Point cloud geometric information compression system, method and computer system

Abstract

The present invention discloses a point cloud geometric information compression system; it includes a point cloud data preprocessing module, an encoding module, a decoding module and a point cloud reconstruction module; the point cloud data preprocessing module is used to divide the point cloud into blocks and normalize each point cloud block; the encoding module encodes the original center point coordinates of the point cloud block through an octree to form an octree code stream; at the same time, the normalized point cloud block is subjected to feature extraction, and then the encoding probability of the corresponding feature vector is predicted according to the original center point of each point cloud block through a context entropy model, and the feature vector is arithmetically encoded according to the predicted probability value; the decoding module performs decoding based on the encoding module; the point cloud reconstruction module processes the received decoded data to obtain the final point cloud reconstruction result. The present invention also provides a point cloud geometric information compression method. The present invention not only reduces information redundancy, but also effectively improves encoding efficiency.

Description

Technical Field

[0001] The present invention relates to the field of multimedia compression technology, and in particular to a point cloud geometric information compression system, method and computer system. Background Art

[0002] Point clouds are raw data obtained by sensors such as lidar. As an important representation of three-dimensional objects, they have been widely used in many fields. With the rise of virtual reality (VR) and mixed reality (MR), point clouds are receiving increasing attention.

[0003] Specifically, a point cloud is a collection of points in three-dimensional space, where each point is specified by (x, y, z) coordinates and optional attributes such as color and normal vector. With advances in point cloud data acquisition methods and computer simulation technology, point cloud data is becoming increasingly accurate and its size is also increasing, making point cloud compression algorithms increasingly important.

[0004] Traditional lossy point cloud compression methods typically perform poorly in low bitrate environments. G-PCC, a standard proposed by the MPEG organization for compressing point cloud geometric information, includes compression algorithms such as octrees and other nested segmentation algorithms. However, this method's compression performance is limited, and at low bitrates, it can produce a blocky effect similar to mosaicking.

[0005] An autoencoder is a data-driven compression model based on deep learning that learns both analytical and synthetic transformations of data. The analytical transformation acts as an encoder, extracting features from a point cloud; the synthetic transformation acts as a decoder, reconstructing the point cloud. Existing autoencoders for point cloud geometric information compression are mostly based on voxelization and 3D convolution, which can perform poorly on more common non-voxelized point clouds.

[0006] Recently proposed deep learning models such as PointNet and PointNet++ can directly take point cloud points as input and extract point cloud features without requiring voxelization. Currently, methods using PointNet for block-by-block compression of point clouds have achieved state-of-the-art performance. However, existing methods fail to consider the global information of the point cloud and instead focus solely on block-by-block reconstruction, using local reconstruction optimization to approximate global reconstruction optimization. This approach causes the compression model to over-consider block-to-block edge information, resulting in information redundancy. Furthermore, these methods use the same probability distribution to encode features for all blocks, which does not achieve optimal coding efficiency. Summary of the Invention

[0007] Purpose of the invention: To address the problems existing in the prior art, the present invention proposes a point cloud geometric information compression system with good compression effect and reduced information redundancy.

[0008] Technical solution: To achieve the above-mentioned purpose, the present invention provides a point cloud geometric information compression system; comprising a point cloud data preprocessing module, an encoding module, a decoding module and a point cloud reconstruction module; the point cloud data preprocessing module sends the preprocessed data to the encoding module, the decoding module and the point cloud reconstruction module for processing; the encoding module sends the encoded code stream to the decoding module for decoding; the decoding module sends the decoded data to the point cloud reconstruction module;

[0009] Among them, the point cloud data preprocessing module is used to divide the point cloud into blocks and normalize each point cloud block;

[0010] The encoding module includes a point cloud block original center point encoding submodule and a point cloud block encoding submodule;

[0011] The point cloud block original center point encoding submodule encodes the original center point coordinates of the point cloud block through an octree to form an octree code stream;

[0012] The point cloud block encoding submodule extracts features from each normalized point cloud block through the encoder of the autoencoder to obtain a corresponding number of feature vectors; at the same time, a context entropy model is used to predict the encoding probability of the corresponding feature vector based on the original center point of each point cloud block to obtain the probability distribution of each feature vector in each dimension for the quantization level; the feature vector is arithmetically encoded using the predicted probability value to form a feature vector bit stream;

[0013] The decoding module decompresses the octree code stream into the original center point coordinates of each point cloud block; decompresses the feature vector bit stream using the context entropy model shared with the parameters in the encoding module to obtain a reconstructed feature vector set; and decodes the reconstructed feature vector set using the decoder in the autoencoder to obtain the decompressed point cloud block.

