WO2019114023A1

WO2019114023A1 - Fourier-graph-transform-based point cloud intraframe coding method and apparatus

Info

Publication number: WO2019114023A1
Application number: PCT/CN2017/117856
Authority: WO
Inventors: 马思伟; 徐逸群; 王苫社; 李俊儒; 胡玮
Original assignee: 北京大学
Priority date: 2017-12-13
Filing date: 2017-12-21
Publication date: 2019-06-20
Also published as: CN108171761B; CN108171761A

Abstract

A Fourier-graph-transform-based point cloud intraframe coding method and apparatus, which belong to the field of point cloud digital signal processing. The method comprises: performing voxelization on an original three-dimensional point cloud so as to obtain multiple point cloud voxels (101); clustering the obtained multiple point cloud voxels so as to obtain multiple point cloud voxel sets (102); performing main-direction-weight-based Fourier graph transform on the multiple point cloud voxel sets, respectively (103); and performing uniform quantization and arithmetic coding on each of the transformed point cloud voxel sets so as to generate a corresponding code stream (104). In the method, each of the point cloud voxel sets obtained by means of clustering is subjected to independent graph construction, so that the complexity of graph construction is lowered; each of the point cloud voxel sets is subjected to independent coding, so that point cloud distribution in each class is more uniform and compact; and a local similarity characteristic is sufficiently utilized, so that the correlation between points is expressed more sufficiently, and the influence of unrelated factors, such as noise, is reduced.

Description

Point cloud intraframe coding method and device based on Fourier transform

Technical field

The invention relates to the field of point cloud digital signal processing, in particular to a point cloud intraframe coding method and device based on Fourier transform.

Background technique

Comparing multi-path texture with depth data format, 3D point cloud is a more efficient data representation, which consists of a large number of three-dimensional unordered points, each of which includes position information (X, Y, Z) and several attribute information. (color, normal vector, etc.). With the development of computer hardware and algorithms, the acquisition of 3D point cloud data is more and more convenient, and the amount of data in the point cloud is also increasing. In order to facilitate the storage and transmission of point cloud data, point cloud compression technology has gradually become the focus of attention.

Existing research on point cloud compression technology, including: MPEG (Moving Pictures Experts Group/Motion Pictures Experts Group) in the following document 1 established working group 3DG, standardized preparation for point cloud compression, and MP3DG-PCC point cloud coding software was introduced; a region-adaptive hierarchical transformation method (RAHT) proposed in the following literature 2, based on the idea of wavelet, performs multi-layer decomposition coding on point cloud color attributes; In this paper, a point cloud coding method based on Fourier transform is proposed. For the 3D point cloud, the graph model is constructed, and each 3D point is regarded as a node in the graph, and the color information is abstracted as a signal above the node. And using the distance between the point, that is, the Euclidean distance as a feature, assigning a weight to the edge, thereby obtaining a Fourier transform coefficient, and then encoding the color information. However, in the existing research, it is usually based on the correlation between point cloud data, and the spatial uniform division is performed, so that the point cloud in each class after division is unevenly distributed and not compact; and when the composition is performed, the matrix dimension is too large. , will bring a huge amount of calculations, increase the complexity.

Document 1: "Draft call for proposals for point cloud compression," in ISO/IECJTC1/SC29/WG11 (MPEG) output document N16538, Oct.2016.

Document 2:: Ricardo Lde Queiroz and Philip A Chou, "Compression of 3d point clouds using a region-adaptive hierarchical transform," IEEE Transactions on Image Processing, vol. 25, no. 8, pp. 3947–3956, 2016.

Document 3: Cha Zhang, Dinei Florencio, and Charles Loop, "Pointcloud attribute compression with graph transform," in IEEE International Conference on Image Processing (ICIP), 2014, pp. 2066 - 2070.

Summary of the invention

In order to solve the deficiencies of the prior art, the present invention provides a point cloud intraframe coding method and apparatus based on Fourier transform.

In one aspect, the present invention provides a point cloud intraframe coding method based on Fourier transform, comprising:

Step S1: performing voxelization on the original three-dimensional point cloud to obtain a plurality of point cloud voxels;

Step S2: clustering the plurality of point cloud voxels to obtain a plurality of point cloud voxel sets;

Step S3: performing a Fourier graph transformation based on the main direction weights on the plurality of point cloud voxel sets respectively;

Step S4: performing uniform quantization and arithmetic coding on the transformed set of point cloud voxels to generate a corresponding code stream.

