Point cloud data registration method and device

Abstract

The invention discloses a point cloud data registration method and device. The method comprises the following steps: acquiring a source point cloud data set and a target point cloud data set; and processing the source point cloud data set and the target point cloud data set by using a point cloud data registration model to obtain a point cloud data registration transformation parameter set. According to the method, a point cloud data registration model is utilized to carry out identification processing on local geometric position features and a global structure of the point cloud, so that discriminative features can be captured and a reliable corresponding relation can be established even under the conditions of significant partial overlapping and data sparsity; the problems of low overlapping rate, incomplete point cloud and point cloud fuzzy matching in a large-scale scene are solved, and the precision and efficiency of point cloud data registration are improved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision, and in particular, to a method and device for point cloud data registration. Background Art

[0002] Point cloud registration is a fundamental task in 3D computer vision and plays a crucial role in various applications such as autonomous driving, SLAM, 3D reconstruction, and object recognition. Its goal is to find the best rigid transformation (rotation and translation) to align multiple point clouds acquired from different viewpoints or sensors into a common coordinate system.

[0003] To develop accurate and robust point cloud registration algorithms, extensive research has been conducted. Among them, the Iterative Closest Point (ICP) algorithm and its variants are classic methods. However, these methods often face some challenges, such as sensitivity to the initial pose, noise, and partial overlap between point clouds, which can significantly affect the registration accuracy. In recent years, leveraging the powerful capabilities of deep learning, learning-based point cloud registration methods have emerged, such as PointNet, PointNet++, PointNetLK, DCP, PRNet, which have made significant progress in registration performance. These methods generally include two stages: establishing correspondences between point clouds through feature extraction and matching, and then estimating the best transformation based on the established correspondences. Therefore, establishing accurate and reliable correspondences is crucial for successful registration. Some methods, such as 3DMatch, D3Feat, 3DFeat-Net, PPFNet, Predator, spinNet, attempt to detect repeatable key points in the point cloud and establish correspondences based on these key points. However, directly detecting repeatable key points in the point cloud, especially for point clouds with limited overlapping regions, can be challenging and usually results in a low inlier rate.

[0004] Recently emerged key-point-free methods, such as CofiNet, PTT, GeoTransformer, have shown promising results in addressing the limitations of key-point-based methods. These methods adopt a coarse-to-fine registration strategy and use the Transformer architecture to learn the global geometric structure from the point cloud. They typically operate on downsampled point clouds (superpoints) to establish correspondences and then propagate these correspondences to a denser point set to obtain fine-grained correspondences.

[0005] Generally, the same point cloud scene contains multiple similar local patches. When using the attention mechanism to learn global geometric correlations, these similar patches in non-overlapping regions can lead to the establishment of incorrect correspondences, and thus a large number of non-overlapping superpoint features will disrupt the learning process of global geometric consistency, thereby affecting the registration accuracy and efficiency. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a point cloud data registration method and device, which overcomes the influence of similar patches on point cloud matching in non-overlapping regions and improves the accuracy and efficiency of point cloud registration.

[0007] To solve the above technical problem, a first aspect of an embodiment of the present invention discloses a point cloud data registration method, the method includes:

[0008] S1, obtaining a source point cloud data set and a target point cloud data set;

[0009] S2, using a point cloud data registration model to process the source point cloud data set and the target point cloud data set to obtain a point cloud data registration transformation parameter set.

[0010] Optionally, the point cloud data registration model includes: a feature extraction module, a local feature fusion module, a global feature learning module, a feature data matching module, and a fine registration module;

[0011] The feature extraction module, the local feature fusion module, the global feature learning module, the feature data matching module, and the fine registration module are sequentially connected by data;

[0012] The feature extraction module is used to perform feature extraction and downsampling processing on the source point cloud data set and the target point cloud data set to obtain a dense point cloud superpoint data set; the dense point cloud superpoint data set includes a plurality of superpoint data and their corresponding superpoint geometric feature data;

[0013] The local feature fusion module is used to perform first encoding and feature fusion processing on the source point cloud data set to obtain a point cloud local fusion feature data set;

[0014] The global feature learning module is used to perform second encoding and feature extraction and matching processing on the point cloud local fusion feature data set to obtain a point cloud global fusion feature data set;

[0015] The feature data matching module is used to perform feature matching processing on the point cloud global fusion feature data set to obtain a superpoint correspondence information set;

[0016] The fine registration module is used to process the superpoint correspondence information set to obtain a point cloud data registration transformation parameter set.

[0017] Optionally, the using a point cloud data registration model to process the source point cloud data set and the target point cloud data set to obtain a point cloud data registration transformation parameter set includes:

[0018] S21. Use the feature extraction module to process the source point cloud dataset and the target point cloud dataset to obtain a dense point cloud superpoint dataset;

[0019] S22. Use the local feature fusion module to process the dense point cloud superpoint dataset to obtain a point cloud local fusion feature dataset;

[0020] S23. Use the global feature learning module to process the dense point cloud superpoint dataset to obtain a point cloud global fusion feature dataset;

[0021] The point cloud global fusion feature dataset includes a source point cloud global feature dataset and a target point cloud global feature dataset;

[0022] S24. Use the feature data matching module to process the point cloud global fusion feature dataset to obtain a superpoint correspondence information set;

[0023] S25. Based on the superpoint correspondence information set, use the fine registration module to process the point cloud local fusion feature dataset to obtain a point cloud data registration transformation parameter set.

[0024] Optionally, the step of using the feature extraction module to process the source point cloud dataset and the target point cloud dataset to obtain a dense point cloud superpoint dataset includes:

[0025] S211. Downsample the source point cloud dataset to obtain a source dense point set

[0026] Downsample the target point cloud dataset to obtain a target dense point set

[0027] S212. Perform discrimination processing on the source dense point set and the target dense point set to obtain a superpoint dataset;

[0028] The superpoint dataset includes a source superpoint dataset and a target superpoint dataset

[0029] The discrimination processing expression is:

[0030]

[0031] where represents the dataset in the source superpoint dataset that conforms to the source superpoint data processing rule; represents the dataset in the target superpoint dataset that conforms to the target superpoint data processing rule; Represents the source data set; Represents the target data set; Represents the source superpoint data set; Represents the target superpoint data set; Represents the feature dimension;

[0032] S213. Process the superpoint data set to obtain a dense point cloud superpoint data set;

[0033] The processing expression is:

[0034]

[0035] Wherein, Represents the local point cloud block unit index in the source point cloud data set; Represents the source dense point set; i represents the superpoint index value that satisfies the minimum distance condition; j represents the superpoint cloud space point index; Represents the superpoint with index j in the source superpoint data set; argmin j () represents selecting the j value that makes the Euclidean distance between the dense point and the superpoint with index j the smallest.