[0014] The point cloud reconstruction module processes the received data to obtain the final point cloud reconstruction result.

[0015] Furthermore, the encoder of the autoencoder uses a local feature extraction operation for each point; then PointNet is used to extract a high-dimensional feature vector for the entire block.

[0016] Furthermore, the decoder of the autoencoder uses a multi-layer perceptron to increase the dimension of the feature vector of the point cloud block. It then uses a deformation operation to transform the increased dimension of the point cloud block feature vector, segmenting each point in the hidden layer dimension. Finally, a shared multi-layer perceptron is used to reconstruct the features of each point into coordinates. This structure improves the decoding ability of the decoder, enabling it to better reconstruct the original shape of the point cloud block.

[0017] Furthermore, the contextual entropy model first performs a PointNet operation on the point cloud composed of the original center points of each point cloud block, extracting the global shape of the point cloud composed of the original center points of each point cloud block. After concatenating the coordinates of the original center points of each point cloud block with the global shape, a shared multi-layer perceptron is used to calculate the probability score of the feature vector of each point cloud block, resulting in a probability score matrix for the feature vector of each point cloud block. Finally, a Softmax operation is used to obtain the probability distribution matrix of the quantization level for the value of the feature vector in each dimension. This significantly improves coding efficiency.

[0018] Furthermore, the normalization involves scaling each point cloud block to a unit space and moving the original center point of each point cloud block to the center of the three-dimensional coordinate system. This improves the versatility of the block-based point cloud compression method across different point cloud datasets.

[0019] The present invention also provides a point cloud geometric information compression method, comprising the following steps:

[0020] Step 1: Preprocess the point cloud data, divide the point cloud into blocks, and normalize each point cloud block;

[0021] Step 2: Use the encoder in the autoencoder to perform feature extraction on the normalized point cloud blocks obtained in step 1 to obtain a feature vector matrix consisting of the feature vectors of all point cloud blocks; and quantize each eigenvector in the eigenvector matrix separately;

[0022] Step 3: Input the original center points of all point cloud blocks into the context entropy model to obtain the probability distribution matrix of the quantization level for each dimension of the feature vector obtained in step 2;

[0023] Step 4: Perform octree encoding on the original center point coordinates of all point cloud blocks to form an octree code stream. At the same time, use the entropy coding probability prediction result obtained in step 3 to perform arithmetic encoding on the eigenvector matrix obtained in step 2. The generated eigenvector bit stream is transmitted to the decoding end respectively.

[0024] Step 5: The decoder decompresses the octree code stream into the original center point coordinates of each point cloud block; and decompresses the feature vector bit stream using the context entropy model shared with the parameters in step 3 to obtain a reconstructed feature vector set;

[0025] Step 6: Use the decoder in the autoencoder to decode the reconstructed feature vector set obtained in step 5 to obtain the decompressed point cloud block;

[0026] Step 7: Use the original center point coordinates and scaling of each point cloud block to perform inverse normalization and merge operations on the decompressed point cloud blocks obtained in step 6 to obtain the final point cloud reconstruction result.

[0027] It also includes a loss function constructed by measuring the reconstruction error between the reconstructed point cloud and the original point cloud and combining it with the bit rate R: Loss = D0 + λR; D0 is the Wasserstein distance between the input point cloud and the reconstructed point cloud; λ is the Lagrange multiplier of the loss function; R represents the bit rate; according to the formula: The bit rate is calculated, where n represents the total number of points in the original point cloud, m represents the total number of point cloud blocks after the block operation, i represents the number of point cloud blocks, d represents the number of dimensions of the feature vector of each point cloud block, and j represents the number of dimensions of the feature vector of each point cloud block. The jth element of the feature vector of the i-th point cloud block is The probability of Represents the result of quantizing the jth element on the feature vector of the i-th point cloud block. This overall reconstruction error measurement allows the system to pay less attention to the redundant edge information of the point cloud block, further reducing information redundancy.

