WO2021022621A1 - 一种基于邻居的权重优化的点云帧内预测方法及设备 - Google Patents

一种基于邻居的权重优化的点云帧内预测方法及设备 Download PDF

Info

Publication number
WO2021022621A1
WO2021022621A1 PCT/CN2019/105308 CN2019105308W WO2021022621A1 WO 2021022621 A1 WO2021022621 A1 WO 2021022621A1 CN 2019105308 W CN2019105308 W CN 2019105308W WO 2021022621 A1 WO2021022621 A1 WO 2021022621A1
Authority
WO
WIPO (PCT)
Prior art keywords
point
current point
attribute
value
nearest neighbor
Prior art date
Application number
PCT/CN2019/105308
Other languages
English (en)
French (fr)
Inventor
李革
张琦
邵薏婷
王静
Original Assignee
北京大学深圳研究生院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京大学深圳研究生院 filed Critical 北京大学深圳研究生院
Priority to US17/631,333 priority Critical patent/US20230196625A1/en
Publication of WO2021022621A1 publication Critical patent/WO2021022621A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T9/00—Image coding
    • G06T9/001—Model-based coding, e.g. wire frame
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
    • H04N19/102—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or selection affected or controlled by the adaptive coding
    • H04N19/124—Quantisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T9/00—Image coding
    • G06T9/004—Predictors, e.g. intraframe, interframe coding
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
    • H04N19/102—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or selection affected or controlled by the adaptive coding
    • H04N19/13—Adaptive entropy coding, e.g. adaptive variable length coding [AVLC] or context adaptive binary arithmetic coding [CABAC]
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/10—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding
    • H04N19/134—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using adaptive coding characterised by the element, parameter or criterion affecting or controlling the adaptive coding
    • H04N19/146—Data rate or code amount at the encoder output
    • H04N19/147—Data rate or code amount at the encoder output according to rate distortion criteria
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/50—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using predictive coding
    • H04N19/587—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using predictive coding involving temporal sub-sampling or interpolation, e.g. decimation or subsequent interpolation of pictures in a video sequence
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
    • H04N19/60—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using transform coding
    • H04N19/61—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using transform coding in combination with predictive coding
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
    • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
    • H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
    • H04N21/43—Processing of content or additional data, e.g. demultiplexing additional data from a digital video stream; Elementary client operations, e.g. monitoring of home network or synchronising decoder's clock; Client middleware
    • H04N21/44—Processing of video elementary streams, e.g. splicing a video clip retrieved from local storage with an incoming video stream or rendering scenes according to encoded video stream scene graphs
    • H04N21/4402—Processing of video elementary streams, e.g. splicing a video clip retrieved from local storage with an incoming video stream or rendering scenes according to encoded video stream scene graphs involving reformatting operations of video signals for household redistribution, storage or real-time display

