A point cloud registration method and system based on keypoint-free detection

Abstract

The present invention discloses a point cloud registration method and system based on keypoint-free detection. The method includes: downsampling the point cloud to be registered to obtain superpoint data; performing superpoint matching processing based on the superpoint data to obtain superpoint correspondence data; performing point matching processing based on the superpoint correspondence data to obtain global point correspondence data; and performing local-to-global registration based on the global point correspondence data to obtain a target estimated transformation, thereby completing point cloud registration. Based on the superpoint data, the superpoint matching through geometric self-attention processing, feature cross-attention processing, and Gaussian correlation matching processing fully captures the geometric structure information of the point cloud, enabling the superpoint correspondence to obtain higher accuracy and inner layer ratio. Furthermore, it does not rely on a robust pose estimator to perform the estimated transformation, greatly improving the rate of point cloud registration. It is suitable for point cloud registration tasks with fast point cloud registration and can be widely applied to the field of point cloud registration technology.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud registration, and in particular, to a point cloud registration method and system based on key-point-free detection. Background Art

[0002] With the rapid development of lidar and three-dimensional reconstruction technologies, the difficulty of obtaining point clouds has been continuously reduced. However, the point clouds obtained by current sensors or three-dimensional reconstruction technologies are only limited to a limited range. Therefore, an accurate and fast point cloud registration algorithm is required to register the point clouds of different parts of a scene into a single point cloud, thereby representing a more complete three-dimensional scene. Among them, the essence of point cloud registration is to find the transformation matrix of the point clouds measured in different coordinate systems and perform coordinate transformation on it.

[0003] The mainstream methods of point cloud registration can be divided into registration methods based on correspondences, direct registration methods, and registration methods based on deep robust estimators. The registration methods based on correspondences can be further divided into methods with key-point detection and key-point-free detection. Among them, the methods with key-point detection are difficult to apply to low-overlap scenes, while the key-point-free registration methods based on correspondences face problems such as long time consumption and low inner layer ratio in low-overlap scenes. Moreover, most of the existing methods often only capture the context information of the point cloud, resulting in low registration accuracy. Therefore, how to obtain a higher inner layer ratio and accuracy in point cloud registration is a technical problem that needs to be solved urgently. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a point cloud registration method and system based on key-point-free detection, which can obtain a higher inner layer ratio and accuracy in point cloud registration.

[0005] On the one hand, an embodiment of the present invention provides a point cloud registration method based on key-point-free detection, including:

[0006] Downsampling the point cloud to be registered to obtain superpoint data;

[0007] wherein the superpoint data includes first superpoint data and second superpoint data; the superpoint data includes context features;

[0008] Performing superpoint matching processing based on the superpoint data to obtain superpoint correspondence data;

[0009] wherein the superpoint matching includes geometric self-attention processing, feature cross-attention processing, and Gaussian correlation matching processing;

[0010] Performing point matching processing based on the superpoint correspondence data to obtain global point correspondence data;

[0011] Based on the global point correspondence data, perform local-to-global registration to obtain the target estimated transformation and complete point cloud registration.

[0012] Optionally, the downsampling of the point cloud to be registered to obtain superpoint data includes:

[0013] Perform downsampling on the point cloud to be registered respectively through the KpConv-FPN algorithm to obtain the first superpoint and the second superpoint;

[0014] Based on the first superpoint, learn the context features of the dense points through the KpConv-FPN algorithm to obtain the first context feature set;

[0015] Based on the second superpoint, learn the context features of the dense points through the KpConv-FPN algorithm to obtain the second context feature set;

[0016] Among them, the first superpoint data includes the first superpoint and the first context feature set, and the second superpoint data includes the second superpoint and the second context feature set.

[0017] Optionally, in the step of performing superpoint matching processing based on the superpoint data to obtain superpoint correspondence data, the geometric self-attention processing includes:

[0018] Based on the superpoint data, encode and determine geometric features;

[0019] Among them, the geometric features include point pair distance, point pair trihedral angle, point cloud curvature, and point pair normal vector angle;

[0020] Integrate the geometric features through geometric structure embedding to obtain geometric structure features;

[0021] Based on the context features and the geometric features, process through the self-attention mechanism to obtain the self-attention feature matrix;

[0022] Among them, the self-attention feature matrix includes the first self-attention feature matrix and the second self-attention feature matrix.

