Unsupervised point cloud registration method and system based on reversible shape-space decoupling, computer readable storage medium and computer program product
By employing a reversible shape-space decoupling method, utilizing the alignment features of IFDM and Gaussian similarity matrix, and combining the least squares method of singular value decomposition to calculate the rigid transformation matrix, the problems of unstable matching and poor robustness in unsupervised point cloud registration are solved, achieving higher registration accuracy and robustness.
Patent Information
- Application Number
- CN202511025442.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-23
- Publication Date
- 2025-11-14
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In existing unsupervised point cloud registration methods, the mixed coupling of shape and spatial features leads to unstable matching, and the corresponding points have poor resistance to disturbances due to their dependence on the original coordinates.
A reversible shape-space decoupling method is adopted, which decouples the point cloud into shape-space features through IFDM, aligns the features using Gaussian similarity matrix, and calculates the rigid transformation matrix by combining the least squares method of singular value decomposition for registration.
Explicitly decoupling transformation-invariant shape and spatial features improves matching accuracy and robustness, avoids direct dependence on the original spatial coordinates, and enhances the expressiveness and reversibility of features.
Smart Images
Figure CN120953330A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to an unsupervised point cloud registration method, system, computer-readable storage medium, and computer program product based on reversible shape-space decoupling. Background Technology
[0002] The 3D point cloud registration task aims to calculate the rigid transformation matrix between two point clouds that partially overlap, for applications such as target recognition, map building, and augmented reality.
[0003] Existing 3D point cloud registration methods include supervised and unsupervised methods. Generally, supervised methods rely on the true registration pose, that is, extracting transformation-invariant features by forcing feature similarity through spatial distance. However, in unsupervised scenarios, the lack of true pose labels makes it difficult to extract stable features.
[0004] Existing unsupervised methods typically model local geometry using graph neural networks, estimating direct corresponding points through relative coordinates. However, these methods still face two main problems: first, the mixed coupling of shape and spatial features leads to unstable matching; second, the corresponding points depend on the original coordinates, resulting in poor robustness to disturbances.
[0005] Therefore, for those skilled in the art, how to extract transformation-invariant shape features and recoverable spatial features in unsupervised point cloud registration tasks, and improve the accuracy and robustness of corresponding point matching, has become a technical problem that urgently needs to be solved. Summary of the Invention
[0006] The purpose of this invention is to provide an unsupervised point cloud registration method, system, computer-readable storage medium, and computer program product based on reversible shape-space decoupling, so as to solve or at least partially solve the technical problems mentioned in the background art.
[0007] To achieve this objective, the present invention adopts the following technical solution:
[0008] In a first aspect, the present invention provides an unsupervised point cloud registration method based on reversible shape-space decoupling, comprising:
[0009] The source point cloud and the target point cloud are decoupled into shape-space features using IFDM; wherein, the shape-space features include shape features and spatial features.
[0010] Calculate the Gaussian similarity matrix using the shape features of the source point cloud and the target point cloud;
[0011] Based on the Gaussian similarity matrix, the spatial features of the source point cloud and the target point cloud are aligned to obtain the transformed point cloud;
[0012] Based on the source point cloud and the transformed point cloud, the rigid transformation matrix is calculated using the least squares method based on singular value decomposition.
[0013] The source point cloud and the target point cloud are registered using a rigid transformation matrix.
[0014] Optionally, the step of decoupling the source point cloud and the target point cloud into shape-space features using IFDM specifically includes:
[0015] Let the source point cloud be P∈R m×3 The target point cloud is Q∈R n×3 ;
[0016] P and Q are mapped to the latent space using a trained invertible neural network, as follows:
[0017]
[0018] Among them, f θ (·) represents an invertible neural network, Z P Represents the potential space of P. Let P be the shape feature to be decoupled. Z represents the spatial characteristics of P to be decoupled; Q Represents the potential space of Q. The shape features of Q to be decoupled The spatial characteristics of Q to be decoupled;
[0019] IFDM uses structure-aware map convolution to decouple the latent spaces mapped by P and Q respectively, and obtain the shape-space features of P and Q.
[0020] Optionally, the IFDM uses structure-aware map convolution to decouple the latent spaces mapped by P and Q respectively, obtaining the shape-space features of P and Q, including:
[0021] IFDM uses structure-aware graph convolution based on MLP to construct a static KNN graph and introduces edge structure embedding, as follows:
[0022] Let G = (V, E) be a KNN graph, where V is the set of points in the point cloud, and E is the set of edges connecting each point in the point cloud to its neighbors; the elements in E are the edges. The calculation method is as follows:
[0023] in, edge vectors modulus, As the reference direction, Θ(·) represents the MLP. edge vectors The angle between the reference direction and the reference direction.