[0036] Optionally, using the local feature fusion module to process the dense point cloud superpoint data set to obtain a point cloud local fusion feature data set, including:

[0037] S221. Based on the first coding model, process the dense point cloud superpoint data set to obtain a first superpoint coding data set;

[0038] The first superpoint coding data set includes a source first superpoint coding data set and a target first superpoint coding data set;

[0039] The expression of the first coding model is:

[0040]

[0041] Wherein, S i Represents the i-th coding data of the local point cloud block coding data set; Represents the S i Corresponding geometric feature; Represents the S i And the Corresponding position coding matrix; Θ(S i ) represents a diagonal coding matrix composed of block matrices; θ k Represents the feature coding method of the k-th channel;

[0042] S222. Perform feature fusion processing on the first superpoint coding data set to obtain a point cloud local fusion feature data set;

[0043] The point cloud local fusion feature dataset includes a source point cloud local fusion feature dataset and a target point cloud local fusion feature dataset;

[0044] The feature fusion processing expression is:

[0045]

[0046] Where represents the fusion feature vector of the i-th superpoint encoded data; MLP() represents a fully connected projection network; cat[·] represents the operation of concatenating features; W k represents the vector key value in the feature projection matrix; W q represents the vector query value in the feature projection matrix; W v represents the vector value in the feature projection matrix; j represents the dense point cloud data index in the local point cloud block; i represents the superpoint encoded data index; Θ(S j ) represents a diagonal encoding matrix composed of block matrices; k j represents the vector key value of the dense point S j calculated according to the feature projection; v j represents the vector value of the dense point S j calculated according to the feature projection; a i，j represents the attention score.

[0047] Optionally, processing the dense point cloud superpoint dataset by using the global feature learning module to obtain a point cloud global fusion feature dataset includes:

[0048] S231. Based on the second encoding model, processing the dense point cloud superpoint dataset to obtain a second superpoint encoding dataset;

[0049] The second superpoint encoding dataset includes a source second superpoint encoding dataset and a target second superpoint encoding dataset;

[0050] S232. Based on the first self-attention global feature extraction model, processing the source second superpoint encoding dataset to obtain a source point cloud global feature dataset;

[0051] The point cloud global feature dataset includes a source point cloud global feature dataset and a target point cloud global feature dataset;

[0052] S233. Based on the second self-attention global feature extraction model, processing the target second superpoint encoding dataset to obtain a target point cloud global feature dataset;

[0053] S234. Use the cross-attention matching model and the first loss function to process the source point cloud global feature dataset and the target point cloud global feature dataset to obtain a point cloud global fusion feature dataset.

[0054] Optionally, using the feature data matching module to process the point cloud global fusion feature dataset to obtain a superpoint correspondence information set includes:

[0055] S241. Normalize the source point cloud global feature dataset and the target point cloud global feature dataset to obtain a normalized source global feature dataset and a normalized target global feature dataset;

[0056] S242. Transform the normalized source global feature dataset and the normalized target global feature dataset to obtain a global correlation feature dataset;

[0057] The transformation processing expression is:

[0058]

[0059] where s i，j represents the correlation between two superpoint features with encoding index i in the source superpoint dataset and encoding index j in the target superpoint dataset; exp() represents performing exponential function operation processing; represents the normalized feature of the superpoint with encoding index i in the source superpoint dataset; represents the normalized feature of the superpoint with encoding index j in the target superpoint dataset;

[0060] S243. Perform double normalization processing on the global correlation feature dataset to obtain a normalized correlation feature dataset;

[0061] The double normalization processing expression is:

[0062]

[0063] where represents the normalized correlation value; s i，k represents the sum of all correlation values in the i-th row; s l，j represents the sum of all correlation values in the j-th column; k represents the index of the first round of normalization; l represents the index of the second round of normalization;

[0064] S244. Perform confidence processing on the normalized correlation feature dataset to obtain a point cloud global relationship information set;

[0065] The confidence processing expression is:

[0066]

[0067] Among them, represents the matching set after confidence processing; represents the superpoint set the x i -th superpoint in; represents the superpoint set the y i -th superpoint in; (x i , y i ) represents the index of a pair of superpoint matches; represents selecting the top k pairs of superpoint matches with the strongest correlation from the normalized correlation matrix ;

[0068] S245, extract information from the point cloud global relationship information set to obtain a global mixed feature information set;

[0069] S246, perform calculation processing on the global mixed feature information set to obtain a superpoint scoring matrix information set;

[0070] The calculation processing expression is:

[0071]

[0072] Among them, Z i represents the scoring matrix between the source local dense point cloud block and the target local dense point cloud block; represents the mixed feature of the x i -th point in the source point dataset; represents the mixed feature of the y i -th point in the target point dataset; W S represents the weight matrix of the source point dataset; W T represents the weight matrix of the target point dataset; represents the x i -th point in the source point set; represents the y i -th point in the target point set; n i represents the number of points in the source local dense point cloud block; m i represents the number of points in the target local dense point cloud block; Θ() represents the diagonal encoding matrix composed of block matrices; represents the dimension constant for normalization; represents the real number field;

[0073] S247, perform conversion processing on the superpoint scoring matrix information set to obtain a superpoint correspondence confidence information set;

[0074] The conversion processing expression is;

[0075]

[0076] Among them, represents the confidence information set of the dense point correspondence; Z(x i , ) represents all columns in the scoring matrix related to the source point x i ; Z(, y i ) represents all rows in the scoring matrix related to the source point y i .

[0077] S248, process the confidence information set of the superpoint correspondence to obtain the superpoint correspondence information set;

[0078]

[0079] Among them, C represents the dense point correspondence information set;

[0080] Among them, U represents the operation of taking the union; represents the points in the local dense point cloud block set represented by x in the source point cloud data set, represented by x i ; j represents; represents the points in the local dense point cloud block set represented by y in the target point cloud data set, represented by y i ; mutual_topk j () represents selecting the k pairs of matches that are the strongest for each other from the confidence information set. x，y () represents selecting the k pairs of matches that are the strongest for each other from the confidence information set.

[0081] In the second aspect of the embodiments of the present invention, a point cloud data registration device is disclosed. The device includes:

[0082] A data acquisition module that acquires a source point cloud data set and a target point cloud data set;

[0083] A data processing module that uses a point cloud data registration model to process the source point cloud data set and the target point cloud data set to obtain a point cloud data registration transformation parameter set.

[0084] In the third aspect of the present invention, another point cloud data registration device is disclosed. It is characterized in that the device includes:

[0085] A memory storing executable program code;

[0086] A processor coupled to the memory;

[0087] The processor calls the executable program code stored in the memory and executes some or all of the steps in the point cloud data registration method disclosed in the first aspect of the embodiments of the present invention.

[0088] A fourth aspect of the present invention discloses a computer-readable storage medium storing computer instructions, which are used to execute some or all steps of the point cloud data registration method disclosed in the first aspect of the embodiments of the present invention when called.

[0089] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0090] The present invention proposes a point cloud data registration method and device, which can be applied to challenging environments such as low overlap rate, incomplete point cloud, and large-scale scenarios. The present invention adopts a point cloud data registration model based on geometric position attention, which effectively integrates local geometric position features and global structure perception, enabling it to capture discriminative features and establish reliable correspondences even in the case of significant partial overlap and data sparsity, having significant robustness and generalization ability, and can be applied to point cloud registration processing on different data sets such as 3DMatch, ModelNet, and KITTI. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] 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 following drawings are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings without creative efforts based on these drawings.