[0028] The present invention also provides a computer system, comprising:

[0029] one or more processors;

[0030] A memory stores operable instructions, wherein when the instructions are executed by the one or more processors, the one or more processors are caused to perform operations, wherein the operations include the process of the aforementioned point cloud geometric information compression method.

[0031] Beneficial effects: Compared with the existing technology, the present invention uses block normalization and inverse normalization operations to improve the versatility of block-based point cloud compression methods on different point cloud datasets; applies a global reconstruction optimization loss function based on the idea of ​​block reconstruction to further reduce information redundancy; uses an octree to entropy encode the original center point coordinates of the point cloud block, improving coding efficiency; at the same time, the present invention proposes a context entropy model for point cloud block compression. This entropy model uses the coordinate information of each point and the global shape information composed of other points as context to infer the local shape information of each point, thereby independently predicting the probability distribution of each block, further improving coding efficiency. The present invention significantly improves point cloud compression and reconstruction performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Schematic diagram of the compression process of the present invention;

[0033] Figure 2 Schematic diagram of the structure of the encoder in the autoencoder of the present invention;

[0034] Figure 3 Schematic diagram of the context entropy model structure of the present invention;

[0035] Figure 4 Schematic diagram of the structure of the decoder in the autoencoder of the present invention;

[0036] Figure 5 The point cloud diagram of the "seat" input for the present invention;

[0037] Figure 6 Schematic diagram of the point cloud of the “seat” reconstructed by the present invention;

[0038] Figure 7 Schematic diagram of the “seat” point cloud reconstructed by other methods;

[0039] Figure 8 This is the point cloud diagram of the “water cup” input by the present invention;

[0040] Figure 9 This is a schematic diagram of the point cloud of the "water cup" reconstructed by the present invention;

[0041] Figure 10 Schematic diagram of the “water cup” point cloud reconstructed by other methods. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the examples of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0043] like Figure 1 As shown, this embodiment provides a point cloud geometric information compression method, including the following steps:

[0044] Step 1: Point cloud data preprocessing. Preprocessing mainly includes block operations and normalization operations.

[0045] The blocking operation downsamples a point cloud containing n points using the farthest point sampling FPS, resulting in a sparse point cloud containing m points. Based on the sparse point cloud, the original point cloud is then blocked using KNN, resulting in m point cloud blocks, each containing K points. n represents the total number of points in the processed point cloud, m represents the total number of point cloud blocks after the blocking operation, and K represents the total number of points in each point cloud block.

[0046] Since each point cloud block is distributed at different positions in a point cloud, the size and position of the point cloud block are relatively complex. Therefore, this embodiment adds a normalization operation to unify the size and position of each point cloud block. The normalization operation is to scale each point cloud block to the size of a unit space and move the original center point of each point cloud block to the center of the three-dimensional coordinate. At the same time, the coordinates and scaling ratio of the original center point of each point cloud block are recorded. Among them, the coordinates of the original center point of the i-th point cloud block are (x 0i ,y 0i , z 0i ), i represents the number of the point cloud block, x 0i Indicates the x-axis coordinate value of the original center point of the i-th point cloud block, y 0i Indicates the y-axis coordinate value of the original center point of the i-th point cloud block, z 0i Represents the z-axis coordinate value of the original center point of the i-th point cloud block. The original center point coordinates of all point cloud blocks are given in the form of a sparse point cloud with a shape of (m, 3); the scaling ratio of all point cloud blocks is given in the form of a matrix with a size of (m, 1).

[0047] Step 2: Use the encoder in the autoencoder to perform feature extraction on the m normalized point cloud blocks obtained in step 1, obtaining m feature vectors. Each feature vector has d dimensions, thus obtaining a feature vector matrix of size (m, d). Quantize the feature vector matrix of size (m, d).

[0048] like Figure 2 As shown in the figure, the encoder used in this embodiment first uses a local feature extraction (SetAbstraction) operation on each point to capture the detailed features of each point cloud block in a smaller range. Then, PointNet is used to extract the high-dimensional feature vector of the entire block. Where D represents the dimension of the hidden layer features in the encoder.