[0023] Optionally, the integrating the geometric features through geometric structure embedding to obtain geometric structure features includes:

[0024] Integrate the geometric features through geometric structure embedding using the geometric structure embedding formula to obtain geometric structure features;

[0025] Among them, the geometric structure embedding formula is:

[0026]

[0027] In the formula, r i，j represents the geometric structure feature, Represents the point pair distance, W D Projection matrix representing the point pair distance Represents the point pair trihedral angle, W A Projection matrix representing the point pair trihedral angle Represents the point cloud curvature, W C Projection matrix representing the point cloud curvature Represents the included angle between the normal vectors of the point pair, W N Projection matrix representing the included angle between the normal vectors of the point pair

[0028] Optionally, in the step of performing superpoint matching processing based on the superpoint data to obtain superpoint correspondence data, the feature cross-attention processing includes:

[0029] Performing feature exchange based on the correlation between the first self-attention feature matrix and the second self-attention feature matrix to obtain a feature cross-attention matrix;

[0030] Among them, the feature cross-attention matrix includes a first feature cross-attention matrix and a second feature cross-attention matrix.

[0031] Optionally, in the step of performing superpoint matching processing based on the superpoint data to obtain superpoint correspondence data, the Gaussian correlation matching processing includes:

[0032] Performing cross-learning on the feature cross-attention matrix for a first preset number of times to extract a mixed feature;

[0033] Among them, the mixed feature includes a first mixed feature and a second mixed feature;

[0034] Normalizing the first mixed feature and the second mixed feature to a unit hypersphere and calculating a first Gaussian correlation matrix;

[0035] Performing double normalization on the first Gaussian correlation matrix to obtain a second Gaussian correlation matrix;

[0036] Based on the second Gaussian correlation matrix, screening to obtain superpoint correspondence data based on preset conditions.

[0037] Optionally, the performing point matching processing based on the superpoint correspondence data to obtain global point correspondence data includes:

[0038] According to the superpoint correspondence data, learning feature matching through an optimal transport layer graph neural network to extract dense point correspondences;

[0039] Determining a cost matrix according to the superpoint data;

[0040] Based on the cost matrix, calculating a confidence matrix through the Sinkhorn algorithm;

[0041] Determine the target dense point correspondence based on the dense point correspondence and in combination with the confidence matrix.

[0042] Determine the global point correspondence data based on the target dense point correspondence.

[0043] Optionally, the performing local-to-global registration based on the global point correspondence data to obtain a target estimated transformation and complete point cloud registration includes:

[0044] Calculate the sub-transformation for each superpoint through weighted singular value decomposition to determine the local transformation.

[0045] Based on the local transformation, determine the target sub-transformation as the global transformation based on the negative log-likelihood loss function.

[0046] Based on the global transformation, perform transformation iterations for a second preset number of times to obtain a target estimated transformation and complete point cloud registration.

[0047] On the other hand, an embodiment of the present invention provides a point cloud registration system based on key-point-free detection, including:

[0048] A first module for downsampling the point cloud to be registered to obtain superpoint data.

[0049] Wherein, the superpoint data includes first superpoint data and second superpoint data; the superpoint data includes context features.

[0050] A second module for performing superpoint matching processing based on the superpoint data to obtain superpoint correspondence data.

[0051] Wherein, the superpoint matching includes geometric self-attention processing, feature cross-attention processing, and Gaussian correlation matching processing.

[0052] A third module for performing point matching processing based on the superpoint correspondence data to obtain global point correspondence data.

[0053] A fourth module for performing local-to-global registration based on the global point correspondence data to obtain a target estimated transformation and complete point cloud registration.

[0054] On the other hand, an embodiment of the present invention provides a point cloud registration system based on key-point-free detection, including a processor and a memory:

[0055] The memory is used to store programs.

[0056] The processor executes the programs to implement the method described above.

