A robust registration method for cross-scale heterogeneous three-dimensional point clouds
By optimizing the phased annealing similarity transformation corresponding to skeleton-guided coarse registration and feature-guided probability, the problem of inaccurate correspondence in cross-scale point cloud registration is solved, and the stability and accuracy of point cloud registration are improved.
Patent Information
- Application Number
- CN202611409073.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-11
- Publication Date
- 2026-11-06
AI Technical Summary
Existing point cloud registration techniques struggle to establish accurate and reliable correspondences when faced with cross-scale, density inconsistencies, local missing data, noise, and outliers, thus affecting the estimation of scale, rotation, and translation parameters.
We employ a method of skeleton-guided coarse registration, feature-guided probabilistic correspondence, and staged annealing similarity transformation optimization. Initial parameters are extracted through the topological skeleton, a feature-guided probabilistic correspondence model is constructed, and staged optimization is performed to gradually update the scale, rotation, and translation parameters.
It achieves stable establishment of correspondence in heterogeneous point clouds across scales, improves the stability and accuracy of point cloud registration, and reduces the impact of noise and outliers.
Smart Images
Figure CN123052010A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and 3D point cloud processing, and specifically to a robust registration method for cross-scale heterogeneous 3D point clouds. Background Technology
[0002] With the rapid development of 3D vision and intelligent sensing technologies, 3D point clouds have been widely used in fields such as 3D reconstruction, autonomous driving, robot perception, and industrial automation. Point cloud registration can unify point clouds from different viewpoints or data sources into the same coordinate system by estimating the spatial transformation relationship between different point clouds, and is a fundamental step in 3D data processing.
[0003] In practical applications, point clouds reconstructed using methods such as structure-of-motion, multi-view stereo vision, and 3D Gaussian splashing often exhibit scale differences compared to real-scale point clouds acquired by LiDAR or depth cameras. Furthermore, they may be accompanied by issues such as inconsistent point cloud density, local missing values, noise, and outliers. Therefore, cross-scale point cloud registration requires not only estimating the rotation and translation parameters between point clouds but also simultaneously recovering scale relationships. However, when point clouds exhibit significant scale differences, inconsistent density distributions, insufficient common areas, or numerous noise and outliers, current point cloud registration techniques often struggle to establish accurate and reliable correspondences, thus affecting the estimation of scale, rotation, and translation parameters. Summary of the Invention
[0004] The purpose of this invention is to provide a robust registration method for heterogeneous 3D point clouds across scales, which can stably establish point cloud correspondence under conditions such as scale differences, density inconsistencies, local missing points, and outliers, and achieve accurate estimation of scale, rotation, and translation parameters between the source point cloud and the target point cloud.
[0005] To achieve the above objectives, the present invention employs the following technical solution: A robust registration method for heterogeneous 3D point clouds across scales includes: Topological skeletons are extracted from the source point cloud and the target point cloud respectively. Based on the topological skeletons, the initial scale parameters and initial translation parameters between the source point cloud and the target point cloud are estimated, and coarse registration is performed on the source point cloud. Local geometric features are extracted from the source and target point clouds after coarse registration. Based on the spatial distance between points and the differences in local geometric features, a feature-guided probabilistic correspondence model is constructed to obtain the soft correspondence between the source and target point clouds. Based on the soft correspondence, a phased annealing similarity transformation optimization is performed, which sequentially performs feature-driven rigid transformation optimization, spatial variance annealing correspondence optimization, and scale-released similarity transformation optimization. The scale, rotation, and translation parameters between the source point cloud and the target point cloud are updated step by step to obtain the final cross-scale 3D point cloud registration result.