[0024] Optionally, the method for calculating the Gaussian similarity matrix using the shape features of the source point cloud and the target point cloud is as follows:
[0025] Let the Gaussian similarity matrix be S∈R m×n The method for calculating S is as follows:
[0026]
[0027] Among them, s ij Let S be a matrix element, representing the i-th point P in P. i And the j-th point Q in Q j Similarity; For P i The corresponding shape features For Q j The corresponding shape features;
[0028] After calculating the Gaussian similarity matrix using the shape features of the source and target point clouds, the method further includes:
[0029] Normalizing and smoothing the Gaussian similarity matrix S using neighborhood averaging yields the optimized similarity matrix.
[0030] Optionally, the step of aligning the spatial features of the source point cloud and the target point cloud based on the Gaussian similarity matrix to obtain the transformed point cloud specifically includes:
[0031] Based on the optimized similarity matrix and the spatial features of the target point cloud, the approximate spatial correspondence features are... The method is as follows:
[0032] splicing and get
[0033] according to The transformed point cloud P′ is obtained by calculating the inverse process of IFDM, as follows:
[0034]
[0035] Optionally, the method for calculating the rigid transformation matrix using the least squares method based on singular value decomposition based on the source point cloud and the transformed point cloud is as follows:
[0036] Let the rigid transformation matrix be... Among them, R * ∈SO(3) is a rotation matrix, t * ∈R 3 R is a translation matrix; * , t *The calculation method is as follows:
[0037]
[0038] Where R is R * The matrix elements, t is t * Matrix elements; C * For the corresponding points in (P, P′) A set of.
[0039] Optionally, before decoupling the source point cloud and the target point cloud into shape-space features using IFDM, the method further includes:
[0040] Prepare multiple sets of training samples in advance, each set of samples containing one source point cloud sample and one target point cloud sample;
[0041] A reversible neural network is constructed, and the multiple sets of training samples are input into the constructed reversible neural network for training.
[0042] Optionally, in the process of constructing a reversible neural network and inputting the multiple sets of training samples into the constructed reversible neural network for training, the loss function is designed as follows:
[0043] Total loss L = L g +aL p +bL n +cL s , where L g For global registration loss, L p For spatial consistency loss, L n For the loss of regional consistency, L s For distributed convergence loss, a, b, and c are all greater than 0 and less than or equal to 1;
[0044] L g The calculation method is as follows:
[0045]
[0046] Where μ is a variable, β is a user-defined real number, P′ is the transformed point cloud, and Q is the target point cloud;
[0047] L p The calculation method is as follows:
[0048]
[0049] Where K(·) represents the KNN aggregation operator based on spatial features, and m is the number of points in P;
[0050] L n The calculation method is as follows:
[0051] N is the set of k nearest neighbors;
[0052] L s The calculation method is as follows:
[0053] in, for The matrix elements are: I, an indicator function operator, and m, the number of points in P.
[0054] Optionally, a, b, and c are 1.0, 0.8, and 0.6, respectively.
[0055] Optionally, β is defined as 0.1.
[0056] Secondly, the present invention provides an unsupervised point cloud registration system based on reversible shape-space decoupling, comprising:
[0057] A reversible feature decoupling module is used to decouple the source point cloud and the target point cloud into shape-space features respectively through IFDM; wherein, the shape-space features include shape features and spatial features;
[0058] The feature alignment module, electrically connected to the reversible feature decoupling module, is used to calculate a Gaussian similarity matrix using the shape features of the source point cloud and the target point cloud; and to align the spatial features of the source point cloud and the target point cloud based on the Gaussian similarity matrix to obtain a transformed point cloud.
[0059] The registration attitude estimation module is electrically connected to the feature alignment module. It is used to calculate the rigid transformation matrix using the least squares method based on singular value decomposition based on the source point cloud and the transformed point cloud. It is also used to register the source point cloud and the target point cloud using the rigid transformation matrix.
[0060] Optionally, the reversible feature decoupling module is specifically used to map P and Q to latent spaces respectively through a trained reversible neural network, and to decouple the latent spaces mapped by P and Q respectively to obtain the shape-space features of P and Q;
[0061] The latent space mapping method is as follows:
[0062] Let the source point cloud be P∈R m×3 The target point cloud is Q∈R n×3 ;
[0063]
[0064] Among them, f θ (·) represents an invertible neural network, Z P Represents the potential space of P. Let P be the shape feature to be decoupled. Z represents the spatial characteristics of P to be decoupled; Q Represents the potential space of Q. The shape features of Q to be decoupled The spatial characteristics of Q to be decoupled;
[0065] The decoupling method for shape-space features is as follows:
[0066] IFDM uses structure-aware map convolution to decouple the latent spaces mapped by P and Q respectively, and obtain the shape-space features of P and Q.
