# US 2020/0394450 A1 · An enhanced graph transformation-based point cloud attribute compression method

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2020/0394450 A1
- Document ID: US20200394450A1
- Status: Published application
- Kind: A1
- Application number: 16/957,187
- Filed: 2018-02-12
- Published: 2020-12-17
- Earliest claimed: 2018-02-11
- Assignees: Peking University Shenzhen Graduate School
- Applicants: Peking University Shenzhen Graduate School
- Inventors: Ge LI; Yiting SHAO
- Web page: https://patentgrep.com/patent/US20200394450A1

## CPC

- H04N 19/186: the unit being a colour or a chrominance component
- G06F 17/16: Matrix or vector computation {, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization (matrix transposition G06F7/78)}
- G06F 18/213: Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F 18/2323: based on graph theory, e.g. minimum spanning trees [MST] or graph cuts
- G06T 7/30: Determination of transform parameters for the alignment of images, i.e. image registration
- G06T 7/60: Analysis of geometric attributes
- G06T 7/90: Determination of colour characteristics
- H04N 19/96: Tree coding, e.g. quad-tree coding
- G06T 2207/10028: Range image; Depth image; 3D point clouds
- H03M 7/405: {Tree adaptation}
- H03M 7/70: {Type of the data to be coded, other than image and sound}
- H04N: PICTORIAL COMMUNICATION, e.g. TELEVISION
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- H03M: CODING; DECODING; CODE CONVERSION IN GENERAL (using fluidic means F15C4/00; optical analogue/digital converters G02F7/00; coding, decoding or code conversion, specially adapted for particular applications, see the relevant subclasses, e.g. G01D, G01R, G06F, G06T, G09G, G10L, G11B, G11C, H04B, H04L, H04M, H04N; ciphering or deciphering for cryptography or other purposes involving the need for secrecy G09C)

## Abstract

An enhanced graph transformation-based point cloud attribute compression method. For point cloud attribute information, a point cloud is first subjected to airspace division by using a K-dimension (KD) tree; a new graph transformation processing method in combination with spectral analysis is provided; the point cloud is then subjected to spectral clustering on graphs in coded blocks of the point cloud; expansion is performed on the basis of existing graph transformation to implement a local graph transformation scheme; enhanced graph transformation with two transformation modes is formed; the compression performance of graph transformation is improved. The method comprises: performing color space transformation of point cloud attributes; dividing the point cloud by using the KD tree to obtain the coded blocks; performing spectral clustering-based enhanced graph transformation; performing transformation mode decision; and performing uniform quantization and entropy coding. Provided is a new spectral analysis-based enhanced graph transformation scheme, wherein two transformation modes are comprised, and the optimal mode is selected by the mode decision; after the point cloud is divided with the tree, a graph is created in each coded block and the graph transformation is used as transformation mode I; on this basis, graph spectral clustering is implemented; the graph is divided into two local graphs and then local graph transformation is performed to serve as transformation mode II; in the enhanced graph transformation scheme supporting the two transformation modes, the optimal mode is selected by the mode decision to achieve the optimal performance of point cloud attribute compression.