[0057] On the other hand, an embodiment of the present invention provides a computer-readable storage medium storing a program, which when executed by a processor implements the method described above.

[0058] An embodiment of the present invention also discloses a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. The processor of the computer device can read the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to execute the method described above.

[0059] In an embodiment of the present invention, the point cloud to be registered is first downsampled to obtain superpoint data; wherein, the superpoint data includes first superpoint data and second superpoint data; the superpoint data includes context features; based on the superpoint data, superpoint matching processing is performed to obtain superpoint correspondence data; wherein, the superpoint matching includes geometric self-attention processing, feature cross-attention processing, and Gaussian correlation matching processing; based on the superpoint correspondence data, point matching processing is performed to obtain global point correspondence data; based on the global point correspondence data, local-to-global registration is performed to obtain a target estimated transformation, completing point cloud registration. Based on the superpoint data, the superpoint matching through geometric self-attention processing, feature cross-attention processing, and Gaussian correlation matching processing in the present invention fully captures the geometric structure information of the point cloud, enabling the superpoint correspondence to obtain higher accuracy and inlier ratio. Furthermore, it is not necessary to rely on a robust pose estimator to perform the estimated transformation, greatly improving the rate of point cloud registration and being applicable to the point cloud registration task of fast point cloud registration. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.

[0061] Figure 1 It is a schematic flowchart of a point cloud registration method based on keypoint-free detection provided by an embodiment of the present invention;

[0062] Figure 2 It is a schematic flowchart of a superpoint matching module provided by an embodiment of the present invention;

[0063] Figure 3 It is a schematic flowchart of the training of a geometric self-attention module provided by an embodiment of the present invention;

[0064] Figure 4 It is a schematic diagram of the effect before and after the registration of two local point clouds provided by an embodiment of the present invention. Detailed implementation manners

[0065] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0066] On the one hand, referring to Figure 1 , an embodiment of the present invention provides a point cloud registration method based on keypoint-free detection, including:

[0067] S100. Downsample the point cloud to be registered to obtain supervoxel data;

[0068] It should be noted that the supervoxel data includes first supervoxel data and second supervoxel data; the supervoxel data includes context features; in some embodiments, the point cloud to be registered is downsampled respectively by the KpConv-FPN algorithm to obtain the first supervoxel and the second supervoxel; based on the first supervoxel, the context features of dense points are learned by the KpConv-FPN algorithm to obtain the first context feature set; based on the second supervoxel, the context features of dense points are learned by the KpConv-FPN algorithm to obtain the second context feature set; wherein, the first supervoxel data includes the first supervoxel and the first context feature set, and the second supervoxel data includes the second supervoxel and the second context feature set. Among them, the context features are the context features of dense points of the first supervoxel and the second supervoxel learned by the KpConv-FPN algorithm, including the first context feature set and the second context feature set.

[0069] Specifically, downsampling the point cloud by the KpConv-FPN algorithm into individual supervoxels, where each point is assigned to the nearest supervoxel in the geometric space, which is essentially a voronoi decomposition. And for the first supervoxel and the second supervoxel the context features of dense points are learned by the KpConv-FPN algorithm to obtain the first context feature set or the second context feature set

[0070] S200. Perform supervoxel matching processing based on the supervoxel data to obtain supervoxel correspondence data;

[0071] It should be noted that the supervoxel matching includes geometric self-attention processing, feature cross-attention processing and Gaussian correlation matching processing.

[0072] In some embodiments, the geometric self-attention processing includes: determining geometric features based on superpoint data; wherein the geometric features include point pair distance, point pair trihedral angle, point cloud curvature, and point pair normal vector; integrating the geometric features through geometric structure embedding to obtain geometric structure features; processing the geometric features and context features through a self-attention mechanism to obtain a self-attention feature matrix; wherein the self-attention feature matrix includes a first self-attention feature matrix and a second self-attention feature matrix.