Optionally, the step S1 is specifically: performing voxelization on the original three-dimensional point cloud to obtain coordinates and attribute information of the plurality of point cloud voxels and each point cloud voxel;

Optionally, the step S2 includes:

Step S2-1: predicting the number of point cloud voxel sets according to the obtained number of point cloud voxels and the average number of points of the preset point cloud voxel set;

Step S2-2: According to the predicted number of point cloud voxel sets and the coordinates of each point, the plurality of point cloud voxels are clustered by a K-means algorithm to obtain a corresponding number of point cloud voxel sets.

Optionally, the step S3 specifically includes:

Step S3-1: arbitrarily selecting one point cloud body element set in the plurality of point cloud body element sets, and determining a first adjacent point cloud body element set of any point cloud body element in the selected point cloud body element set;

Step S3-2: respectively determining a second set of neighboring cloud cloud elements of each point cloud voxel in the first adjacent point cloud meta-set;

Step S3-3: According to the K-proximity algorithm, find a preset number of neighbors of the any point cloud voxel in the first adjacent point cloud voxel set to form a first neighborhood, and in each second phase A neighboring cloud body element set respectively finds a preset number of neighbors of each point cloud body element in the first adjacent point cloud cloud element set, and constitutes a corresponding second neighborhood;

Step S3-4: calculating a main direction vector of the first neighborhood and each of the second neighborhoods, and calculating a weight between any two main direction vectors to form a weight matrix;

Step S3-5: transforming the weight matrix to obtain a Fourier transform coefficient;

Step S3-6: transform attribute information of the selected point cloud voxel set according to the Fourier transform coefficient;

Step S3-7: The above operation is repeatedly performed until the processing of the plurality of point cloud voxel sets is completed.

Optionally, the step S3-4 includes:

Step S3-4-1: calculating the covariance between any two point cloud voxels in each neighborhood according to the coordinates of each point cloud voxel in each neighborhood, and forming each covariance matrix for the respective coordination The variance matrix performs eigenvalue decomposition to obtain each feature vector, and the feature vectors are used as corresponding main direction vectors of each neighborhood;

Step S3-4-2: Calculate a sine value of an angle between any two main direction vectors, calculate a weight between the corresponding two main direction vectors according to the sine value, and form a weight matrix.

Optionally, the step S3-5 includes:

Step S3-5-1: adding each element in each row of the weight matrix to obtain each calculation result;

Step S3-5-2: forming each of the calculation results as a diagonal element constituting degree matrix;

Step S3-5-3: calculating the weight matrix and the degree matrix to obtain a Laplacian matrix;

Step S3-5-4: Calculate the feature vector of the Laplacian matrix, and form the calculated feature vector into a matrix to obtain a Fourier transform coefficient.

In another aspect, the present invention provides a point cloud intraframe coding apparatus based on Fourier transform, comprising:

a voxelization module for performing voxelization on the original three-dimensional point cloud to obtain a plurality of point cloud voxels;

a clustering module, configured by the plurality of point cloud voxels obtained by the voxelization module to obtain a plurality of point cloud voxel sets;

a transform module, configured to respectively perform a Fourier graph transformation based on a primary direction weight on the plurality of point cloud voxel sets obtained by the clustering module;

And a generating module, configured to perform uniform quantization and arithmetic coding on each set of point cloud meta-elements transformed by the transform module, to generate a corresponding code stream.

Optionally, the voxelization module is specifically configured to: perform voxelization on the original three-dimensional point cloud, and obtain coordinate and attribute information of the plurality of point cloud voxels and each point cloud voxel;

Optionally, the clustering module specifically includes: a prediction submodule and a clustering submodule;

The prediction submodule is configured to predict the number of point cloud voxel sets according to the number of point cloud voxels obtained by the voxelization module and the average number of points of the preset point cloud voxel set;

The clustering sub-module is configured to: according to the number of point cloud voxel sets predicted by the prediction sub-module and the coordinates of each point obtained by the voxelization module, the plurality of point clouds by a K-means algorithm The voxels are clustered to obtain a corresponding number of point cloud voxel sets.