[0006] Furthermore, the process of coarsely registering the source point cloud based on the topological skeleton is as follows: 1. Calculate the centroids of the source point cloud and the target point cloud respectively, and center and normalize the source point cloud and the target point cloud according to the maximum radial distance from each point in the point cloud to the centroid, so as to reduce the impact of the original scale difference on the topology skeleton extraction process. 2. Apply Laplace shrinkage to the normalized source and target point clouds respectively. Iterate the shrinkage of the point clouds by using the discrete Laplace operator, shrinkage weights and position constraints, so that the point clouds gradually shrink towards their main structure. 3. For the shrunken point set, a graph structure is constructed and topology is refined to obtain the source point cloud topology skeleton and the target point cloud topology skeleton in the normalized coordinate system, respectively. The topology skeleton is then restored to the original point cloud coordinate system according to the centroid and scale parameters corresponding to the normalization process. 4. Calculate the skeleton edge lengths in the topological skeletons of the target point cloud and the source point cloud respectively, determine the maximum skeleton edge lengths of the two, and determine the initial scale parameters of the source point cloud relative to the target point cloud based on the ratio between the maximum skeleton edge length of the target point cloud and the maximum skeleton edge length of the source point cloud. 5. Determine the initial translation parameters based on the centroids of the source point cloud and the target point cloud and the initial scale parameters, and transform the source point cloud using the initial scale parameters and the initial translation parameters to obtain the coarsely registered source point cloud.
[0007] Furthermore, the construction of a feature-guided probabilistic correspondence model based on the spatial distance between points and the differences in local geometric features includes: The local geometric description features of each point's neighborhood in the source point cloud and target point cloud after coarse registration are calculated, and the differences in local geometric features are determined based on the differences between the corresponding local geometric description features of the source point and the target point. The source point after coarse registration is transformed based on the current residual scale parameters, residual rotation parameters, and residual translation parameters, and the spatial distance between the transformed source point and the target point is calculated. The spatial distance and local geometric feature differences are jointly introduced into the probabilistic correspondence model, where the spatial distance is used to describe the positional consistency between the source point and the target point, and the local geometric feature differences are used to describe the consistency of their neighborhood structures. The influence of local geometric features on the corresponding probability is adjusted by feature-guided weights. The correspondence probabilities between the source and target points are normalized, and an outlier term is introduced to reduce the impact of outliers on the correspondence estimation, resulting in a soft correspondence matrix between the source and target point clouds. The elements in the soft correspondence matrix are used to characterize the matching probability between the corresponding source and target points, providing a correspondence for subsequent similarity transformation parameter optimization.
[0008] Furthermore, the phased annealing similarity transformation optimization includes feature-driven rigid transformation optimization, spatial variance annealing correspondence optimization, and scale-released similarity transformation optimization.
[0009] Furthermore, the feature-driven rigid transformation optimization process is as follows: The residual scale parameter is fixed at 1, and a large initial spatial variance is set to reduce the constraint of spatial distance on the correspondence in the early stage of optimization. At the same time, a large feature guidance weight is set so that local geometric features play a major guiding role in establishing the correspondence between the source point cloud and the target point cloud in the early stage of optimization. Based on the current soft correspondence matrix, the weighted centroids of the source point cloud and the target point cloud are calculated respectively. The coordinates of the source point and the target point are centered and a weighted cross-covariance matrix is constructed according to the soft correspondence probability. Singular value decomposition is performed on the weighted cross-covariance matrix. The residual rotation parameters are obtained based on the singular value decomposition results. The residual translation parameters are then updated based on the residual rotation parameters and the weighted centroids of the source and target point clouds. Iterative updates of soft correspondence, residual rotation parameters, and residual translation parameters are used to reduce rotation and translation errors that still exist after coarse registration.
[0010] Furthermore, the optimization process for the correspondence of the spatial variance annealing is as follows: Keeping the residual scale parameter at 1, calculate the target spatial variance based on the current soft correspondence and the spatial distribution between the source point cloud and the target point cloud; Within a preset number of iterations, the spatial variance is gradually reduced from the initial spatial variance to the target spatial variance, thereby gradually strengthening the constraint effect of spatial distance in the corresponding probability estimation, and thus gradually concentrating the soft correspondence probability between the source point cloud and the target point cloud towards point pairs that are consistent in both spatial location and local structure. As the spatial variance gradually decreases, the soft correspondence matrix is continuously updated, and the residual rotation and translation parameters are iteratively calculated based on the updated soft correspondence matrix to gradually improve the accuracy of point cloud correspondences. The spatial variance is updated using linear annealing in this stage.