[0073] Wherein, in some embodiments, integrating the geometric features through geometric structure embedding to obtain geometric structure features includes: integrating the geometric features through geometric structure embedding using a geometric structure embedding formula to obtain geometric structure features; wherein the geometric structure embedding formula is:

[0074]

[0075] In the formula, r i，j represents the geometric structure feature, represents the point pair distance, and W D represents the projection matrix of the point pair distance, represents the point pair trihedral angle, and W A represents the projection matrix of the point pair trihedral angle, represents the point cloud curvature, and W C represents the projection matrix of the point cloud curvature, represents the point pair normal vector angle, and W N represents the projection matrix of the point pair normal vector angle. Among them, the projection matrices of the point pair distance, point pair trihedral angle, and point cloud curvature are all trainable parameter matrices.

[0076] In some embodiments, the feature cross-attention processing includes: performing feature exchange based on the correlation between the first self-attention feature matrix and the second self-attention feature matrix to obtain a feature cross-attention matrix; wherein the feature cross-attention matrix includes a first feature cross-attention matrix and a second feature cross-attention matrix.

[0077] In some embodiments, the Gaussian correlation matching processing includes: performing cross-learning on the feature cross-attention matrix for a first preset number of times to extract mixed features; wherein the mixed features include a first mixed feature and a second mixed feature; normalizing the first mixed feature and the second mixed feature to a unit hypersphere, calculating to obtain a first Gaussian correlation matrix; performing double normalization on the first Gaussian correlation matrix to obtain a second Gaussian correlation matrix; screening superpoint corresponding data based on the second Gaussian correlation matrix according to a preset condition.

[0078] Specifically, as Figure 2As shown, the superpoint matching module performs superpoint matching processing to obtain a superpoint correspondence set (i.e., superpoint correspondence data). The superpoint matching module mainly includes a geometric self-attention module, a feature cross-attention module, and a Gaussian correlation matching module.

[0079] 1. Geometric self-attention module

[0080] First, encode the pairwise distances of each dense point for each superpoint Pairwise trihedral angles Curvature of the point cloud And the angle between the pairwise normal vectors Then, integrate the above four geometric features through geometric structure embedding to obtain the geometric structure feature r of each dense point i，j , and finally, the geometric structure features r of all dense points of this superpoint i，j constitute the geometric structure feature matrix of this superpoint Similarly, the geometric structure feature matrix of the second superpoint can be obtained

[0081] For two dense points in a given superpoint The distance of the point pair based on odd-even position encoding is:

[0082]

[0083]

[0084] In the formula, d i，j is the Euclidean distance between two dense points , σ d is a factor controlling distance sensitivity, d t is the vector dimension, and x = 0, 1, 2…dt / 2. The calculation of each pairwise distance in

[0085] For two dense points in a given superpoint The pairwise trihedral angle based on odd-even position encoding is:

[0086]

[0087]

[0088] In the formula, is The angle between the vectors of two points and the vectors of its three nearest points respectively, is a factor controlling angle sensitivity, d t is the vector dimension, and x = 0, 1, 2…d t / 2. The calculation of the three - dimensional angles for each point is the same by analogy.

[0089] A dense point in the given super - point Its curvature based on odd - even position encoding is:

[0090]

[0091]

[0092] In the formula, c i is the curvature of the dense point σ c is the factor controlling the curvature sensitivity, d t is the vector dimension, x = 0, 1, 2…d t / 2. The calculation of each curvature in

[0093] is the same by analogy. The normal vector angle of the point pair based on odd - even position encoding is:

[0094]

[0095]

[0096] In the formula, n i,j is the included angle of the normal vectors of the two dense points, σ n is the factor controlling the sensitivity of the normal vector angle, d t is the vector dimension, x = 0, 1, 2…d t / 2, The calculation of each point - pair normal vector angle in

[0097] The geometric structure embedding is:

[0098]

[0099] The W D 、W A 、W C 、 in the above formula are the projection matrices corresponding to the geometric feature embeddings respectively.

