A TLS and ULS point cloud automatic registration method based on multi-scale voxel feature analysis

Abstract

This invention discloses an automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis, belonging to the field of forest resource survey and monitoring technology. It aims to solve the problem of automatic point cloud registration in complex natural scenes such as forests. The invention includes constructing an adaptive voxel-centric downsampling method to downsample TLS point cloud data and UAV lidar ULS point cloud data, performing initial alignment with full degrees of freedom to obtain overall aligned data; constructing a multi-scale voxel iterative registration method to obtain the optimal result at the current scale; using the optimal result at the current scale as the initial value for the next scale, performing voxel iterative registration at the next scale, recording the registration RMSE at each scale, and selecting the transformation corresponding to the global minimum as the final result, thus completing the automatic registration of TLS and ULS point clouds based on multi-scale voxel feature analysis. This invention ultimately achieves complete registration with high accuracy, high robustness, and strong generalization ability.

Description

Technical Field

[0001] This invention belongs to the field of forest resource survey and monitoring technology, specifically relating to an automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis. Background Technology

[0002] LiDAR has made significant progress in forest resource surveys in recent years. ULS (Unmanned Aerial Vehicle Laser Scanning) can quickly acquire large-scale canopy information, while TLS (Terrestrial Laser Scanning) can finely depict understory details. By fusing the two, comprehensive and high-precision three-dimensional data of the forest scene can be obtained, thereby improving the efficiency and accuracy of forest resource surveys.

[0003] However, current technologies are generally registration methods based on tree canopy voxel features, which have the following main drawbacks and limitations:

[0004] 1. Strong dependence on specific, single structural features leads to poor generalization ability: Although the canopy voxel-based method expands the features from "points" to "voxels", its feature source is still limited to the point cloud of the upper canopy, which needs to be extracted through height-frequency filtering. This is difficult to achieve in complex forests with unclear canopy stratification, continuous canopy or damaged structure. The single feature source limits its scene adaptability.

[0005] 2. Simple feature representation, limited discriminative power and robustness: The tree canopy voxel-based method mainly uses the spatial location of the voxel's core for template matching, which is essentially a simplified location feature matching. It does not fully utilize the multi-dimensional information such as the geometric shape (e.g., linear, planar, scattering), density distribution, and vertical structure of the point cloud inside the voxel, which is prone to matching ambiguity or errors in forest scenes with repetitive structures.

[0006] 3. Insufficient preprocessing and adaptability for data characteristics (such as density differences and missing data): The significant density differences between TLS and ULS point clouds are not addressed. High-density TLS point clouds impose a huge computational burden and may introduce noise interference. There is a lack of robust design for partial data loss due to occlusion and viewpoint; the canopy voxel method experiences a sharp performance drop when the canopy layer is not fully extracted.

[0007] 4. The initial alignment strategy is simple but struggles to handle large initial deviations: The tree canopy voxel-based method uses point cloud bounding box main direction alignment for initialization. This method heavily relies on the accuracy of the overall point cloud distribution. When the point cloud coverage is uneven, has missing parts, or is asymmetrically distributed, the main direction estimation is prone to errors, leading to initialization failure and failing to provide a good starting point for subsequent steps.

[0008] 5. The registration process lacks an incremental optimization mechanism, making it prone to getting trapped in local optima or insufficient optimization: the tree canopy voxel-based method uses a single fixed-scale voxelization and template matching. A fixed scale makes it difficult to balance global convergence and local accuracy: if the scale is too coarse, the matching becomes blurry; if the scale is too fine, initial bias directly leads to matching failure. Its single process lacks the ability for incremental optimization from coarse to fine, making it susceptible to interference from local optima. Summary of the Invention

[0009] The problem to be solved by this invention is to realize automatic registration of ULS and TLS point clouds for complex natural scenes such as forests, and to propose an automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] An automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis includes the following steps:

[0012] S1. Collect ground-based lidar (TLS) point cloud data and UAV lidar (ULS) point cloud data in the forest area, construct an adaptive volumetric centroid downsampling method, downsample the TLS point cloud data, and obtain downsampled TLS point cloud data;

[0013] S2. Perform initial alignment of the downsampled TLS point cloud data and UAV LiDAR ULS point cloud data obtained in step S1 to obtain overall aligned data;

[0014] S3. For the overall aligned data obtained in step S2, construct a multi-scale voxel iterative registration method. For the current scale, first calculate the voxel feature descriptor, then use KD tree to perform nearest neighbor matching and use double constraints to filter effective matching point pairs, then use SVD to solve the optimal rigid body transformation matrix for the effective matching point pairs, then apply the optimal rigid body transformation matrix to the overall aligned TLS point cloud to complete the voxel feature-based registration process, and then use the dynamic threshold ICP algorithm to perform ICP fine registration to obtain the optimal result for the current scale.