[0011] Furthermore, the similarity transformation optimization process for the scale release is as follows: After the soft correspondence stabilizes, the fixation of the residual scale parameter is lifted, and the residual scale parameter is updated according to the current soft correspondence and the weighted geometric statistics of the source point cloud and the target point cloud. The spatial variance is re-estimated based on the updated residual scale parameters, and the feature guidance weights are gradually reduced according to the preset decay method, so that the probabilistic correspondence model is gradually transformed from being guided by local geometric features to being dominated by spatial distance. Based on the updated soft correspondence matrix, the residual scale parameters, residual rotation parameters, and residual translation parameters are iteratively updated to achieve local fine similarity transformation registration between the source point cloud and the target point cloud.
[0012] Furthermore, the initial scale parameters and initial translation parameters obtained in the coarse registration stage are combined with the residual scale parameters, residual rotation parameters, and residual translation parameters obtained from the staged annealing similarity transformation optimization to obtain the final scale parameters, final rotation parameters, and final translation parameters from the source point cloud to the target point cloud. A similarity transformation is then performed on the source point cloud based on these final scale parameters, final rotation parameters, and final translation parameters to obtain the final 3D point cloud registration result. In this paper, the final transformation relationship is obtained by combining coarse registration and residual transformation.
[0013] A 3D point cloud registration system includes a skeleton-guided coarse registration module, a feature-guided probabilistic correspondence module, and an annealed similarity transformation optimization module. The skeleton-guided coarse registration module estimates initial scale parameters and initial translation parameters based on the topological skeletons of the source and target point clouds. The feature-guided probabilistic correspondence module establishes a soft correspondence between the source and target point clouds based on spatial distance and local geometric feature differences. The annealed similarity transformation optimization module optimizes the residual scale parameters, residual rotation parameters, and residual translation parameters in stages based on the soft correspondence and outputs the final point cloud registration result.
[0014] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor; when the processor executes the computer program, it implements the robust registration method for cross-scale heterogeneous 3D point clouds.
[0015] A computer-readable storage medium storing a computer program; when executed by a processor, the computer program implements the robust registration method for cross-scale heterogeneous 3D point clouds.
[0016] Compared with the prior art, the present invention has the following technical features: 1. Skeleton-guided initialization: The initial scale and translation parameters are estimated by using the scale-normalized topological skeleton, which reduces the impact of large scale differences on subsequent point cloud registration.
[0017] 2. Feature-guided probabilistic correspondence: Local geometric features are introduced into the probabilistic correspondence model, and soft correspondence is established by combining spatial location and local structural information, reducing erroneous matching caused by density differences, local missing data and outliers.
[0018] 3. Staged annealing optimization: Rigid optimization with fixed scale, spatial variance annealing, and similarity transformation optimization with scale release are performed sequentially to reduce the instability caused by simultaneous optimization of scale, rotation, and translation parameters.
[0019] 4. Progressive fine registration: Through an optimization process that gradually shifts from feature-guided to spatial distance-driven, a gradual transition from establishing reliable correspondences to local fine registration is achieved, improving the stability and accuracy of cross-scale heterogeneous point cloud registration. Attached Figure Description
[0020] Figure 1 This is a schematic flowchart of the method of the present invention; Figure 2 This is a visualization of the initial states of the source point cloud and the target point cloud in one embodiment of the present invention; Figure 3 This is a visualization of the topological skeleton extraction results of the source point cloud and the target point cloud in one embodiment of the present invention; Figure 4 This is a visualization of the skeleton-guided coarse registration result in one embodiment of the present invention; Figure 5 This is a visualization of the final registration result of a cross-scale 3D point cloud in one embodiment of the present invention. Detailed Implementation
[0021] This invention provides a robust registration method for heterogeneous 3D point clouds across scales. This method achieves joint estimation of scale, rotation, and translation parameters between the source and target point clouds through skeleton-guided coarse registration, feature-guided probabilistic correspondence, and staged annealing similarity transformation optimization. The overall process of this invention is as follows: Figure 1 As shown.
[0022] Step 1: Perform skeleton-guided coarse registration on the source point cloud and target point cloud to be registered.