[0100] As Figure 3 shown, the context features of each super - point or and the geometric structure feature r i，j are processed based on the self - attention mechanism to obtain the self - attention feature matrices x P 、X Q The encoded geometric self - attention module is:

[0101]

[0102]

[0103] In the formula, e i,j is the geometric self-attention score, is the context feature vector of point , d t is the dimension of the vector, W Q , W K , W V , are the projection matrices of queries, keys, values, and geometric structure embeddings respectively, where the projection matrix is a trainable parameter matrix; a i，j is the score after softmax normalization of e i，j , z i is the feature vector output by geometric self-attention encoding, corresponding to the geometric self-attention encoding z of the output superpoint in Figure 3 . Furthermore, the self-attention feature matrix X of the superpoint (i.e., the feature matrix X P , X Q ) is obtained through the training process of the neural network, where the neural network training process is implemented based on linear layers, residual connections and normalization, and forward propagation.

[0104] 2. Feature Cross-Attention Module

[0105] Cross-attention is a typical module in the point cloud registration task, which is used to exchange features between two point clouds. Given the self-attention feature matrices X P , X Q corresponding to respectively, then the feature cross-attention matrix Z P of is calculated by using the features. Similarly, the cross-attention feature matrix of is calculated in the same way.

[0106]

[0107]

[0108] In the formula, a i,j is calculated from the cross-attention score e i,j , and e i,j is calculated from the correlation between X P and X Q . Finally, the feature cross-attention matrix is obtained, where are the self-attention feature matrices XP , a subset of X Q , is a subset of . Similarly, the subset of can be calculated.

[0109] 3. Gaussian correlation matching module

[0110] First, cross-learn the above two attention modules N t times (3 times can be used) to extract the mixed features and

[0111]

[0112]

[0113] Then, normalize the mixed features to a unit hypersphere and calculate the first Gaussian correlation matrix

[0114]

[0115] In the formula, are the subsets of the mixed features respectively, and S is the first Gaussian correlation matrix.

[0116] Then, perform double normalization on the first Gaussian correlation matrix to further suppress ambiguous matching:

[0117]

[0118] In the formula, is the second Gaussian correlation matrix, is the number of dense points of the superpoint , is the number of dense points of the superpoint , and S is the first Gaussian correlation matrix.

[0119] Finally, select the largest 256 items in the second Gaussian correlation matrix as the matching correspondence of the superpoints (i.e., superpoint correspondence data)

[0120]

[0121] In the formula, is a pair of superpoint correspondences, are the (x i , y i ) coordinates of the k items with the largest values in the second Gaussian correlation matrix.

[0122] S300. Perform point matching processing based on the superpoint correspondence data to obtain global point correspondence data;

[0123] It should be noted that in some embodiments, according to the superpoint correspondence data, feature matching is learned through an optimal transport graph neural network to extract dense point correspondences; according to the superpoint data, a cost matrix is determined; based on the cost matrix, a confidence matrix is calculated through the Sinkhorn algorithm; according to the dense point correspondences, combined with the confidence matrix, target dense point correspondences are determined; based on the target dense point correspondences, global point correspondence data is determined.

[0124] Specifically, point matching processing is performed through a point matching module. The point matching module is based on the superpoint correspondence For each superpoint correspondence, an optimal transport graph neural network (OptimalTransport Graph Neural Networks) is used to learn feature matching and extract dense point correspondences. A cost matrix is calculated through the feature matrices of two superpoints where n and m are the numbers of dense points of the two superpoints respectively.

[0125]

[0126] Then a learnable parameter is used to i augment a new row and a new column for C, and then use the Sinkhorn algorithm to calculate it. Finally, the last row and the last column are removed to become the confidence matrix Z i , and the 256 point pairs with the largest confidence scores are selected as dense point correspondences. After selecting the 256 point correspondences with the largest confidence scores for each superpoint, the final global point correspondence is formed

[0127] S400. Perform local-to-global registration based on the global point correspondence data to obtain the target estimated transformation and complete point cloud registration;

[0128] It should be noted that in some embodiments, the sub-transformations of each superpoint are calculated through weighted singular value decomposition to determine the local transformation; according to the local transformation, the target sub-transformation is determined as the global transformation based on the negative log-likelihood loss function; based on the global transformation, a second preset number of transformation iterations are performed to obtain the target estimated transformation and complete point cloud registration.

