scoring function for this point. , representing the structural complexity between this point and its neighboring points:
[0110]
[0111] A higher score indicates that the point is more informational and representative of its neighbors in terms of normal, curvature, and space.
[0112] S35, tally all ratings mean with standard deviation Points that meet the following conditions are selected as key structural points:
[0113] in The recommended value is 1.0.
[0114] The provided sampling method based on structural entropy preserves morphological representativeness while eliminating redundancy and outliers. For details, see [link to relevant documentation]. Figure 2 and Figure 3 contrast( Figure 2 To perform uniform downsampling, Figure 3 (For cross-entropy sampling).
[0115] S4, robust point-to-point registration. A robust point-to-point registration method is used to calculate the optimal rigid transformation matrix between the source and target point clouds obtained through S3 structural entropy sampling, achieving highly robust point cloud registration. The specific operation is as follows:
[0116] Introducing robust optimization functions such as Figure 4 Define the residual cost between point pairs to enhance the robustness of registration to outliers.
[0117] in, To determine the Euclidean distance between registration points. The scaling parameter is adjustable. Robust optimization function value. The penalty coefficient is the number of points paired. The larger the penalty coefficient, the smaller the proportion of the point pair in the subsequent calculation of the error function minimization process, and vice versa.
[0118] By combining the Anderson Acceleration method, the registration point pairs are updated iteratively, with the Euclidean distance of the registration point pairs representing the error. The optimal rigid transformation matrix is then solved iteratively. The goal is to minimize the cumulative error of the error function. During the iteration process, the rigid transformation matrix from the source point cloud to the target point cloud is continuously solved and updated. :
[0119]
[0120] in Let be a rotation matrix. It is a translation vector.
[0121] The iterative process is as follows: Apply the current rigid transformation matrix A rigid transformation is performed on the source point cloud to generate intermediate registration results. New registration point pairs are established between the transformed source and target point clouds. The robust optimization function value is recalculated to confirm the proportion of the error in the error function for the new registration point pair. The Euclidean distance between the current registration point pairs is calculated; if the error function value satisfies the minimization condition, the iteration terminates; otherwise, the next iteration begins. The rigid transformation matrix is updated using Anderson Acceleration technology, which integrates historical iteration information with the current estimate. This accelerates convergence; finally, the source point cloud after the last transformation is output as the registration result, and the optimal rigid transformation matrix is retained.
[0122] S5, Registration result restoration: The registration result obtained in S4 is multiplied again by the original point cloud scaling factor. And add the geometric center obtained by S2 This maintains the point cloud structure consistent with the original coordinate system, completing the spatial restoration of the registration result. The translation vector in the rigid transformation matrix obtained in S4... Restored to the original scale: Construct the rigid transformation matrix in the original coordinates: .
[0123] S6, Output the results. The final output includes the transformation matrix. The point cloud is then registered and spatially restored. The transformation matrix can be used for subsequent attitude analysis or map stitching, and the point cloud data is saved in a standard format file for easy system integration.
[0124] Furthermore, the method outputs The transformation matrix, including rotation and translation components, is suitable for air-to-ground fusion point cloud data with inconsistent sources, uneven density, and different viewpoints. It can also be used for subsequent point cloud registration, attitude estimation, pose tracking, and other tasks.
[0125] Applying the algorithm of this invention to actual air-to-ground fusion data, the effects of the first set of air-to-ground fusion point cloud data before and after registration using the method of this invention are as follows: Figure 5 , Figure 6 As shown, the effects of the second set of air-to-ground fused point cloud data before and after registration using the method of this invention are respectively as follows: Figure 7 , Figure 8 As shown in the figure, the method of this invention can register small-error spatial-to-ground fusion data.
Claims
1. A robust registration method for multi-source point clouds based on structural entropy sampling, characterized in that, include: S1: Read the source point cloud data and target point cloud data respectively, and extract their spatial coordinates, normal vectors, and curvatures to obtain the coordinate set of the source point cloud. The set of coordinates of the target point cloud; S2, standardizes the source point cloud and target point cloud obtained from S1, including scale normalization and decentralization, to eliminate scale bias and spatial translation differences of point clouds from different sources. S3: Perform structural entropy sampling on the standardized source point cloud and target point cloud obtained in S2, filter out key points, and construct a set of sampling points for registration; S4 employs a robust point-to-point registration method to calculate the optimal rigid transformation matrix between the source point cloud and the target point cloud obtained by the structural entropy sampling in S3, thereby achieving highly robust point cloud registration. S5, multiply the registration result obtained in S4 by the original point cloud scaling factor again. And add the geometric center obtained by S2 This ensures that the point cloud structure remains consistent with the original coordinate system, thus completing the spatial restoration of the registration results; The translation vector in the well-rigid transformation matrix obtained by S4 Restored to the original scale: Construct the rigid transformation matrix in the original coordinates: ,in, It is a rotation matrix; S6, the final output includes the transformation matrix. The point cloud is registered and spatially restored; the transformation matrix can be used for subsequent attitude analysis or map stitching, and the point cloud data is saved as a standard format file for easy system integration.