Definitions

  • the invention belongs to the field of point cloud compression, and relates to a point cloud data compression method, in particular to a point cloud intra-frame prediction method and device based on neighbor weight optimization.
  • Three-dimensional point cloud is an important form of digital representation of the real world. With the rapid development of 3D scanning equipment (laser, radar, etc.), the accuracy and resolution of point clouds are higher. High-precision point clouds are widely used in the construction of urban digital maps, and play a technical support role in many popular researches such as smart cities, unmanned driving, and cultural relics protection.
  • the point cloud is obtained by sampling the surface of an object by a three-dimensional scanning device.
  • the number of points in a frame of point cloud is generally in the order of one million. Each point contains geometric information, color, reflectivity and other attribute information, and the amount of data is very large.
  • the huge data volume of 3D point clouds brings huge challenges to data storage and transmission, so point cloud compression is very necessary.
  • Point cloud compression is mainly divided into geometric compression and attribute compression.
  • the test platform TMC13v6 (Test Model for Category 1&3 version 6) provided by the International Standards Organization (Moving Picture Experts Group, referred to as MPEG) is the point cloud attribute compression framework described in There are:
  • LOD Lifting Transform
  • LOD Level of Detail
  • This method first uses the point cloud sorted by Morton code to construct LOD, that is, according to the preset number of LOD levels , Down-sampling the points that have been sorted, and the points obtained after each sampling constitute a layer of LOD, and the sampling distance is from large to small until the entire LOD is constructed. Then find neighbors for the points in the point cloud in the order of LOD, and use the weighted average of all neighbors as the predicted value. The weight of each neighbor is the reciprocal of the Euclidean distance squared. Finally, the actual value of the current point is subtracted from the predicted value to obtain the residual Difference value, use the residual value of the LOD point of this layer to update the weight of the point of the next layer of LOD, and complete the Lifting operation.
  • LOD-based intra prediction mode decision-making method The process of constructing LOD in this method is the same as described above. After constructing LOD, according to the K-Nearest Neighbor (KNN) algorithm to find at most K neighbors for each point, There are K+1 prediction modes, which are: the first, second, ..., and K-th neighbors are used as prediction reference values, and the weighted average of K neighbors is used as the prediction reference value. Then perform rate-distortion optimization (RDO) to determine which prediction mode is finally selected as the intra-frame prediction reference value at the current point.
  • KNN K-Nearest Neighbor
  • Point cloud attribute compression method based on RAHT (Region-Adaptive Hierarchical Transform): This method mainly uses a new transformation algorithm. Different from the traditional discrete cosine transform DCT (Discrete Cosine Transform), RAHT performs a three-dimensional (x, y, z) merging operation on the attribute values in the block generated by the octree, and generates a set of high-frequency The coefficients and a set of low-frequency coefficients are combined to the root node. The final signal to be encoded and transmitted is all high-frequency coefficients and a low-frequency coefficient generated at the root node.
  • DCT Discrete Cosine Transform
  • the present invention proposes an intra-frame prediction mode, when calculating the weight of neighbors, the distance is between x, y, and The method of optimizing the coefficients of the components in the three directions of z, thereby improving the compression performance of the point cloud attributes.
  • the present invention provides a point cloud intra prediction method based on neighbor weight optimization, including: performing steps 1) to 5) at the encoding end, wherein: step 1): traverse the point cloud Point, add them to their respective LOD; step 2): determine the K nearest neighbors of the current point; step 3): calculate the current point according to the coordinates of the current point and the coordinates of the K nearest neighbors The optimization weight of each nearest neighbor point of a point; step 4): Use the optimization weight of the K nearest neighbor points of the current point to perform a weighted summation of the reconstruction attribute values of the K nearest neighbor points to calculate the current point Point attribute prediction value; Step 5): Perform encoding processing according to the attribute prediction value of the current point; and perform step 6) to step 10) at the decoding end, where: step 6) to step 9) are the same as step 1) to Step 4); Step 10): Rebuild the attributes of the point cloud, and perform decoding processing according to the predicted values of the attributes of the current point.
  • the step 2) includes: determining the K nearest neighbors of the current point according to the spatial Euclidean distance from the point in the point cloud to the current point.
  • the method for obtaining the optimal weight of one of the K nearest neighbor points in step 3) includes: calculating the current point and the nearest neighbor point at x, y, z
  • the difference in the three coordinate components is calculated and squared respectively to obtain the square difference of the current point and the nearest neighbor in the three coordinate components of x, y, and z; for the current point and the nearest neighbor
  • the square difference of the three coordinate components of x, y, z is multiplied by the corresponding coefficients ⁇ , ⁇ , and ⁇ to obtain the weighted square of the current point and the nearest neighbor on the three coordinate components of x, y, and z.
  • Difference, where ⁇ , ⁇ , and ⁇ are all constants, and the default values are 1, 1, 1.
  • the three are calculated Sum to obtain the weighted distance square of the current point and the nearest neighbor point; calculate the reciprocal of the weighted distance square of the current point and the nearest neighbor point to obtain the optimized weight of the nearest neighbor point.
  • the step 4) includes: obtaining the optimized weight and attribute reconstruction value of each of the K nearest neighbors of the current point;
  • the optimization weight of each nearest neighbor point in the two nearest neighbor points is divided by the sum of the optimization weights of the K nearest neighbor points to obtain the relative weight of the K nearest neighbor points of the current point; for the current point
  • the relative weight and the attribute reconstruction value of each of the K nearest neighbor points are multiplied in a one-to-one correspondence, and the K products are summed to obtain the attribute prediction value of the current point.
  • step 5 performs encoding processing according to the attribute predicted value of the current point, including: calculating the difference between the attribute value of the current point and the attribute predicted value to determine the current point Prediction residual value; the prediction residual value is coded, and the code stream is obtained through transformation, quantization, and entropy coding.
  • the step 10) performs decoding processing according to the attribute prediction value of the current point, including: entropy decoding, inverse quantization, and inverse transformation on the code stream to obtain the prediction residual of the current point Value; the attribute value of the current point is determined according to the sum of the attribute predicted value of the current point and the predicted residual value.
  • a device for point cloud intra prediction including a point cloud encoding device and a point cloud decoding device
  • the point cloud encoding device includes: a first determining module, To determine the K nearest neighbors of the current point according to the spatial Euclidean distance between the point in the point cloud and the current point; the second determining module is used to determine the K nearest neighbors of the current point according to the coordinates of the current point and the K nearest neighbors Coordinates, calculate the optimal weight of each nearest neighbor point of the current point; the third determining module is used to use the optimal weight of the K nearest neighbor points of the current point to reconstruct the attribute value of the K nearest neighbor points Performing weighted summation to calculate the attribute prediction value of the current point; an encoding module for performing encoding processing according to the attribute prediction value of the current point; and the point cloud decoding device, including: a first determining module for Determining the K nearest neighbors of the current point; a second determining module, configured to calculate the optimal weight of
  • the second determining module is specifically configured to obtain optimized weights for all K nearest neighbors of the current point, wherein for obtaining the optimized weight of one of the K nearest neighbors , Including: calculating the difference between the current point and the nearest neighbor point in the three coordinate components of x, y, and z, and squaring them to obtain the current point and the nearest neighbor point at x, y, z
  • the square difference between the three coordinate components; for the square difference between the current point and the nearest neighbor point in the three coordinate components of x, y, and z, the corresponding coefficients ⁇ , ⁇ , and ⁇ are respectively multiplied to obtain the current point
  • the third determining module is specifically configured to: obtain the optimized weight and the attribute reconstruction value of each of the K nearest neighbors of the current point; for the current point The optimized weights of each of the K nearest neighbors of the K nearest neighbors are divided by the sum of the optimized weights of the K nearest neighbors to obtain the relative weights of the K nearest neighbors of the current point; for the current The relative weight and attribute reconstruction value of each of the K nearest neighbors of a point are multiplied in a one-to-one correspondence, and the K products are summed to obtain the attribute prediction value of the current point.
  • the encoding module is specifically configured to: calculate the difference between the attribute value of the current point and the attribute prediction value, determine the prediction residual value of the current point and the prediction residual value
  • the value is encoded, and the code stream is obtained through transformation, quantization, and entropy coding
  • the decoding module is specifically configured to: perform entropy decoding, inverse quantization, and inverse transformation on the code stream to obtain the prediction residual value of the current point; The sum of the attribute predicted value of the current point and the predicted residual value determines the attribute value of the current point.
  • the present invention provides a point cloud intra prediction and device based on neighbor weight optimization.
  • the method is based on the point cloud distribution in the x, y, and z directions when the point cloud is compressed for intra prediction.
  • the density difference is a solution to optimize the weights of neighbors referenced by points in the point cloud.
  • the coefficients of the distance components in the x, y, and z directions are optimized when calculating the weight of the neighbors.
  • the present invention can improve the accuracy of intra-frame prediction by strengthening the utilization of the overall geometric information of the point cloud, and then perform transformation, quantization and entropy coding processing on the prediction residual to achieve better point cloud attribute compression performance.
  • Fig. 1 is a schematic flowchart of an attribute compression encoding end of an embodiment.
  • Fig. 2 is a schematic flowchart of an attribute compression decoding end of an embodiment.
  • FIG. 3A is a performance comparison diagram of the benchmark results of the embodiment and the test platform TMC13v6 under the conditions of lossless geometry and lossy attributes.
  • FIG. 3B is a performance comparison diagram of the benchmark results of the embodiment and the test platform TMC13v6 under the condition of lossy geometry and lossy attributes.