[0067] Optionally, the IFDM uses structure-aware map convolution to decouple the latent spaces mapped by P and Q respectively, obtaining the shape-space features of P and Q, including:
[0068] IFDM uses structure-aware graph convolution based on MLP to construct a static KNN graph and introduces edge structure embedding, as follows:
[0069] Let G = (V, E) be a KNN graph, where V is the set of points in the point cloud, and E is the set of edges connecting each point in the point cloud to its neighbors; the elements in E are the edges. The calculation method is as follows:
[0070] in, edge vectors modulus, As the reference direction, Θ(·) represents the MLP. edge vectors The angle between the reference direction and the reference direction.
[0071] Optionally, the method for calculating the Gaussian similarity matrix using the shape features of the source point cloud and the target point cloud is as follows:
[0072] Let the Gaussian similarity matrix be S∈R m×n The method for calculating S is as follows:
[0073]
[0074] Among them, s ij Let S be a matrix element, representing the i-th point P in P. i And the j-th point Q in Q j Similarity; For P i The corresponding shape features For Q j The corresponding shape features;
[0075] The feature alignment module is also used to normalize and smooth the Gaussian similarity matrix S using neighborhood averaging to obtain an optimized similarity matrix.
[0076] Optionally, the feature alignment module is specifically used to approximate spatially corresponding features based on the optimized similarity matrix and the spatial features of the target point cloud.
[0077] Also used for splicing and get
[0078] Also used according to The transformed point cloud P′ is obtained by inverse IFDM calculation; wherein, the calculation method of P′ is as follows:
[0079]
[0080] Optionally, the method for calculating the rigid transformation matrix is as follows:
[0081] Let the rigid transformation matrix be... Among them, R * ∈SO(3) is a rotation matrix, t * ∈R 3 R is a translation matrix; * , t * The calculation method is as follows:
[0082]
[0083] Where R is R * The matrix elements, t is t * Matrix elements; C * For the corresponding points in (P, P′) A set of.
[0084] Optionally, the unsupervised point cloud registration system further includes a neural network construction unit;
[0085] The neural network building unit is used to build a reversible neural network and input multiple pre-prepared training samples into the built reversible neural network for training; wherein, each set of samples includes a source point cloud sample and a target point cloud sample.
[0086] Optionally, the neural network building unit is further configured to design the loss function of the invertible neural network, wherein the loss function is specifically designed as follows:
[0087] Total loss L = L g +aL p +bL n +cL s , where L g For global registration loss, L p For spatial consistency loss, Ln For the loss of regional consistency, L s For distributed convergence loss, a, b, and c are all greater than 0 and less than or equal to 1;
[0088] L g The calculation method is as follows:
[0089]
[0090] Where μ is a variable, β is a user-defined real number, P′ is the transformed point cloud, and Q is the target point cloud;
[0091] L p The calculation method is as follows:
[0092]
[0093] Where K(·) represents the KNN aggregation operator based on spatial features, and m is the number of points in P;
[0094] L n The calculation method is as follows:
[0095] N is the set of k nearest neighbors;
[0096] L s The calculation method is as follows:
[0097] in, for The matrix elements are: I, an indicator function operator, and m, the number of points in P.
[0098] Optionally, a, b, and c are 1.0, 0.8, and 0.6, respectively.
[0099] Optionally, β is defined as 0.1.
[0100] Thirdly, the present invention also provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement an unsupervised point cloud registration method based on reversible shape-space decoupling as described above.
[0101] Fourthly, the present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the above-described unsupervised point cloud registration method based on reversible shape-space decoupling.
[0102] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0103] The unsupervised point cloud registration method based on reversible shape-space decoupling provided in this application applies IFDM to point cloud registration, explicitly decoupling transformation-invariant shape features and spatial features, thereby enhancing the expressiveness and reversibility of features; the feature alignment process is based on shape correlation, avoiding direct reliance on the original spatial coordinates and improving matching robustness. Attached Figure Description
[0104] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0105] Figure 1 The flowchart illustrates an unsupervised point cloud registration method based on reversible shape-space decoupling, as provided in an embodiment of the present invention.
[0106] Figure 2 This is a schematic diagram of the source point cloud and the target point cloud provided in an embodiment of the present invention.
[0107] Figure 3 This is a schematic diagram of the decoupled shape features and spatial features provided in an embodiment of the present invention.
[0108] Figure 4 This is a schematic diagram illustrating the alignment of spatial features using a similarity matrix, as provided in an embodiment of the present invention.