[0023] Let the source point cloud be
[0024] Target point cloud
[0025] Where M and N represent the number of points in the source point cloud and the target point cloud, respectively. The objective of this invention is to estimate the similarity transformation parameters between the source and target point clouds, including the scale parameter s, the rotation matrix R, and the translation vector t, so that the source and target point clouds are spatially aligned after the similarity transformation. The similarity transformation parameters are expressed as follows:
[0026] The coordinates of the j-th point in the source point cloud after the final similarity transformation are represented as follows:
[0027] Where s is the scale parameter, R is the rotation matrix, and t is the translation vector.
[0028] To reduce the impact of large-scale differences on the subsequent probabilistic registration process, this invention divides the overall registration process into two parts: skeleton-guided coarse registration and residual similarity transformation optimization. First, the initial scale parameters are estimated using a topological skeleton. and initial translation parameters The source point after coarse registration is obtained:
[0029] Based on this, the residual scale parameters are then estimated. Residual rotation matrix and residual translation vector :
[0030] The final similarity transformation parameters are obtained by combining the coarse registration parameters and the residual similarity transformation parameters:
[0031] Step 1.1: Normalize the source point cloud and the target point cloud.
[0032] For any input point cloud P, first calculate its centroid. And calculate the maximum radial distance from each point in the point cloud to the centroid, with the centroid as the center. :
[0033] Where p represents any point in the point cloud P.
[0034] The point cloud is centered and scale-normalized based on the centroid and the maximum radial distance. The normalized point coordinates are represented as follows:
[0035] Through the above processing, the input point clouds of different scales are normalized to a similar scale range, reducing the impact of the original scale difference between the source point cloud and the target point cloud on the subsequent skeleton extraction process.
[0036] The source point cloud Y and the target point cloud X are normalized according to the above method to obtain the normalized source point cloud and target point cloud.
[0037] Step 1.2: Extract the topological skeleton from the normalized source point cloud and target point cloud.
[0038] A skeleton extraction method based on Laplacian contraction is used for both the normalized source point cloud and the target point cloud to gradually shrink the original point cloud toward its main structure.
[0039] Let the coordinate matrix of the normalized point cloud be... The coordinate matrix of the shrunk point set is The Laplace contraction process is represented as:
[0040] Where L represents the discrete Laplace operator, The shrinkage weight matrix represents the degree to which the point cloud shrinks towards the skeleton structure. This represents the position constraint weight matrix, used to prevent the shrunk points from deviating excessively from the original normalized point cloud.
[0041] By using Laplacian contraction, the point cloud is gradually concentrated into a one-dimensional structure describing its main geometric features. After obtaining the contracted point set, graph structure construction and topological refinement are performed on the point set to obtain the topological skeleton in the normalized coordinate system:
[0042] in, This represents the set of skeleton vertices in a normalized coordinate system. This represents the set of skeleton edges.
[0043] Step 1.3: Restore the topological skeleton to the original point cloud coordinate system.
[0044] Since the topological skeleton is extracted from the normalized point cloud, it is necessary to restore the skeleton vertices to the original point cloud coordinate system by following the reverse process of normalization.
[0045] For any skeleton vertex in the normalized coordinate system The restored coordinates v are represented as follows:
[0046] This yields the topological skeleton in the original coordinate system:
[0047] in, This represents the set of skeleton vertices in the original coordinate system. This represents the corresponding skeleton edge set.
[0048] Process the source point cloud and target point cloud according to the above process to obtain the topological skeleton of the source point cloud. and target point cloud topology skeleton The topological skeleton extraction results of the source and target point clouds are as follows: Figure 3 As shown.
[0049] Step 1.4: Estimate the initial scale parameters using the topological skeleton.
[0050] Calculate the Euclidean length of each skeleton edge in the target point cloud topology skeleton and the source point cloud topology skeleton respectively, and determine the length of the maximum skeleton edge in each case.
[0051] Maximum skeleton side length of the target point cloud topology skeleton Represented as:
[0052] Maximum skeleton side length of source point cloud topology skeleton Represented as:
[0053] in, and This represents the two endpoints of the corresponding skeleton edge.