[0092] Figure 1 Scene schematic diagram of the image processing system provided by the embodiments of the present invention;

[0093] Figure 2 is a flowchart of a point cloud data registration method disclosed in the embodiments of the present invention;

[0094] Figure 3 is a schematic diagram of a point cloud data registration model of a point cloud data registration method disclosed in the embodiments of the present invention;

[0095] Figure 4 is a schematic diagram of the structure of a point cloud data registration device disclosed in the embodiments of the present invention;

[0096] Figure 5 is a schematic diagram of the structure of another point cloud data registration device disclosed in the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0097] To enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0098] The terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, device, product or equipment that includes a series of steps or units is not limited to the listed steps or units, but may optionally further include steps or units not listed, or may optionally further include other steps or units inherent to these processes, methods, products or equipment.

[0099] The mention of "embodiment" in this article means that the specific features, structures or characteristics described in connection with the embodiment may be included in at least one embodiment of the present invention. The appearance of this phrase in various positions in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art will explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.

[0100] In this application, the term "exemplary" is used to mean "serving as an example, illustration, or description". Any embodiment described as "exemplary" in this application is not necessarily construed as being more preferred or having more advantages than other embodiments. In order to enable any person skilled in the art to implement and use this application, the following description is given. In the following description, details are set forth for the purpose of explanation. It should be understood that those of ordinary skill in the art can recognize that this application can be implemented without using these specific details. In other instances, well-known structures and processes will not be elaborated in detail to avoid unnecessary details from obscuring the description of this application. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope that conforms to the principles and features disclosed in this application.

[0101] It should be noted that since the method of the embodiment of this application is executed in a computer device, the processing objects of each computer device exist in the form of data or information. For example, time, which is actually time information. It can be understood that in subsequent embodiments, if dimensions, quantities, positions, etc. are mentioned, they are all corresponding data existences for the computer device to process, and specific details will not be elaborated here.

[0102] It should be noted that a brief introduction to the artificial intelligence related technologies that may be involved in this application is provided. Artificial Intelligence (AI) uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, including theories, methods, technologies, and application systems that can perceive the environment, acquire knowledge, and use knowledge to achieve the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a way similar to human intelligence. Artificial intelligence also studies the design principles and implementation methods of various intelligent machines to enable machines to have the functions of perception, reasoning, and decision-making.

[0103] Artificial intelligence technology is an interdisciplinary subject with a wide range of fields, including both hardware-level and software-level technologies. Artificial intelligence basic technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, and mechatronics. Artificial intelligence software technologies mainly include several major directions such as computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0104] Computer Vision Technology (CV) Computer vision is a science that studies how to enable machines to "see". More specifically, it refers to using cameras and computers to replace human eyes for tasks such as object recognition and measurement in machine vision, and further performing graphic processing to make the images processed by the computer more suitable for human eyes to observe or be transmitted to instruments for detection. As a scientific discipline, computer vision studies related theories and technologies and attempts to build artificial intelligence systems that can obtain information from images or multi-dimensional data. Computer vision technology usually includes technologies such as image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content / behavior recognition, three-dimensional object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping, and also includes common biometric recognition technologies such as face recognition and fingerprint recognition.

[0105] Single-modal information is data of only one type, such as one of the data information types like text, image, audio, video, electromagnetic signals, etc. Multi-modal information is data information that includes at least two types of single-modal information. Further, multi-modal information is applicable to complex tasks that require integrating multiple information sources, such as sentiment analysis, robot interaction, and autonomous driving. By integrating information of multiple modalities, higher performance and accuracy can usually be achieved in tasks.

[0106] A large model refers to an artificial neural network model with a very large number of parameters. In the field of artificial intelligence, a large model usually refers to a model with hundreds of millions to trillions of parameters. The model usually needs to be trained on a large-scale dataset and requires a large amount of computing resources for optimization and adjustment. Large models are usually used to solve complex natural language processing, computer vision, speech recognition and other tasks. Generative AI is a type of AI that can create new content and ideas, including conversations, stories, images, videos and music. In the embodiments of this application, the large model can be large language models such as ChatGPT, BERT, XLNet, Zhipu Model, Claude, Moonshot AI Model, ChatGLM Model, Qianyitongwen Model, MiniMax Model, Spark Model, Llama Model, 360GPT Model, Qwen Model, Baichuan Model, Lark Model, vivoLM Model and Wenxin Yiyan, and the embodiments of this application do not make limitations.

[0107] The embodiments of this application provide a point cloud data registration method, system, device, computer device and computer-readable storage medium, which will be described in detail below respectively.

[0108] Please refer to Figure 1 , Figure 1 which is a schematic diagram of the scenario of the image processing system provided by the embodiments of this application. The system may include a computer device 100, and a point cloud data registration device is integrated in the computer device 100, such as Figure 1 the computer device in

[0109] In the embodiments of this application, the computer device 100 may be an independent server, or a server network or server cluster composed of servers. For example, the computer device 100 described in the embodiments of this application includes, but is not limited to, a computer, a network host, a single network server, a set of multiple network servers or a cloud server composed of multiple servers. Among them, the cloud server is composed of a large number of computers or network servers based on cloud computing.

[0110] It can be understood that the computer device 100 used in the embodiments of this application may be a device that includes both receiving and transmitting hardware, that is, a device with receiving and transmitting hardware capable of performing two-way communication on a two-way communication link. Such a device may include: a cellular or other communication device, which has a single-line display or a multi-line display or a cellular or other communication device without a multi-line display. Specifically, the computer device 100 may specifically be a desktop terminal or a mobile terminal, and the computer device 100 may specifically also be one of a mobile phone, a tablet computer, a laptop computer, etc.

[0111] Those skilled in the art can understand, Figure 1The application environment shown is merely one application scenario of the solution of this application, and does not constitute a limitation on the application scenarios of the solution of this application. Other application environments may also include more or fewer computer devices than those shown in Figure 1 For example, Figure 1 In

[0112] Figure 1 Figure 1 As shown, the image processing system may further include a memory 200 for storing result data and sample data, such as simulation result data, etc.

[0113] It should be noted that Figure 1 The scene schematic diagram of the image processing system shown is merely an example. The point cloud registration system and the scene described in the embodiments of this application are for more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. Those of ordinary skill in the art can know that with the evolution of the classification control management system and the emergence of new business scenarios, the technical solutions provided by the embodiments of this application are equally applicable to similar technical problems.

[0114] The present invention discloses a point cloud data registration method and device, which overcomes the problems of low overlap rate, incomplete point cloud, and large-scale scene point cloud registration. By adopting a point cloud data registration model based on geometric position attention, it effectively integrates local geometric position features and global structure perception, enabling it to capture discriminative features and establish reliable correspondence relationships even in the case of significant partial overlap and data sparsity, having remarkable robustness and generalization ability, and can be applied to point cloud registration processing on different data sets such as 3DMatch, ModelNet, KITTI, etc.