[0109] Figure 5 This is a partial point cloud registration result diagram on the dataset provided in this embodiment of the invention.
[0110] Figure 6 This is a schematic diagram of the architecture of an unsupervised point cloud registration system based on reversible shape-space decoupling, provided for an embodiment of the present invention. Detailed Implementation
[0111] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0112] Example 1:
[0113] For details, please refer to Figure 1 , Figure 1A flowchart of an unsupervised point cloud registration method based on reversible shape-space decoupling provided in this embodiment of the invention is shown. The method specifically includes:
[0114] Step 110: Decouple the source point cloud and the target point cloud into shape-space features using IFDM.
[0115] Among them, IFDM (Invertible Feature Decoupling Model) is a generative model that combines invertible transformation and feature decoupling. It aims to decompose the implicit representation of data into independent and interpretable factors through bidirectional mapping, supporting lossless reconstruction and controllable generation. In this embodiment, IFDM is used to decouple the point cloud (all points) into shape-space features, i.e., shape features and spatial features; the method is as follows:
[0116] Let the source point cloud be P∈R m×3 The target point cloud is Q∈R n×3 ;
[0117] P and Q are mapped to the latent space using a trained invertible neural network (i.e., the backbone network of IFDM), as follows:
[0118]
[0119] Among them, f θ (·) represents an invertible neural network, Z P Represents the potential space of P. Let P be the shape feature to be decoupled. Z represents the spatial characteristics of P to be decoupled; Q Represents the potential space of Q. The shape features of Q to be decoupled The spatial characteristics of Q to be decoupled;
[0120] In this step, IFDM uses SAGC (Structure-Aware Graph Convolution) to decouple the latent spaces mapped by P and Q, respectively, to obtain the shape-space features of P and Q, specifically including:
[0121] IFDM uses structure-aware graph convolution based on MLP (Multilayer Perceptron) to construct a static KNN (K-Nearest Neighbors) graph and introduces edge structure embedding, as follows:
[0122] The neighborhood relationships of each point are obtained by calculating the Euclidean distance between points in the point cloud. Let the KNN graph be G = (V, E), where V is the set of points in the point cloud and E is the set of edges connecting each point in the point cloud to its neighbors.
[0123] Specifically, the elements in E, i.e., the edges The calculation method is as follows:
[0124] in, edge vectors modulus, As the reference direction, Θ(·) represents the MLP. edge vectors The angle between the reference direction and the reference direction.
[0125] In this step, the KNN graph significantly improves the quality and interpretability of feature decoupling by modeling the local structural relationships between data points. Since static graphs are computationally efficient, a static KNN graph is used in this step. In the context of reversible feature decoupling, introducing edge structure embedding can enhance the model's ability to model local structures, thereby better decoupling features. SAGC can more precisely utilize the structural information of the graph to guide the neighbor aggregation process, thereby improving the learning ability of the graph structure.
[0126] It should be noted that MLP, structure-aware graph convolution, and static KNN graph are all conventional techniques in this field, so their principles will not be elaborated in this embodiment.
[0127] Step 120: Calculate the Gaussian similarity matrix using the shape features of the source point cloud and the target point cloud.
[0128] The specific method for calculating the Gaussian similarity matrix is as follows:
[0129] Let the Gaussian similarity matrix be S∈R m×n The method for calculating S is as follows:
[0130]
[0131] Among them, s ij Let S be a matrix element, representing the i-th point P in P. i And the j-th point Q in Q j Similarity; For P i The corresponding shape features For Q j The corresponding shape features.
[0132] After step 120, the following is also included:
[0133] Step 121: Normalize and smooth the Gaussian similarity matrix S using neighborhood averaging to obtain the optimized similarity matrix.
[0134] This step improves the quality of data representation and enhances the robustness and generalization ability of the model by normalizing the Gaussian similarity matrix and applying neighborhood average smoothing. Since normalization and neighborhood average smoothing are conventional smoothing filtering methods in this field, their implementation process will not be described in detail in this embodiment.
[0135] Step 130: Based on the Gaussian similarity matrix, align the spatial features of the source point cloud and the target point cloud to obtain the transformed point cloud.
[0136] Step 130 includes:
[0137] Step 131: Based on the optimized similarity matrix and the spatial features of the target point cloud, approximate the spatial corresponding features. The method is as follows:
[0138] Step 132, splicing and get Where cat is the data concatenation operator;
[0139] Step 133, according to The coordinates of the corresponding points are recovered through the inverse process of IFDM to obtain the transformed point cloud P′. The method is as follows:
[0140]
[0141] Step 140: Based on the source point cloud and the transformed point cloud, calculate the rigid transformation matrix using the least squares method based on singular value decomposition.