## Claims

### Claim 1

An enhanced graph transformation-based point cloud attribute compression method, comprising, for point cloud attribute information, first subjecting a point cloud to airspace division by using a K-dimension (KD) tree; providing a new graph transformation processing method in combination with spectral analysis; then subjecting the point cloud to spectral clustering on graphs in coded blocks of the point cloud; performing expansion on the basis of existing graph transformation to implement a local graph transformation scheme so as to form enhanced graph transformation with two transformation modes to improve the compression performance of graph transformation; the method comprises the steps of:
1) color space transformation of point cloud attributes
reading in point cloud attribute information to be processed, and transforming a point cloud color space from an RGB space to a YUV space according to the visual characteristics of human eyes and the difficulty degree of compression processing;
2) dividing the point cloud by using a KD tree to obtain coded blocks, and numbering the coded blocks according to a breadth traversal sequence
reading in geometric information of the point cloud, performing KD tree division on the point cloud according to the geometric information, selecting a coordinate axis with a largest distribution variance in the position coordinate of the point cloud as a division axis each time, selecting a point with the coordinate being a median value as a division point, and performing iterative division until reaching a set KD tree depth; a block obtained from a last layer divided by the KD tree being a coded block of the point cloud, numbering the coded blocks according to a breadth traversal sequence, and the number being used as a sequence for later processing of the coded blocks;
3) constructing a graph within the coded block by using enhanced graph transformation, with two transformation modes
in the coded block, connecting every two points ni and nj by an edge εij to construct an entire points graph G, wherein each point on the graph has color information, and the weight of the edge is determined by the geometric relative position of the two points; the weight ωij of the edge εij reflects geometric correlation between the two points ni and nj, and the weights ω of all the edges form an adjacent matrix W of the graph to further obtain a feature vector matrix A;
transformation mode I: taking the feature vector A of the entire points graph as a transformation matrix, and transforming the color information of the coded block;
transformation mode II: the feature vector matrix A of a graph G reflects the spectrum distribution of the graph, so that spectral clustering is performed on the basis of the matrix A to realize local graph segmentation; clustering all points in the block into two types by using a plus-minus sign of the second dimension column vector of the matrix A; if the number of the two types of points reaches more than 40% of the total number of the points in the block, dividing the graph G into two local graphs G1, G2 according to the clustering condition, so as to respectively obtain corresponding feature vector matrix A1 and A2; transforming the color information in the two local graphs G1 and G2 by using a corresponding transformation matrix;
4) transformation mode decision:
providing two modes for transforming the color information of the coded block, and performing a mode decision by estimating the transformation performance to select an optimal transformation mode; calculating the proportion of the sum of absolute values of the previous k maximum coefficients in the transformed coefficients in the sum of the absolute values of the transformed coefficients as the score of the transformation mode; the higher the score is, the higher the proportion of the selected transformation coefficient in the total transformation coefficient is, representing higher transformation efficiency and better performance in the mode, wherein the mode with the highest score is selected as the transformation mode of the current block;
5) generation of a code stream with point cloud attribute compression: processing all the coded blocks according to the coding sequence, quantizing the transformed coefficients, and performing entropy coding by combining the transformation mode information to obtain a final code stream with point cloud attribute compression.

### Claim 2 (depends on 1)

The point cloud attribute compression method of claim 1, wherein the specific process of the color space transformation in the step 1) is as follows: a point pi in the point cloud has color values gi and bi in an RGB color space; the RGB is transformed into a YUV color space by a formula 1, and the color values thereof are yi, ui and vi.
[
y
i
,
u
i
,
v
i
]
=
[
r
i
,
g
i
,
b
i
]
×
{
0
.
2

1

2

6
0
.
7

1

5

2
0
.
0

7

2

2
-
0
.
1

1

4

6
-
0
.
3

8

5

4
0
.
5
0
.
5
-
0
.
4

5

4

2
-
0
.
0

4

5

8
}
(
Formula


1
)

### Claim 3 (depends on 1)

The point cloud attribute compression method of claim 1, wherein the KD tree division method in the step 2) is a binary division method; let the point cloud to be processed have N points, and the division depth set by the KD tree be d, 2d coded blocks are obtained after the point cloud is divided ford times; and all coded blocks are numbered with b1, b2, . . . , bi, . . . , b2d in the breadth traversal sequence.