[0054] The initial scale parameter is calculated based on the ratio between the maximum skeleton side lengths of the target point cloud and the source point cloud. :
[0055] The main scale relationship between the source point cloud and the target point cloud is obtained by the proportional relationship between the maximum skeleton structure lengths, so that the source point cloud and the target point cloud have similar overall scales before entering the subsequent probabilistic registration.
[0056] Step 1.5: Estimate the initial translation parameters and complete the coarse registration.
[0057] Calculate the centroids of the target point cloud and the source point cloud respectively. and According to the initial scale parameters Calculate the initial translation parameters :
[0058] The source point cloud is then transformed based on the initial scale parameters and initial translation parameters:
[0059] The source point cloud after coarse registration is obtained .
[0060] After the skeleton-guided coarse registration described above, the significant scale differences and overall positional deviations between the source and target point clouds are preliminarily corrected, providing initial conditions for subsequent local geometric feature extraction and probabilistic correspondence establishment. The results of the skeleton-guided coarse registration are as follows: Figure 4 As shown.
[0061] Step 2: Establish a probabilistic correspondence based on features for the source point cloud and target point cloud after coarse registration.
[0062] After step 1, the source and target point clouds have been preliminarily corrected in terms of overall scale and spatial location. However, residual scale, rotation, and translation errors may still exist between them. In addition, there may be problems such as density inconsistency, local missing points, and outliers in cross-source point clouds. It is difficult to obtain a stable and reliable correspondence based solely on the spatial distance between points.
[0063] Therefore, based on spatial distance constraints, this invention further introduces local geometric features and establishes a soft correspondence between the source point cloud and the target point cloud by jointly utilizing the spatial positional relationship of points and the similarity of local geometric structures.
[0064] Step 2.1: Extract local geometric features from the source point cloud and the target point cloud.
[0065] For the source point cloud after coarse registration Calculate the local geometric description features of the neighborhood of each point in the target point cloud X.
[0066] The set of local geometric features corresponding to the source point cloud after coarse registration is represented as follows:
[0067] The set of local geometric features corresponding to the target point cloud is represented as follows:
[0068] in, This represents the local geometric description features of the j-th source point. This represents the local geometric description features of the i-th target point.
[0069] In this embodiment, the FPFH local geometric descriptor is used to describe the neighborhood structure of the point cloud.
[0070] Step 2.2: Calculate the local geometric feature differences between the source point and the target point.
[0071] For the j-th source point and the i-th target point, the feature difference is calculated based on the Euclidean distance between their local geometric description features:
[0072] in, It is used to characterize the degree of inconsistency between the local neighborhood structures of the source point and the target point. The smaller the value, the more similar the local geometry of the two points.
[0073] Step 2.3: Calculate the spatial distance between the source point and the target point.
[0074] Based on the current residual scale parameters Residual rotation matrix and residual translation vector Residual similarity transformation is performed on the source points after coarse registration.
[0075] The spatial distance between the j-th source point and the i-th target point is represented as:
[0076] in, Used to describe the spatial difference between the source point and the target point after the current similarity transformation.
[0077] Step 2.4: Construct a feature-guided probabilistic correspondence model.
[0078] To ensure that the probability of correspondence between point pairs is constrained by both spatial location and local geometry, spatial distance is used. and differences in local geometric features A probability model is introduced together.
[0079] The unnormalized log probability between the source and target points is expressed as:
[0080] in, This represents the spatial variance, used to control the impact of the spatial distance term on the corresponding probability; This represents the feature-guided weight, used to adjust the impact of local geometric feature differences on the corresponding probability.
[0081] When the source point and the target point are relatively close in spatial location and have similar local geometric structures, their correspondence probability is high; conversely, when the spatial distance is large or the local structures are significantly different, their correspondence probability decreases.
[0082] Step 2.5: Normalize the corresponding probabilities and introduce outlier terms.
[0083] The unnormalized probability is further normalized to obtain the soft correspondence probability between the j-th source point and the i-th target point. :
[0084] Here, the summation term represents the probability normalization between the i-th target point and all source points, and c represents the outlier term.
[0085] The outlier term is represented as follows:
[0086] Where D represents the dimension of the point cloud data, and in this embodiment, the three-dimensional point cloud D=3; w represents the weight of outliers; M and N represent the number of points in the source point cloud and the target point cloud, respectively.