  • the point cloud intra-frame prediction method based on neighbor weight optimization of the present invention optimizes the coefficients of the components in the x, y, and z directions when calculating neighbor weights in the point cloud attribute compression module.
  • Intra-frame prediction can better capture the overall geometric information of the point cloud, make intra-frame prediction more accurate, and improve the compression performance of point cloud attributes.
  • the point cloud intra prediction method based on neighbor weight optimization of the present invention includes the following steps A and B:
  • steps (1) to (5) on the encoding side including:
  • Construct LOD Traverse the points in the PointCloud and add them to the respective LOD.
  • the specific construction process is: sort all the points in the point cloud according to the Morton code order, that is, according to the preset LOD layer number, the sorted points are down-sampled, and the points obtained after each sampling constitute a Layer LOD, sampling distance from large to small, until the entire LOD construction is completed, the point cloud after LOD sorting is obtained, expressed as point cloud LOD .
  • Finding neighbors determining the K nearest neighbors of the current point, which includes determining the K nearest neighbors of the current point according to the spatial Euclidean distance from the point in the point cloud to the current point. Specifically, according to the order of the points in the point cloud LOD , the entire point cloud space is traversed, and when the current point is processed, K nearest neighbor points (neighbors) are searched for the current point O in the previously processed point set according to the Euclidean distance.
  • a nearest neighbor point calculates the difference between the current point and the nearest neighbor point in the three coordinate components of x, y, and z, and square them respectively to obtain the current point and the nearest neighbor point at x
  • the reconstructed attribute value of each neighbor is The weight of each neighbor is w i , then the attribute prediction value of the current point
  • (5) Generation of point cloud attribute compression code stream encoding according to the attribute prediction value of the current point, including: calculating the difference between the attribute value of the current point and the attribute prediction value to determine the prediction residual of the current point Difference; encode the prediction residual value, and obtain a bitstream after transformation, quantization, and entropy encoding.
  • the point cloud attribute residual value is calculated, and then the residual is transformed, quantized and entropy coded to obtain the point cloud attribute code stream, that is, the compressed code stream of the point cloud attribute.
  • steps (6) to (10) on the decoder side including:
  • the equipment involved in the method of the present invention includes point cloud coding equipment and point cloud decoding equipment:
  • point cloud coding equipment includes:
  • the first determining module is configured to determine the K nearest neighbors of the current point, and is specifically used to determine the K nearest neighbors of the current point according to the spatial Euclidean distance from the point in the point cloud to the current point.
  • the second determining module is configured to calculate the optimal weight of each nearest neighbor point of the current point according to the coordinates of the current point and the coordinates of the K nearest neighbor points; specifically used to: take a nearest neighbor point as For example, calculate the difference between the current point and the nearest neighbor point in the three coordinate components of x, y, and z, and square them respectively to obtain the three coordinates of the current point and the nearest neighbor point in x, y, and z.
  • the square difference in the coordinate components for the square difference between the current point and the nearest neighbor in the three coordinate components of x, y, and z, multiply the corresponding coefficients ⁇ , ⁇ , and ⁇ to obtain the current point and The weighted square difference of the nearest neighbor point in the three coordinate components of x, y, z, where ⁇ , ⁇ , and ⁇ are all constants, and the default values are 1, 1, 1, which can be manually modified and set; for the current point The weighted square difference of the three coordinate components of x, y, and z from the nearest neighbor point, and the sum of the three to obtain the weighted distance square of the current point and the nearest neighbor point; for the current point and the nearest neighbor point Calculate the reciprocal of the square of the weighted distance to obtain the optimal weight of the nearest neighbor point; for all K nearest neighbor points of the current point, find the above-mentioned optimization weight.
  • the third determining module is configured to use the optimized weights of the K nearest neighbor points of the current point to perform a weighted summation of the reconstructed attribute values of the K nearest neighbor points to calculate the attribute predicted value of the current point; ⁇ : Obtain the optimized weight and attribute reconstruction value of each of the K nearest neighbors of the current point; the optimized weight of each of the K nearest neighbors of the current point , Divided by the sum of the optimized weights of the K nearest neighbors to obtain the relative weights of the K nearest neighbors of the current point; the relative weight of each of the K nearest neighbors of the current point The weight and the attribute reconstruction value are multiplied in a one-to-one correspondence, and the K products are summed to obtain the attribute prediction value of the current point.
  • the encoding module is used to perform encoding processing according to the predicted attribute value of the current point. Specifically used to: calculate the difference between the attribute value of the current point and the attribute prediction value to determine the prediction residual value of the current point; encode the prediction residual value, and obtain the code after transformation, quantization, and entropy coding flow.
  • point cloud decoding equipment includes:
  • the first determining module is configured to determine the K nearest neighbors of the current point; specifically, it is used to determine the K nearest neighbors of the current point according to the spatial Euclidean distance from the point in the point cloud to the current point.
  • the second determining module is configured to calculate the optimal weight of each nearest neighbor point of the current point according to the coordinates of the current point and the coordinates of the K nearest neighbor points; the second determining module is specifically configured to: Taking a nearest neighbor point as an example, calculate the difference between the current point and the nearest neighbor point in the three coordinate components of x, y, and z, and square them respectively to obtain the current point and the nearest neighbor point at x The square difference of the three coordinate components of, y and z; for the square difference of the current point and the nearest neighbor in the three coordinate components of x, y, z, multiply the corresponding coefficients ⁇ , ⁇ , ⁇ , Obtain the weighted squared difference between the current point and the nearest neighbor point in the three coordinate components of x, y, and z, where ⁇ , ⁇ , and ⁇ are all constants, and the default values are 1, 1, 1, which can be modified manually ; For the weighted square difference between the current point and the nearest neighbor point in the three coordinate components of x,
  • the third determining module is configured to use the optimized weights of the K nearest neighbor points of the current point to perform a weighted summation of the reconstructed attribute values of the K nearest neighbor points to calculate the attribute predicted value of the current point;
  • the third determining module is specifically configured to: obtain the optimized weight and attribute reconstruction value of each of the K nearest neighbors of the current point; for each of the K nearest neighbors of the current point
  • the optimized weights of the nearest neighbors are divided by the sum of the optimized weights of the K nearest neighbors to obtain the relative weights of the K nearest neighbors of the current point; for each of the K nearest neighbors of the current point
  • the relative weights of the nearest neighbor points and the attribute reconstruction value are multiplied in a one-to-one correspondence, and the K products are summed to obtain the attribute prediction value of the current point.
  • the decoding module is used to perform decoding processing according to the predicted attribute value of the current point. It is specifically used to: perform entropy decoding, inverse quantization, and inverse transformation on the code stream to obtain the prediction residual value of the current point; determine the current point's value according to the sum of the attribute prediction value of the current point and the prediction residual value Attribute value.
  • Steps (1) to (5) are performed on the encoding end, as shown in Figure 1 is a schematic diagram of the structure of the attribute compression encoder of the embodiment:
  • the current point is the first point in the point cloud LOD , it has no neighbors; if the current point is the second point in the point cloud LOD , it can find at most one neighbor; the same is true for the third point at most two Neighbors, the fourth point can find at most three neighbors and all subsequent points can find all three neighbors.
  • Steps (6) to (10) are performed on the decoding end, as shown in Figure 2 is a schematic diagram of the structure of the attribute compression decoder of the embodiment:
  • the present invention tested the distance in the point cloud attribute compression module when calculating the weight of neighbors in the third type of data set according to the standard test condition (Common Test Condition, referred to as CTC) provided by TMC13v6.
  • CTC Common Test Condition
  • Figure 3A and Figure 3B are the experimental results applicable to "LOD-based Lifting Transform".
  • the experimental conditions in Figure 3A are lossless geometry and lossy attributes
  • the experimental conditions in Figure 3B are lossy geometry and lossy attributes. Attributes.
  • the method of the present invention can achieve gains under two CTCs: for the reflectivity attribute, under the two conditions of lossless geometry, lossy genus and lossy geometry, and lossy attributes respectively. Rate-distortion gains of 5.8% and 5.5% are obtained; for color attributes, 0.4% to 1.3% rate-distortion gains are obtained under lossless geometry and lossy conditions, and 1.9 under lossy geometry and lossy attributes are obtained % To 3.0% rate distortion gain.
  • the present invention provides a method for performing intra-frame prediction based on the difference in the density of the point cloud in the x, y, and z directions when compressing the attributes of the point cloud to perform calculation on the weights of neighbors referenced by points in the point cloud.
  • the coefficients are artificially set according to the characteristics of the data set, which is the density of points in the point cloud in the three directions of x, y, and z, to obtain better intra-frame attribute prediction values and improve compression performance.
  • the coefficients of the components of the distance in the three directions of x, y, and z are optimized when calculating the weight of the neighbor.
  • the present invention can improve the accuracy of intra-frame prediction by strengthening the utilization of the overall geometric information of the point cloud, and then perform transformation, quantization and entropy coding processing on the prediction residual to achieve better point cloud attribute compression performance.
  • a point cloud intra prediction method and device based on neighbor weight optimization can be widely used in the real-world digital technology field. With the rapid development of 3D scanning equipment (laser, radar, etc.), the accuracy and resolution of point clouds are higher.
  • the high-precision point cloud of the present invention is widely used in the construction of urban digital maps, and plays a technical support role in many popular researches such as smart cities, unmanned driving, and cultural relics protection.

