Implicit feature coding method, device and equipment of three-dimensional point cloud, medium and product
By building standardized local sampling spheres for three-dimensional point clouds and performing implicit feature encoding, the problem of implicit representations not matching the point cloud data structure is solved, and the performance of point cloud deep learning network is improved.
Patent Information
- Application Number
- CN202510152814.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-12
- Publication Date
- 2025-06-10
AI Technical Summary
When applying implicit representations to point cloud analysis tasks, the prior art faces the problem of mismatch between the data structures of continuous implicit fields and discrete irregular point clouds, which leads to incompatibility with existing deep learning architectures.
An implicit feature encoding method for three-dimensional point clouds is proposed. By pre-constructing a global sampling sphere, an initial local sampling sphere is constructed for each data point in the input three-dimensional point cloud, and its main direction is aligned with the coordinate axis of the Cartesian coordinate system to obtain a standardized local sampling sphere. Then, based on these specifications, the implicit fields of the three-dimensional point cloud are sampled in field values, and the implicit eigenvectors of the corresponding data points are encoded.
By introducing rich geometric information in implicit representations, the performance of existing point cloud deep learning networks in point cloud analysis tasks is improved, and the problem of implicit representations not matching the point cloud data structure is solved.
Smart Images
Figure CN120125832A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of surveying and mapping and geographic information technology, and in particular to a method, device, equipment, medium and product for implicit feature coding of a three-dimensional point cloud. Background Art
[0002] Similar to the development of 2D image processing, the field of 3D point cloud analysis is currently dominated by deep learning technology. As the cornerstone of point cloud deep learning, PointNet solves the challenge of point cloud disorder and creates a new trend of learning directly from irregular point cloud data. Since then, point cloud deep learning strategies have developed rapidly, and various novel network architectures capable of directly processing point cloud data have emerged, and have achieved year-on-year accuracy improvements on major benchmark datasets. The idea of point cloud deep learning can be summarized as: using a carefully designed neural network to learn the potential mapping relationship between the input point cloud and the true value label.
[0003] In addition to explicit representations such as point clouds, geometric shapes can also be expressed as a zero-order isosurface described by a continuous implicit function. This implicit representation can approximate the shape itself at arbitrary resolution and contains detailed and important geometric information. Although implicit representation is very effective in reconstruction tasks, its potential in point cloud analysis tasks including semantic segmentation remains to be further explored. The challenge of applying implicit representation to point cloud analysis tasks is that there is a data structure mismatch between the continuous implicit field and the discrete and irregular point cloud, which makes it incompatible with existing deep learning architectures designed specifically for point cloud analysis tasks. Summary of the invention
[0004] In order to solve the above problems, the present invention proposes an implicit feature encoding method, device, equipment, medium and product for three-dimensional point cloud, which can introduce rich geometric information contained in the implicit representation and improve the performance of existing point cloud deep learning networks for point cloud analysis tasks.
[0005] An embodiment of the present invention provides a method for implicit feature encoding of a three-dimensional point cloud, comprising:
[0006] According to the pre-constructed global sampling sphere, an initial local sampling sphere is constructed for each data point in the input 3D point cloud;
[0007] Aligning the main direction of the initial local sampling sphere of each data point with the coordinate axis of the rectangular coordinate system where the three-dimensional point cloud is located to obtain a standard local sampling sphere of each data point;
[0008] According to the canonical local sampling sphere of each data point, the implicit field of the three-dimensional point cloud is sampled, and the implicit feature vector of the corresponding data point is encoded.
[0009] Preferably, the process of pre - constructing a global sampling sphere specifically includes:
[0010] Taking the coordinate origin (0, 0, 0) as the center of the sphere and constructing a sphere domain range with a preset radius;
[0011] Taking the center of the sphere as one sampling position, randomly determining M sampling positions within the determined sphere domain range, and updating the randomly determined M sampling positions with a preset optimization objective;
[0012] Constructing a global sampling sphere with the updated M sampling positions;
[0013] Wherein, M is an integer not less than 1.