[0015] S4. Using the optimal result of the current scale obtained in step S3 as the initial value of the next scale, perform voxel iterative registration at the next scale based on the method in step S3, record the registration RMSE at each scale, select the transformation corresponding to the global minimum as the final result, and complete the automatic registration of TLS and ULS point clouds based on multi-scale voxel feature analysis.

[0016] Furthermore, the specific implementation method of step S1 includes the following steps:

[0017] S1.1. Collect TLS point cloud data from ground-based lidar and ULR lidar point cloud data from the forest area. Then, use fast KD-tree queries to calculate the minimum point spacing and point density of the ULR and TLS point clouds respectively, and then calculate the density ratio. ,in For the point cloud density of the TLS point cloud, Point cloud density for ULS point cloud;

[0018] S1.2. Minimum spacing of the ULS point cloud obtained in step S1.1 And density ratio, calculate TLS adaptive voxel size, the calculation formula is:

[0019]

[0020] in, The density adaptive function is defined as follows:

[0021] ;

[0022] S1.3. Calculation of TLS point cloud P based on TLS adaptive voxel size TLS The voxel index of each point is then used to generate the TLS point cloud P using a hash function. TLS Voxel bond value at each point;

[0023] S1.4. The TLS point cloud P obtained in step S1.3... TLS The voxel key values ​​of each point are sorted and grouped, and the centroid of each voxel is calculated. As downsampled TLS point cloud data:

[0024]

[0025] in, Let p be the set of all points belonging to voxel k. w Let be the three-dimensional coordinate vector of the w-th point of voxel k.

[0026] Furthermore, the specific implementation method of step S2 includes the following steps:

[0027] S2.1. First, robust spatial block sampling is performed on the downsampled TLS point cloud data and ULS point cloud data, dividing the point cloud XY plane into a regular grid of 20m×20m; multiple points are randomly selected within each grid cell. If the total number of sampling points is less than the preset value, additional random sampling is performed to obtain the source point cloud sampling point set P. s,sample and target point cloud sampling point set P t,sample ;

[0028] S2.2. In angular steps within the range of 0-360°. For P s,sample Perform a system rotation search, and construct a rotation matrix around the z-axis for the i-th candidate rotation angle. as follows:

[0029]

[0030]

[0031] in, It is the i-th candidate rotation angle, and n is the total number of candidate rotation angles;

[0032] Then, the translation vector t based on centroid alignment is calculated. i :

[0033]

[0034] in, and These are the centroids of the target point cloud sampling point set and the source point cloud sampling point set, respectively.

[0035] S2.3. Construct candidate matrix T using rotation matrix and translation vector. i It also quickly evaluates the alignment quality of the sample point set and selects the optimal angle. :

[0036]

[0037] Among them, RMSE i Let be the root mean square error corresponding to the i-th candidate transformation matrix, and NN represent finding the nearest neighbor in the target point cloud;

[0038] S2.4. At the optimal angle Perform a detailed angle search in the vicinity, that is... A fine-grained search is performed within the range using smaller step sizes of 1°, and local translational fine-tuning is carried out in the XY plane to finally obtain the initial transformation matrix T. init ;

[0039] S2.5. Multiply the downsampled TLS point cloud data by the initial transformation matrix T. init Obtain the overall alignment data.

[0040] Furthermore, the specific implementation method of step S3 includes the following steps:

[0041] S3.1. Based on the overall aligned data, let the scale set be... Where L is the total number of scales, S L For the Lth scale, S1>S2>…>S L ;

[0042] S3.2. At each scale, a point cloud voxel feature descriptor D is generated for the overall aligned data. The feature descriptor includes 12-dimensional features.