Optionally, the transformation module specifically includes: a selection submodule, a first determining submodule, a second determining submodule, a constituent submodule, a first computing submodule, a second computing submodule, a first transform submodule, and a second transform submodule;

The selecting sub-module is configured to arbitrarily select one set of point cloud body elements in the plurality of point cloud voxel sets obtained by the clustering module;

The first determining sub-module is configured to determine a first neighboring point cloud body element set of any point cloud body element in the point cloud body element set selected by the selecting sub-module;

The second determining submodule is configured to determine a second neighboring point of each point cloud voxel in the first neighboring point cloud meta-set determined by the first determining sub-module;

The constructing submodule is configured to find, according to the K proximity algorithm, a preset number of neighbors of the any point cloud body element in the first neighboring point cloud body element set determined by the first determining submodule, and form a first a neighborhood, and respectively determining, in each second neighboring point cloud meta-set determined by the second determining sub-module, a preset of each point cloud voxel in the first adjacent point cloud meta-collection The number of neighbors constitutes the corresponding second neighborhoods;

The first calculation submodule is configured to use a primary direction vector of the first neighborhood and each second neighborhood formed by the constituent submodules;

The second calculation submodule is configured to calculate a weight between any two main direction vectors obtained by the first calculation submodule, and constitute a weight matrix;

The first transform submodule is configured to transform a weight matrix formed by the second calculating submodule to obtain a Fourier transform coefficient;

The second transform submodule is configured to transform, according to a Fourier transform coefficient obtained by the first transform submodule, attribute information of a set of point cloud voxels selected by the selected submodule.

Optionally, the first calculating sub-module is specifically configured to: calculate covariance between any two point cloud voxels in each neighborhood according to coordinates of each point cloud voxel in each neighborhood, and form a co-coordination a variance matrix, performing eigenvalue decomposition on each covariance matrix to obtain each feature vector, and using each feature vector as a main direction vector of each corresponding neighborhood;

Optionally, the second calculating sub-module is specifically configured to: calculate a sine value of an angle between any two main direction vectors, calculate a weight between the corresponding two main direction vectors according to the sine value, and form Weight matrix.

Optionally, the first transform submodule specifically includes: a first calculating unit, a forming unit, a second calculating unit, and a third calculating unit;

The first calculating unit is configured to add each element in each row of the weight matrix obtained by the second calculating submodule to obtain each calculation result;

The structuring unit is configured to use each calculation result obtained by the first calculating unit as a diagonal element constituting degree matrix;

The second calculating unit is configured to calculate a weight matrix obtained by the second calculating submodule and a degree matrix obtained by the forming unit to obtain a Laplacian matrix;

The third calculating unit is configured to calculate a feature vector of the Laplacian matrix obtained by the second calculating unit, and form the calculated feature vector into a matrix to obtain a Fourier transform coefficient.

The advantages of the invention are:

In the present invention, on the one hand, by pre-processing the point cloud, using the location information-based clustering method, the overall point cloud is divided into a plurality of point cloud meta-collections (ie, sub-point clouds), and for each point The cloud element set is independent of the composition, which reduces the complexity of the composition; at the same time, each point cloud voxel set is independently coded. Compared with the spatial uniform division, the positional distribution of the point cloud is considered in the present invention, so that in each class The point cloud distribution is more uniform and compact. On the other hand, the weight assignment based on the neighborhood main direction vector is performed, and the local similarity feature is fully utilized in the present invention compared to the discrete weight assignment based on the Euclidean distance, which can more fully express the point-to-point relationship. Relevance. On the other hand, the Fourier transform based on the principal direction similarity is more robust, and the influence of unrelated factors such as noise can be reduced compared to the Fourier transform of the point-to-point feature.

DRAWINGS

Various other advantages and benefits will become apparent to those skilled in the art from a The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting. Throughout the drawings, the same reference numerals are used to refer to the same parts. In the drawing:

1 is a flowchart of a method for encoding a point cloud intraframe based on Fourier transform according to the present invention;

2 is a schematic view showing an angle between an adjacent main direction vector and an Euclidean distance according to the present invention;

FIG. 3 is a schematic diagram of an application example of a point cloud intraframe coding method based on Fourier transform according to the present invention; FIG.

4 is a performance comparison diagram of different encoding methods provided by the present invention;

FIG. 5 is a block diagram of a module of a point cloud intraframe coding apparatus based on Fourier transform according to the present invention.

Detailed ways

Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While the exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the invention may be embodied in various forms and not limited by the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be more fully understood, and the scope of the disclosure can be fully conveyed to those skilled in the art.

Embodiment 1

According to an embodiment of the present invention, a point cloud intra-frame coding method based on Fourier transform is provided. As shown in FIG. 1, the method includes:

Step 101: performing voxelization on the original three-dimensional point cloud to obtain a plurality of point cloud voxels;

Specifically, a three-dimensional grid of a preset size is constructed, and the original three-dimensional point cloud is placed in the constructed three-dimensional grid to obtain coordinates of each point, and the three-dimensional grid containing the points is used as a point cloud voxel to obtain multiple points. The coordinates and attribute information of the cloud element and each point cloud element. The attribute information of the point cloud meta-collection, such as intensity, color, and the like; for the sake of universality, the color is taken as an example in the present invention.