2. The robust registration method for multi-source point clouds based on structural entropy sampling according to claim 1, characterized in that... The normalization and decentralization operations described in S2 include: S21, calculate the input source point cloud and target point cloud respectively. Find the maximum and minimum values of the boundary in three directions, and calculate their maximum norm as a scaling factor. : in, , , These are the scaling factor, maximum boundary value, and minimum boundary value of the source point cloud, respectively; , , These are the scaling factor, maximum boundary value, and minimum boundary value of the target point cloud, respectively; S22, scales the input source and target point cloud coordinates according to their corresponding scaling factors, so that all coordinates fall within the unit space range: in, This indicates that the source point cloud is scaled according to the scaling factor. The scaled set of coordinates; This indicates that the target point cloud is scaled according to the scaling factor. The scaled set of coordinates; S23, calculate the geometric center position of all points in the scaled source point cloud. Geometric center position of all points in the target point cloud : in, The total number of points in the source point cloud. The total number of points in the target point cloud. The first point in the source cloud The position coordinates of each point For the target point cloud The position coordinates of each point; S24, mean removal processing: using all the geometric center positions obtained in S23, the source point cloud and the target point cloud are translated as a whole, so that their geometric centers are moved to the origin: After the above processing, the source point cloud and the target point cloud are transformed to a uniform scale and their centers of gravity are aligned, which helps to improve the numerical stability of subsequent sampling and registration.
3. The robust registration method for multi-source point clouds based on structural entropy sampling according to claim 1, characterized in that, The structural entropy sampling described in S3 includes the following steps: S31, for each point, construct a structure containing its... The local neighborhood of the nearest neighbor. It can be set to 12~64; S32, construct the covariance matrix based on the local neighborhood of each point, perform eigenvalue decomposition, and define the curvature of the point as the ratio of the minimum eigenvalue to the sum of eigenvalues: in, are eigenvalues, and It is the smallest eigenvalue; S33, for each point Its neighboring points Calculate the following three types of difference information respectively: (1) Difference in the angle between normal vectors: used to measure the degree of change in the surface orientation of two points: in , Let these be the normal vectors of the two points; (2) Curvature difference squared term: reflects the degree of difference between the local curvature structures of two points: (3) Spatial distance term: Euclidean distance between two points, which can be used to assess the density of the distribution between the two points: in Represents the coordinates of a point; S34, Using the above three types of difference information as structural entropy values, construct the structural entropy scoring function for this point. , representing the structural complexity between this point and its neighboring points: The higher the score, the richer and more representative the point is in terms of normal, curvature, and space compared to its neighboring points; S35, tally all ratings mean with standard deviation Points that meet the following conditions are selected as key structural points: The proposed sampling method based on structural entropy preserves morphological representativeness while eliminating redundancy and outliers.
4. The robust registration method for multi-source point clouds based on structural entropy sampling according to claim 3, characterized in that, recommend K The value is 32. The value is 1.
0.
5. The robust registration method for multi-source point clouds based on structural entropy sampling according to claim 1, characterized in that, The robust point-to-point registration method described in S4 includes: A robust optimization function is introduced to define the residual cost between point pairs to enhance the robustness of registration to outliers, as shown in the following formula: in, To determine the Euclidean distance between registration points. Adjustable scaling parameter; robust optimization function value The penalty coefficient is the number of points paired. The larger the penalty coefficient, the smaller the proportion of the point pair in the subsequent calculation of the error function, and vice versa. By combining the Anderson Acceleration optimization method, the registration point pairs are updated iteratively, with the Euclidean distance of the registration point pairs representing the error. The optimal rigid transformation matrix is then solved iteratively. The goal is to minimize the cumulative error of the error function; During the iteration process, the rigid transformation matrix from the source point cloud to the target point cloud is continuously solved and updated. : ， The iterative process is as follows: Apply the current rigid transformation matrix A rigid transformation is performed on the source point cloud to generate intermediate registration results. New registration point pairs are established between the transformed source and target point clouds. The robust optimization function value is recalculated to confirm the proportion of the error in the error function for the new registration point pair. The Euclidean distance between the current registration point pairs is calculated; if the error function value satisfies the minimization condition, the iteration terminates; otherwise, the next iteration begins. The rigid transformation matrix is updated using Anderson Acceleration technology, which integrates historical iteration information with the current estimate. This accelerates convergence; finally, the source point cloud after the last transformation is output as the registration result, and the optimal rigid transformation matrix is retained.