[0043] Including height characteristics, with average height Z mean Standard deviation of height Z std Height range Z range ;

[0044] The horizontal feature is the range of the XY plane. This describes the distribution range of voxels in the horizontal direction;

[0045] The local geometric features are calculated using principal component analysis: linearity L, flatness P, and scattering S. C And verticality V:

[0046] Density features reflect the sampling density of the point cloud and are used to match regions with similar densities. To address the density differences between ULS and TLS, logarithmic compression is used to maintain comparability, resulting in:

[0047]

[0048] in, The number of valid point clouds in the current voxel; the normalized positional feature is the voxel center coordinates. ;

[0049] S3.3. Perform feature normalization on the feature descriptor D of TLS and ULS.

[0050]

[0051] KD-trees are used for nearest neighbor matching, and double constraints are employed to filter valid matching pairs. M is the total number of valid matching point pairs;

[0052] The dual constraints include the distance ratio constraint and the spatial consistency constraint. The formula for calculating the distance ratio constraint is:

[0053]

[0054] Where d1 and d2 are the nearest and second nearest distances, respectively. The distance ratio threshold;

[0055] The formula for calculating spatial consistency constraints is:

[0056]

[0057] Among them, c s For the matching source voxel center, c t For the target voxel center to be matched;

[0058] S3.4. Based on valid matching point pairs The optimal rigid body transformation matrix T is solved by SVD. SVD ;

[0059] S3.5. The dynamic threshold ICP algorithm is used to apply the T... SVD The point cloud is finely registered; the distance threshold is dynamically adjusted based on the matching results of the current iteration. :

[0060]

[0061] Where, N valid To effectively match the number of points, N total The total number of points is given. B1 and B2 are threshold coefficients, which must satisfy B1>1>B2. β1 and β2 are effective matching point coefficients, which also satisfy β1<β2. Finally, the registration result at this voxel scale is obtained and used as the initial value for the next voxel iteration.

[0062] The beneficial effects of this invention are:

[0063] This invention presents an automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis, achieving a leap from "relying on a single structure" to "utilizing multi-dimensional features across the entire scene," significantly improving generalization ability and scene robustness. This invention employs multi-scale voxelization of point clouds across the entire scene and extracts a 12-dimensional feature descriptor that integrates height statistics, local geometry (linear / planar / scattering / verticality), density, and location information. It overcomes the strong dependence of the "canopy voxel method" on clearly layered canopy structures and remains effective in scenes where canopy layers are difficult to separate, such as broadleaf forests and mixed forests. This invention does not rely on any specific or variable forest structure components and can adapt to diverse forest scenes, including coniferous forests, broadleaf forests, mixed forests, and forests with complex terrain, achieving unprecedented generalization ability and scene robustness.

[0064] This invention presents an automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis. Through adaptive downsampling, it significantly improves computational efficiency. Adaptive voxel downsampling balances data density, reducing the amount of TLS data by approximately 90% compared to existing techniques that directly process high-density raw data (especially TLS), thus greatly reducing computational load and memory consumption. While maintaining or even improving registration accuracy, it achieves an order-of-magnitude improvement in computational efficiency.

[0065] This invention presents an automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis. It provides robust initial alignment through a systematic global search, ensuring registration convergence even with large deviations. The invention proposes a full-degree-of-freedom initial alignment strategy, employing a 0-360° systematic rotation search combined with fine-tuning. This addresses the problem in the "canopy voxel method," where initialization based on the bounding box's principal direction is sensitive to the overall point cloud distribution and prone to errors when point cloud coverage is uneven or incomplete. This invention can reliably handle arbitrarily large initial translation and rotation deviations (>80° or even larger), providing a high-quality starting point close to global optimum for subsequent fine registration, fundamentally preventing registration failures due to poor initial values.

[0066] This invention presents an automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis. It establishes a progressive registration process with multi-scale iterative optimization, ensuring both accuracy and stability. The core of this invention is a multi-scale voxel iterative registration framework from coarse to fine, incorporating dynamic threshold ICP refinement. Unlike the "tree canopy voxel method," which uses a fixed single-scale voxel matching and directly leads to fine-scale matching failure when coarse-scale mismatch occurs, this invention's multi-scale strategy achieves progressive optimization: coarse-scale capture of the global structure achieves stable convergence, while fine-scale optimization of local details achieves centimeter-level accuracy. This mechanism effectively avoids local optima, ensuring the stability and high-precision output of the algorithm throughout the optimization process, ultimately achieving centimeter-level translation errors and sub-degree-level rotation errors.