[0014] Furthermore, the optimization objective is
[0015] The global sampling sphere
[0016] Wherein, x i , x j respectively represent the i - th sampling position and the j - th sampling position.
[0017] As a preferred solution, according to the pre - constructed global sampling sphere, constructing an initial local sampling sphere for each data point in the input three - dimensional point cloud, including:
[0018] Adjusting each sampling position in the global sampling sphere according to the specific position of each data point in the three - dimensional point cloud to obtain the corresponding initial local sampling sphere;
[0019] Wherein, the initial local sampling sphere of data point q x i is the i - th sampling position among the M sampling positions, and M is an integer not less than 1.
[0020] As a preferred solution, aligning the principal direction of the initial local sampling sphere of each data point with the coordinate axes of the rectangular coordinate system where the three - dimensional point cloud is located to obtain the canonical local sampling sphere of each data point, including:
[0021] Retrieving the data points located inside the initial local sampling sphere of each data point, and subtracting the coordinate value of the data point from the coordinate values of the retrieved data points to obtain the neighborhood point set of each data point;
[0022] Using the singular value decomposition method to decompose the neighborhood point set of each data point into an orthogonal matrix composed of left singular vectors, an orthogonal matrix composed of right singular vectors, and a diagonal matrix;
[0023] Based on predefined anchor points, updating each right singular vector, and each updated right singular vector represents a principal direction of the initial local sampling sphere;
[0024] Align the three principal directions of the initial local sampling spheres of each data point with the three coordinate axes of the rectangular coordinate system where the three-dimensional point cloud is located to obtain the canonical local sampling spheres of each data point.
[0025] As a preferred solution, according to the canonical local sampling spheres of each data point, sample the field values of the implicit field of the three-dimensional point cloud and encode them to form the implicit feature vectors corresponding to the data points, including:
[0026] Establish a k-dimensional tree structure based on the three-dimensional point cloud;
[0027] For each canonical local sampling sphere, retrieve the data point in the three-dimensional point cloud that is closest to each sampling position, and divide the obtained closest distance by the radius of the canonical local sampling sphere as the eigenvalue corresponding to the sampling position;
[0028] According to the order of the sampling positions in each canonical local sampling sphere, concatenate the eigenvalues of each obtained sampling position to form a multi-dimensional vector as the implicit feature vector of the data point corresponding to the canonical local sampling sphere.
[0029] An embodiment of the present invention further provides an implicit feature encoding device for a three-dimensional point cloud, and the device includes:
[0030] An initial module, configured to construct an initial local sampling sphere for each data point in the input three-dimensional point cloud according to a pre-constructed global sampling sphere;
[0031] A canonical module, configured to align the principal directions of the initial local sampling spheres of each data point with the coordinate axes of the rectangular coordinate system where the three-dimensional point cloud is located to obtain the canonical local sampling spheres of each data point;
[0032] A vector module, configured to sample the field values of the implicit field of the three-dimensional point cloud according to the canonical local sampling spheres of each data point and encode them to form the implicit feature vectors corresponding to the data points.
[0033] An embodiment of the present invention further provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the implicit feature encoding method for a three-dimensional point cloud as described in any one of the above embodiments.
[0034] An embodiment of the present invention further provides a computer-readable storage medium, and the computer-readable storage medium includes a stored computer program. Wherein, when the computer program runs, it controls the device where the computer-readable storage medium is located to execute the implicit feature encoding method for a three-dimensional point cloud as described in any one of the above embodiments.
[0035] An embodiment of the present invention further provides a computer program product, including a computer program / instructions, which when executed by a processor, implement the steps of any of the above methods.