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Signal Processing (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Compression Or Coding Systems Of Tv Signals (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Compression, Expansion, Code Conversion, And Decoders (AREA)

Abstract

一种基于邻居的权重优化的点云帧内预测及设备，该方法是在点云的属性压缩做帧内预测时，基于点云在x、y、z三个方向上分布的密集程度差异，对点云中的点所参考的邻居的权重进行优化的方案，具体是在计算邻居的权重时对距离在x、y、z三个方向上的分量的系数进行优化。该方法能够通过加强对点云整体几何信息的利用，提高帧内预测的准确性，然后再对预测残差进行变换、量化和熵编码处理，以达到更好的点云属性压缩性能。

Description

一种基于邻居的权重优化的点云帧内预测方法及设备 技术领域
本发明属于点云压缩领域，涉及点云数据压缩方法，尤其涉及一种基于邻居的权重优化的点云帧内预测方法及设备。
背景技术
三维点云是现实世界数字化的重要表现形式。随着三维扫描设备(激光、雷达等)的快速发展，点云的精度、分辨率更高。高精度点云广泛应用于城市数字化地图的构建，在如智慧城市、无人驾驶、文物保护等众多热门研究中起技术支撑作用。点云是三维扫描设备对物体表面采样所获取的，一帧点云的点数一般是百万级别，其中每个点包含几何信息和颜色、反射率等属性信息，数据量十分庞大。三维点云庞大的数据量给数据存储、传输等带来巨大挑战，所以点云压缩十分必要。
点云压缩主要分为几何压缩和属性压缩，目前由国际标准组织(Moving Picture Experts Group，简称为MPEG)所提供的测试平台TMC13v6(Test Model for Category 1&3 version 6)中描述的点云属性压缩框架主要有：
一、基于渐近层次表达(Level of Detail，简称为LOD)的升降变换(Lifting Transform)策略：该方法首先用已按照莫顿码排序的点云构建LOD，即根据预设好的LOD层数，对已经排好序的点进行下采样，每采样一次后已经得到的点构成一层LOD，采样距离由大到小，直至整个LOD构建完成。然后以LOD顺序对点云中的点寻找邻居，以所有邻居的加权平均作为预测值，其中每个邻居的权重是欧氏距离平方的倒数，最后并用当前点的实际值减去预测值得到残差值，用本层LOD点的残差值更新下一层LOD的点的权重，完成了Lifting操作。
二、基于LOD的帧内预测模式决策方法：本方法构建LOD的过程同上面所述，构建完LOD之后，按照K最近邻(K-NearestNeighbor，KNN)算法为每一个点寻找最多K个邻居，则共有K+1种预测模式，分别是：以第一个、第二个、……、第K个邻居作为预测参考值，以及以K个邻居的加权平均作为预测参考值。然后做率失真优化(RDO)，确定最终选择哪种预测模式作为当前点的帧内预测参考值。
三、基于区域自适应层次变换RAHT(Region-Adaptive Hierarchical Transform) 的点云属性压缩方法：此方法主要在于使用了一种新的变换算法。不同于传统的离散余弦变换DCT(Discrete Cosine Transform)，RAHT对由八叉树生成的块内属性值进行三个维度(x、y、z)上的合并操作，每合并一次生成一组高频系数和一组低频系数，直至合并至根节点为止，最终需要编码和传输的信号为所有高频系数和最后在根节点处生成的一个低频系数。
发明的公开
针对上述基于渐近层次表达(Level of Detail，简称为LOD)的升降变化策略(Lifting Transform)的属性编码方案，为了提高属性帧内预测的准确性，更好地利用空间位置关系所反映在属性上的相关度，根据点云中的点在x、y、z三个方向上分布的密集程度，本发明提出一种基于帧内预测模式，在计算邻居的权重时对距离在x、y、z三个方向上的分量的系数进行优化的方法，从而提高点云属性的压缩性能。
本发明提供的技术方案描述如下：
根据本发明的一方面，本发明提供了一种基于邻居的权重优化的点云帧内预测方法，包括：在编码端执行步骤1)至5)，其中：步骤1)：遍历点云中的点，将它们分别加入所属LOD中；步骤2)：确定当前点的K个最近邻点；步骤3)：根据所述当前点的坐标和所述K个最近邻点的坐标，计算所述当前点的每个最近邻点的优化权重；步骤4)：利用所述当前点的K个最近邻点的优化权重，对K个最近邻点的重构属性值进行加权求和，计算所述当前点的属性预测值；步骤5)：根据所述当前点的属性预测值进行编码处理；以及在解码端执行步骤6)至步骤10)，其中：步骤6)～步骤9)同步骤1)～步骤4)；步骤10)：重建点云属性，根据所述当前点的属性预测值进行解码处理。
优选的，在上述方法中，所述步骤2)包括：根据点云中的点到所述当前点的空间欧式距离大小确定所述当前点的K个最近邻点。
优选的，在上述方法中，所述步骤3)中对于获取所述K个最近邻点中的一个的优化权重的方法，包括：计算所述当前点与该最近邻点在x、y、z三个坐标分量上的差值，并分别求平方，得到所述当前点与该最近邻点在x、y、z三个坐标分量上的平方差；对于所述当前点与该最近邻点在x、y、z三个坐标分量上的平方差， 分别乘上相应的系数α、β、γ，得到所述当前点与该最近邻点在x、y、z三个坐标分量上的加权平方差，其中α、β、γ均为常数，默认值为1、1、1；对于所述当前点与该最近邻点在x、y、z三个坐标分量上的加权平方差，三者求和得到所述当前点与该最近邻点的加权距离平方；对于所述当前点与该最近邻点的加权距离平方求倒数，得到该最近邻点的优化权重。
优选的，在上述方法中，所述步骤4)，包括：获得所述当前点的K个最近邻点中的每个最近邻点的优化权重和属性重构值；对于所述当前点的K个最近邻点中的每个最近邻点的优化权重，都除以K个最近邻点的优化权重的和，得到所述当前点的K个最近邻点的相对权重；对于所述当前点的K个最近邻点中的每个最近邻点的相对权重和属性重构值，一一对应相乘，K个乘积再求和，得到所述当前点的属性预测值。
优选的，在上述方法中，其中所述步骤5)根据所述当前点的属性预测值进行编码处理，包括：计算所述当前点的属性值与属性预测值之间的差值确定当前点的预测残差值；对所述预测残差值进行编码，经过变换、量化、熵编码，得到码流。
优选的，在上述方法中，所述步骤10)根据所述当前点的属性预测值进行解码处理，包括：对码流进行熵解码、反量化、反变换，得到所述当前点的预测残差值；根据所述当前点的属性预测值与预测残差值的和确定所述当前点的属性值。
根据本发明的另一方面，还提供了一种用于点云帧内预测的设备，包括点云编码设备和点云解码设备，其中，所述点云编码设备包括：第一确定模块，用于根据点云中的点到所述当前点的空间欧式距离大小确定当前点的K个最近邻点；第二确定模块，用于根据所述当前点的坐标和所述K个最近邻点的坐标，计算所述当前点的每个最近邻点的优化权重；第三确定模块，用于利用所述当前点的K个最近邻点的优化权重，对K个最近邻点的重构属性值进行加权求和，计算所述当前点的属性预测值；编码模块，用于根据所述当前点的属性预测值进行编码处理；以及所述点云解码设备，包括：第一确定模块，用于确定当前点的K个最近邻点；第二确定模块，用于根据所述当前点的坐标和所述K个最近邻点的坐标，计算所述当前点的每个最近邻点的优化权重；第三确定模块，用于利用所述当前点的K个最近邻点的优化权重，对K个最近邻点的重构属性值进行加权求和，计算所述当前点的属性预测值；解码模块，用于根据所述当前点的属性预测值进行解码处理。