[0067] This invention presents an automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis. Through a synergistic approach of adaptive downsampling, full-degree-of-freedom initial alignment, and multi-scale iterative registration, it systematically addresses the inherent shortcomings of existing technologies in terms of feature dependency, initialization robustness, optimization stability, computational efficiency, and generalization ability. It provides a reliable solution for automatic registration of ULS and TLS point clouds in complex forest environments, offering high accuracy (centimeter-to-sub-degree level), high robustness (coping with large biases and missing data), high efficiency (significantly reducing resource consumption), and strong generalization (adapting to various forest types). This lays a superior foundation for data fusion in precise forestry surveys and ecological research.

[0068] The present invention discloses an automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis, which can overcome the strong dependence on single and easily missing specific structural features; adopt multi-dimensional features that can fully express the diversity of three-dimensional structures; provide a robust initialization strategy that can cope with arbitrarily large initial pose deviations; establish a multi-scale progressive optimization process from global to local to ensure convergence stability; and effectively handle point cloud density differences and partial missing problems, ultimately achieving complete registration with high accuracy, high robustness, and strong generalization ability. Attached Figure Description

[0069] Figure 1 The flowchart is a method for automatic registration of TLS and ULS point clouds based on multi-scale voxel feature analysis as described in this invention.

[0070] Figure 2 This is a structural block diagram of an automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis as described in this invention.

[0071] Figure 3 The point cloud registration effect of the automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis described in this invention. Figure 1 ;

[0072] Figure 4 The point cloud registration effect of the automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis described in this invention. Figure 2 . Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The components of the specific embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations, and the invention may also have other embodiments.

[0074] Therefore, the following detailed description of specific embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected specific embodiments of the invention. All other specific embodiments obtained by those skilled in the art based on these specific embodiments without inventive effort are within the scope of protection of this invention.

[0075] To further understand the invention's content, features, and effects, the following specific embodiments are provided, along with accompanying drawings. Figure 1 - Appendix Figure 4 Detailed explanation is as follows:

[0076] Example 1:

[0077] An automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis includes the following steps:

[0078] S1. Collect ground-based lidar (TLS) point cloud data and UAV lidar (ULS) point cloud data in the forest area, construct an adaptive volumetric centroid downsampling method, downsample the TLS point cloud data, and obtain downsampled TLS point cloud data;

[0079] Furthermore, the specific implementation method of step S1 includes the following steps:

[0080] S1.1. Collect TLS point cloud data from ground-based lidar and ULR lidar point cloud data from the forest area. Then, use fast KD-tree queries to calculate the minimum point spacing and point density of the ULR and TLS point clouds respectively, and then calculate the density ratio. ,in For the point cloud density of the TLS point cloud, Point cloud density for ULS point cloud;

[0081] Furthermore, regarding point clouds The minimum point spacing d is calculated using a fast KD-tree query. min and point cloud density :

[0082]

[0083] Where, p j,nn Point p j The nearest neighbor is M, the number of sampling points used for statistics, and V is the bounding box volume of the point cloud.

[0084] S1.2. Minimum spacing of the ULS point cloud obtained in step S1.1 And density ratio, calculate TLS adaptive voxel size, the calculation formula is:

[0085]

[0086] in, The density adaptive function is defined as follows:

[0087] ;

[0088] When the point cloud density difference is large, a larger density adaptive coefficient is used to expand the adaptive voxel size, which can ensure downsampling efficiency. Finally, the voxel size is constrained within the range of [0.02, 0.05] meters to adapt to the characteristics of forest point clouds.

[0089] S1.3. Calculation of TLS point cloud P based on TLS adaptive voxel size TLS The voxel index of each point is then used to generate the TLS point cloud P using a hash function. TLS Voxel bond value at each point;

[0090] Furthermore, the TLS point cloud P is calculated. TLS The voxel index of each point:

[0091] ;

[0092] Use a hash function to generate a unique voxel key value K. i :

[0093]

[0094] A, B, and C are large prime numbers used to reduce hash collisions.

[0095] S1.4. The TLS point cloud P obtained in step S1.3... TLS The voxel key values ​​of each point are sorted and grouped, and the centroid of each voxel is calculated. As downsampled TLS point cloud data:

[0096]

[0097] in, Let p be the set of all points belonging to voxel k. w Let be the three-dimensional coordinate vector of the w-th point of voxel k. This process reduces the size of the TLS point cloud by approximately 90% while preserving the spatial distribution characteristics of the forest structure.