[0142] Specifically, the method for calculating the rigid transformation matrix is as follows:
[0143] Let the rigid transformation matrix be... Among them, R * ∈SO(3) is a rotation matrix, t * ∈R 3 R is a translation matrix; * , t * The calculation method is as follows:
[0144]
[0145] Where R is R * The matrix elements, t is t * Matrix elements; C * For the corresponding points in (P, P′) A set of.
[0146] Therefore, according to the above calculation formula, if the corresponding points in (P, P′) are known... By finding the coordinates, we can calculate R and t, and then calculate the rigid transformation matrix.
[0147] Step 150: Register the source point cloud and the target point cloud using a rigid transformation matrix.
[0148] Using a rigid transformation matrix for registration ensures that the distance between points remains unchanged before and after the transformation; that is, there is only rotation and translation, without scaling.
[0149] For example, to facilitate understanding of the above unsupervised point cloud registration method based on reversible shape-space decoupling, please refer to... Figures 2-5 , Figure 2 This is a schematic diagram of the source point cloud and the target point cloud provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the decoupled shape and spatial features provided in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the alignment of spatial features using a similarity matrix, as provided in an embodiment of the present invention.
[0150] Furthermore, prior to step 110, the following steps are also included:
[0151] Pre-set step 100: Prepare multiple sets of training samples in advance.
[0152] Each sample group contains one source point cloud sample and one target point cloud sample.
[0153] Preliminary step 101: Construct a reversible neural network and input multiple sets of training samples into the constructed reversible neural network for training.
[0154] The construction of a reversible neural network involves the design of a loss function, as detailed below:
[0155] Let the total loss be L = L g +aL p +bL n +cL s , where L g For global registration loss, L p For spatial consistency loss, L n For the loss of regional consistency, L s For distributed convergence loss, a, b, and c are all greater than 0 and less than or equal to 1;
[0156] L g The calculation method is as follows:
[0157]
[0158] Where μ is a variable, β is a user-defined real number, P′ is the transformed point cloud, and Q is the target point cloud;
[0159] L p The calculation method is as follows:
[0160]
[0161] Where K(·) represents the KNN aggregation operator based on spatial features, and m is the number of points in P;
[0162] L n The calculation method is as follows:
[0163] N is the set of k nearest neighbors;
[0164] L s The calculation method is as follows:
[0165] in, for The matrix elements are: I, an indicator function operator, and m, the number of points in P.
[0166] In this embodiment, a, b, and c are 1.0, 0.8, and 0.6, respectively.
[0167] In this embodiment, β is defined as 0.1.
[0168] After the network is constructed, it needs to be trained. The training process follows the unsupervised point cloud registration method based on reversible shape-space decoupling described above, and the optimal network parameters of the reversible neural network are obtained through training. Specific training data can use point cloud data from multiple datasets such as ModelNet40, 7Scenes, ICL-NUIM, and KITTI as training samples. For example, this embodiment uses a portion of the point cloud data from the ModelNet40 dataset as samples.
[0169] To verify the accuracy and robustness of the registration, this embodiment also uses another portion of the point cloud data from the ModelNet40 dataset as a verification sample. Please refer to [link / reference needed]. Figure 5 , Figure 5 This is a partial point cloud registration result illustration from the dataset provided in this embodiment of the invention. Figure 5 It can be seen that the registration method has high accuracy and robustness, and has achieved the expected results.
[0170] In summary, compared with existing technologies, the unsupervised point cloud registration method based on reversible shape-space decoupling provided in this embodiment has at least the following beneficial effects:
[0171] 1) For the first time, reversible neural networks are applied to point cloud registration, explicitly decoupling transformation-invariant shape features and spatial features, thereby enhancing the expressiveness and reversibility of features;
[0172] 2) The feature alignment process is based on shape correlation, which avoids direct reliance on the original spatial coordinates and improves the robustness of matching;
[0173] 3) The proposed structure-aware graph convolution utilizes a dual structure of edge angle and edge length to enhance local topological expressiveness;
[0174] 4) Spatial consistency loss ensures that the local structure of corresponding points is consistent, further improving registration accuracy.
[0175] 5) It achieves high registration accuracy on datasets such as ModelNet40, verifying its practicality.