[0036] The present invention provides a method, apparatus, device, medium, and product for implicit feature encoding of 3D point clouds. According to a pre-constructed global sampling sphere, an initial local sampling sphere is constructed for each data point in the input 3D point cloud; the main direction of the initial local sampling sphere of each data point is aligned with the coordinate axes of the rectangular coordinate system where the 3D point cloud is located to obtain a canonical local sampling sphere for each data point; according to the canonical local sampling spheres of each data point, field value sampling is performed on the implicit field of the 3D point cloud, and an implicit feature vector corresponding to the data point is encoded. The solution of the present application can introduce the rich geometric information contained in the implicit representation and improve the performance of the existing point cloud deep learning network for point cloud analysis tasks. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 is a schematic flowchart of the method for implicit feature encoding of 3D point clouds provided by an embodiment of the present invention;
[0038] Figure 2 is a schematic diagram of the effects of global sampling spheres under different M value settings;
[0039] Figure 3 is a schematic diagram of the effects of partial point cloud data provided by an embodiment of the present invention;
[0040] Figure 4 is a schematic diagram of the visualization effect of the implicit features of partial "person" category point cloud data provided by an embodiment of the present invention;
[0041] Figure 5 is a schematic diagram of the visualization effect of the implicit features of point cloud data of several categories provided by an embodiment of the present invention;
[0042] Figure 6 is a schematic diagram of the visualization effect of the implicit features of a certain field point cloud data provided by an embodiment of the present invention;
[0043] Figure 7 is a schematic structural diagram of an apparatus for implicit feature encoding of 3D point clouds provided by an embodiment of the present invention;
[0044] Figure 8 is a schematic structural diagram of a terminal device provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention 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 belong to the scope of protection of the present invention.
[0046] In view of the above technical problems, this case provides an implicit feature encoding method for three-dimensional point clouds. Refer to Figure 1 , which is a schematic flowchart of the implicit feature encoding method for three-dimensional point clouds provided by the embodiments of the present invention. The method includes the following steps:
[0047] Step S1: Construct an initial local sampling sphere for each data point in the input three-dimensional point cloud according to the pre-constructed global sampling sphere;
[0048] Step S2: Align the main direction of the initial local sampling sphere of each data point with the coordinate axes of the rectangular coordinate system where the three-dimensional point cloud is located to obtain the canonical local sampling sphere of each data point;
[0049] Step S3: Sample the field values of the implicit field of the three-dimensional point cloud according to the canonical local sampling sphere of each data point, and encode to form an implicit feature vector corresponding to the data point.
[0050] In the specific implementation of this embodiment, when performing implicit feature encoding of three-dimensional point clouds, first construct a global sampling sphere;
[0051] Construct an initial local sampling sphere for each data point in the input three-dimensional point cloud according to the constructed global sampling sphere. That is, according to the positions of different data points in the three-dimensional point cloud, construct the global sampling sphere of the sampling point, and then obtain the initial local sampling sphere of the data point relative to all three-dimensional point clouds.
[0052] Align the main direction of the initial local sampling sphere of each data point with the coordinate axes of the rectangular coordinate system where the three-dimensional point cloud is located, and convert each initial local sampling sphere to its canonical posture to obtain the canonical local sampling sphere of each data point.
[0053] Based on the canonical local sampling spheres of each data point in the canonical posture, sample the field values of the implicit field of the three-dimensional point cloud, and encode to form an implicit feature vector corresponding to the data point.
[0054] The solution of this application discloses an implicit feature encoding method for 3D point clouds, which can enhance the input point cloud into a data representation that combines explicit point position information and implicit shape geometry, thereby improving the performance of point cloud deep neural networks. Different from existing normalization methods, the local normalization method proposed in this embodiment calculates the normalization pose based on the local neighborhood rather than the entire point cloud instance. Therefore, it is applicable not only to independent objects but also to large scenes containing multiple objects. Secondly, this method transforms the sampling sphere rather than the original point cloud data, thus ensuring the continuity of the encoded features in the scene.