[0129] Specifically, most of the methods in the prior art need to rely on a robust pose estimator (such as RANSAC) to estimate the transformation because the assumed correspondences are usually dominated by outliers. However, the method of this embodiment adds better geometric structure encoding in the superpoint matching stage, and the superpoint correspondence has a better inner ratio, so it does not need to rely on a robust pose estimator. Therefore, this embodiment uses a local-to-global registration scheme.

[0130] For the local part, a sub - transformation (rigid transformation) \(T\) i =\(\{R\) i , \(t\) i \}\), where \(R\) i is the rotation matrix corresponding to the \(i\) - th pair of super - points, and \(t\) i is the translation vector corresponding to the \(i\) - th pair of super - points. The weight values will be provided by the confidence matrix \(Z\) i .

[0131] For the global part, an optimal sub - transformation (judged by minimizing the negative log - likelihood loss function as the best sub - transformation) will be selected as the global transformation, and the optimal sub - transformation will achieve the best effect at the global dense points.

[0132] Among them, the sub - transformation refers to the rigid transformation corresponding to a pair of super - points, not the rigid transformation of the entire point - cloud registration. The role of the sub - transformation is to select an optimal sub - transformation as the rigid transformation of the entire point - cloud registration in the global registration. Because the matching of super - points in the super - point matching module has a relatively high accuracy, an optimal sub - transformation can be selected as the global transformation. Finally, the entire transformation is iterated 5 times to re - estimate the transformation and obtain the best estimated transformation.

[0133] In summary, the technical solution of the present invention not only captures the context information of the point cloud, but also fully captures the geometric structure information of the point cloud, enabling higher accuracy and inlier ratio for super - point correspondence. Therefore, it is not necessary to use the RANSAC algorithm adapted to a large number of outliers to estimate the transformation, and the speed is increased by more than 100 times. Therefore, it is applicable to the point - cloud registration task of fast point - cloud registration. As Figure 4 shown, the registration effect of the technical solution of the present invention is good. Moreover, due to the relatively high inlier ratio of super - point matching and different weightings according to the overlap rate, the present invention is more applicable to the point - cloud registration task in low - overlap scenarios. Finally, since the geometric structure information encoded by the present invention is rigid - body invariant, it is applicable to the point - cloud registration task of rigid - body invariant.

[0134] On the other hand, an embodiment of the present invention provides a point - cloud registration system based on key - point - free detection, including: a first module for down - sampling the point cloud to be registered to obtain super - point data; wherein, the super - point data includes first super - point data and second super - point data; a second module for performing super - point matching processing based on the super - point data to obtain super - point correspondence data; wherein, the super - point matching includes geometric self - attention processing, feature cross - attention processing, and Gaussian correlation matching processing; a third module for performing point matching processing based on the super - point correspondence data to obtain global point correspondence data; a fourth module for performing local - to - global registration based on the global point correspondence data to obtain the target estimated transformation and complete the point - cloud registration.

[0135] The content of the method embodiments of the present invention is applicable to the system embodiments of the present invention. The functions specifically implemented by the system embodiments of the present invention are the same as those of the above-mentioned method embodiments, and the beneficial effects achieved are also the same as those achieved by the above-mentioned methods.

[0136] On the other hand, an embodiment of the present invention further provides a point cloud registration system based on keypoint-free detection, including a processor and a memory;

[0137] The memory is used to store programs;

[0138] The processor executes the program to implement the method as described above.

[0139] The content of the method embodiments of the present invention is applicable to the system embodiments of the present invention. The functions specifically implemented by the system embodiments of the present invention are the same as those of the above-mentioned method embodiments, and the beneficial effects achieved are also the same as those achieved by the above-mentioned methods.

[0140] On the other hand, an embodiment of the present invention further provides a computer-readable storage medium, where the storage medium stores a program, and the program is executed by a processor to implement the method as described above.

[0141] The content of the method embodiments of the present invention is applicable to the computer-readable storage medium embodiments of the present invention. The functions specifically implemented by the computer-readable storage medium embodiments of the present invention are the same as those of the above-mentioned method embodiments, and the beneficial effects achieved are also the same as those achieved by the above-mentioned methods.