[0098] S2. Perform initial alignment of the downsampled TLS point cloud data and UAV LiDAR ULS point cloud data obtained in step S1 to obtain overall aligned data;

[0099] Furthermore, the specific implementation method of step S2 includes the following steps:

[0100] S2.1. First, robust spatial block sampling is performed on the downsampled TLS point cloud data and ULS point cloud data, dividing the point cloud XY plane into a regular grid of 20m×20m; multiple points are randomly selected within each grid cell. If the total number of sampling points is less than the preset value, additional random sampling is performed to obtain the source point cloud sampling point set P. s,sample and target point cloud sampling point set P t,sample ;

[0101] S2.2. In angular steps within the range of 0-360°. For P s,sample Perform a system rotation search, and construct a rotation matrix around the z-axis for the i-th candidate rotation angle. as follows:

[0102]

[0103]

[0104] in, It is the i-th candidate rotation angle, and n is the total number of candidate rotation angles;

[0105] Then, the translation vector t based on centroid alignment is calculated. i :

[0106]

[0107] in, and These are the centroids of the target point cloud sampling point set and the source point cloud sampling point set, respectively.

[0108] S2.3. Construct candidate matrix T using rotation matrix and translation vector. i It also quickly evaluates the alignment quality of the sample point set and selects the optimal angle. :

[0109]

[0110] Among them, RMSE i Let be the root mean square error corresponding to the i-th candidate transformation matrix, and NN represent finding the nearest neighbor in the target point cloud;

[0111] S2.4. At the optimal angle Perform a detailed angle search in the vicinity, that is... A fine-grained search is performed within the range using smaller step sizes of 1°, and local translational fine-tuning is carried out in the XY plane to finally obtain the initial transformation matrix T. init ;

[0112] S2.5. Multiply the downsampled TLS point cloud data by the initial transformation matrix T. init Obtain the overall alignment data.

[0113] S3. For the overall aligned data obtained in step S2, construct a multi-scale voxel iterative registration method. For the current scale, first calculate the voxel feature descriptor, then use KD tree to perform nearest neighbor matching and use double constraints to filter effective matching point pairs, then use SVD to solve the optimal rigid body transformation matrix for the effective matching point pairs, then apply the optimal rigid body transformation matrix to the overall aligned TLS point cloud to complete the voxel feature-based registration process, and then use the dynamic threshold ICP algorithm to perform ICP fine registration to obtain the optimal result for the current scale.

[0114] Furthermore, the specific implementation method of step S3 includes the following steps:

[0115] S3.1. Based on the overall aligned data, let the scale set be... Where L is the total number of scales, S L For the Lth scale, S1>S2>…>S L ;

[0116] S3.2. At each scale, a point cloud voxel feature descriptor D is generated for the overall aligned data. The feature descriptor includes 12-dimensional features.

[0117] Including height characteristics, with average height Z mean Standard deviation of height Z std Height range Z range ;

[0118] The horizontal feature is the range of the XY plane. This describes the distribution range of voxels in the horizontal direction;

[0119] The local geometric features are calculated using principal component analysis: linearity L, flatness P, and scattering S. C And verticality V:

[0120] Density features reflect the sampling density of the point cloud and are used to match regions with similar densities. To address the density differences between ULS and TLS, logarithmic compression is used to maintain comparability, resulting in:

[0121]

[0122] in, The number of valid point clouds in the current voxel; the normalized positional feature is the voxel center coordinates. ;

[0123] Furthermore, the local geometric features are calculated using principal component analysis, including linearity L, flatness P, and scattering S. C The formula for perpendicularity V is:

[0124]

[0125] in, e represents the eigenvalues ​​of the covariance matrix, and e3 represents the eigenvector corresponding to the smallest eigenvalue. These features are designed to leverage the complementary properties of forest point clouds, effectively capturing structural information and suppressing noise.

[0126] S3.3. Perform feature normalization on the feature descriptor D of TLS and ULS.