Further, in this embodiment, the coordinates of the point cloud body element are specifically the coordinates of the center point of each point in the point cloud body element; the color information of the point cloud body element, specifically the color of each point in the point cloud body element The average of the information.

Further, in some embodiments, the original three-dimensional point cloud may be voxelized by using an octree method or the like to obtain a plurality of point cloud voxels, which are not detailed in the present invention.

Step 102: Clustering the obtained plurality of point cloud voxels to obtain a plurality of point cloud voxel sets;

According to an embodiment of the present invention, step 102 specifically includes:

Step 102-1: predict the number of point cloud voxel sets according to the obtained number of point cloud voxels and the average number of points of the preset point cloud voxel set;

Specifically, according to the obtained number of point cloud voxels and the average number of points of the preset point cloud voxel set, the number of point cloud voxel sets is predicted by the following formula 1;

Formula 1: K=N/n, where K is the number of predicted point cloud voxel sets, N is the number of point cloud voxels, and n is the average number of points cloud voxel sets (ie, point cloud voxel sets) The number of midpoint cloud voxels).

It should be pointed out that n is only the number of point cloud voxels in the set of point cloud voxels, which is used to predict the number of point cloud voxel sets, and the number of point cloud voxels in the cluster of point cloud voxels obtained by clustering. Not necessarily n.

Step 102-2: According to the predicted number of point cloud voxel sets and the coordinates of each point cloud voxel, cluster the obtained plurality of point cloud voxels by K-means algorithm to obtain a corresponding number of point cloud voxels. set.

Step 103: Perform a Fourier graph transformation based on the main direction weights on the obtained plurality of point cloud voxel sets respectively.

According to an embodiment of the present invention, step 103 includes:

Step 103-1: arbitrarily select one point cloud body element set in the obtained plurality of point cloud body element sets, and determine a first adjacent point cloud body element set of any point cloud body element in the selected point cloud body element set;

Specifically, the point cloud element i in the point cloud body element set is taken as the center, and the adjacent area of the cloud element i is fixed by the preset length as the radius, and each point cloud element in the adjacent area is defined. That is, the first adjacent point cloud meta-collection of the point cloud element i.

Among them, the preset length can be set according to the needs.

Step 103-2: respectively determining a second adjacent point cloud element set of each point cloud body element in the first adjacent point cloud body element set;

Specifically, each point cloud element j in the first adjacent point cloud element set is centered, and each point cloud element j in the first adjacent point cloud element set is delimited by a preset length In the adjacent area, each point cloud element f located in the circled adjacent area is the second adjacent point cloud element set of the corresponding point cloud element j.

Step 103-3: According to the K proximity algorithm, find a preset number of neighbors of the any point cloud body element in the first adjacent point cloud body element set to form a first neighborhood, and at each second adjacent point. A preset number of neighbors of the point cloud voxels in the corresponding first neighboring point cloud meta-collections are respectively found in the cloud meta-collection, and the corresponding second neighbor domains are formed;

Among them, the preset number can be set according to the needs.

Step 103-4: Calculate a primary direction vector of the first neighborhood and each second neighborhood, and calculate a weight between any two main direction vectors to form a weight matrix;

According to an embodiment of the present invention, step 103-4 specifically includes:

Step 103-4-1: Calculate the covariance between any two point cloud voxels in each neighborhood according to the coordinates of each point cloud voxel in each neighborhood, and form each covariance matrix for each covariance matrix. Performing eigenvalue decomposition to obtain each feature vector, and taking each feature vector as a main direction vector of each corresponding neighborhood;

Step 103-4-2: Calculate the sine value of the angle between any two main direction vectors, calculate the weight between the corresponding two main direction vectors according to the sine value, and form a weight matrix.

Wherein, the weight between the corresponding two main direction vectors is calculated according to the sine value, specifically: according to the sine value, the weight between the corresponding two main direction vectors is calculated by the following formula 2;

Formula 2:

Where W _ij is the weight between the main direction vectors of the neighborhood where the neighboring point cloud elements i and j are located, and θ is the angle between the main direction vectors of the neighboring point cloud element i and the neighborhood where j is located, σ is In order to find the optimal value of W _ij , set the adjustment variable by yourself.

Further, in this embodiment, a schematic diagram of the angle between the main direction vectors of the neighborhoods of two adjacent point cloud elements i and j is given, as shown in FIG. 2; It is indicated and not used for limitation.