[0142] An embodiment of the present invention also discloses a computer program product or a computer program, which includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device can read the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to enable the computer device to execute the method as described above.

[0143] In some alternative embodiments, the functions / operations mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the functions / operations involved, two consecutive blocks shown may actually be executed substantially simultaneously or the blocks can sometimes be executed in the reverse order. In addition, the embodiments presented and described in the flowcharts of the present invention are provided by way of example for a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logical flows presented herein. Alternative embodiments are expected, where the order of various operations is changed and the sub-operations described as part of a larger operation are executed independently.

[0144] In addition, although the present invention has been described in the context of functional modules, it should be understood that, unless otherwise stated to the contrary, one or more of the described functions and / or features may be integrated in a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It should also be understood that a detailed discussion of the actual implementation of each module is not necessary for understanding the present invention. Rather, given the attributes, functions, and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the modules will be understood within the ordinary skills of an engineer. Thus, those skilled in the art can implement the present invention as set forth in the claims without undue experimentation. It should also be understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present invention, which is determined by the full scope of the appended claims and their equivalents.

[0145] If the described function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this 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 for causing 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 such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0146] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a defined sequence list of executable instructions for implementing a logical function, and can be specifically implemented in any computer-readable medium for use by an instruction execution device, apparatus, or equipment (such as a computer-based device, a device including a processor, or other devices that can fetch and execute instructions from the instruction execution device, apparatus, or equipment), or in combination with these instruction execution devices, apparatuses, or equipment. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by or in combination with an instruction execution device, apparatus, or equipment.

[0147] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection (electronic device) having one or more wirings, a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable media can even be paper or other suitable media on which the program can be printed, as the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or otherwise processing as appropriate, and then stored in a computer memory.

[0148] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution device. For example, if implemented in hardware, as in another embodiment, any one or a combination of the following techniques well-known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0149] In the description of this specification, the descriptions referring to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0150] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the claims and their equivalents.

[0151] The above has specifically described the preferred embodiments of the present invention, but the present invention is not limited to the described embodiments. Those skilled in the art can also make various equivalent deformations or substitutions without departing from the spirit of the present invention, and these equivalent deformations or substitutions are all included within the scope defined by the claims of the present invention.

Claims

1. A point cloud registration method based on keypoint-free detection, characterized in that, Including: Downsample the point cloud to be registered to obtain superpoint data; Among them, the superpoint data includes first superpoint data and second superpoint data; the superpoint data includes context features; Based on the superpoint data, perform superpoint matching processing to obtain superpoint correspondence data; Among them, the superpoint matching includes geometric self-attention processing, feature cross-attention processing, and Gaussian correlation matching processing; the geometric self-attention processing includes: Based on the superpoint data, encode and determine geometric features; Among them, the geometric features include point pair distance, point pair trihedral angle, point cloud curvature, and point pair normal vector angle; Integrate the geometric features through geometric structure embedding to obtain geometric structure features; Based on the context features and the geometric features, process through a self-attention mechanism to obtain a self-attention feature matrix; Among them, the self-attention feature matrix includes a first self-attention feature matrix and a second self-attention feature matrix; Based on the superpoint correspondence data, perform point matching processing to obtain global point correspondence data; Based on the global point correspondence data, perform local-to-global registration to obtain a target estimated transformation and complete point cloud registration.

2. The point cloud registration method based on keypoint-free detection according to claim 1, wherein The step of downsampling the point cloud to be registered to obtain superpoint data includes: Respectively downsample the point cloud to be registered through the KpConv-FPN algorithm to obtain a first superpoint and a second superpoint; Based on the first superpoint, learn the context features of dense points through the KpConv-FPN algorithm to obtain a first context feature set; Based on the second superpoint, learn the context features of dense points through the KpConv-FPN algorithm to obtain a second context feature set; Among them, the first superpoint data includes the first superpoint and the first context feature set, and the second superpoint data includes the second superpoint and the second context feature set.