[0127]

[0128] KD-trees are used for nearest neighbor matching, and double constraints are employed to filter valid matching pairs. M is the total number of valid matching point pairs;

[0129] The dual constraints include the distance ratio constraint and the spatial consistency constraint. The formula for calculating the distance ratio constraint is:

[0130]

[0131] Where d1 and d2 are the nearest and second nearest distances, respectively. The distance ratio threshold;

[0132] The formula for calculating spatial consistency constraints is:

[0133]

[0134] Among them, c s For the matching source voxel center, c t For the target voxel center to be matched;

[0135] S3.4. Based on valid matching point pairs The optimal rigid body transformation matrix T is solved by SVD. SVD ;

[0136] Furthermore, SVD is solved and applied to the TLS point cloud to complete the voxel-based feature registration process. First, the centroid is calculated:

[0137] ;

[0138] Calculate the covariance matrix H:

[0139] ;

[0140] Perform SVD decomposition:

[0141] ;

[0142] Where UV is an orthogonal matrix and is a singular value diagonal matrix.

[0143] Calculate the rotation matrix R:

[0144] ;

[0145] Calculate the translation vector t:

[0146] ;

[0147] S3.5. The dynamic threshold ICP algorithm is used to apply the T... SVD The point cloud is finely registered; the distance threshold is dynamically adjusted based on the matching results of the current iteration. :

[0148]

[0149] Where, N valid To effectively match the number of points, N total The total number of points is given. B1 and B2 are threshold coefficients, which must satisfy B1>1>B2. β1 and β2 are effective matching point coefficients, which also satisfy β1<β2. Finally, the registration result at this voxel scale is obtained and used as the initial value for the next voxel iteration.

[0150] S4. Using the optimal result of the current scale obtained in step S3 as the initial value of the next scale, perform voxel iterative registration at the next scale based on the method in step S3, record the registration RMSE at each scale, select the transformation corresponding to the global minimum as the final result, and complete the automatic registration of TLS and ULS point clouds based on multi-scale voxel feature analysis.

[0151] Furthermore, a two-layer iteration approach is adopted: intra-scale iteration: Voxel feature matching and ICP registration are performed sequentially within each scale, combining the advantages of global matching and local convergence. Inter-scale iteration: The optimal result at the current scale is used as the initial value for the next scale to achieve incremental optimization and ensure that the optimization process does not get trapped in local optima. Optimal result selection: The registration RMSE at each scale is recorded, and the transformation corresponding to the global minimum is selected as the final result. This mechanism can avoid matching failures at a certain scale, ultimately achieving robust and high-precision registration in complex forest 3D structures.

[0152] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0153] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. An automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis, characterized in that, Includes the following steps: S1. Collect ground-based lidar (TLS) point cloud data and UAV lidar (ULS) point cloud data in the forest area, construct an adaptive volumetric centroid downsampling method, downsample the TLS point cloud data, and obtain downsampled TLS point cloud data; S2. Perform initial alignment of the downsampled TLS point cloud data and UAV LiDAR ULS point cloud data obtained in step S1 to obtain overall aligned data; S3. For the overall aligned data obtained in step S2, construct a multi-scale voxel iterative registration method. For the current scale, first calculate the voxel feature descriptor, then use KD tree to perform nearest neighbor matching and use double constraints to filter effective matching point pairs, then use SVD to solve the optimal rigid body transformation matrix for the effective matching point pairs, then apply the optimal rigid body transformation matrix to the overall aligned TLS point cloud to complete the voxel feature-based registration process, and then use the dynamic threshold ICP algorithm to perform ICP fine registration to obtain the optimal result for the current scale. S4. Using the optimal result of the current scale obtained in step S3 as the initial value of the next scale, perform voxel iterative registration at the next scale based on the method in step S3, record the registration RMSE at each scale, select the transformation corresponding to the global minimum as the final result, and complete the automatic registration of TLS and ULS point clouds based on multi-scale voxel feature analysis.

2. The automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis according to claim 1, characterized in that, The specific implementation method of step S1 includes the following steps: S1.

1. Collect TLS point cloud data from ground-based lidar and ULR lidar point cloud data from the forest area. Then, use fast KD-tree queries to calculate the minimum point spacing and point density of the ULR and TLS point clouds respectively, and then calculate the density ratio. ,in For the point cloud density of the TLS point cloud, Point cloud density for ULS point cloud; S1.

2. Minimum spacing of the ULS point cloud obtained in step S1.1 And density ratio, calculate TLS adaptive voxel size, the calculation formula is: in, The density adaptive function is defined as follows: ； S1.