[0115] Embodiment 1

[0116] Please refer to Figure 2 , Figure 2 which is a schematic flowchart of a point cloud data registration method disclosed in an embodiment of the present invention. Among them, Figure 2 The described point cloud data registration method is applied to an image processing system, such as a local server or a cloud server managed by the image processing system, etc., which is not limited in the embodiments of the present invention. As Figure 2 shown, the point cloud data registration method may include the following operations:

[0117] The method includes:

[0118] S1, obtaining a source point cloud data set and a target point cloud data set;

[0119] It should be noted that the source point cloud dataset and the target point cloud dataset include, but are not limited to, 3DMatch, ModelNet, and KITTI datasets;

[0120] It should be noted that the source point cloud dataset and the target point cloud dataset partially overlap;

[0121] S2. Use the point cloud data registration model to process the source point cloud dataset and the target point cloud dataset to obtain a point cloud data registration transformation parameter set.

[0122] It can be seen that when implementing the point cloud data registration method described in the embodiments of the present invention, by using the point cloud data registration model based on geometric position attention, the robustness and generalization ability of data processing are significantly improved, and it can be applied to point cloud data registration processing on different datasets such as 3DMatch, ModelNet, and KITTI.

[0123] In an optional embodiment, in the above step S2, as Figure 3 shown, the point cloud data registration model includes: a feature extraction module, a local feature fusion module, a global feature learning module, a feature data matching module, and a fine registration module;

[0124] The feature extraction module, the local feature fusion module, the global feature learning module, the feature data matching module, and the fine registration module are sequentially connected by data;

[0125] The feature extraction module is used to perform feature extraction and downsampling processing on the source point cloud dataset and the target point cloud dataset to obtain a dense point cloud superpoint dataset; the dense point cloud superpoint dataset includes a number of superpoint data and their corresponding superpoint geometric feature data;

[0126] The local feature fusion module is used to perform first encoding and feature fusion processing on the source point cloud dataset to obtain a point cloud local fusion feature dataset;

[0127] The global feature learning module is used to perform second encoding and feature extraction and matching processing on the point cloud local fusion feature dataset to obtain a point cloud global fusion feature dataset;

[0128] The feature data matching module is used to perform feature matching processing on the point cloud global fusion feature dataset to obtain a superpoint correspondence information set;

[0129] The fine registration module is used to process the superpoint correspondence information set to obtain a point cloud data registration transformation parameter set.

[0130] It can be seen that when implementing the point cloud data registration method described in the embodiments of the present invention, by adopting a point cloud data registration model based on geometric position attention, local geometric position features and global structure perception are effectively integrated, enabling it to capture discriminative features and establish reliable correspondences even in the case of significant partial overlap and sparse data, and having remarkable robustness and generalization ability.

[0131] In an alternative embodiment, in step S2 above, the using the point cloud data registration model to process the source point cloud data set and the target point cloud data set to obtain a point cloud data registration transformation parameter set includes:

[0132] S21, using the feature extraction module to process the source point cloud data set and the target point cloud data set to obtain a dense point cloud superpoint data set;

[0133] S22, using the local feature fusion module to process the dense point cloud superpoint data set to obtain a point cloud local fusion feature data set;

[0134] S23, using the global feature learning module to process the dense point cloud superpoint data set to obtain a point cloud global fusion feature data set;

[0135] The point cloud global fusion feature data set includes a source point cloud global feature data set and a target point cloud global feature data set;

[0136] S24, using the feature data matching module to process the point cloud global fusion feature data set to obtain a superpoint correspondence information set;

[0137] S25, based on the superpoint correspondence information set, using the fine registration module to process the point cloud local fusion feature data set to obtain a point cloud data registration transformation parameter set.

[0138] It can be seen that when implementing the point cloud data registration method described in the embodiments of the present invention, by adopting a point cloud data registration model based on geometric position attention, local geometric position features and global structure features of the data are effectively obtained, significantly improving the robustness and generalization ability of data processing.

[0139] In an alternative embodiment, in step S21 above, the using the feature extraction module to process the source point cloud data set and the target point cloud data set to obtain a dense point cloud superpoint data set includes:

[0140] S211, performing downsampling processing on the source point cloud data set to obtain a source dense point set

[0141] Downsample the target point cloud dataset to obtain a target dense point set

[0142] It should be noted that the downsampling process refers to performing a convolution operation on the source point cloud dataset and the target point cloud dataset; the convolution operation has good translational invariance and robustness to changes in point cloud density;

[0143] S212. Perform discrimination processing on the source dense point set and the target dense point set to obtain a superpoint dataset;

[0144] The superpoint dataset includes a source superpoint dataset and a target superpoint dataset

[0145] The discrimination processing expression is:

[0146]

[0147] where, represents the dataset in the source superpoint dataset that conforms to the source superpoint data processing rules; represents the dataset in the target superpoint dataset that conforms to the target superpoint data processing rules; represents the source dataset; represents the target dataset; represents the source superpoint dataset; represents the target superpoint dataset; represents the feature dimension;

[0148] S213. Process the superpoint dataset to obtain a dense point cloud superpoint dataset;

[0149] The processing expression is:

[0150]

[0151] where, represents the local point cloud block unit index in the source point cloud dataset; represents the source dense point set; i represents the superpoint index value that satisfies the minimum distance condition; j represents the superpoint cloud space point index; represents the superpoint with index j in the source superpoint dataset; argmin j () represents selecting the j value that makes the Euclidean distance between the dense point and the superpoint with index j the smallest;

[0152] It should be noted that this processing method ensures that each local point cloud block has a unique superpoint as the center and contains rich local geometric information.

[0153] It can be seen that by implementing the point cloud data registration method described in the embodiments of the present invention, downsampling and regular limiting processing are performed on the target point cloud data set and the target point cloud data set to obtain a dense point cloud superpoint data set, which provides support for obtaining local geometric position features and global structure features of the point cloud, enabling it to capture discriminative features and establish reliable correspondence relationships even in the case of significant partial overlap and sparse data.

[0154] In an alternative embodiment, in the above step S22, the use of the local feature fusion module to process the dense point cloud superpoint data set to obtain a point cloud local fusion feature data set includes:

[0155] S221, based on the first encoding model, process the dense point cloud superpoint data set to obtain a first superpoint encoding data set;

[0156] The first superpoint encoding data set includes a source first superpoint encoding data set and a target first superpoint encoding data set;

[0157] The expression of the first encoding model is:

[0158]

[0159] where S i represents the i-th encoded data of the local point cloud block encoding data set; represents the geometric feature corresponding to S i ; represents the position encoding matrix corresponding to S i and ; Θ(S i ) represents a diagonal encoding matrix composed of block matrices; θ k represents the feature encoding method of the k-th channel;

[0160] It should be noted that Θ(S i ) specifies the encoding calculation method of S i ;

[0161] S222, perform feature fusion processing on the first superpoint encoding data set to obtain a point cloud local fusion feature data set;

[0162] The point cloud local fusion feature data set includes a source point cloud local fusion feature data set and a target point cloud local fusion feature data set;

[0163] The expression of the feature fusion processing is:

[0164]

[0165] Among them, represents the fusion feature vector of the i-th superpoint coding data; MLP() represents a fully connected projection network; cat[·] represents the operation of splicing features; W k represents the vector key value in the feature projection matrix; W q represents the vector query value in the feature projection matrix; W v represents the vector value in the feature projection matrix; j represents the index of the dense point cloud data in the local point cloud block; i represents the superpoint coding data index; Θ(S j ) represents a diagonal coding matrix composed of block matrices; k j represents the vector key value of the dense point S j calculated according to the feature projection; v j represents the vector value of the dense point S j calculated according to the feature projection; a i，j represents the attention score;

[0166] It should be noted that the Θ(S j ) stipulates the encoding calculation method of S j .