[0055] In another embodiment provided by the present invention, the process of constructing the global sampling sphere in step S1 specifically includes the following steps:
[0056] Establish a spherical domain range with the coordinate origin (0, 0, 0) as the center of the sphere and a radius of r;
[0057] Take the center of the sphere as 1 sampling position, and randomly determine M sampling positions within the determined spherical domain range, and update these M sampling positions with a preset optimization target;
[0058] Construct a global sampling sphere with the updated M sampling positions;
[0059] Wherein, M is an integer not less than 1.
[0060] See Figure 2 , which is a schematic diagram of the effects of global sampling spheres under different M value settings. For a dataset, only one global sampling sphere is constructed for the implicit feature encoding of all point cloud instances in the entire dataset. In another embodiment provided by the present invention, when performing random sampling position update, the optimization target for solving the optimization problem is
[0061] Construct a global sampling sphere according to the obtained updated M sampling positions
[0062] Wherein, x i , x j respectively represent the i-th sampling position and the j-th sampling position.
[0063] In another embodiment provided by the present invention, the process of constructing an initial local sampling sphere for each data point in the 3D point cloud in the above step S1 specifically includes the following steps:
[0064] Adjust each sampling position in the global sampling sphere according to the specific positions of the data points in the 3D point cloud to obtain the corresponding initial local sampling sphere;
[0065] The initial local sampling sphere of data point q x iis the i-th sampling position among M sampling positions, where M is an integer not less than 1.
[0066] In another embodiment provided by the present invention, the step S2 specifically includes the following steps:
[0067] Retrieve the data points located inside their initial local sampling spheres, and subtract the coordinate values of the retrieved data points from the coordinate values of these data points to obtain the neighborhood point sets of each data point;
[0068] Using the singular value decomposition method, decompose the neighborhood point set of each data point into an orthogonal matrix composed of left singular vectors, an orthogonal matrix composed of right singular vectors, and a diagonal matrix;
[0069] Based on predefined anchor points, update each right singular vector, and each updated right singular vector represents a principal direction of the initial local sampling sphere;
[0070] Align the three principal directions of the initial local sampling sphere of each data point with the three coordinate axes of the rectangular coordinate where the input point cloud is located to obtain the canonical local sampling sphere of each data point.
[0071] When specifically implementing this embodiment, for each data point q, retrieve the data points located inside the local sampling sphere and subtract their coordinate values from the coordinate value of q to obtain the neighborhood point set of the data point q
[0072] Using the singular value decomposition method, decompose into an orthogonal matrix U composed of left singular vectors and an orthogonal matrix V composed of right singular vectors, and a diagonal matrix ∑:
[0073] That is
[0074] wherein, each column vector of the orthogonal matrix V = {v k , 1 ≤ k ≤ 3} is a right singular vector, and the three right singular vectors are perpendicular to each other;
[0075] Based on the predefined anchor point p a , update each right singular vector, and each updated right singular vector is representing a principal direction of the initial local sampling sphere;
[0076] Align the three principal directions of the initial local sampling sphere with the three coordinate axes of the rectangular coordinate system where the input three-dimensional point cloud is located to obtain the local sampling sphere in the canonical pose
[0077] For the measured point clouds in most scenarios, the Z-axis of the data is usually vertically upward. In this case, this prior information can be utilized during implementation, and the singular value decomposition calculation is only performed using the planar coordinates of the points. The anchor point is defined as the nearest neighbor point farthest from the query point in a fixed direction, that is, the local highest point. In the case where the above prior information is not available, the singular value decomposition calculation is performed using the three-dimensional coordinates of the points, and the anchor point is defined as the nearest neighbor point farthest from the query point.
[0078] In another embodiment provided by the present invention, the step S3 specifically includes the following steps:
[0079] Establish a k-dimensional tree structure based on the three-dimensional point cloud;
[0080] For each canonical local sampling sphere, retrieve the data point in the three-dimensional point cloud that is closest to each sampling position, and divide the obtained closest distance by the radius of the canonical local sampling sphere as the eigenvalue corresponding to the sampling position.