[0176] Example 2:
[0177] Please refer to Figure 6 , Figure 5 This is a schematic diagram illustrating the architecture of an unsupervised point cloud registration system based on reversible shape-space decoupling, provided for an embodiment of the present invention. The system specifically includes: a registration unit 10 and a neural network construction unit 20;
[0178] The registration unit 10 specifically includes:
[0179] The reversible feature decoupling module 11 is used to decouple the source point cloud and the target point cloud into shape-space features through IFDM; wherein, the shape-space features include shape features and spatial features;
[0180] The feature alignment module 12 and the electrically connected reversible feature decoupling module 11 are used to calculate the Gaussian similarity matrix using the shape features of the source point cloud and the target point cloud; and to align the spatial features of the source point cloud and the target point cloud based on the Gaussian similarity matrix to obtain the transformed point cloud.
[0181] The registration attitude estimation module 13 and the electrical connection feature alignment module 12 are used to calculate the rigid transformation matrix using the least squares method based on singular value decomposition based on the source point cloud and the transformed point cloud; they are also used to register the source point cloud and the target point cloud using the rigid transformation matrix.
[0182] Specifically, the reversible feature decoupling module 11 is used to map P and Q to latent spaces through the trained reversible neural network, and to decouple the latent spaces mapped by P and Q to obtain the shape-space features of P and Q.
[0183] It should be noted that since the latent space mapping method, the shape-space feature decoupling method, and the Gaussian similarity matrix calculation method have all been described in detail in Example 1, they will not be repeated in this example.
[0184] Feature alignment module 12 is also used to normalize and smooth the Gaussian similarity matrix S by neighborhood averaging to obtain the optimized similarity matrix.
[0185] Specifically, the feature alignment module 12 is used to approximate spatially corresponding features based on the optimized similarity matrix and the spatial features of the target point cloud.
[0186] Also used for splicing and get
[0187] Also used according to The transformed point cloud P′ is obtained by inverse IFDM calculation; wherein, the calculation method of P′ is as follows:
[0188]
[0189] It should be noted that the calculation method of the rigid transformation matrix has been described in detail in Example 1, so it will not be repeated in this example.
[0190] Specifically, the unsupervised point cloud registration system also includes neural network building units 20;
[0191] The neural network building unit 20 is used to build a reversible neural network and input multiple pre-prepared training samples into the built reversible neural network for training; wherein, each set of samples contains a source point cloud sample and a target point cloud sample.
[0192] Specifically, the neural network building unit 20 is also used to design the loss function of the invertible neural network;
[0193] It should be noted that the specific design of the loss function has been described in detail in Example 1, so it will not be repeated in this example.
[0194] In summary, the unsupervised point cloud registration system based on reversible shape-space decoupling provided in this application applies IFDM to point cloud registration, explicitly decoupling transformation-invariant shape features and spatial features, thereby enhancing the expressiveness and reversibility of features; the feature alignment process is based on shape correlation, avoiding direct reliance on the original spatial coordinates and improving matching robustness.
[0195] Example 3:
[0196] This embodiment also provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement an unsupervised point cloud registration method based on reversible shape-space decoupling as described in Embodiment 1.
[0197] Since Example 1 has already described the unsupervised point cloud registration method based on reversible shape-space decoupling in detail, it will not be repeated in this example.
[0198] Example 4:
[0199] The present invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements an unsupervised point cloud registration method based on reversible shape-space decoupling as described in Embodiment 1.
[0200] Since Example 1 has already described the unsupervised point cloud registration method based on reversible shape-space decoupling in detail, it will not be repeated in this example.
[0201] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware, or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0202] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 the present invention.
Claims
1. An unsupervised point cloud registration method based on IFDM, characterized in that, include: The source point cloud and the target point cloud are decoupled into shape-space features using IFDM; wherein, the shape-space features include shape features and spatial features. Calculate the Gaussian similarity matrix using the shape features of the source point cloud and the target point cloud; Based on the Gaussian similarity matrix, the spatial features of the source point cloud and the target point cloud are aligned to obtain the transformed point cloud; Based on the source point cloud and the transformed point cloud, the rigid transformation matrix is calculated using the least squares method based on singular value decomposition. The source point cloud and the target point cloud are registered using a rigid transformation matrix.
2. The unsupervised point cloud registration method based on IFDM according to claim 1, characterized in that, The step of decoupling the source point cloud and the target point cloud into shape-space features using IFDM specifically includes: Let the source point cloud be P∈R m×3 The target point cloud is Q∈R n×3 ; P and Q are mapped to the latent space using a trained invertible neural network, as follows: Among them, f θ (·) represents an invertible neural network, Z P Represents the potential space of P. Let P be the shape feature to be decoupled. Z represents the spatial characteristics of P to be decoupled; Q Represents the potential space of Q. The shape features of Q to be decoupled The spatial characteristics of Q to be decoupled; IFDM uses structure-aware map convolution to decouple the latent spaces mapped by P and Q respectively, and obtain the shape-space features of P and Q.