[0049] Quantization is to map each element value in each feature vector to an integer range of (0, L-1). Specifically according to the formula:

[0050]

[0051] Carry out the quantization process. Among them, y i,j represents the jth element (1≤i≤m and 1≤j≤d) on the feature vector of the i-th point cloud block, L represents the quantization level, Indicates y i,j The quantization result of , σ represents the Sigmoid function.

[0052] Since the quantization process will make the deep learning method unable to perform backpropagation, this embodiment uses an identity function to calculate the gradient:

[0053]

[0054] This gradient is used in the subsequent gradient descent process of the neural network.

[0055] Step 3: Input the sparse point cloud (m, 3) composed of the original center point coordinates of all point cloud blocks obtained in step 1 into the context entropy model to obtain the probability distribution matrix for the quantization level of the feature vector obtained in step 2 in each dimension.

[0056] The context entropy model predicts the encoding probability of the corresponding feature vector based on the original center point of each point cloud block, and mainly uses the position of each center point relative to the overall shape to infer the value distribution of the feature vector of the corresponding point cloud block. Figure 3 As shown, the context entropy model used in this embodiment first performs a PointNet operation on the sparse point cloud of shape (m, 3) to extract the global shape of the sparse point cloud and obtain a global feature vector of (1, C), where C represents the dimension of the global feature vector. The global feature vector (1, C) is copied m times and then spliced ​​with the sparse point cloud (m, 3). After splicing the point cloud coordinates with the global shape, a shared multi-layer perceptron is used to calculate the probability score of the feature vector of each point cloud block to obtain a probability score matrix (m, d×L) of the feature vector of each point cloud block. Among them, m represents the total number of point cloud blocks, d represents the number of dimensions of the feature vector of each point cloud block, and L represents the quantization level, that is, the range of possible values ​​in each dimension. Finally, we use the Softmax operation in the last dimension to obtain the probability distribution matrix (m, d, L) for the quantization level of the feature vector obtained in step 2 in each dimension.

[0057] Among them, the probability distribution predicted by the context entropy model can be expressed as:

[0058]

[0059] Among them, p i,j (l) represents the probability that the jth element on the feature vector of the i-th point cloud block takes the value l, satisfying l = 0, 1, 2, ..., L-1. The probability distribution of the value of each eigenvector in each dimension corresponds to a distribution law, and each distribution law represents the probability distribution of the corresponding dimension of the corresponding eigenvector at the quantization level L.

[0060] Step 4: Perform octree encoding on the original center point coordinates of all point cloud blocks obtained in step 1 to form an octree code stream; at the same time, use the entropy coding probability prediction results obtained in step 3 to perform arithmetic encoding on the m feature vectors obtained in step 2 to form a feature vector bit stream; the octree code stream and the feature vector bit stream are transmitted to the decoding end respectively.

[0061] Step 5: The decoder decompresses the octree code stream into the original center point coordinates of each point cloud block; and uses the context entropy model shared with the parameters in step 3 to decompress the feature vector bit stream to obtain the reconstructed feature vector set.

[0062] Step 6: Use the decoder in the autoencoder to decode the reconstructed feature vector set obtained in step 5 to obtain the decompressed point cloud block.

[0063] The decoder of the autoencoder reconstructs m feature vectors into m point cloud blocks, each of which contains k points. Figure 4 As shown, in the decoder used in this embodiment, a multi-layer perceptron is first used to increase the dimension of the block feature vector. Then, a deformation operation is used to transform the block feature vector after the dimension increase into the form of (k, D') to segment each point in the hidden layer dimension. Finally, a shared multi-layer perceptron is used to reconstruct the features of each point into coordinates. Where D' represents the dimension of the hidden layer feature in the decoder.

[0064] Step 7: Use the original center point coordinates and scaling scale of each point cloud block to perform inverse normalization and merge operations on the compressed point cloud blocks obtained in step 6 to obtain the final point cloud reconstruction result.