[0167] It can be seen that by implementing the point cloud data registration method described in the embodiments of the present invention, encoding and feature fusion processing are performed on the target point cloud data set and the target point cloud data set to obtain a point cloud local geometric position feature set, providing a basis for subsequent point cloud data registration, enabling it to capture discriminative features and establish reliable correspondence relationships even in the case of significant partial overlap and data sparsity.

[0168] In an optional embodiment, in the above step S23, the use of the global feature learning module to process the dense point cloud superpoint data set to obtain a point cloud global fusion feature data set includes:

[0169] S231, based on the second encoding model, process the dense point cloud superpoint data set to obtain a second superpoint coding data set;

[0170] The second superpoint coding data set includes a source second superpoint coding data set and a target second superpoint coding data set;

[0171] S232, based on the first self-attention global feature extraction model, process the source second superpoint coding data set to obtain a source point cloud global feature data set;

[0172] The point cloud global feature data set includes a source point cloud global feature data set and a target point cloud global feature data set;

[0173] S233. Process the target second superpoint encoded dataset based on the second self-attention global feature extraction model to obtain a target point cloud global feature dataset;

[0174] S234. Use the cross-attention matching model and the first loss function to process the source point cloud global feature dataset and the target point cloud global feature dataset to obtain a point cloud global fusion feature dataset.

[0175] It can be seen that by implementing the point cloud data registration method described in the embodiments of the present invention, encoding, feature extraction, and feature fusion processing are performed on the dense point cloud superpoint dataset to obtain a point cloud global fusion feature dataset, providing a basis for subsequent point cloud data registration, enabling it to capture discriminative features and establish reliable correspondences even in the case of significant partial overlap and data sparsity.

[0176] In an optional embodiment, in the above step S231, the expression of the second encoding model is:

[0177]

[0178]

[0179] Among them, δ i，j represents the overall encoding embedding data of the superpoint; W D represents the superpoint distance embedding learning matrix; represents the first superpoint angle embedding learning matrix; represents the second superpoint angle embedding learning matrix; represents the distance encoding corresponding to two superpoints; ρ i，j represents the Euclidean distance between the i-th superpoint and the j-th superpoint; σ d represents the sensitivity coefficient of the distance encoding to the Euclidean distance; d t represents the dimension of the feature; represents the i-th source superpoint data; represents the j-th source superpoint data; represents the local point cloud block represented by the superpoint in the source point cloud; represents the superpoint and the normal vector of the local plane formed by the two nearest neighbor points and , that is, the first normal vector; represents the angle between the superpoint and the first local plane, that is, the first angle; represents the encoding of the first angle, that is, the first encoding; σ α represents the sensitivity coefficient of the first encoding to the first angle change; Represents the local point cloud block represented by the superpoint in the source point cloud ; Represents the said superpoint And the normal vector of the local plane composed of two nearest neighbor points And That is, the second normal vector; Represents the superpoint And the included angle between the said second local plane, that is, the second included angle; Represents the encoding of the said second included angle, that is, the second encoding; σ b Represents the sensitivity coefficient of the said second encoding to the angle change of the said second included angle;

[0180] It can be seen that by implementing the point cloud data registration method described in the embodiments of the present invention, the second encoding model is used to encode the dense point cloud superpoint data set to obtain the second superpoint encoding data set, laying a foundation for the subsequent acquisition of the point cloud global fusion feature data set, enabling it to capture discriminative features and establish reliable corresponding relationships even in the case of significant partial overlap and data sparsity.

[0181] In an optional embodiment, in the above step S232, the expression of the self-attention global feature extraction model of the source point cloud is:

[0182]

[0183] Wherein, Represents the self-attention feature data of the i-th superpoint in the source point cloud; a i，j Represents the self-attention score between the i-th superpoint and the j-th superpoint in the source point cloud; W s Q Represents the query vector matrix in the source learning feature projection matrix; W s K Represents the key-value vector matrix in the source learning feature projection matrix; W s V Represents the value vector matrix in the source learning feature projection matrix; W s δ Represents the learnable projection matrix of the position encoding embedding; Represents the feature representation of the source superpoint i; δ j，i Represents the overall encoding embedding data of the superpoint; d t Represents the feature dimension; i represents the index of the current source superpoint; j represents the index of other source superpoints;

[0184] It can be seen that when implementing the point cloud data registration method described in the embodiments of the present invention, the first self-attention global feature extraction model is used to perform feature extraction processing on the second superpoint encoded data set to obtain the source point cloud global feature data set, providing a basis for subsequent point cloud data registration, enabling it to capture discriminative features and establish reliable correspondence relationships even in the case of significant partial overlap and data sparsity.

[0185] In an alternative embodiment, in the above step S232, the expression of the second self-attention global feature extraction model is:

[0186]

[0187] Wherein, represents the self-attention feature data of the j-th superpoint in the target point cloud; b i，j represents the self-attention score between the i-th superpoint and the j-th superpoint in the target point cloud; W T Q represents the query vector matrix in the target learning feature projection matrix; W T K represents the key-value vector matrix in the target learning feature projection matrix; W T V represents the value vector matrix in the target learning feature projection matrix; W T δ represents the learnable projection matrix of the position encoding embedding; represents the feature representation of the target superpoint i; β i，j represents the superpoint overall encoding embedding data; d t represents the feature dimension; i represents the index of the current target superpoint; j represents the index of other source superpoints;

[0188] It can be seen that when implementing the point cloud data registration method described in the embodiments of the present invention, the second self-attention global feature extraction model is used to perform feature extraction processing on the second superpoint encoded data set to obtain the target point cloud global feature data set, providing a basis for subsequent point cloud data registration, enabling it to capture discriminative features and establish reliable correspondence relationships even in the case of significant partial overlap and data sparsity.