优选的，在上述设备中，所述第二确定模块具体用于对所述当前点的所有K个最近邻点求优化权重，其中，对于获取所述K个最近邻点中的一个的优化权重，包括：计算所述当前点与该最近邻点在x、y、z三个坐标分量上的差值，并分别求平方，得到所述当前点与该最近邻点在x、y、z三个坐标分量上的平方差；对于所述当前点与该最近邻点在x、y、z三个坐标分量上的平方差，分别乘上相应的系数α、β、γ，得到所述当前点与该最近邻点在x、y、z三个坐标分量上的加权平方差，其中α、β、γ均为常数，默认值为1、1、1，可以人为修改设定；对于所述当前点与该最近邻点在x、y、z三个坐标分量上的加权平方差，三者求和得到所述当前点与该最近邻点的加权距离平方；对于所述当前点与该最近邻点的加权距离平方求倒数，得到该最近邻点的优化权重。
优选的，在上述设备中，所述第三确定模块具体用于：获得所述当前点的K个最近邻点中的每个最近邻点的优化权重和属性重构值；对于所述当前点的K个最近邻点中的每个最近邻点的优化权重，都除以K个最近邻点的优化权重的和，得到所述当前点的K个最近邻点的相对权重；对于所述当前点的K个最近邻点中的每个最近邻点的相对权重和属性重构值，一一对应相乘，K个乘积再求和，得到所述当前点的属性预测值。
优选的，在上述设备中，其中，所述编码模块具体用于：计算所述当前点的属性值与属性预测值之间的差值确定当前点的预测残差值和对所述预测残差值进行编码，经过变换、量化、熵编码，得到码流；以及所述解码模块具体用于：对码流进行熵解码、反量化、反变换，得到所述当前点的预测残差值；根据所述当前点的属性预测值与预测残差值的和确定所述当前点的属性值。
本发明的有益效果：
本发明提供了一种基于邻居的权重优化的点云帧内预测及设备，该方法是在点云的属性压缩做帧内预测时，基于点云在x、y、z三个方向上分布的密集程度差异，对点云中的点所参考的邻居的权重进行优化的方案，具体是在计算邻居的权重时对距离在x、y、z三个方向上的分量的系数进行优化。本发明能够通过加强对点云整体几何信息的利用，提高帧内预测的准确性，然后再对预测残差进行变换、量化和熵编码处理，以达到更好的点云属性压缩性能。
附图说明
图1是实施例的属性压缩编码端的流程示意图。
图2是实施例的属性压缩解码端的流程示意图。
图3A是在无损几何、有损属性条件下，实施例与测试平台TMC13v6的基准结果的性能对比图。
图3B是在有损几何、有损属性条件下，实施例与测试平台TMC13v6的基准结果的性能对比图。
具体实施方式
下面结合附图，通过实施例进一步描述本发明，但不以任何方式限制本发明的范围。
本发明的基于邻居的权重优化的点云帧内预测方法，在点云属性压缩模块中在计算邻居的权重时对x、y、z三个方向上的分量的系数进行优化，基于点云的帧内预测，更好地捕捉点云整体的几何信息，使得帧内预测更加准确，提高点云属性的压缩性能。
本发明的基于邻居的权重优化的点云帧内预测方法包括以下步骤A和B：
A.在编码端执行步骤(1)至(5)，包括：
(1)构建LOD：遍历点云(PointCloud)中的点，将它们分别加入所属LOD中。具体构建过程为：按照莫顿码顺序点云中所有的点进行排序，即根据预设好的LOD层数，对已经排好序的点进行下采样，每采样一次后已经得到的点构成一层LOD，采样距离由大到小，直至整个LOD构建完成，得到LOD排序后的点云，表示为点云 LOD。
(2)寻找邻居：确定当前点的K个最近邻点，其包括根据点云中的点到所述当前点的空间欧式距离大小确定所述当前点的K个最近邻点。具体地，按照点云 LOD中点的顺序，遍历整个点云空间，处理当前点时，为当前点O在其之前已处理的点集中根据欧氏距离寻找K个最近邻点(邻居)。
(3)计算邻居的权重：根据所述当前点的坐标和所述K个最近邻点的坐标，计算所述当前点的每个最近邻点的优化权重，具体地，设定当前点的几何位置是(x，y，z)，K个邻居中的每个邻居的几何位置为(x i，y i，z i) i＝1，2，...K，则每个邻居的 权重为
Figure PCTCN2019105308-appb-000001
其中α、β、γ是x、y、z三个方向上的系数，需要根据数据集本身的特点人为设定。
以一个最近邻点为例，计算所述当前点与该最近邻点在x、y、z三个坐标分量上的差值，并分别求平方，得到所述当前点与该最近邻点在x、y、z三个坐标分量上的平方差；对于所述当前点与该最近邻点在x、y、z三个坐标分量上的平方差，分别乘上相应的系数α、β、γ，得到所述当前点与该最近邻点在x、y、z三个坐标分量上的加权平方差，其中α、β、γ均为常数，默认值为1、1、1，可以人为修改设定；对于所述当前点与该最近邻点在x、y、z三个坐标分量上的加权平方差，三者求和得到所述当前点与该最近邻点的加权距离平方；对于所述当前点与该最近邻点的加权距离平方求倒数，得到该最近邻点的优化权重；对于所述当前点的所有K个最近邻点求上述的优化权重。
(4)计算属性预测值：利用所述当前点的K个最近邻点的优化权重，对K个最近邻点的重构属性值进行加权求和，计算所述当前点的属性预测值，包括：获得所述当前点的K个最近邻点中的每个最近邻点的优化权重和属性重构值；对于所述当前点的K个最近邻点中的每个最近邻点的优化权重，都除以K个最近邻点的优化权重的和，得到所述当前点的K个最近邻点的相对权重；对于所述当前点的K个最近邻点中的每个最近邻点的相对权重和属性重构值，一一对应相乘，K个乘积再求和，得到所述当前点的属性预测值。
具体地，每个邻居的重构后的属性值为
Figure PCTCN2019105308-appb-000002
每个邻居的权重为w i，则当前点的属性预测值
Figure PCTCN2019105308-appb-000003
(5)点云属性压缩码流的生成：根据所述当前点的属性预测值进行编码处理，包括：计算所述当前点的属性值与属性预测值之间的差值确定当前点的预测残差值；对所述预测残差值进行编码，经过变换、量化、熵编码，得到码流。
具体地，计算点云属性残差值，然后对残差进行变换、量化和熵编码，得到点云属性码流，即点云属性的压缩码流。
B.在解码端执行步骤(6)至(10)，包括：
(6)构建LOD：同步骤(1)；
(7)寻找邻居：同步骤(2)；
(8)计算邻居的权重：同步骤(3)；
(9)计算属性预测值：同步骤(4)；
(10)重建点云属性：根据所述当前点的属性预测值进行解码处理，包括对码流进行熵解码、反量化、反变换，得到所述当前点的预测残差值；根据所述当前点的属性预测值与预测残差值的和确定所述当前点的属性值。
本发明方法涉及的设备，包括点云编码设备和点云解码设备：
其中，点云编码设备包括：
第一确定模块，用于确定当前点的K个最近邻点，具体用于：根据点云中的点到所述当前点的空间欧式距离大小确定所述当前点的K个最近邻点。
第二确定模块，用于根据所述当前点的坐标和所述K个最近邻点的坐标，计算所述当前点的每个最近邻点的优化权重；具体用于：以一个最近邻点为例，计算所述当前点与该最近邻点在x、y、z三个坐标分量上的差值，并分别求平方，得到所述当前点与该最近邻点在x、y、z三个坐标分量上的平方差；对于所述当前点与该最近邻点在x、y、z三个坐标分量上的平方差，分别乘上相应的系数α、β、γ，得到所述当前点与该最近邻点在x、y、z三个坐标分量上的加权平方差，其中α、β、γ均为常数，默认值为1、1、1，可以人为修改设定；对于所述当前点与该最近邻点在x、y、z三个坐标分量上的加权平方差，三者求和得到所述当前点与该最近邻点的加权距离平方；对于所述当前点与该最近邻点的加权距离平方求倒数，得到该最近邻点的优化权重；对于所述当前点的所有K个最近邻点求上述的优化权重。