[0065] The point cloud geometric information compression method provided in this embodiment requires training before compressing the point cloud. Through continuous training, the loss function is minimized. In this embodiment, the reconstruction error between the reconstructed point cloud and the original point cloud is measured, and the loss function is constructed in combination with the bit rate R: Loss = D0 + λR. Where D0 is the Wasserstein distance between the input point cloud and the reconstructed point cloud; λ is the Lagrange multiplier of the loss function; and R represents the bit rate. According to the formula: The bit rate is calculated. This overall reconstruction error measurement allows the system to pay less attention to the redundant edge information of the point cloud block, further reducing information redundancy.

[0066] This example sets the size of each block (K) when dividing the point cloud and the depth of the octree for encoding the sparse point cloud (L oc ) to control the compression ratio of the point cloud, that is, different models are trained for different compression ratios. During testing, by specifying K and L oc The point cloud can be compressed at different compression ratios.

[0067] Example

[0068] The present invention will be further described in detail below with reference to a specific embodiment. For ease of description and without loss of generality, the following assumptions are made:

[0069] The method proposed in this paper is intended to be trained and tested using the ModelNet40 dataset. ModelNet40 is the most widely used benchmark for point cloud analysis and is popular for its wide variety, clear shapes, and well-structured data. The original ModelNet40 consists of 12,311 CAD-generated meshes from 40 categories (such as airplanes, cars, factories, and lamps), of which 9,843 are used for training and the remaining 2,468 are used for testing. The point cloud used in this paper is point cloud data uniformly sampled from the ModelNet40 mesh surface, and then further preprocessed by moving to the origin and scaling to the unit sphere.

[0070] This example uses Python 3.9 and PyTorch 1.9 to encode the neural network and compression process.

[0071] Other parameter settings for this example are as follows:

[0072] Feature vector quantization level L = 7; block feature vector dimension d = 16; loss function Lagrange multiplier λ = 10 -6 In addition, the number of blocks m of the point cloud segmentation is determined by the number of point cloud points n and the block size K, specifically: m = 2×n ÷ K; the number of points k of the point cloud block reconstructed by the decoder is determined by the number of points K of the input block, specifically: k = K ÷ 2.

[0073] Taking the compression of a point cloud with 8192 points at a medium bit rate as an example, K=256 is set, that is, the size of each point cloud block is set to 256. This results in 64 point cloud blocks, each containing 256 points. After block normalization, the point cloud blocks are obtained, with the original center point set of size (64, 3) and the scale set of size (64, 1) of the 64 point cloud blocks. After the normalized point cloud blocks pass through the encoder of the autoencoder, they are mapped to 64 16-dimensional feature vectors. After the original center point set of the point cloud block passes through the context entropy model, a 64-dimensional (16, 7) matrix is ​​obtained, where each row of the matrix represents the probability distribution of the feature vector of the corresponding block in each dimension.

[0074] The parameters of the encoder implemented in this embodiment are as follows:

[0075] For the collective feature extraction layer, we set the number of grouped points = 16 and the multi-layer shared perceptron size = (32, 64, 128); for PointNet, we set the multi-layer shared perceptron size = (128, 256, 512, 16).

[0076] The parameters of the context entropy model implemented in this embodiment are as follows:

[0077] For PointNet, set the multi-layer shared perceptron size to (64, 128, 256); for the following multi-layer shared perceptron, set its size to (256, 512, 16×7).

[0078] The probability distribution predicted by the context entropy model is used to perform arithmetic encoding on the 64 16-dimensional feature vectors, and the original center point coordinates of the point cloud block are encoded using an octree.

[0079] On the decoder side, we first use a context entropy model, whose parameters are shared with the encoder side, to predict the probability distribution of the feature vector based on the decoded point cloud block center point coordinates to complete the decoding of the feature vector. After arithmetic decoding, we recover 64 16-dimensional feature vectors, and then use the decoder to reconstruct the point cloud block based on these feature vectors.

[0080] The parameters of the decoder implemented in this embodiment are as follows:

[0081] For the multi-layer perceptron, set its size to (256, 1024, k×128); for the subsequent multi-layer shared perceptron, set its size to (128, 256, 512, 3).

[0082] After the decoder decodes and obtains 64 reconstructed point cloud blocks with 128 points each, we perform inverse normalization on these blocks. Specifically, we use the coordinates of the center point of the original point cloud block decoded from the octree to move the center of the reconstructed point cloud block to the original center point position, and use the transmitted scaling factor to scale the reconstructed point cloud block back to its original size. Finally, we merge the 64 inverse normalized point cloud blocks with a shape of (128, 3) into a (64×128, 3) point cloud, thus obtaining the final point cloud reconstruction result.