[0189] In an alternative embodiment, in the above step S234, the expression of the cross-attention matching model is:

[0190]

[0191] Wherein, represents the cross-attention mixed feature of the source point cloud; a i，jRepresents the attention score between the source point cloud i and the target superpoint j; WQ represents the query vector matrix in the learned feature projection matrix; W K Represents the key-value vector matrix in the learned feature projection matrix; W V Represents the value vector matrix in the learned feature projection matrix; W δ Represents the learnable projection matrix of the position encoding embedding; Represents the target point cloud super dataset; Represents the feature representation of the target superpoint j; Represents the feature representation of the source superpoint i;

[0192] In an optional embodiment, in the above step S234, the first loss function expression is:

[0193]

[0194] Where, Represents the distance in the feature space; Represents The overlap ratio of two local point cloud patches; Represents the positive weight coefficient; Represents the negative weight coefficient; Represents the local point cloud patch where the superpoint i is located in the source superpoints; Represents The set composed of the corresponding positive point cloud patches in the target point cloud; Represents The set composed of the corresponding negative point cloud patches in the target point cloud; Y represents the first hyperparameter; Δ p Represents the second hyperparameter; Δ n Represents the third hyperparameter; A represents the anchor point cloud patch, that is, the set of local point cloud patches in the source point cloud where there are corresponding positive point cloud patches in the target point cloud; Represents The local point cloud patch with index k in; Represents The local point cloud patch with index j in; Represents The local point cloud patch with index i in; Represents The local point cloud patch with index k in; Represents the local point cloud patch with index i in the anchor point cloud patch in the source point cloud; Represents the local point cloud patch with index j in the anchor point cloud patch in the target point cloud; Represents the distance in the feature space; Represents the distance in the feature space;

[0195] It should be noted that in this embodiment, Δ pRepresents a negative margin value, Δ p = 0.1; Δ n Represents a positive margin value, Δ n = 1.4;

[0196] In an optional embodiment, in the above step S24, the use of the feature data matching module to process the point cloud global fusion feature data set to obtain a superpoint correspondence information set includes:

[0197] S241, normalize the source point cloud global feature data set and the target point cloud global feature data set to obtain a normalized source global feature data set and a normalized target global feature data set;

[0198] S242, transform the normalized source global feature data set and the normalized target global feature data set to obtain a global correlation feature data set;

[0199] The transformation processing expression is:

[0200]

[0201] where s i，j represents the correlation between two superpoint features with encoding index i in the source superpoint data set and encoding index j in the target superpoint data set; exp() represents performing exponential function operation processing; represents the normalized feature of the superpoint with encoding index i in the source superpoint data set; represents the normalized feature of the superpoint with encoding index j in the target superpoint data set;

[0202] S243, perform double normalization processing on the global correlation feature data set to obtain a normalized correlation feature data set;

[0203] The double normalization processing expression is:

[0204]

[0205] where represents the normalized correlation value; s i，k represents the sum of all correlation values in the i-th row; s l，j represents the sum of all correlation values in the j-th column; k represents the index of the first round of normalization; l represents the index of the second round of normalization;

[0206] It should be noted that the represents the element at the index position i, j in the correlation matrix :

[0207] It should be noted that the s k，j, representing the element at indices i and k in the correlation matrix, which is the denominator part for the first-round normalization process;

[0208] S244. Perform confidence processing on the normalized correlation feature dataset to obtain a point cloud global relationship information set;

[0209] The expression for the confidence processing is:

[0210]

[0211] where, represents the matching set after confidence processing represents the set of super points the x i -th super point in; represents the set of super points the y i -th super point in; (x i , y i ) represents the index of a pair of super point matches; represents selecting the top k pairs of super point matches with the strongest correlation from the normalized correlation matrix ;

[0212] It should be noted that each element in the represents a pair of super point matches after confidence screening;

[0213] S245. Extract information from the point cloud global relationship information set to obtain a global mixed feature information set;

[0214] It should be noted that the information extraction means;

[0215] S246. Perform calculation processing on the global mixed feature information set to obtain a super point scoring matrix information set;

[0216] The expression for the calculation processing is:

[0217]

[0218] where, Z i represents the scoring matrix between the source local dense point cloud block and the target local dense point cloud block; represents the mixed feature of the x i -th point in the source point dataset; represents the mixed feature of the y i -th point in the target point dataset; W S represents the weight matrix of the source point dataset; W T represents the weight matrix of the target point dataset; represents the x i -th point in the source point set; Represents the y-th point in the set of target points; i ; n i Represents the number of points in the source local dense point cloud block; m i Represents the number of points in the target local dense point cloud block; Θ() represents the diagonal encoding matrix composed of block matrices; Represents the dimensional constant for normalization; Represents the real number field;

[0219] It should be noted that the Θ() stipulates the encoding calculation method of the point cloud data;

[0220] It should be noted that the Is the dimensionality of the input feature;

[0221] S247, perform conversion processing on the superpoint scoring matrix information set to obtain the superpoint correspondence confidence information set;

[0222] The conversion processing expression is;

[0223]

[0224] Among them, Represents the dense point correspondence confidence information set; Z(x i , ) represents all columns in the scoring matrix related to the source point x i ; Z(, y i ) represents all rows in the scoring matrix related to the source point y i (i.e., the scoring of the source point);

[0225] It should be noted that the Represents the confidence of the dense point correspondence;

[0226] It should be noted that the Z(x i , ) represents the scoring of the target point;

[0227] It should be noted that the Z(, y i ) represents the scoring of the source point;

[0228] S248, process the superpoint correspondence confidence information set to obtain the superpoint correspondence information set;

[0229]

[0230] Among them, C represents the dense point correspondence information set;

[0231] It should be noted that the C is obtained by performing a union operation on each C i ;

[0232] Among them, U represents performing the union operation; represents the points in the set of local dense point cloud blocks represented by x i in the source point cloud dataset, and is represented by x j ; represents the points in the set of local dense point cloud blocks represented by y i in the target point cloud dataset, and is represented by y j ; mutual_topk x，y () represents selecting the k pairs of matches that are the strongest among each other from the confidence information set.

[0233] It can be seen that by implementing the point cloud data registration method described in the embodiments of the present invention, the feature data matching module is used to perform normalization and data transformation processing on the point cloud global fusion feature dataset, obtaining a confidence information set of superpoint correspondence relationships, providing a basis for subsequent point cloud data registration, enabling it to capture discriminative features and establish reliable correspondence relationships even in the case of significant partial overlap and data sparsity.

[0234] In an optional embodiment, in the above step S241, the normalization processing expression is:

[0235]

[0236] Among them, x iN represents the normalized value of the i-th original data; x i represents the i-th original data in the original dataset; x max represents the maximum value in the original dataset; x min represents the minimum value in the original dataset; i represents the data number index;

[0237] It should be noted that the original dataset is the source point cloud global feature dataset and the target point cloud global feature dataset;

[0238] It can be seen that by implementing the point cloud data registration method described in the embodiments of the present invention, normalization processing is performed on the source point cloud global feature dataset and the target point cloud global feature dataset, obtaining a normalized source global feature dataset and a normalized target global feature dataset, providing a basis for subsequent point cloud data registration, enabling it to capture discriminative features and establish reliable correspondence relationships even in the case of significant partial overlap and data sparsity.