[0087] By setting an outlier term, points that cannot reliably correspond to another point cloud can be considered outliers, thereby reducing the interference of local missing points, non-overlapping regions, and outliers on similarity transformation estimation.
[0088] After calculating the soft correspondence probabilities for all source and target points, a soft correspondence matrix is constructed:
[0089] The soft correspondence matrix is used to describe the probabilistic correspondence between the source point cloud and the target point cloud, and serves as the basis for subsequent optimization of residual similarity transformation parameters.
[0090] Step 3: Perform phased annealing similarity transformation optimization based on soft correspondence.
[0091] While skeleton-guided coarse registration can reduce significant scale and positional differences between the source and target point clouds, the soft correspondence may still exhibit considerable uncertainty in the early stages of optimization. If scale, rotation, and translation parameters are updated simultaneously from the initial optimization phase, erroneous correspondences may lead to unreasonable changes in scale parameters, further disrupting subsequent correspondences.
[0092] Therefore, this invention adopts a phased annealing similarity transformation optimization strategy, which divides the residual similarity transformation optimization process into three stages: feature-driven rigid transformation optimization, spatial variance annealing correspondence optimization, and scale-released similarity transformation optimization.
[0093] Step 3.1: Initialize spatial variance and feature-guided weights.
[0094] Based on the source point cloud after coarse registration Calculate the initial spatial variance with the target point cloud X :
[0095] Where D represents the point cloud dimension, and N and M represent the number of points in the target point cloud and the source point cloud, respectively. This represents the maximum skeleton edge length in the target point cloud topology skeleton. This represents the initial coefficients of spatial variance.
[0096] In this embodiment Take 10.0.
[0097] Simultaneously set the initial feature guidance weights:
[0098] in, This represents the initial feature guidance coefficient, in this embodiment... Take 10.0.
[0099] A larger initial spatial variance can reduce the limitation of spatial distance on the corresponding probability in the early stage of optimization, while a larger feature-guided weight allows the local geometry to play a major role in the early correspondence establishment process.
[0100] Step 3.2: Perform feature-driven rigid transformation optimization.
[0101] At this stage, the residual scale parameter is fixed as follows:
[0102] This means that no further adjustments will be made to the scale; only the residual rotation matrix will be updated. and residual translation vector .
[0103] Calculate the weighted centroids of the target point cloud and the coarsely registered source point cloud based on the current soft correspondence matrix P, and then center the source point coordinates and the target point coordinates respectively.
[0104] Construct a weighted cross-covariance matrix A based on the soft correspondence probability:
[0105] in, This represents the centralized coordinates of the target point relative to the weighted centroid of the target point cloud. This represents the centered coordinates of the coarse registration source point relative to the weighted centroid of the source cloud.
[0106] Perform singular value decomposition on the weighted cross-covariance matrix A:
[0107] Update the residual rotation matrix based on the singular value decomposition results:
[0108] in:
[0109] Matrix C is used to avoid reflection transformations during rotation estimation.
[0110] After obtaining the residual rotation matrix, update the residual translation vector based on the weighted centroids of the source and target point clouds:
[0111] in, This represents the weighted centroid of the target point cloud. This represents the weighted centroid of the coarsely registered source point cloud.
[0112] In this stage, the soft correspondence matrix P is repeatedly calculated, and the results are iteratively obtained based on the updated soft correspondence. and This is to gradually reduce the remaining rotation and translation errors after coarse registration of the skeleton.
[0113] Step 3.3: Optimize the correspondence of spatial variance annealing.
[0114] After feature-driven rigid transformation optimization, the correspondence between the source point cloud and the target point cloud gradually stabilizes. At this point, the residual scale parameter is still maintained. And gradually reduce spatial variance This gradually strengthens the role of spatial distance in the probability estimation process.
[0115] Calculate the target space variance based on the current soft correspondence. :
[0116] in, This indicates the number of valid correspondences under the current soft correspondence. , and This represents the weighted second-order statistic calculated based on the current soft correspondence probability.
[0117] Within a preset number of iterations, the spatial variance is gradually reduced from the initial spatial variance at the beginning of this stage to the target spatial variance. This embodiment uses linear annealing, which gradually reduces the spatial variance with a fixed step size to avoid abrupt changes in the corresponding probability matrix.