[0083] Examples of 3D point cloud compression reconstruction results are as follows: Figures 5 to 10 As shown. Figure 7 and Figure 10 Represents the reconstruction results of existing block-based and deep learning point cloud compression methods, Figure 6 and Figure 9 The reconstruction results obtained by the embodiment of the present invention are shown in Figure 2. The results of the two methods are obtained at similar bit rates. It can be seen that the present invention has a significant improvement in compression performance compared to existing block-based and deep learning methods.

[0084] The present invention also provides a computer system comprising: one or more processors; a memory storing operable instructions, wherein when the instructions are executed by the one or more processors, the one or more processors perform operations, wherein the operations include the process of the aforementioned point cloud geometric information compression method.

[0085] It should be understood that the examples of the aforementioned point cloud geometric information compression method of the present invention can be implemented in any computer system that includes data storage and data processing capabilities. The aforementioned computer system can be at least one electronic processing system or electronic device including a processor and memory, such as a personal computer, whether a personal computer, a commercial computer, a graphics processing computer, or a server-class computer. These PCs can implement wired and / or wireless data transmission, especially image data, by having data interfaces and / or network interfaces.

[0086] In other embodiments, the computer system may also be a server, in particular a cloud server, having data storage, processing, and network communication functions.

[0087] As an example, a computer system generally includes at least one processor, a memory, and a network interface connected by a system bus. The network interface is used to communicate with other devices / systems.

[0088] The processor is used to provide calculation and control of the system.

[0089] The memory includes non-volatile memory and cache.

[0090] Non-volatile memory generally has a massive storage capacity and can store operating systems and computer programs. These computer programs may include operable instructions that, when executed by one or more processors, enable one or more processors to execute the process of the point cloud geometric information compression method of the aforementioned embodiment of the present invention.

[0091] In necessary or reasonable implementations, the aforementioned computer system, whether a PC device or a server, may also include more or fewer components than shown in the figure, or a combination, or adopt different hardware, software and other components or different deployment methods.

Claims

1. A point cloud geometric information compression system; characterized by: It includes a point cloud data preprocessing module, an encoding module, a decoding module and a point cloud reconstruction module; the point cloud data preprocessing module sends the preprocessed data to the encoding module, the decoding module and the point cloud reconstruction module for processing; the encoding module sends the encoded code stream to the decoding module for decoding; the decoding module sends the decoded data to the point cloud reconstruction module; Among them, the point cloud data preprocessing module is used to divide the point cloud into blocks and normalize each point cloud block; The encoding module includes a point cloud block original center point encoding submodule and a point cloud block encoding submodule; The point cloud block original center point encoding submodule encodes the original center point coordinates of the point cloud block through an octree to form an octree code stream; The point cloud block encoding submodule extracts features from each normalized point cloud block through the encoder of the autoencoder to obtain a corresponding number of feature vectors; at the same time, a context entropy model is used to predict the encoding probability of the corresponding feature vector based on the original center point of each point cloud block to obtain the probability distribution of each feature vector in each dimension for the quantization level; the feature vector is arithmetically encoded using the predicted probability value to form a feature vector bit stream; The decoding module decompresses the octree code stream into the original center point coordinates of each point cloud block; decompresses the feature vector bit stream using the context entropy model shared with the parameters in the encoding module to obtain a reconstructed feature vector set; and decodes the reconstructed feature vector set using the decoder in the autoencoder to obtain the decompressed point cloud block. The point cloud reconstruction module processes the received data to obtain the final point cloud reconstruction result; The decoder of the autoencoder uses a multi-layer perceptron to increase the dimension of the feature vector of the point cloud block, then uses a deformation operation to deform the feature vector of the point cloud block after the dimension increase, segments each point in the hidden layer dimension, and finally uses a shared multi-layer perceptron to reconstruct the features of each point into coordinates; The context entropy model first performs a PointNet operation on the point cloud composed of the original center points of each point cloud block, extracts the global shape of the point cloud composed of the original center points of each point cloud block, concatenates the coordinates of the original center points of each point cloud block with the global shape, and uses a shared multi-layer perceptron to calculate the probability score of the eigenvector of each point cloud block to obtain a probability score matrix of the eigenvector of each point cloud block; finally, a Softmax operation is used to obtain a probability distribution matrix for the quantization level of the eigenvector values ​​in each dimension.