[0081] According to the order of the sampling positions in each canonical local sampling sphere, concatenate the eigenvalues of each obtained sampling position to form a multi-dimensional vector as the implicit feature vector of the data point corresponding to the canonical local sampling sphere.
[0082] During the specific implementation of this embodiment, based on the input three-dimensional point cloud, a k-dimensional tree structure is established, and the previously established k-dimensional tree structure is used to accelerate the process of feature encoding.
[0083] For the local sampling sphere in the canonical pose of each data point, calculate the eigenvalue and determine the implicit feature vector in sequence.
[0084] For each sampling position, retrieve the data point in the input point cloud that is closest to it, and divide the closest distance corresponding to the retrieved data point by the radius r of the sampling sphere as the eigenvalue of the sampling position.
[0085] According to the order of the sampling positions in each canonical local sampling sphere, concatenate the eigenvalues of the M obtained sampling positions to form an M-dimensional vector as the implicit feature vector of the data point.
[0086] In another embodiment provided by the present invention, the method proposed in this application is used to perform implicit feature encoding on the independent object point clouds in the ModelNet dataset, and taking PointNet as an example, the effect of implicit features on improving the performance of point cloud neural networks for point cloud object classification tasks is demonstrated. ModelNet contains simulation CAD models of 40 categories, with a total of 12,311 independent objects. The data can be divided into pre-aligned data (objects of the same category have the same orientation) and randomly rotated data. See Figure 3, which is a schematic diagram of the effect of partial point cloud data provided by an embodiment of the present invention. Five categories of pre-aligned data and randomly rotated data are shown in the figure.
[0087] Set the parameter values of the implicit feature encoding as follows: the sampling sphere radius is 0.35 m, and the number of sampling positions is 32. See Figure 4 , which is a schematic diagram of the visualization effect of the implicit features of partial point cloud data of the "person" category provided by an embodiment of the present invention. See Figure 5 , which is a schematic diagram of the visualization effect of the implicit features of point cloud data of several categories provided by an embodiment of the present invention. Using the original point cloud as the input directly, the overall accuracies of the PointNet object classification tasks for pre-aligned data and randomly rotated data are 90.8% and 71.0% respectively; using the enhanced point cloud with concatenated encoded implicit features as the input, the overall accuracies of the PointNet object classification tasks for pre-aligned data and randomly rotated data are increased by 2.5% and 15.3% respectively, reaching 93.3% and 86.3%.
[0088] In another embodiment provided by the present invention, the method proposed in this application is used to perform implicit feature encoding on the scene point cloud in the SensatUrban dataset, and taking RandLA-Net as an example, the effect of implicit features on improving the performance of the point cloud neural network for the scene point cloud semantic segmentation task is shown. The scene point cloud in the SensatUrban dataset is divided into 13 categories including ground, vegetation, building, wall, bridge, parking lot, railway track, traffic road, street attachment, car, sidewalk, bicycle, and water area, with a total of about 2.8 billion points.
[0089] Set the parameter values of the implicit feature encoding as follows: the sampling sphere radius is 1 m, and the number of sampling positions is 32. See Figure 6 , which is a schematic diagram of the visualization effect of the implicit features of a certain scene point cloud data provided by an embodiment of the present invention. Compared with using the original point cloud as the input, using the enhanced point cloud with concatenated encoded implicit features as the input, RandLA-Net produces more accurate segmentation results in 9 out of 13 categories, and the mean intersection over union score for measuring the overall performance is increased from 54.2% to 56.3%.
[0090] Through experimental tests, the method provided by the solution of this application improves the performance of multiple representative point cloud deep neural networks in the point cloud object classification and scene point cloud semantic segmentation tasks. The average running speed of the implicit feature encoding is about 290,000 points per second, with good applicability and good running efficiency.