3. The unsupervised point cloud registration method based on IFDM according to claim 2, characterized in that, The IFDM uses structure-aware map convolution to decouple the latent spaces mapped by P and Q, respectively, to obtain the shape-space features of P and Q, including: IFDM uses structure-aware graph convolution based on MLP to construct a static KNN graph and introduces edge structure embedding, as follows: Let G = (V, E) be a KNN graph, where V is the set of points in the point cloud, and E is the set of edges connecting each point in the point cloud to its neighbors; the elements in E are the edges. The calculation method is as follows: in, edge vectors modulus, As the reference direction, Θ(·) represents the MLP. edge vectors The angle between the reference direction and the reference direction.
4. The unsupervised point cloud registration method based on IFDM according to claim 3, characterized in that, The method for calculating the Gaussian similarity matrix using the shape features of the source and target point clouds is as follows: Let the Gaussian similarity matrix be S∈R m×n The method for calculating S is as follows: Among them, s ij Let S be a matrix element, representing the i-th point P in P. i And the j-th point Q in Q j Similarity; For P i The corresponding shape features For Q j The corresponding shape features; After calculating the Gaussian similarity matrix using the shape features of the source and target point clouds, the method further includes: Normalizing and smoothing the Gaussian similarity matrix S using neighborhood averaging yields the optimized similarity matrix.
5. The unsupervised point cloud registration method based on IFDM according to claim 4, characterized in that, The process of aligning the spatial features of the source point cloud and the target point cloud based on the Gaussian similarity matrix to obtain the transformed point cloud specifically includes: Based on the optimized similarity matrix and the spatial features of the target point cloud, the approximate spatial correspondence features are... The method is as follows: splicing and get according to The transformed point cloud P′ is obtained by calculating the inverse process of IFDM, as follows:
6. The unsupervised point cloud registration method based on IFDM according to claim 5, characterized in that, The rigid transformation matrix is calculated using the least squares method based on singular value decomposition, based on the source point cloud and the transformed point cloud. The method is as follows: Let the rigid transformation matrix be... Among them, R * ∈SO(3) is a rotation matrix, t * ∈R 3 R is a translation matrix; * , t * The calculation method is as follows: Where R is R * The matrix elements, t is t * Matrix elements; C * For the corresponding points in (P, P′) A set of.
7. The unsupervised point cloud registration method based on IFDM according to claim 6, characterized in that, Before decoupling the source point cloud and target point cloud into shape-space features using IFDM, the process also includes: Prepare multiple sets of training samples in advance, each set of samples containing one source point cloud sample and one target point cloud sample; A reversible neural network is constructed, and the multiple sets of training samples are input into the constructed reversible neural network for training.
8. The unsupervised point cloud registration method based on IFDM according to claim 7, characterized in that, In the process of constructing a reversible neural network and inputting the multiple sets of training samples into the constructed reversible neural network for training, the loss function is designed as follows: Total loss L = L g +aL p +bL n +cL s , where L g For global registration loss, L p For spatial consistency loss, L n For the loss of regional consistency, L s For distributed convergence loss, a, b, and c are all greater than 0 and less than or equal to 1; L g The calculation method is as follows: Where μ is a variable, β is a user-defined real number, P′ is the transformed point cloud, and Q is the target point cloud; L p The calculation method is as follows: Where K(·) represents the KNN aggregation operator based on spatial features, and m is the number of points in P; L n The calculation method is as follows: N is the set of k nearest neighbors; L s The calculation method is as follows: in, for The matrix elements are: I, an indicator function operator, and m, the number of points in P.
9. The unsupervised point cloud registration method based on IFDM according to claim 8, characterized in that, a, b, and c are 1.0, 0.8, and 0.6 respectively.
10. The unsupervised point cloud registration method based on IFDM according to claim 8, characterized in that, β is defined as 0.
1.
11. An unsupervised point cloud registration system based on reversible shape-space decoupling, characterized in that, include: A reversible feature decoupling module is used to decouple the source point cloud and the target point cloud into shape-space features respectively through IFDM; wherein, the shape-space features include shape features and spatial features; The feature alignment module, electrically connected to the reversible feature decoupling module, is used to calculate a Gaussian similarity matrix using the shape features of the source point cloud and the target point cloud; and to align the spatial features of the source point cloud and the target point cloud based on the Gaussian similarity matrix to obtain a transformed point cloud. The registration attitude estimation module is electrically connected to the feature alignment module. It is used to calculate the rigid transformation matrix using the least squares method based on singular value decomposition based on the source point cloud and the transformed point cloud. It is also used to register the source point cloud and the target point cloud using the rigid transformation matrix.