第三确定模块，用于利用所述当前点的K个最近邻点的优化权重，对K个最近邻点的重构属性值进行加权求和，计算所述当前点的属性预测值；具体用于：获得所述当前点的K个最近邻点中的每个最近邻点的优化权重和属性重构值；对于所述当前点的K个最近邻点中的每个最近邻点的优化权重，都除以K个最近邻点的优化权重的和，得到所述当前点的K个最近邻点的相对权重；对于所述当前点的K个最近邻点中的每个最近邻点的相对权重和属性重构值，一一对应相乘，K个乘积再求和，得到所述当前点的属性预测值。
编码模块，用于根据所述当前点的属性预测值进行编码处理。具体用于：计算所述当前点的属性值与属性预测值之间的差值确定当前点的预测残差值；对所述预测残差值进行编码，经过变换、量化、熵编码，得到码流。
其中，点云解码设备，包括：
第一确定模块，用于确定当前点的K个最近邻点；具体用于：根据点云中的点到所述当前点的空间欧式距离大小确定所述当前点的K个最近邻点。
第二确定模块，用于根据所述当前点的坐标和所述K个最近邻点的坐标，计算所述当前点的每个最近邻点的优化权重；所述第二确定模块具体用于：以一个最近邻点为例，计算所述当前点与该最近邻点在x、y、z三个坐标分量上的差值，并分别求平方，得到所述当前点与该最近邻点在x、y、z三个坐标分量上的平方差；对于所述当前点与该最近邻点在x、y、z三个坐标分量上的平方差，分别乘上相应的系数α、β、γ，得到所述当前点与该最近邻点在x、y、z三个坐标分量上的加权平方差，其中α、β、γ均为常数，默认值为1、1、1，可以人为修改设定；对于所述当前点与该最近邻点在x、y、z三个坐标分量上的加权平方差，三者求和得到所述当前点与该最近邻点的加权距离平方；对于所述当前点与该最近邻点的加权距离平方求倒数，得到该最近邻点的优化权重
第三确定模块，用于利用所述当前点的K个最近邻点的优化权重，对K个最近邻点的重构属性值进行加权求和，计算所述当前点的属性预测值；所述第三确定模块具体用于：获得所述当前点的K个最近邻点中的每个最近邻点的优化权重和属性重构值；对于所述当前点的K个最近邻点中的每个最近邻点的优化权重，都除以K个最近邻点的优化权重的和，得到所述当前点的K个最近邻点的相对权重；对于所述当前点的K个最近邻点中的每个最近邻点的相对权重和属性重构值，一一对应相乘，K个乘积再求和，得到所述当前点的属性预测值。
解码模块，用于根据所述当前点的属性预测值进行解码处理。具体用于：对码流进行熵解码、反量化、反变换，得到所述当前点的预测残差值；根据所述当前点的属性预测值与预测残差值的和确定所述当前点的属性值。
具体实施步骤描述如下：
在编码端执行步骤(1)至(5)，如图1所示为实施例的属性压缩编码器的结构示意图：
(1)构建LOD：配置文件中规定名称为Ford_01_vox1mm-0100.ply的点云文件的LOD层数为8。
第一次遍历时，对整个点云进行下采样，构成集合O 0，则LOD 0＝O 0，剩余的 点被放入R 0中；
第二次遍历时，对R 0进行下采样得到的点构成集合O 1，则LOD 1＝LOD 0∪O 1，剩余的点被放入R 1中；
……
第七次遍历时，从R 5进行下采样得到的点构成集合O 6，则LOD 6＝LOD 5∪O 6，剩余的点构成O 7，LOD 7＝LOD 6∪O 7。
这样，就完成了点云Ford_01_vox1mm-0100.ply的8层LOD构建，生成LOD排序后的点云，表示为点云 LOD。
(2)寻找邻居：对于点云 LOD中的每个点，配置文件中规定点云Ford_01_vox1mm-0100.ply的K近邻数量最多为3。在当前点之前的点中寻找邻居，以欧氏距离最近的3个点作为当前点的邻居。
如果当前点是点云 LOD中的第一个点，则它没有邻居；如果当前点是点云 LOD中的第二个点，则它最多找到一个邻居；同理第三个点最多找到两个邻居，第四个点最多找到三个邻居及之后的点均能找全三个邻居。
(3)计算每个邻居的权重：引入α、β、γ，计算邻居的权重，当前点的几何位置是(x，y，z)，三邻居中的每个邻居的几何位置为(x i，y i，z i) i＝1，2，3，人为设定x、y、z三个方向上的系数为α＝1、β＝1、γ＝3则每个邻居的权重为
Figure PCTCN2019105308-appb-000004
Figure PCTCN2019105308-appb-000005
(4)计算属性预测值：每个邻居的重构后的属性值为
Figure PCTCN2019105308-appb-000006
每个邻居的权重为w i，则当前点的属性预测值
Figure PCTCN2019105308-appb-000007
其中重构属性是通过已经计算得到的属性预测值进行反量化、反变换得到的。
(5)点云属性压缩码流的生成：计算点云属性残差值，然后对残差进行变换、量化和熵编码，得到点云属性的压缩码流。
在解码端执行步骤(6)至(10)，如图2所示为实施例的属性压缩解码器的结构示意图：
(6)构建LOD：同步骤(1)；
(7)寻找邻居：同步骤(2)；
(8)计算邻居的权重：同步骤(3)；
(9)计算属性预测值:同步骤(4)；
(10)重建点云属性：解析点云属性码流，得到当前点的属性值，随后进行反量化、反变换操作，得到当前点的残差，将残差值与预测值相加，得到点云属性重建值。
为了验证本发明的效果，本发明按照TMC13v6提供的标准测试条件(Common Test Condition，简称为CTC)在第三类数据集上测试了在点云属性压缩模块中在计算邻居的权重时对距离在x、y、z三个方向上的分量的系数进行优化之后的效果，结果如图3所示。
图3A和图3B是适用于“基于LOD的升降变化策略(Lifting Transform)”的在实验结果，图3A的实验条件是无损几何、有损属性，图3B的实验条件是有损几何、有损属性。
从图3A和图3B可以看出，在两种CTC下，本发明的方法均能取得增益：对于反射率属性，在无损几何，有损属和有损几何，有损属性两个条件下分别获得了5.8％和5.5％的率失真增益；对于颜色属性，在无损几何，有损属条件下获得了0.4％到1.3％的率失真增益，在有损几何，有损属性条件下获得了1.9％到3.0％的率失真增益。
本发明提供了一种在点云的属性压缩做帧内预测时，基于点云在x、y、z三个方向上分布的密集程度差异，对点云中的点所参考的邻居的权重进行优化的方案，根据点云中的点在x、y、z三个方向上分布的密集程度这一数据集自身特点，人为设定系数，进而得到更好帧内属性预测值，提高压缩性能。具体是在计算邻居的权重时对距离在x、y、z三个方向上的分量的系数进行优化。本发明能够通过加强对点云整体几何信息的利用，提高帧内预测的准确性，然后再对预测残差进行变换、量化和熵编码处理，以达到更好的点云属性压缩性能。
需要注意的是，公布实施例的目的在于帮助进一步理解本发明，但是本领域的技术人员可以理解：在不脱离本发明及所附权利要求的精神和范围内，各种替换和修改都是可能的。因此，本发明不应局限于实施例所公开的内容，本发明要求保护的范围以权利要求书界定的范围为准。
工业应用性
一种基于邻居的权重优化的点云帧内预测方法及设备可以广泛使用于现实世界数字化技术领域。随着三维扫描设备(激光、雷达等)的快速发展，点云的精度、分辨率更高。本发明的高精度点云广泛应用于城市数字化地图的构建，在如智慧城市、无人驾驶、文物保护等众多热门研究中起技术支撑作用。