[0091] See Figure 7 , which is a schematic diagram of the structure of an implicit feature encoding device for three-dimensional point cloud provided by an embodiment of the present invention. The device includes:
[0092] An initial module for constructing an initial local sampling sphere for each data point in the input three-dimensional point cloud according to a pre-constructed global sampling sphere;
[0093] A specification module for aligning the principal direction of the initial local sampling sphere of each data point with the coordinate axes of the rectangular coordinate system where the three-dimensional point cloud is located to obtain a canonical local sampling sphere for each data point;
[0094] A vector module for sampling the field value of the implicit field of the three-dimensional point cloud according to the canonical local sampling sphere of each data point and encoding to form an implicit feature vector corresponding to the data point.
[0095] The implicit feature encoding device for three-dimensional point cloud provided in this embodiment can execute all steps and functions of the implicit feature encoding method for three-dimensional point cloud in any of the above embodiments, and the specific functions of this device will not be elaborated here.
[0096] See Figure 8 , which is a schematic structural diagram of a terminal device provided in an embodiment of the present invention. The terminal device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor, such as an implicit feature encoding program for three-dimensional point cloud. When the processor executes the computer program, the steps in the above embodiments of the implicit feature encoding method for three-dimensional point cloud are implemented, such as Figure 1 The steps S1 to S3 shown. Alternatively, when the processor executes the computer program, the functions of each module in the above device embodiments are implemented.
[0097] Exemplarily, the computer program can be divided into one or more modules. The one or more modules are stored in the memory and executed by the processor to complete the present invention. The one or more modules can be a series of computer program instruction segments capable of performing specific functions, and these instruction segments are used to describe the execution process of the computer program in the implicit feature encoding device for three-dimensional point cloud. For example, the computer program can be divided into several modules, and the specific functions of each module have been described in detail in the implicit feature encoding method for three-dimensional point cloud provided in any of the above embodiments, and the specific functions of this device will not be elaborated here.
[0098] The terminal device can be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The terminal device may include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the schematic diagram is only an example of the terminal device and does not constitute a limitation on the implicit feature encoding device for three-dimensional point cloud, and may include more or fewer components than shown, or combine certain components, or different components. For example, the terminal device may further include input / output devices, network access devices, a bus, etc.
[0099] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The processor is the control center of the implicit feature encoding device for three-dimensional point clouds, connecting various parts of the entire implicit feature encoding device for three-dimensional point clouds through various interfaces and circuits.
[0100] The memory can be used to store the computer programs and / or modules. By running or executing the computer programs and / or modules stored in the memory, and calling the data stored in the memory, the processor realizes various functions of the implicit feature encoding device for three-dimensional point clouds. The memory mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the mobile phone (such as audio data, phone book, etc.). In addition, the memory can include high-speed random access memory, and can also include non-volatile memory, such as a hard disk, memory, plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, at least one magnetic disk storage device, flash device, or other volatile solid-state storage devices.
[0101] Among them, if the modules integrated in the implicit feature encoding device for three-dimensional point clouds are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-described embodiment methods of the present invention, it can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-described various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.
[0102] An embodiment of the present invention further provides a computer program product, including a computer program / instructions. When the computer program / instructions are executed by a processor, the steps of any one of the above-described methods are implemented.
[0103] The computer program product provided in this embodiment can execute all the steps and functions of the implicit feature encoding method for three-dimensional point clouds provided in any one of the above embodiments. The specific functions of this product will not be elaborated here.
[0104] It should be noted that for those of ordinary skill in the art in this technical field, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.
Claims
1. An implicit feature encoding method for three-dimensional point clouds, characterized in that: The method comprises: According to the pre-constructed global sampling sphere, an initial local sampling sphere is constructed for each data point in the input 3D point cloud; Aligning the main direction of the initial local sampling sphere of each data point with the coordinate axis of the rectangular coordinate system where the three-dimensional point cloud is located to obtain a standard local sampling sphere of each data point; According to the canonical local sampling sphere of each data point, the implicit field of the three-dimensional point cloud is sampled, and the implicit feature vector of the corresponding data point is encoded.