12. The unsupervised point cloud registration system based on IFDM according to claim 11, characterized in that, The reversible feature decoupling module is specifically used to map P and Q to latent spaces through a trained reversible neural network, and to decouple the latent spaces mapped by P and Q to obtain the shape-space features of P and Q. The latent space mapping method is as follows: Let the source point cloud be P∈R m×3 The target point cloud is Q∈R n×3 ; Among them, f θ (·) represents an invertible neural network, Z P Represents the potential space of P. Let P be the shape feature to be decoupled. Z represents the spatial characteristics of P to be decoupled; Q Represents the potential space of Q. The shape features of Q to be decoupled The spatial characteristics of Q to be decoupled; The decoupling method for shape-space features is as follows: IFDM uses structure-aware map convolution to decouple the latent spaces mapped by P and Q respectively, and obtain the shape-space features of P and Q.
13. The unsupervised point cloud registration system based on IFDM according to claim 12, characterized in that, The IFDM uses structure-aware map convolution to decouple the latent spaces mapped by P and Q, respectively, to obtain the shape-space features of P and Q, including: IFDM uses structure-aware graph convolution based on MLP to construct a static KNN graph and introduces edge structure embedding, as follows: Let G = (V, E) be a KNN graph, where V is the set of points in the point cloud, and E is the set of edges connecting each point in the point cloud to its neighbors; the elements in E are the edges. The calculation method is as follows: in, edge vectors modulus, As the reference direction, Θ(·) represents the MLP. edge vectors The angle between the reference direction and the reference direction.
14. The unsupervised point cloud registration system based on IFDM according to claim 13, characterized in that, The method for calculating the Gaussian similarity matrix using the shape features of the source and target point clouds is as follows: Let the Gaussian similarity matrix be S∈R m×n The method for calculating S is as follows: Among them, s ij Let S be a matrix element, representing the i-th point P in P. i And the j-th point Q in Q j Similarity; For P i The corresponding shape features For Q j The corresponding shape features; The feature alignment module is also used to normalize and smooth the Gaussian similarity matrix S using neighborhood averaging to obtain an optimized similarity matrix.
15. The unsupervised point cloud registration system based on IFDM according to claim 14, characterized in that, The feature alignment module is specifically used to approximate spatially corresponding features based on the optimized similarity matrix and the spatial features of the target point cloud. Also used for splicing and get Also used according to The transformed point cloud P′ is obtained by inverse IFDM calculation; wherein, the calculation method of P′ is as follows:
16. The unsupervised point cloud registration system based on IFDM according to claim 15, characterized in that, The method for calculating the rigid transformation matrix is as follows: Let the rigid transformation matrix be... Among them, R * ∈SO(3) is a rotation matrix, t * ∈R 3 R is a translation matrix; * , t * The calculation method is as follows: Where R is R * The matrix elements, t is t * Matrix elements; C * For the corresponding points in (P, P′) A set of.
17. The unsupervised point cloud registration system based on IFDM according to claim 16, characterized in that, It also includes neural network building blocks; The neural network building unit is used to build a reversible neural network and input multiple pre-prepared training samples into the built reversible neural network for training; wherein, each set of samples includes a source point cloud sample and a target point cloud sample.
18. The unsupervised point cloud registration system based on IFDM according to claim 17, characterized in that, The neural network building unit is also used to design the loss function of the invertible neural network, and the loss function is specifically designed as follows: Total loss L = L g +aL p +bL n +cL s , where L g For global registration loss, L p For spatial consistency loss, L n For the loss of regional consistency, L s For distributed convergence loss, a, b, and c are all greater than 0 and less than or equal to 1; L g The calculation method is as follows: Where μ is a variable, β is a user-defined real number, P′ is the transformed point cloud, and Q is the target point cloud; L p The calculation method is as follows: Where K(·) represents the KNN aggregation operator based on spatial features, and m is the number of points in P; L n The calculation method is as follows: N is the set of k nearest neighbors; L s The calculation method is as follows: in, for The matrix elements are: I, an indicator function operator, and m, the number of points in P.
19. An unsupervised point cloud registration system based on IFDM according to claim 18, characterized in that, a, b, and c are 1.0, 0.8, and 0.6 respectively.
20. The unsupervised point cloud registration system based on IFDM according to claim 18, characterized in that, β is defined as 0.
1.
21. A computer-readable storage medium storing at least one instruction, characterized in that, The instructions are loaded and executed by the processor to implement an unsupervised point cloud registration method based on reversible shape-space decoupling as described in any one of claims 1-10.
22. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the unsupervised point cloud registration method based on reversible shape-space decoupling as described in any one of claims 1-10.