### Claim 4 (depends on 1)

The point cloud attribute compression method of claim 1, wherein the specific process of the enhanced graph transformation in the step 3) comprises:
(4-1) constructing a graph G in each transformation block, wherein each two points ni and nj are connected by an edge εij and the weight ωij of the edge εij is related to the Euclidean distance between the two points and generally calculated by a formula 2:
ω
i
,
j
=
{
e
-

n
i
-
n
j

2
2
σ
2
,
if

||
n
i
-
n
j

2
2
≤
τ
;
0
,
else
,
(
Formula


2
)
wherein the parameter σ reflects the distribution variance of the current point cloud; the parameter ti is a distance threshold for judging whether a connection line is formed between the two points, which determines the sparsity of a graph Laplacian matrix, and an experience set value is generally used;
(4-2) the adjacent matrix W of the graph G is a set of edge weights ωij and reflects the correlation between each point in the block; the density matrix D of the graph G is a diagonal matrix, wherein Di is the sum of non-zero elements in the i row of the adjacent matrix and represented by a formula 3, reflecting the density of correlation between the ith point and other points; the transformation operator of graph G adopts a Laplacian matrix L, and is represented by a formula 4:
Di=Σjωi,j (Formula 3)
L=D−W (Formula 4)
(4-3) performing characteristic decomposition on the Laplacian matrix L, which is represented by a formula 5 to obtain a feature vector matrix A, which serves as a global graph transformation matrix in the transformation mode 1 for compressing the point cloud attribute information:
L=AΛA′ (Formula 5)
wherein A is a feature vector matrix;
(4-4) performing spectral clustering on the second dimension column vector V2 of the feature vector matrix A, and clustering the points into two types C1 and C2 according to the plus-minus sign of the vector parameter value, which is represented by a formula 6; dividing the global graph G into two local graphs G1 and G2 according to the clustering condition, wherein the two local graphs are composed of corresponding points n and edges e and are represented by a formula 7:
{
C
1
=
(
p
i
|
V
2

(
p
i
)
>
0
)
,

if


count

(
C
1
)
count

(
block
)
>
40

%
;
C
2
=
(
p
i
|
V
2

(
p
i
)
≤
0
)
,

if


count

(
C
2
)
count

(
block
)
>
40

%
,
(
Formula


6
)
{
G
1
=
(
n

(
C
1
)
,
ɛ

(
C
1
)
)
,
G
2
=
(
n

(
C
2
)
,
ɛ

(
C
2
)
)
,
(
Formula


7
)
in the formula 6, count(C1) is used for calculating the number of points in C1, count(C2) for calculating the number of points in C2, and count(block) for calculating the total number of points in the current coded block;
(4-5) respectively performing characteristic decomposition on the two local graphs G1 and G2 to obtain transformation matrices A1 and A2, which are used as local graph transformation matrices in the transformation mode II and used for transformation of color information corresponding to a point in the graph;
(4-6) the graph transformation with two transformation modes forms an enhanced graph transformation scheme.

### Claim 5 (depends on 1)

The point cloud attribute compression method of claim 1, wherein in the transformation mode decision in the step 4), the key is to calculate the score J of the transformation mode; firstly, the absolute values of the transformed coefficients Trans are arranged in a descending order, and then the sum of the absolute values of the largest coefficients in the previous k-dimension selected is calculated for the proportion in the sum of the absolute values of the transformed coefficients, which is represented by a formula 8:
J
=
∑
i
=
1
k


Trans
(
i
)

∑
i
=
1
all


Trans
(
i
)

.
(
Formula


8
)

### Claim 6 (depends on 1)

The point cloud attribute compression method of claim 1, wherein the specific details in the step 5) are as follows:
(6-1) the code stream of point cloud attribute information is mainly composed of compressed header information and coded block information; wherein the header information mainly comprises quantization step length and the like; the coded block information stream is arranged in the order of coded blocks in units of coded blocks, and each block mainly comprises transformation mode information and color residual information of the coded blocks;
(6-2) the performance of point cloud attribute compression is measured by a bitrate and a Peak Signal to Noise Ratio (PSNR), wherein the ode rate is in units of bits per point (bpp), and PSNR is in decibel dB; the smaller the ode rate is, the larger the PSNR is, and the better the performance of point cloud attribute compression is.