Step 103-5: Transform the obtained weight matrix to obtain a Fourier transform coefficient;

According to an embodiment of the present invention, the step 103-5 specifically includes:

Step 103-5-1: adding each element in each row of the weight matrix to obtain each calculation result;

Step 103-5-2: The calculation result is used as a diagonal element to form a degree matrix;

Specifically, each calculation result is taken as a diagonal element, and other elements are filled with 0 to form a degree matrix.

Step 103-5-3: Calculating the weight matrix and the degree matrix to obtain a Laplacian matrix;

Specifically, the weight matrix and the degree matrix are calculated according to the following formula 3 to obtain a Laplacian matrix;

Equation 3: L = D-W, where L is a Laplacian matrix, D is a degree matrix, and W is a weight matrix.

Step 103-5-4: Calculate the eigenvectors of the Laplacian matrix, and form the matrix of the calculated eigenvectors to obtain Fourier transform coefficients.

Step 103-6: Transform attribute information of the selected point cloud voxel set according to the obtained Fourier transform coefficient;

Specifically, the attribute information of the selected point cloud voxel set is transformed according to the obtained Fourier transform coefficient by the following formula 4;

Formula 4:

Where T is the result of the transformation,

As the device matrix of the Fourier transform coefficients, Q is the attribute vector of the selected set of point cloud voxels.

In the present invention, color is taken as an example for description. Specifically, the color of the selected set of point cloud voxels is organized into three m*1 column vectors (Y component, U component, and V component, respectively), and the Y component is For example, according to the formula 4, the Y component is transformed, then

Step 103-7: Repeat the above operation until the obtained plurality of point cloud voxel sets are processed.

Step 104: Perform uniform quantization and arithmetic coding on the transformed set of point cloud voxels to generate a corresponding code stream.

The process of uniform quantization and arithmetic coding is a technical means well known to those skilled in the art, and will not be described in detail in the present invention.

For a better understanding of the technical solution of the present invention, a specific application example is given in this embodiment. As shown in FIG. 3, for a human body point cloud in a certain frame, clustering based on location information is performed. Each set obtained by clustering performs independent composition and calculates the main direction vector weight by the neighborhood formed by the texture, and finally performs uniform quantization and arithmetic coding to generate a corresponding code stream.

In the present invention, according to the position information of the point cloud, the point cloud data is clustered, and the whole point cloud is divided into a plurality of point cloud body element sets (ie, a sub-point cloud); and then, for each point cloud body element set Using the distance as the standard, the neighboring points are screened out, the point cloud distribution information is fully utilized, the point cloud is clustered, and each point cloud body element set is independently composed, which reduces the composition complexity. At the same time, the main direction similarity in the neighborhood of adjacent points is used to assign values to the edges between the two points, and the features of the points and their neighborhoods are fully utilized, and the weight matrix is modified to improve the overall coding effect.

Further, in order to embody the advantages of the technical solution of the present invention, as shown in FIG. 4, the method using the present invention (corresponding to OURS in FIG. 4) and the existing methods RAHT, DCT, MP3DG-PCC are respectively given names. Performance comparison map for the point cloud frames of Andrew, Boy, David, Dimitris, Phil, Ricardo, and Sarah, where the horizontal axis Color Byte per Voxel (B/V) is the code rate, vertical The axis PSNR-Y (dB) is the peak signal to noise ratio. Overall, the performance of the encoding method in the present invention is much better than other methods.

Embodiment 2

According to an embodiment of the present invention, a point cloud intra-frame coding apparatus based on Fourier transform is provided. As shown in FIG. 4, the method includes:

The voxelization module 201 is configured to perform voxelization on the original three-dimensional point cloud to obtain a plurality of point cloud voxels;

The clustering module 202 is configured to cluster a plurality of point cloud voxels obtained by the voxelization module 201 to obtain a plurality of point cloud voxel sets;

The transforming module 203 is configured to perform a Fourier graph transform based on the main direction weights on the plurality of point cloud voxel sets obtained by the clustering module 202, respectively;

The generating module 204 is configured to perform uniform quantization and arithmetic coding on each set of point cloud meta-contracts transformed by the transform module 203 to generate a corresponding code stream.