2. The point cloud geometric information compression system according to claim 1, characterized in that: The autoencoder’s encoder uses a local feature extraction operation for each point; PointNet is then used to extract a high-dimensional feature vector for the entire block.

3. The point cloud geometric information compression system according to claim 1, characterized in that: The normalization is to scale each point cloud block to a unit space and move the original center point of each point cloud block to the center of the three-dimensional coordinate.

4. A point cloud geometric information compression method, characterized by: The following steps are involved: Step 1: Preprocess the point cloud data, divide the point cloud into blocks, and normalize each point cloud block; Step 2: Use the encoder in the autoencoder to perform feature extraction on the normalized point cloud blocks obtained in step 1 to obtain a feature vector matrix consisting of the feature vectors of all point cloud blocks; and quantize each eigenvector in the eigenvector matrix separately; Step 3: Input the original center points of all point cloud blocks into the context entropy model to obtain the probability distribution matrix of the quantization level for each dimension of the feature vector obtained in step 2; Step 4: Perform octree encoding on the original center point coordinates of all point cloud blocks to form an octree code stream. At the same time, use the entropy coding probability prediction result obtained in step 3 to perform arithmetic encoding on the eigenvector matrix obtained in step 2. The generated eigenvector bit stream is transmitted to the decoding end respectively. Step 5: The decoder decompresses the octree code stream into the original center point coordinates of each point cloud block; and decompresses the feature vector bit stream using the context entropy model shared with the parameters in step 3 to obtain a reconstructed feature vector set; Step 6: Use the decoder in the autoencoder to decode the reconstructed feature vector set obtained in step 5 to obtain the decompressed point cloud block; Step 7: Use the original center point coordinates and scaling of each point cloud block to perform inverse normalization and merge operations on the decompressed point cloud blocks obtained in step 6 to obtain the final point cloud reconstruction result; The decoder of the autoencoder uses a multi-layer perceptron to increase the dimension of the feature vector of the point cloud block, then uses a deformation operation to deform the feature vector of the point cloud block after the dimension increase, segments each point in the hidden layer dimension, and finally uses a shared multi-layer perceptron to reconstruct the features of each point into coordinates; The context entropy model first performs a PointNet operation on the point cloud composed of the original center points of each point cloud block, extracts the global shape of the point cloud composed of the original center points of each point cloud block, concatenates the coordinates of the original center points of each point cloud block with the global shape, and uses a shared multi-layer perceptron to calculate the probability score of the eigenvector of each point cloud block to obtain a probability score matrix of the eigenvector of each point cloud block; finally, a Softmax operation is used to obtain a probability distribution matrix for the quantization level of the eigenvector values ​​in each dimension.

5. The point cloud geometric information compression method according to claim 4, characterized in that: It also includes a loss function constructed by measuring the reconstruction error between the reconstructed point cloud and the original point cloud and combining it with the bit rate R: Loss = D0 + λR; where D0 is the Wasserstein distance between the input point cloud and the reconstructed point cloud; λ is the Lagrange multiplier of the loss function; R represents the bit rate; according to the formula: The bit rate is calculated, where n represents the total number of points in the original point cloud, m represents the total number of point cloud blocks after the block operation, i represents the number of point cloud blocks, d represents the number of dimensions of the feature vector of each point cloud block, and j represents the number of dimensions of the feature vector of each point cloud block. The jth element of the feature vector of the i-th point cloud block is The probability of Represents the result of quantizing the jth element on the feature vector of the i-th point cloud block.

6. The point cloud geometric information compression method according to claim 4, characterized in that: The normalization in step 1 is to scale each point cloud block to a unit space and move the original center point of each point cloud block to the center of the three-dimensional coordinate.

7. A computer system, characterized in that: include: one or more processors; A memory storing operable instructions, wherein when the instructions are executed by the one or more processors, the one or more processors are caused to perform operations, wherein the operations include the process of the point cloud geometric information compression method according to any one of claims 4 to 6.

Images