Claims (10)

  1. 一种基于邻居的权重优化的点云帧内预测方法，其特征在于，包括：
    在编码端执行步骤1)至5)，其中：
    步骤1)：遍历点云中的点，将它们分别加入所属LOD中；
    步骤2)：确定当前点的K个最近邻点；
    步骤3)：根据所述当前点的坐标和所述K个最近邻点的坐标，计算所述当前点的每个最近邻点的优化权重；
    步骤4)：利用所述当前点的K个最近邻点的优化权重，对K个最近邻点的重构属性值进行加权求和，计算所述当前点的属性预测值；
    步骤5)：根据所述当前点的属性预测值进行编码处理；以及在解码端执行步骤6)至步骤10)，
    其中：步骤6)～步骤9)同步骤1)～步骤4)；
    步骤10)：重建点云属性，根据所述当前点的属性预测值进行解码处理。
  2. 根据权利要求1所述的方法，其特征在于，所述步骤2)包括：根据点云中的点到所述当前点的空间欧式距离大小确定所述当前点的K个最近邻点。
  3. 根据权利要求1所述的方法，其特征在于，所述步骤3)中对于获取所述K个最近邻点中的一个的优化权重的方法，包括：
    计算所述当前点与该最近邻点在x、y、z三个坐标分量上的差值，并分别求平方，得到所述当前点与该最近邻点在x、y、z三个坐标分量上的平方差；
    对于所述当前点与该最近邻点在x、y、z三个坐标分量上的平方差，分别乘上相应的系数α、β、γ，得到所述当前点与该最近邻点在x、y、z三个坐标分量上的加权平方差，其中α、β、γ均为常数，默认值为1、1、1；
    对于所述当前点与该最近邻点在x、y、z三个坐标分量上的加权平方差，三者 求和得到所述当前点与该最近邻点的加权距离平方；
    对于所述当前点与该最近邻点的加权距离平方求倒数，得到该最近邻点的优化权重。
  4. 根据权利要求1及3所述的方法，其特征在于，所述步骤4)，包括：
    获得所述当前点的K个最近邻点中的每个最近邻点的优化权重和属性重构值；
    对于所述当前点的K个最近邻点中的每个最近邻点的优化权重，都除以K个最近邻点的优化权重的和，得到所述当前点的K个最近邻点的相对权重；
    对于所述当前点的K个最近邻点中的每个最近邻点的相对权重和属性重构值，一一对应相乘，K个乘积再求和，得到所述当前点的属性预测值。
  5. 根据权利要求1所述的方法，其特征在于，其中所述步骤5)根据所述当前点的属性预测值进行编码处理，包括：计算所述当前点的属性值与属性预测值之间的差值确定当前点的预测残差值；对所述预测残差值进行编码，经过变换、量化、熵编码，得到码流。
  6. 根据权利要求1所述的方法，其特征在于，所述步骤10)根据所述当前点的属性预测值进行解码处理，包括：
    对码流进行熵解码、反量化、反变换，得到所述当前点的预测残差值；
    根据所述当前点的属性预测值与预测残差值的和确定所述当前点的属性值。
  7. 一种用于点云帧内预测的设备，其特征在于，包括点云编码设备和点云解码设备，
    其中，所述点云编码设备包括：
    第一确定模块，用于根据点云中的点到所述当前点的空间欧式距离大小确定当前点的K个最近邻点；
    第二确定模块，用于根据所述当前点的坐标和所述K个最近邻点的坐标，计 算所述当前点的每个最近邻点的优化权重；
    第三确定模块，用于利用所述当前点的K个最近邻点的优化权重，对K个最近邻点的重构属性值进行加权求和，计算所述当前点的属性预测值；
    编码模块，用于根据所述当前点的属性预测值进行编码处理；以及
    所述点云解码设备，包括：
    第一确定模块，用于确定当前点的K个最近邻点；
    第二确定模块，用于根据所述当前点的坐标和所述K个最近邻点的坐标，计算所述当前点的每个最近邻点的优化权重；
    第三确定模块，用于利用所述当前点的K个最近邻点的优化权重，对K个最近邻点的重构属性值进行加权求和，计算所述当前点的属性预测值；
    解码模块，用于根据所述当前点的属性预测值进行解码处理。
  8. 根据权利要求7所述的设备，其特征在于，所述第二确定模块具体用于对所述当前点的所有K个最近邻点求优化权重，其中，对于获取所述K个最近邻点中的一个的优化权重，包括：
    计算所述当前点与该最近邻点在x、y、z三个坐标分量上的差值，并分别求平方，得到所述当前点与该最近邻点在x、y、z三个坐标分量上的平方差；对于所述当前点与该最近邻点在x、y、z三个坐标分量上的平方差，分别乘上相应的系数α、β、γ，得到所述当前点与该最近邻点在x、y、z三个坐标分量上的加权平方差，其中α、β、γ均为常数，默认值为1、1、1，可以人为修改设定；
    对于所述当前点与该最近邻点在x、y、z三个坐标分量上的加权平方差，三者求和得到所述当前点与该最近邻点的加权距离平方；对于所述当前点与该最近邻点的加权距离平方求倒数，得到该最近邻点的优化权重。
  9. 根据权利要求7所述的设备，其特征在于，所述第三确定模块具体用于： 获得所述当前点的K个最近邻点中的每个最近邻点的优化权重和属性重构值；对于所述当前点的K个最近邻点中的每个最近邻点的优化权重，都除以K个最近邻点的优化权重的和，得到所述当前点的K个最近邻点的相对权重；对于所述当前点的K个最近邻点中的每个最近邻点的相对权重和属性重构值，一一对应相乘，K个乘积再求和，得到所述当前点的属性预测值。
  10. 根据权利要求7所述的设备，其特征在于，其中，
    所述编码模块具体用于：计算所述当前点的属性值与属性预测值之间的差值确定当前点的预测残差值和对所述预测残差值进行编码，经过变换、量化、熵编码，得到码流；以及
    所述解码模块具体用于：对码流进行熵解码、反量化、反变换，得到所述当前点的预测残差值；根据所述当前点的属性预测值与预测残差值的和确定所述当前点的属性值。
PCT/CN2019/105308 2019-08-05 2019-09-11 一种基于邻居的权重优化的点云帧内预测方法及设备 WO2021022621A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/631,333 US20230196625A1 (en) 2019-08-05 2019-09-11 Point cloud intra prediction method and device based on weights optimization of neighbors

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910717002.5A CN110418135B (zh) 2019-08-05 2019-08-05 一种基于邻居的权重优化的点云帧内预测方法及设备
CN201910717002.5 2019-08-05

Publications (1)

Publication Number Publication Date
WO2021022621A1 true WO2021022621A1 (zh) 2021-02-11

Family

ID=68365801

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/105308 WO2021022621A1 (zh) 2019-08-05 2019-09-11 一种基于邻居的权重优化的点云帧内预测方法及设备

Country Status (3)

Country Link
US (1) US20230196625A1 (zh)
CN (1) CN110418135B (zh)
WO (1) WO2021022621A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115412716A (zh) * 2021-05-26 2022-11-29 荣耀终端有限公司 一种点云坐标转换残差的编解码方法及装置
CN115412716B (zh) * 2021-05-26 2024-06-04 荣耀终端有限公司 一种点云坐标转换残差的编解码方法及装置