According to the embodiment of the present invention, the voxelization module 201 is specifically configured to: perform voxelization on the original three-dimensional point cloud, and obtain coordinate and attribute information of the plurality of point cloud voxels and each point cloud voxel;

According to an embodiment of the present invention, the clustering module 202 specifically includes: a prediction submodule and a clustering submodule, wherein:

a prediction submodule, configured to predict, according to the number of point cloud voxels obtained by the voxelization module 201 and the average number of points of the preset point cloud voxel set, the number of point cloud voxel sets;

The clustering sub-module is configured to obtain, according to the number of the point cloud voxel set predicted by the prediction sub-module and the coordinates of each point cloud voxel obtained by the voxelization module 201, the K-means algorithm obtains the voxelization module 201 The point cloud voxels are clustered to obtain a corresponding number of point cloud voxel sets.

More specifically, the prediction submodule is configured to predict the number of point cloud voxel sets according to the number of point cloud voxels obtained by the voxelization module 201 and the average number of points of the preset point cloud voxel set by the following formula 1. ;

According to an embodiment of the present invention, the transformation module 203 specifically includes: a selection submodule, a first determining submodule, a second determining submodule, a constituent submodule, a first computing submodule, a second computing submodule, and a first transform Module and second transform submodule, wherein:

The sub-module is selected to arbitrarily select one set of point cloud voxels in the plurality of point cloud voxel sets obtained by the clustering module 202;

a first determining sub-module, configured to determine a first neighboring point cloud body element set of any point cloud body element in the set of point cloud body elements selected by the sub-module;

In this embodiment, the first determining sub-module is specifically configured to: select any point cloud element i in the set of point cloud body elements selected by the sub-module as a center, and set a point cloud element i with a preset length as a radius In the adjacent area, each point cloud element j located in the circled adjacent area is the first adjacent point cloud element set of the point cloud element i.

a second determining submodule, configured to determine a second adjacent point of each point cloud voxel in the first adjacent point cloud meta-set determined by the first determining sub-module;

In this embodiment, the second determining sub-module is specifically configured to: respectively, each point cloud element j in the first adjacent point cloud element set obtained by the first determining sub-module is a center of the circle, and the preset length is The radius encircles the adjacent region of each point cloud element j in the first adjacent point cloud body element set, and each point cloud body element f located in the circled adjacent area is the corresponding point cloud element element j Two adjacent point cloud meta-collections.

Constructing a sub-module, configured to find a preset number of neighbors of the any point cloud meta-weight in the first adjacent point cloud meta-set determined by the first determining sub-module according to the K-proximity algorithm, to form a first neighborhood, And respectively finding a preset number of neighbors of each point cloud body element in the corresponding first neighboring point cloud body element set in each second neighboring point cloud body element set determined by the second determining sub-module, and forming corresponding corresponding Second neighborhood;

Among them, the preset number can be set according to the needs.

a first calculation submodule, configured to calculate a primary direction vector of the first neighborhood and each second neighborhood formed by the submodule;

In this embodiment, the first calculation sub-module is specifically configured to: calculate covariance between any two point cloud voxels in each neighborhood according to the coordinates of each point cloud voxel in each neighborhood, and form a co-coordination a variance matrix, performing eigenvalue decomposition on each covariance matrix to obtain each feature vector, and using each feature vector as a main direction vector of each corresponding neighborhood;

a second calculation sub-module, configured to calculate a weight between any two main direction vectors obtained by the first calculation sub-module, to form a weight matrix;

In this embodiment, the second calculation sub-module is specifically configured to: calculate a sine value of an angle between any two main direction vectors obtained by the first calculation sub-module, and calculate a corresponding two main direction vectors according to the sine value. The weights and constitute the weight matrix.

More specifically, the second calculation submodule is configured to calculate the weight between the corresponding two main direction vectors according to the sine value by using the following formula 2;

Formula 2:

a first transform submodule, configured to transform a weight matrix formed by the second computation submodule to obtain a Fourier transform coefficient;

And a second transform submodule, configured to transform attribute information of the set of point cloud voxels selected by the submodule according to the Fourier transform coefficients obtained by the first transform submodule.

Further, according to an embodiment of the present invention, the first transform submodule specifically includes: a first calculating unit, a constituting unit, a second calculating unit, and a third calculating unit, where:

a first calculating unit, configured to add each element in each row of the weight matrix obtained by the second calculating sub-module to obtain each calculation result;

a constituting unit, configured to use each calculation result obtained by the first calculation unit as a diagonal element constituting degree matrix;

a second calculating unit, configured to calculate a weight matrix obtained by the second calculating sub-module and a degree matrix obtained by the forming unit to obtain a Laplacian matrix;

In this embodiment, the second calculating unit is specifically configured to: calculate a weight matrix obtained by the second calculating sub-module and a degree matrix obtained by the constituent unit, and calculate a Laplacian matrix according to the following formula 3;

And a third calculating unit, configured to calculate a feature vector of the Laplacian matrix obtained by the second calculating unit, and form the calculated feature vector into a matrix to obtain a Fourier transform coefficient.