3. Calculation of TLS point cloud P based on TLS adaptive voxel size TLS The voxel index of each point is then used to generate the TLS point cloud P using a hash function. TLS Voxel bond value at each point; S1.

4. The TLS point cloud P obtained in step S1.3... TLS The voxel key values ​​of each point are sorted and grouped, and the centroid of each voxel is calculated. As downsampled TLS point cloud data: in, Let p be the set of all points belonging to voxel k. w Let be the three-dimensional coordinate vector of the w-th point of voxel k.

3. The automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis according to claim 2, characterized in that, The specific implementation method of step S2 includes the following steps: S2.

1. First, robust spatial block sampling is performed on the downsampled TLS point cloud data and ULS point cloud data, dividing the point cloud XY plane into a regular grid of 20m×20m; multiple points are randomly selected within each grid cell. If the total number of sampling points is less than the preset value, additional random sampling is performed to obtain the source point cloud sampling point set P. s,sample and target point cloud sampling point set P t,sample ; S2.

2. In angular steps within the range of 0-360°. For P s,sample Perform a system rotation search, and construct a rotation matrix around the z-axis for the i-th candidate rotation angle. as follows: in, It is the i-th candidate rotation angle, and n is the total number of candidate rotation angles; Then, the translation vector t based on centroid alignment is calculated. i : in, and These are the centroids of the target point cloud sampling point set and the source point cloud sampling point set, respectively. S2.

3. Construct candidate matrix T using rotation matrix and translation vector. i It also quickly evaluates the alignment quality of the sample point set and selects the optimal angle. : Among them, RMSE i Let be the root mean square error corresponding to the i-th candidate transformation matrix, and NN represent finding the nearest neighbor in the target point cloud; S2.

4. At the optimal angle Perform a detailed angle search in the vicinity, that is... A fine-grained search is performed within the range using smaller step sizes of 1°, and local translational fine-tuning is carried out in the XY plane to finally obtain the initial transformation matrix T. init ; S2.

5. Multiply the downsampled TLS point cloud data by the initial transformation matrix T. init Obtain the overall alignment data.

4. The automatic registration method for TLS and ULS point clouds based on multi-scale voxel feature analysis according to claim 3, characterized in that, The specific implementation method of step S3 includes the following steps: S3.

1. Based on the overall aligned data, let the scale set be... Where L is the total number of scales, S L For the Lth scale, S1>S2>…>S L ; S3.

2. At each scale, a point cloud voxel feature descriptor D is generated for the overall aligned data. The feature descriptor includes 12-dimensional features. include The height characteristic is the average height Z. mean Standard deviation of height Z std Height range Z range ; The horizontal feature is the range of the XY plane. This describes the distribution range of voxels in the horizontal direction; The local geometric features are calculated using principal component analysis: linearity L, flatness P, and scattering S. C And verticality V: Density features reflect the sampling density of the point cloud and are used to match regions with similar densities. To address the density differences between ULS and TLS, logarithmic compression is used to maintain comparability, resulting in: in, The number of valid point clouds in the current voxel; the normalized positional feature is the voxel center coordinates. ; S3.

3. Perform feature normalization on the feature descriptor D of TLS and ULS. KD-trees are used for nearest neighbor matching, and double constraints are employed to filter valid matching pairs. M is the total number of valid matching point pairs; The dual constraints include the distance ratio constraint and the spatial consistency constraint. The formula for calculating the distance ratio constraint is: Where d1 and d2 are the nearest and second nearest distances, respectively. The distance ratio threshold; The formula for calculating spatial consistency constraints is: Among them, c s For the matching source voxel center, c t For the target voxel center to be matched; S3.

4. Based on valid matching point pairs The optimal rigid body transformation matrix T is solved by SVD. SVD ; S3.

5. The dynamic threshold ICP algorithm is used to apply the T... SVD The point cloud is finely registered; the distance threshold is dynamically adjusted based on the matching results of the current iteration. : Where, N valid To effectively match the number of points, N total The total number of points is given. B1 and B2 are threshold coefficients, which must satisfy B1>1>B2. β1 and β2 are effective matching point coefficients, which also satisfy β1<β2. Finally, the registration result at this voxel scale is obtained and used as the initial value for the next voxel iteration.

Images