[0239] In an optional embodiment, in the above step S25, based on the superpoint correspondence relationship information set, the fine registration module is used to process the point cloud local fusion feature dataset, obtaining a point cloud data registration transformation parameter set, including:

[0240] S251. Perform parameter estimation on the local fusion feature dataset of the point cloud to obtain a first set of transformation parameter information;

[0241] The parameter estimation expression is:

[0242]

[0243] where R i , t i represent the transformation parameters between local point cloud blocks for index i, i.e., the rotation matrix and the translation vector; is the weight value of the dense point correspondence in the local point cloud block; represents the point with source point cloud index x j in the matching relationship; represents the point with target point cloud index x j in the matching relationship;

[0244] It should be noted that in this embodiment, the weight value of the superpoint correspondence is determined by the confidence matrix and is set to the confidence value;

[0245] S252. Based on the second loss function, perform inlier matching processing on the first set of transformation parameter information and the superpoint correspondence information set to obtain a set of point cloud data registration transformation parameters;

[0246] The expression of the second loss function is:

[0247]

[0248] where α represents the first empirical parameter; η represents the second empirical parameter; represents the set of true matches in the local point cloud block with index i; represents the loss value of the matching points in the i-th local point cloud block; (x, y) represents the coordinate values of the matching points; C(x, y) represents the confidence value of the matching points; N represents the number of local point cloud blocks;

[0249] It should be noted that the first empirical parameter is set to 0.25, i.e., α = 0.25;

[0250] It should be noted that the second empirical parameter is set to 2, i.e., η = 2;

[0251] The expression of the inlier matching processing is:

[0252]

[0253] where R i , t i represent the i-th candidate transformation matrix parameters obtained in the local generation stage; represents the matching points in the source point cloud; represents the matching points in the target point cloud; C represents the set of matching point pairs; represents performing the Iverson operation, which is 1 when the distance between the transformed point pairs is less than the threshold T, otherwise 0. This operation is used to mark inliers; T represents the inlier matching distance threshold;

[0254] It should be noted that in this embodiment, the inlier matching distance threshold is set to 0.1.

[0255] It can be seen that implementing the point cloud data registration method described in the embodiments of the present invention, using the fine registration module to perform normalization and data transformation processing on the superpoint correspondence information set and the point cloud local fusion feature data set, and obtaining the point cloud data registration transformation parameter set, has significant robustness and generalization ability, and can be applied to point cloud data registration processing on different data sets such as 3DMatch, ModelNet, KITTI, etc.

[0256] Embodiment Two

[0257] Please refer to Figure 4 , Figure 4 which is a schematic structural diagram of a point cloud data registration device disclosed in the embodiments of the present invention. Among them, Figure 4 the described device is applied in an image processing system, such as a local server or a cloud server for image processing system management, etc., which is not limited in the embodiments of the present invention. As Figure 4 shown, the device may include:

[0258] A data acquisition module 101 that acquires a source point cloud data set and a target point cloud data set;

[0259] A data processing module 102 that uses a point cloud data registration model to process the source point cloud data set and the target point cloud data set to obtain a point cloud data registration transformation parameter set.

[0260] It can be seen that implementing the point cloud data registration method described in the embodiments of the present invention, adopting a point cloud data registration model based on geometric position attention, effectively integrates local geometric position features and global structure perception, enabling it to capture discriminative features and establish reliable correspondences even in the case of significant partial overlap and data sparsity, having significant robustness and generalization ability, and can be applied to point cloud registration processing on different data sets such as 3DMatch, ModelNet, KITTI, etc.

[0261] Embodiment Three

[0262] Please refer to Figure 5 , Figure 5It is a schematic structural diagram of another point cloud data registration device disclosed in an embodiment of the present invention. Among them, Figure 5 The described device can be applied to a satellite observation system, such as a local server or a cloud server for satellite observation, etc., which is not limited in the embodiments of the present invention. Such as Figure 5 As shown, the device may include:

[0263] A memory 202 storing executable program code;

[0264] A processor 201 coupled to the memory;

[0265] The processor 201 calls the executable program code stored in the memory 202 to execute the steps in the point cloud data registration method described in Embodiment 1.

[0266] Embodiment 4

[0267] An embodiment of the present invention discloses a computer-readable storage medium storing a computer program for electronic data exchange, wherein the computer program causes a computer to execute the steps in the point cloud data registration method described in Embodiment 1.

[0268] The device embodiments described above are only illustrative. The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed to multiple network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative labor.

[0269] Through the specific descriptions of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution or the part that contributes to the prior art can be embodied in the form of a software product, and this computer software product can be stored in a computer-readable storage medium. The storage medium includes read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc memories, magnetic disk memories, tape memories, or any other medium that can be used to carry or store data and is computer-readable.

[0270] Finally, it should be noted that: What is disclosed in an outbound order splitting method and device according to an embodiment of the present invention is only a preferred embodiment of the present invention, and is only used to illustrate the technical solution of the present invention, rather than limiting it; Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: It is still possible to modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; And these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A point cloud data registration method, characterized in that, The method includes: S1. Obtain a source point cloud data set and a target point cloud data set; S2. Use a point cloud data registration model to process the source point cloud data set and the target point cloud data set to obtain a point cloud data registration transformation parameter set.

2. The point cloud data registration method according to claim 1, wherein The point cloud data registration model includes: a feature extraction module, a local feature fusion module, a global feature learning module, a feature data matching module, and a fine registration module; The feature extraction module, the local feature fusion module, the global feature learning module, the feature data matching module, and the fine registration module are sequentially connected by data; The feature extraction module is used to perform feature extraction and downsampling processing on the source point cloud data set and the target point cloud data set to obtain a dense point cloud superpoint data set; The local feature fusion module is used to perform first encoding and feature fusion processing on the source point cloud data set to obtain a point cloud local fusion feature data set; The global feature learning module is used to perform second encoding and feature extraction and matching processing on the point cloud local fusion feature data set to obtain a point cloud global fusion feature data set; The feature data matching module is used to perform feature matching processing on the point cloud global fusion feature data set to obtain a superpoint correspondence information set; The fine registration module is used to process the superpoint correspondence information set to obtain a point cloud data registration transformation parameter set.

3. The point cloud data registration method according to claim 2, characterized in that The using the point cloud data registration model to process the source point cloud data set and the target point cloud data set to obtain a point cloud data registration transformation parameter set includes: S21. Use the feature extraction module to process the source point cloud data set and the target point cloud data set to obtain a dense point cloud superpoint data set; S22. Use the local feature fusion module to process the dense point cloud superpoint data set to obtain a point cloud local fusion feature data set; S23. Use the global feature learning module to process the dense point cloud superpoint data set to obtain a point cloud global fusion feature data set; The point cloud global fusion feature data set includes a source point cloud global feature data set and a target point cloud global feature data set; S24. Use the feature data matching module to process the point cloud global fusion feature data set to obtain a superpoint correspondence information set; S25. Based on the superpoint correspondence information set, use the fine registration module to process the point cloud local fusion feature data set to obtain a point cloud data registration transformation parameter set.