According to an embodiment of the present invention, the second transform sub-module is specifically configured to: according to the Fourier transform coefficient obtained by the first transform sub-module, perform attribute information of the set of point cloud meta-sets selected by the selected sub-module by using the following formula 4 Transform

Formula 4:

Where T is the result of the transformation,

In the present invention, on the one hand, by pre-processing the point cloud, using the location information-based clustering method, the overall point cloud is divided into a plurality of point cloud meta-collections (ie, sub-point clouds), and for each point The cloud element set is independently coded. Compared with the spatial uniform division, the position distribution of the point cloud is considered in the present invention, so that the point cloud distribution in each class is more uniform and compact. On the other hand, the weight assignment based on the neighborhood main direction vector is performed, and the local similarity feature is fully utilized in the present invention compared to the discrete weight assignment based on the Euclidean distance, which can more fully express the point-to-point relationship. Relevance. On the other hand, the Fourier transform based on the principal direction similarity is more robust, and the influence of unrelated factors such as noise can be reduced compared to the Fourier transform of the point-to-point feature.

The above is only a preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily think of changes or within the technical scope disclosed by the present invention. Alternatives are intended to be covered by the scope of the present invention. Therefore, the scope of the invention should be determined by the scope of the appended claims.

Claims

A point cloud intraframe coding method based on Fourier transform, which is characterized in that it comprises:

Step S1: performing voxelization on the original three-dimensional point cloud to obtain a plurality of point cloud voxels;

Step S2: clustering the plurality of point cloud voxels to obtain a plurality of point cloud voxel sets;

Step S3: performing a Fourier graph transformation based on the main direction weights on the plurality of point cloud voxel sets respectively;

Step S4: performing uniform quantization and arithmetic coding on the transformed set of point cloud voxels to generate a corresponding code stream.
The method of claim 1 wherein

The step S1 is specifically: performing voxelization on the original three-dimensional point cloud, and obtaining coordinate and attribute information of the plurality of point cloud voxels and each point cloud voxel;

The step S2 specifically includes:

Step S2-1: predicting the number of point cloud voxel sets according to the obtained number of point cloud voxels and the average number of points of the preset point cloud voxel set;

Step S2-2: According to the predicted number of point cloud voxel sets and the coordinates of each point, the plurality of point cloud voxels are clustered by a K-means algorithm to obtain a corresponding number of point cloud voxel sets.
The method of claim 2, wherein the step S3 comprises:

Step S3-1: arbitrarily selecting one point cloud body element set in the plurality of point cloud body element sets, and determining a first adjacent point cloud body element set of any point cloud body element in the selected point cloud body element set;

Step S3-2: respectively determining a second set of neighboring cloud cloud elements of each point cloud voxel in the first adjacent point cloud meta-set;

Step S3-3: According to the K-proximity algorithm, find a preset number of neighbors of the any point cloud voxel in the first adjacent point cloud voxel set to form a first neighborhood, and in each second phase A neighboring cloud body element set respectively finds a preset number of neighbors of each point cloud body element in the first adjacent point cloud cloud element set, and constitutes a corresponding second neighborhood;

Step S3-4: calculating a main direction vector of the first neighborhood and each of the second neighborhoods, and calculating a weight between any two main direction vectors to form a weight matrix;

Step S3-5: transforming the weight matrix to obtain a Fourier transform coefficient;

Step S3-6: transform attribute information of the selected point cloud voxel set according to the Fourier transform coefficient;

Step S3-7: The above operation is repeatedly performed until the processing of the plurality of point cloud voxel sets is completed.
The method of claim 3 wherein:

The step S3-4 specifically includes:

Step S3-4-1: calculating the covariance between any two point cloud voxels in each neighborhood according to the coordinates of each point cloud voxel in each neighborhood, and forming each covariance matrix for the respective coordination The variance matrix performs eigenvalue decomposition to obtain each feature vector, and the feature vectors are used as corresponding main direction vectors of each neighborhood;

Step S3-4-2: Calculate a sine value of an angle between any two main direction vectors, calculate a weight between the corresponding two main direction vectors according to the sine value, and form a weight matrix.
The method according to claim 3, wherein the step S3-5 comprises:

Step S3-5-1: adding each element in each row of the weight matrix to obtain each calculation result;