Families Citing this family (34)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021023206A1 (zh) * 2019-08-05 2021-02-11 北京大学深圳研究生院 基于邻居权重优化的点云属性预测、编码和解码方法及设备
CN114667732A (zh) * 2019-12-02 2022-06-24 Oppo广东移动通信有限公司 属性信息的预测方法、编码器、解码器、及存储介质
CN112995758B (zh) * 2019-12-13 2024-02-06 鹏城实验室 点云数据的编码方法、解码方法、存储介质及设备
JP2023514034A (ja) 2020-01-06 2023-04-05 オッポ広東移動通信有限公司 イントラ予測方法、装置、エンコーダ、デコーダ及び記憶媒体
WO2021138785A1 (zh) * 2020-01-06 2021-07-15 Oppo广东移动通信有限公司 划分方法、编码器、解码器及计算机存储介质
WO2021138787A1 (zh) * 2020-01-06 2021-07-15 Oppo广东移动通信有限公司 一种帧内预测方法、编码器、解码器及存储介质
EP4083924A4 (en) * 2020-01-06 2023-02-22 Guangdong Oppo Mobile Telecommunications Corp., Ltd. NEAREST NEIGHBOR SEARCH METHOD, APPARATUS, DEVICE AND INFORMATION MEDIA
CN111242997B (zh) * 2020-01-13 2023-11-10 北京大学深圳研究生院 一种基于滤波器的点云属性预测方法及设备
CN111340899B (zh) * 2020-02-14 2022-09-06 福州大学 一种彩色点云的压缩采样及重构方法
JP2023525207A (ja) 2020-03-24 2023-06-15 オッポ広東移動通信有限公司 イントラ予測方法、装置、エンコーダ、デコーダ、及び記憶媒体
CN111405284B (zh) * 2020-03-30 2022-05-31 北京大学深圳研究生院 一种基于点云密度的属性预测方法及设备
CN111405281A (zh) * 2020-03-30 2020-07-10 北京大学深圳研究生院 一种点云属性信息的编码方法、解码方法、存储介质及终端设备
CN113473127B (zh) * 2020-03-30 2022-09-23 鹏城实验室 一种点云几何编码方法、解码方法、编码设备及解码设备
CN113473153B (zh) * 2020-03-30 2023-04-25 鹏城实验室 一种点云属性预测方法、编码方法、解码方法及其设备
US20230154052A1 (en) * 2020-04-24 2023-05-18 Lg Electronics Inc. Point cloud data transmission device, point cloud data transmission method, point cloud data reception device and point cloud data reception method
KR20230023667A (ko) * 2020-06-24 2023-02-17 베이징 시아오미 모바일 소프트웨어 컴퍼니 리미티드 포인트 클라우드의 인코딩 및 디코딩 방법
CN111953998B (zh) * 2020-08-16 2022-11-11 西安电子科技大学 基于dct变换的点云属性编码及解码方法、装置及系统
CN116744013A (zh) * 2020-09-30 2023-09-12 Oppo广东移动通信有限公司 一种点云分层方法及解码器、编码器、存储介质
WO2022116122A1 (zh) * 2020-12-03 2022-06-09 Oppo广东移动通信有限公司 帧内预测方法及装置、编解码器、设备、存储介质
CN112565794B (zh) * 2020-12-03 2022-10-04 西安电子科技大学 一种点云孤立点编码、解码方法及装置
CN114598883A (zh) 2020-12-07 2022-06-07 腾讯科技（深圳）有限公司 点云属性的预测方法、编码器、解码器及存储介质
CN117710587A (zh) * 2020-12-22 2024-03-15 Oppo广东移动通信有限公司 点云的解码方法、编码方法、解码器以及编码器
WO2022170521A1 (zh) * 2021-02-09 2022-08-18 Oppo广东移动通信有限公司 几何重构方法、解码器以及计算机存储介质
CN116941242A (zh) * 2021-03-04 2023-10-24 Oppo广东移动通信有限公司 帧内预测方法及装置、编解码器、设备、存储介质
CN118042192A (zh) * 2021-03-12 2024-05-14 腾讯科技（深圳）有限公司 点云编码、解码的方法、装置及设备
EP4325857A1 (en) * 2021-04-13 2024-02-21 Guangdong Oppo Mobile Telecommunications Corp., Ltd. Point cloud encoding and decoding methods, encoder, decoder, and computer readable storage medium
JP2024513583A (ja) * 2021-04-16 2024-03-26 オッポ広東移動通信有限公司 符号化・復号化方法およびその関連機器、記憶媒体
CN118075464A (zh) * 2021-06-11 2024-05-24 Oppo广东移动通信有限公司 点云属性的预测方法、装置及编解码器
CN115484462A (zh) * 2021-06-15 2022-12-16 中兴通讯股份有限公司 一种数据处理方法、装置、电子设备和存储介质
CN113534454A (zh) * 2021-07-12 2021-10-22 北京邮电大学 一种多芯光纤信道损伤均衡方法及系统
WO2023023918A1 (zh) * 2021-08-23 2023-03-02 Oppo广东移动通信有限公司 解码方法、编码方法、解码器以及编码器
CN113766229B (zh) * 2021-09-30 2023-04-28 咪咕文化科技有限公司 一种编码方法、解码方法、装置、设备及可读存储介质
WO2023097694A1 (zh) * 2021-12-03 2023-06-08 Oppo广东移动通信有限公司 解码方法、编码方法、解码器以及编码器
CN116781908A (zh) * 2022-03-11 2023-09-19 腾讯科技（深圳）有限公司 点云预测处理方法、装置、计算机、存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101650838A (zh) * 2009-09-04 2010-02-17 浙江工业大学 基于重采样方法和仿射聚类算法的点云简化处理方法
US20180053324A1 (en) * 2016-08-19 2018-02-22 Mitsubishi Electric Research Laboratories, Inc. Method for Predictive Coding of Point Cloud Geometries
CN109257604A (zh) * 2018-11-20 2019-01-22 山东大学 一种基于tmc3点云编码器的颜色属性编码方法
CN109889840A (zh) * 2019-03-20 2019-06-14 北京大学深圳研究生院 点云编码和解码的方法、编码设备和解码设备

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170214943A1 (en) * 2016-01-22 2017-07-27 Mitsubishi Electric Research Laboratories, Inc. Point Cloud Compression using Prediction and Shape-Adaptive Transforms
US10861196B2 (en) * 2017-09-14 2020-12-08 Apple Inc. Point cloud compression
CN108322742B (zh) * 2018-02-11 2019-08-16 北京大学深圳研究生院 一种基于帧内预测的点云属性压缩方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101650838A (zh) * 2009-09-04 2010-02-17 浙江工业大学 基于重采样方法和仿射聚类算法的点云简化处理方法
US20180053324A1 (en) * 2016-08-19 2018-02-22 Mitsubishi Electric Research Laboratories, Inc. Method for Predictive Coding of Point Cloud Geometries
CN109257604A (zh) * 2018-11-20 2019-01-22 山东大学 一种基于tmc3点云编码器的颜色属性编码方法
CN109889840A (zh) * 2019-03-20 2019-06-14 北京大学深圳研究生院 点云编码和解码的方法、编码设备和解码设备

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115412716A (zh) * 2021-05-26 2022-11-29 荣耀终端有限公司 一种点云坐标转换残差的编解码方法及装置
CN115412716B (zh) * 2021-05-26 2024-06-04 荣耀终端有限公司 一种点云坐标转换残差的编解码方法及装置

Also Published As

Publication number Publication date
US20230196625A1 (en) 2023-06-22
CN110418135B (zh) 2022-05-27
CN110418135A (zh) 2019-11-05

Similar Documents

Publication Publication Date Title
WO2021022621A1 (zh) 一种基于邻居的权重优化的点云帧内预测方法及设备
CN110572655B (zh) 一种基于邻居权重的参数选取和传递的点云属性编码和解码的方法及设备
WO2019153326A1 (zh) 一种基于帧内预测的点云属性压缩方法
WO2019213986A1 (zh) 一种基于多角度自适应帧内预测的点云属性压缩方法
WO2020186548A1 (zh) 点云编码和解码的方法、编码设备和解码设备
WO2019210531A1 (zh) 一种基于删除量化矩阵中0元素的点云属性压缩方法
WO2019213985A1 (zh) 一种基于层次划分的点云属性压缩方法
WO2019153342A1 (zh) 一种基于增强图变换的点云属性压缩方法
US11494948B2 (en) Point cloud geometric compression method based on depth auto-encoder
Huang et al. Octree-Based Progressive Geometry Coding of Point Clouds.
US10552989B2 (en) Point cloud attribute compression method based on KD tree and optimized graph transformation
JP7330306B2 (ja) 変換方法、逆変換方法、エンコーダ、デコーダ及び記憶媒体
WO2021023206A1 (zh) 基于邻居权重优化的点云属性预测、编码和解码方法及设备
WO2021142985A1 (zh) 一种基于滤波器的点云属性预测方法及设备
JP7386337B2 (ja) 分割方法、符号器、復号器及びコンピュータ記憶媒体
WO2021062743A1 (zh) 占位信息的预测方法、编码器、解码器、及存储介质
CN106878737A (zh) 高效视频编码中的运动估计加速方法
WO2021062772A1 (zh) 预测方法、编码器、解码器及计算机存储介质
CN113518226A (zh) 一种基于地面分割的g-pcc点云编码改进方法
WO2021138788A1 (zh) 帧内预测方法、装置、编码器、解码器、及存储介质
WO2021062771A1 (zh) 颜色分量预测方法、编码器、解码器及计算机存储介质
Bletterer et al. Point cloud compression using depth maps
Wang et al. Point-voting based point cloud geometry compression
Daribo et al. Point cloud compression for grid-pattern-based 3D scanning system
WO2023123471A1 (zh) 编解码方法、码流、编码器、解码器以及存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19940863

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19940863

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 19940863

Country of ref document: EP

Kind code of ref document: A1