[0118] along with As the distance between the source and target points decreases, the corresponding probability between them becomes more suppressed, causing the probability quality to gradually concentrate on point pairs that are more consistent in spatial location and local geometry.
[0119] During spatial variance annealing, the residual rotation matrix is continuously calculated based on the updated soft correspondence matrix. and residual translation vector Meanwhile, the residual scale parameters remain fixed.
[0120] Step 3.4: Perform similarity transformation optimization for scale release.
[0121] Once the soft correspondence has gradually stabilized after the aforementioned two stages, the restriction on the residual scale parameter is lifted. With the parameters fixed, we begin to simultaneously optimize the scale, rotation, and translation parameters.
[0122] Calculate the residual scale parameters based on the current soft correspondence:
[0123] The target space variance is then re-estimated based on the current residual scale:
[0124] Simultaneously, the influence of local geometric features on the corresponding probability estimation is gradually reduced. Feature-guided weights decay exponentially.
[0125] Where k represents the current iteration number, This indicates the starting position for the feature weight decay iteration. Let represent the feature weight decay coefficient, and satisfy:
[0126] With feature-guided weights As the distance gradually decreases, the influence of local geometric features on soft correspondences gradually weakens, while the role of spatial distance gradually strengthens. This causes the optimization process to gradually transition from relying on local geometric features to establish reliable correspondences over a large area in the early stage to local fine registration dominated by spatial distance.
[0127] In this stage, the residual scale parameters are iteratively obtained based on the updated soft correspondence matrix. Residual rotation matrix and residual translation vector This continues until the residual similarity transformation optimization is completed.
[0128] Step 4: Calculate the final similarity transformation parameters and output the registration results.
[0129] The initial scale parameters are obtained after step 1. and initial translation parameters The residual scale parameters are obtained through step 3. Residual rotation matrix and residual translation vector .
[0130] Combining the two-stage transformation yields the final scale parameters between the source point cloud and the target point cloud:
[0131] Final rotation matrix:
[0132] Final translation vector:
[0133] Based on the final scale parameter, final rotation matrix, and final translation vector, a similarity transformation is performed on each point in the original source point cloud:
[0134] The registered source point cloud is obtained and unified with the target point cloud to the same spatial coordinate system, thus completing the cross-scale heterogeneous 3D point cloud registration.
[0135] The initial states of the source point cloud and the target point cloud in one embodiment of the present invention are as follows: Figure 2 As shown, there are significant differences between the two in terms of scale, rotation, and translation; after skeleton-guided coarse registration, the overall scale and position of the source point cloud and the target point cloud are initially corrected, as shown in the figure. Figure 4 As shown; after further optimization through feature-guided probabilistic correspondence and phased annealing similarity transformation, the final cross-scale 3D point cloud registration result is obtained, as shown. Figure 5 As shown.
[0136] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. A robust registration method for heterogeneous 3D point clouds across scales, characterized in that, include: (1) Extract the topological skeleton of the source point cloud and the target point cloud, estimate the initial scale parameter and the initial translation parameter based on the topological skeleton, and perform coarse registration of the source point cloud based on the initial scale parameter and the initial translation parameter. (2) Extract local geometric features from the source point cloud and target point cloud after coarse registration, and construct a probabilistic correspondence model by combining the spatial distance between the source point and the target point and the difference in local geometric features to obtain the soft correspondence between the source point cloud and the target point cloud; (3) Based on the soft correspondence, perform phased annealing similarity transformation optimization, and sequentially perform rigid transformation optimization with fixed scale, spatial variance annealing optimization, and scale release similarity transformation optimization to obtain the final scale parameters, rotation parameters, and translation parameters, thereby realizing cross-scale registration between the source point cloud and the target point cloud.