2. The implicit feature encoding method of three-dimensional point cloud according to claim 1, characterized in that: The process of pre-building a global sampling sphere includes: Take the coordinate origin (0,0,0) as the center of the sphere and construct the spherical domain with the preset radius; The center of the sphere is taken as one sampling position, and M sampling positions are randomly determined within the determined spherical domain, and the randomly determined M sampling positions are updated with a preset optimization target; Construct a global sampling sphere with the updated M sampling positions; Wherein, M is an integer not less than 1.
3. The implicit feature encoding method of three-dimensional point cloud according to claim 2, characterized in that: The optimization goal is The global sampling sphere Among them, x i , x j represent the i-th sampling position and the j-th sampling position respectively.
4. The implicit feature encoding method of three-dimensional point cloud according to claim 1, characterized in that: According to the pre-constructed global sampling sphere, an initial local sampling sphere is constructed for each data point in the input 3D point cloud, including: Adjust each sampling position in the global sampling sphere according to the specific position of each data point in the three-dimensional point cloud to obtain a corresponding initial local sampling sphere; Among them, the initial local sampling sphere of data point q x i is the i-th sampling position among M sampling positions, where M is an integer not less than 1.
5. The implicit feature encoding method of three-dimensional point cloud according to claim 1, characterized in that: The main direction of the initial local sampling sphere of each data point is aligned with the coordinate axis of the rectangular coordinate system where the three-dimensional point cloud is located to obtain the standard local sampling sphere of each data point, including: Retrieve the data points of each data point that are inside its initial local sampling sphere, and subtract the coordinate value of the data point from the coordinate value of the retrieved data point to obtain a neighborhood point set of each data point; Using the singular value decomposition method, the neighborhood point set of each data point is decomposed into an orthogonal matrix composed of left singular vectors, an orthogonal matrix composed of right singular vectors, and a diagonal matrix; Based on the predefined anchor point, each right singular vector is updated, and each updated right singular vector represents a main direction of the initial local sampling sphere; The three main directions of the initial local sampling sphere of each data point are aligned with the three coordinate axes of the rectangular coordinates of the three-dimensional point cloud to obtain a standard local sampling sphere of each data point.
6. The implicit feature encoding method of three-dimensional point cloud according to claim 1, characterized in that: According to the canonical local sampling sphere of each data point, the implicit field of the three-dimensional point cloud is sampled, and the implicit feature vector of the corresponding data point is encoded, including: Establishing a k-dimensional tree structure according to the three-dimensional point cloud; For each standard local sampling sphere, retrieve the data point closest to each sampling position in the three-dimensional point cloud, and divide the obtained closest distance by the radius of the standard local sampling sphere to obtain the characteristic value of the corresponding sampling position; According to the order of sampling positions in each standard local sampling ball, the eigenvalues of each sampling position are concatenated to form a multidimensional vector as the implicit eigenvector of the corresponding data point of the standard local sampling ball.
7. An implicit feature encoding device for a three-dimensional point cloud, characterized in that: The device comprises: An initial module, used to construct an initial local sampling sphere for each data point in the input three-dimensional point cloud according to a pre-constructed global sampling sphere; A standardization module, used to align the main direction of the initial local sampling sphere of each data point with the coordinate axis of the rectangular coordinate system where the three-dimensional point cloud is located, so as to obtain a standard local sampling sphere of each data point; The vector module is used to sample the implicit field of the three-dimensional point cloud according to the standard local sampling sphere of each data point, and encode and form an implicit feature vector of the corresponding data point.
8. A terminal device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the implicit feature encoding method of the three-dimensional point cloud as described in any one of claims 1 to 6 is implemented.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the implicit feature encoding method for a three-dimensional point cloud as described in any one of claims 1 to 6.
10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.