3. A point cloud registration method based on keypoint-free detection according to claim 1, characterized in that, The step of integrating the geometric features through geometric structure embedding to obtain geometric structure features includes: Integrate the geometric features through geometric structure embedding using a geometric structure embedding formula to obtain geometric structure features; Among them, the geometric structure embedding formula is: In the formula, represents the geometric structure feature, represents the point pair distance, represents the projection matrix of the point pair distance, represents the point pair trihedral angle, represents the projection matrix of the point pair trihedral angle, represents the point cloud curvature, represents the projection matrix of the point cloud curvature, represents the included angle between the point pair normal vectors, represents the projection matrix of the included angle between the point pair normal vectors.

4. A point cloud registration method based on keypoint-free detection according to claim 1, wherein, In the step of performing superpoint matching processing based on the superpoint data to obtain superpoint correspondence data, the feature cross-attention processing includes: Perform feature exchange based on the correlation between the first self-attention feature matrix and the second self-attention feature matrix to obtain a feature cross-attention matrix; Among them, the feature cross-attention matrix includes a first feature cross-attention matrix and a second feature cross-attention matrix.

5. A point cloud registration method based on key-point-free detection according to claim 4, characterized in that, In the step of performing superpoint matching processing based on the superpoint data to obtain superpoint correspondence data, the Gaussian correlation matching processing includes: Perform cross-learning on the feature cross-attention matrix for a first preset number of times to extract mixed features; Among them, the mixed features include first mixed features and second mixed features; Normalize the first mixed features and the second mixed features to a unit hypersphere and calculate a first Gaussian correlation matrix; Perform double normalization on the first Gaussian correlation matrix to obtain a second Gaussian correlation matrix; Based on the second Gaussian correlation matrix, screen to obtain superpoint correspondence data according to preset conditions.

6. A point cloud registration method based on keypoint-free detection according to claim 1, characterized in that Performing point matching processing based on the above-mentioned superpoint correspondence data to obtain global point correspondence data, including: According to the above-mentioned superpoint correspondence data, learning feature matching through an optimal transport layer graph neural network to extract dense point correspondences; Determining a cost matrix according to the above-mentioned superpoint data; Calculating a confidence matrix based on the cost matrix through the Sinkhorn algorithm; Determining target dense point correspondences according to the dense point correspondences in combination with the confidence matrix; Determining global point correspondence data based on the target dense point correspondences.

7. A point cloud registration method based on key-point-free detection according to claim 1, characterized in that Performing local-to-global registration based on the global point correspondence data to obtain a target estimated transformation and complete point cloud registration, including: Calculating a sub-transformation for each superpoint through weighted singular value decomposition to determine a local transformation; Based on the local transformation, determining a target sub-transformation as the global transformation based on the negative log-likelihood loss function; Based on the global transformation, performing transformation iterations a second preset number of times to obtain a target estimated transformation and complete point cloud registration.

8. A point cloud registration system based on keypoint-free detection, characterized in that, Including: A first module for downsampling the point cloud to be registered to obtain superpoint data; Wherein, the superpoint data includes first superpoint data and second superpoint data; the superpoint data includes context features; A second module for performing superpoint matching processing based on the superpoint data to obtain superpoint correspondence data; Wherein, the superpoint matching includes geometric self-attention processing, feature cross-attention processing, and Gaussian correlation matching processing; the geometric self-attention processing includes: Encoding to determine geometric features based on the superpoint data; Wherein, the geometric features include point pair distance, point pair trihedral angle, point cloud curvature, and point pair normal vector angle; Integrating the geometric features through geometric structure embedding to obtain geometric structure features; Processing through a self-attention mechanism based on the context features and the geometric features to obtain a self-attention feature matrix; Wherein, the self-attention feature matrix includes a first self-attention feature matrix and a second self-attention feature matrix; A third module for performing point matching processing based on the superpoint correspondence data to obtain global point correspondence data; A fourth module for performing local-to-global registration based on the global point correspondence data to obtain a target estimated transformation and complete point cloud registration.

9. A point cloud registration system based on keypoint-free detection, characterized in that, Including a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method according to any one of claims 1 to 7.

Images