Step S3-5-2: forming each of the calculation results as a diagonal element constituting degree matrix;

Step S3-5-3: calculating the weight matrix and the degree matrix to obtain a Laplacian matrix;

Step S3-5-4: Calculate the feature vector of the Laplacian matrix, and form the calculated feature vector into a matrix to obtain a Fourier transform coefficient.
A point cloud intra-frame coding device based on Fourier transform, characterized in that it comprises:

a voxelization module for performing voxelization on the original three-dimensional point cloud to obtain a plurality of point cloud voxels;

a clustering module, configured by the plurality of point cloud voxels obtained by the voxelization module to obtain a plurality of point cloud voxel sets;

a transform module, configured to respectively perform a Fourier graph transformation based on a primary direction weight on the plurality of point cloud voxel sets obtained by the clustering module;

And a generating module, configured to perform uniform quantization and arithmetic coding on each set of point cloud meta-elements transformed by the transform module, to generate a corresponding code stream.
The device of claim 6 wherein:

The voxelization module is specifically configured to: perform voxelization on the original three-dimensional point cloud, and obtain coordinate and attribute information of the plurality of point cloud voxels and each point cloud voxel;

The clustering module specifically includes: a prediction submodule and a clustering submodule;

The prediction submodule is configured to predict the number of point cloud voxel sets according to the number of point cloud voxels obtained by the voxelization module and the average number of points of the preset point cloud voxel set;

The clustering sub-module is configured to: according to the number of point cloud voxel sets predicted by the prediction sub-module and the coordinates of each point obtained by the voxelization module, the plurality of point clouds by a K-means algorithm The voxels are clustered to obtain a corresponding number of point cloud voxel sets.
The apparatus according to claim 6, wherein the transformation module comprises: a selection submodule, a first determining submodule, a second determining submodule, a constituent submodule, a first computing submodule, and a second computing a submodule, a first transform submodule, and a second transform submodule;

The selecting sub-module is configured to arbitrarily select one set of point cloud body elements in the plurality of point cloud voxel sets obtained by the clustering module;

The first determining sub-module is configured to determine a first neighboring point cloud body element set of any one of the cloud cloud element elements selected by the selecting sub-module;

The second determining submodule is configured to determine a second neighboring point of each point cloud voxel in the first neighboring point cloud meta-set determined by the first determining sub-module;

The constructing submodule is configured to find, according to the K proximity algorithm, a preset number of neighbors of the any point cloud body element in the first neighboring point cloud body element set determined by the first determining submodule, and form a first a neighborhood, and respectively determining, in each second neighboring point cloud meta-set determined by the second determining sub-module, a preset of each point cloud voxel in the first adjacent point cloud meta-collection The number of neighbors constitutes the corresponding second neighborhoods;

The first calculation submodule is configured to use a primary direction vector of the first neighborhood and each second neighborhood formed by the constituent submodules;

The second calculation submodule is configured to calculate a weight between any two main direction vectors obtained by the first calculation submodule, and constitute a weight matrix;

The first transform submodule is configured to transform a weight matrix formed by the second calculating submodule to obtain a Fourier transform coefficient;

The second transform submodule is configured to transform, according to a Fourier transform coefficient obtained by the first transform submodule, attribute information of a set of point cloud voxels selected by the selected submodule.
The device of claim 8 wherein:

The first calculating sub-module is specifically configured to: calculate a covariance between any two point cloud voxels in each neighborhood according to coordinates of each point cloud voxel in each neighborhood, and form each covariance matrix, Performing eigenvalue decomposition on each covariance matrix to obtain each feature vector, and using each feature vector as a main direction vector of each corresponding neighborhood;

The second calculation sub-module is specifically configured to: calculate a sine value of an angle between any two main direction vectors obtained by the first calculation sub-module, and calculate a weight between the corresponding two main direction vectors according to the sine value And constitute a weight matrix.
The apparatus according to claim 8, wherein the first transform sub-module comprises: a first calculating unit, a constituent unit, a second calculating unit, and a third calculating unit;

The first calculating unit is configured to add each element in each row of the weight matrix obtained by the second calculating submodule to obtain each calculation result;

The structuring unit is configured to use each calculation result obtained by the first calculating unit as a diagonal element constituting degree matrix;

The second calculating unit is configured to calculate a weight matrix obtained by the second calculating submodule and a degree matrix obtained by the forming unit to obtain a Laplacian matrix;

The third calculating unit is configured to calculate a feature vector of the Laplacian matrix obtained by the second calculating unit, and form the calculated feature vector into a matrix to obtain a Fourier transform coefficient.