2. The robust registration method for cross-scale heterogeneous 3D point clouds according to claim 1, characterized in that, The topological skeleton extraction and coarse registration include: (1) The source point cloud and the target point cloud are centered and scaled respectively; (2) Perform Laplacian shrinkage on the normalized source point cloud and target point cloud, and obtain the corresponding topological skeleton through graph structure construction and topological refinement; (3) Restore the topological skeleton to the original point cloud coordinate system; (4) Determine the initial scale parameters based on the ratio of the maximum skeleton side lengths of the source point cloud and the target point cloud; (5) Determine the initial translation parameters based on the initial scale parameters and the centroids of the source point cloud and the target point cloud.
3. The robust registration method for cross-scale heterogeneous 3D point clouds according to claim 2, characterized in that, The source point cloud is transformed according to the initial scale parameters and initial translation parameters to obtain the coarsely registered source point cloud, which provides initial conditions for the subsequent establishment of probability correspondence.
4. The robust registration method for cross-scale heterogeneous 3D point clouds according to claim 1, characterized in that, The construction of the probability correspondence model includes: (1) Extract the local geometric description features of the source point cloud and the target point cloud after coarse registration respectively; (2) Calculate the differences in local geometric features between the source point and the target point; (3) Calculate the spatial distance between the source point and the target point based on the current residual similarity transformation parameters; (4) The spatial distance and local geometric feature differences are introduced into the probability model, and the influence of local geometric features on the corresponding probability is adjusted by feature-guided weights; (5) Normalize the corresponding probabilities and introduce outlier terms to obtain the soft correspondence matrix between the source point cloud and the target point cloud.
5. The robust registration method for cross-scale heterogeneous 3D point clouds according to claim 4, characterized in that, The elements in the soft correspondence matrix represent the matching probability between the corresponding source point and the target point, and are used to estimate the residual similarity transformation parameters.
6. The robust registration method for cross-scale heterogeneous 3D point clouds according to claim 1, characterized in that, The phased annealing similarity transformation optimization includes: (1) In the first stage, the residual scale parameters are fixed, and a large spatial variance and feature-guided weights are set. The residual rotation parameters and residual translation parameters are updated according to the soft correspondence. (2) In the second stage, the residual scale parameters are kept fixed, the spatial variance is gradually reduced, the spatial distance constraint is gradually strengthened, and the residual rotation parameters and residual translation parameters are continuously updated. (3) In the third stage, the residual scale parameters are released and the feature guidance weights are gradually reduced, and the residual scale parameters, residual rotation parameters and residual translation parameters are jointly updated.
7. The robust registration method for cross-scale heterogeneous 3D point clouds according to claim 6, characterized in that: (1) In the first and second stages, the residual scale parameter is fixed at 1; (2) Calculate the weighted centroids of the source point cloud and the target point cloud based on the current soft correspondence matrix, and construct the weighted cross-covariance matrix; (3) Perform singular value decomposition on the weighted cross-covariance matrix and determine the residual rotation parameters based on the singular value decomposition results; (4) Determine the residual translation parameters based on the residual rotation parameters and the weighted centroids of the source point cloud and the target point cloud.
8. The robust registration method for cross-scale heterogeneous 3D point clouds according to claim 6, characterized in that: (1) In the third stage, the residual scale parameters and spatial variance are updated based on the current soft correspondence; (2) The feature guidance weights are gradually reduced according to the preset decay method, so that the optimization process is gradually changed from being guided by local geometric features to being dominated by spatial distance; (3) Combine the residual scale parameters, residual rotation parameters and residual translation parameters with the initial scale parameters and initial translation parameters obtained in the coarse registration stage to obtain the final similarity transformation parameters between the source point cloud and the target point cloud.
9. A three-dimensional point cloud registration system, characterized in that, include: (1) Skeleton-guided coarse registration module, used to extract the topological skeleton of the source point cloud and the target point cloud, and estimate the initial scale parameters and initial translation parameters based on the topological skeleton; (2) Feature-guided probability correspondence module, used to establish a soft correspondence based on the spatial distance and local geometric feature differences between the source point cloud and the target point cloud after coarse registration; (3) Annealing similarity transformation optimization module, used to perform phased annealing similarity transformation optimization according to the soft correspondence, obtain the final similarity transformation parameters and output the three-dimensional point cloud registration result.
10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; characterized in that, When the processor executes the computer program, it implements the robust registration method for cross-scale heterogeneous 3D point clouds according to any one of claims 1-8.