4. The point cloud data registration method according to claim 3, wherein The using the feature extraction module to process the source point cloud data set and the target point cloud data set to obtain a dense point cloud superpoint data set includes: S211, downsample the source point cloud dataset to obtain a source dense point set Downsample the target point cloud dataset to obtain a target dense point set S212, perform discrimination processing on the source dense point set and the target dense point set to obtain a superpoint data set; The superpoint dataset includes a source superpoint dataset and a target superpoint dataset The discrimination processing expression is: Among them, represents the data set in the source superpoint data set that conforms to the source superpoint data processing rules; represents the data set in the target superpoint data set that conforms to the target superpoint data processing rules; represents the source data set; represents the target data set; represents the source superpoint data set; represents the target superpoint data set; represents the feature dimension; S213. Process the superpoint data set to obtain a dense point cloud superpoint data set; The processing expression is: Among them, represents the local point cloud block unit index in the source point cloud dataset; represents the source dense point set; i represents the superpoint index value that satisfies the minimum distance condition; j represents the superpoint cloud space point index; represents the superpoint with index j in the source superpoint dataset; argmin j () represents selecting the j value that minimizes the Euclidean distance between the dense point and the superpoint with index j.

5. The point cloud data registration method according to claim 3, wherein The using the local feature fusion module to process the dense point cloud superpoint data set to obtain a point cloud local fusion feature data set includes: S221. Based on a first encoding model, process the dense point cloud superpoint data set to obtain a first superpoint encoding data set; The first superpoint encoding dataset includes a source first superpoint encoding dataset and a target first superpoint encoding dataset; The first encoding model expression is: Among them, S i represents the i-th encoded data of the local point cloud block encoding dataset; represents the geometric feature corresponding to the above S i ; represents the position encoding matrix corresponding to the above S i and the above ; Θ(S i ) represents a diagonal encoding matrix composed of block matrices; θ k represents the feature encoding method of the k-th channel; S222. Perform feature fusion processing on the first superpoint encoding dataset to obtain a point cloud local fusion feature dataset; The point cloud local fusion feature dataset includes a source point cloud local fusion feature dataset and a target point cloud local fusion feature dataset; The feature fusion processing expression is: Among them, represents the fusion feature vector of the i-th superpoint coding data; MLP() represents a fully connected projection network; cat[·] represents the operation of concatenating features; W k represents the vector key value in the feature projection matrix; W q represents the vector query value in the feature projection matrix; W v represents the vector value in the feature projection matrix; j represents the dense point cloud data index in the local point cloud block; i represents the superpoint coding data index; Θ(S j ) represents a diagonal coding matrix composed of block matrices; kj represents the vector key value of the dense point S j calculated according to the feature projection; v j represents the vector value of the dense point S j calculated according to the feature projection; a i，j represents the attention score.

6. The point cloud data registration method according to claim 3, wherein Using the global feature learning module to process the dense point cloud superpoint dataset to obtain a point cloud global fusion feature dataset, including: S231. Based on the second encoding model, process the dense point cloud superpoint dataset to obtain a second superpoint encoding dataset; The second superpoint encoding dataset includes a source second superpoint encoding dataset and a target second superpoint encoding dataset; S232. Based on the first self-attention global feature extraction model, process the source second superpoint encoding dataset to obtain a source point cloud global feature dataset; The point cloud global feature dataset includes a source point cloud global feature dataset and a target point cloud global feature dataset; S233. Based on the second self-attention global feature extraction model, process the target second superpoint encoding dataset to obtain a target point cloud global feature dataset; S234. Using the cross-attention matching model and the first loss function, process the source point cloud global feature dataset and the target point cloud global feature dataset to obtain a point cloud global fusion feature dataset.

7. The point cloud data registration method according to claim 3, wherein Using the feature data matching module to process the point cloud global fusion feature dataset to obtain a superpoint correspondence information set, including: S241. Normalize the source point cloud global feature dataset and the target point cloud global feature dataset to obtain a normalized source global feature dataset and a normalized target global feature dataset; S242. Perform transformation processing on the normalized source global feature dataset and the normalized target global feature dataset to obtain a global correlation feature dataset; The transformation processing expression is: Among them, s i，j represents the correlation between two hyperpoint features with encoding index i in the source hyperpoint dataset and encoding index j in the target hyperpoint dataset; exp() represents performing exponential function operation processing; represents the normalized feature of the hyperpoint with encoding index i in the source hyperpoint dataset; represents the normalized feature of the hyperpoint with encoding index j in the target hyperpoint dataset; S243. Perform double normalization processing on the global correlation feature dataset to obtain a normalized correlation feature dataset; The double normalization processing expression is: Among them, represents the normalized correlation value; s i，k represents the sum of all correlation values in the i-th row; s l，j represents the sum of all correlation values in the j-th column; k represents the index of the first-round normalization; l represents the index of the second-round normalization; S244. Perform confidence processing on the normalized correlation feature dataset to obtain a point cloud global relationship information set; The confidence processing expression is: Among them, represents the matching set after confidence processing; represents the superpoint set the x i th superpoint; represents the superpoint set the y i th superpoint; (x i , y i ) represents the index of a pair of superpoint matches; represents selecting the top k pairs of superpoint matches with the strongest correlations from the normalized correlation matrix ; S245. Extract information from the point cloud global relationship information set to obtain a global mixed feature information set; S246. Perform calculation processing on the global mixed feature information set to obtain a scoring matrix information set between local dense point cloud blocks; The calculation processing expression is: Among them, Z i represents the scoring matrix between the source local dense point cloud block and the target local dense point cloud block; represents the mixed feature of the x i -th point in the source point dataset; represents the mixed feature of the y i -th point in the target point dataset; W s represents the weight matrix of the source point dataset; W T represents the weight matrix of the target point dataset; represents the x i -th point in the source point set; represents the y i -th point in the target point set; n i represents the number of points in the source local dense point cloud block; m i represents the number of points in the target local dense point cloud block; Θ() represents the diagonal encoding matrix composed of block matrices; represents the dimensional constant for normalization; represents the real number field; S247. Perform conversion processing on the dense point cloud block scoring matrix information set to obtain a local dense point cloud block correspondence confidence information set; The conversion processing expression is; Among them, represents the confidence information set of the dense point correspondence; Z(x i , ) represents all columns in the scoring matrix related to the source point x i ; Z(, y i ) represents all rows in the scoring matrix related to the source point y i ; S248. Process the dense point correspondence confidence information set to obtain a dense point correspondence information set; Where C represents the dense point correspondence information set; Among them, U represents the operation of taking the union; represents the points in the set of local dense point cloud blocks represented by x in the source point cloud dataset, which are represented by x i ; j represents; represents the points in the set of local dense point cloud blocks represented by y in the target point cloud dataset, which are represented by y i ; j represents; mutual_topk x，y () represents selecting the top k matching pairs that are the strongest among each other from the confidence information set.

8. A point cloud data registration device, the device includes: A data acquisition module that acquires a source point cloud data set and a target point cloud data set; A data processing module that uses a point cloud data registration model to process the source point cloud data set and the target point cloud data set to obtain a point cloud data registration transformation parameter set.

9. A point cloud data registration device, characterized in that, The device includes: A memory storing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the point cloud data registration method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when called, are used to execute the point cloud data registration method according to any one of claims 1-7.

Images