A fast point cloud data transmission method and system

Abstract

The present invention provides a fast point cloud data transmission method and system, which includes: collecting original point cloud data through sensors on the client; using a Poisson disk downsampling method based on an adaptive sampling strategy to perform sparse processing on the original point cloud data to generate sparse point cloud data; transmitting the sparse point cloud data to the server through a low-latency communication protocol; using a pre-trained dynamic graph convolutional neural network to upsample the sparse point cloud data to reconstruct the original point cloud data. Not only can it accurately eliminate redundant information and significantly improve the efficiency of point cloud data transmission, but it can also ensure that the captured high-resolution and key structural information is effectively preserved during the transmission process, and is suitable for real-time application scenarios such as autonomous driving and drone collaborative perception.

Description

Technical Field

[0001] The present invention belongs to the field of point cloud data processing and transmission, and in particular relates to a fast point cloud data transmission method and system. Background Art

[0002] As the core achievement of 3D scanning technology (such as LiDAR, laser scanner, etc.), 3D point cloud data contains rich 3D coordinate information and is used in many technical fields such as autonomous driving and cooperative perception of drones. However, its huge data size and complex structural characteristics have brought great challenges to real-time transmission and efficient processing. However, in application scenarios such as autonomous driving and cooperative perception of drones, the efficiency and accuracy of point cloud data transmission are required to be high, and the existing transmission methods are difficult to meet this demand. Traditional compression methods, such as voxelization or grid downsampling, can reduce the amount of data to a certain extent, but still hide a lot of redundant information, such as repeated local structures; at the same time, there is also the risk of information loss, especially when dealing with complex scenes, key structural features will be lost, and it is difficult to accurately restore the complete appearance of the original data. Moreover, the traditional point cloud compression method has weak adaptability and usually requires professional knowledge and manual parameter adjustment. Therefore, a point cloud data transmission method that can not only accurately eliminate redundant information and significantly improve the efficiency of point cloud data transmission, but also ensure that the high resolution and key structural information captured by technologies such as LiDAR are effectively preserved during the transmission process is needed.

[0003] The use of deep learning methods improves the compression rate, but there may be information loss in the encoding and decoding process. Combining the advantages and disadvantages of traditional methods and learning methods, this patent proposes a point cloud data transmission method that integrates adaptive radius Poisson disk downsampling (Poisson Disk Downsampling) and dynamic graph convolutional neural network (Dynamic Graph Convolutional Network, DGCNN) upsampling. This methodology can not only accurately eliminate redundant information and significantly improve the efficiency of point cloud data transmission, but also ensure that the high-resolution and key structural information captured by technologies such as LiDAR are effectively preserved during the transmission process, ensuring that it can be applied to real-time application scenarios such as autonomous driving and drone collaborative perception. Summary of the invention

[0004] The purpose of the present invention is to provide a fast point cloud data transmission method and system, which can not only accurately eliminate redundant information and significantly improve the efficiency of point cloud data transmission, but also ensure that the captured high-resolution and key structural information are effectively preserved during the transmission process. It is suitable for real-time application scenarios such as autonomous driving and drone collaborative perception.

[0005] The technical solution to achieve the purpose of the present invention is:

[0006] A fast point cloud data transmission method, comprising:

[0007] Collect raw point cloud data through sensors on the client side;

[0008] The original point cloud data is sparsely processed by using the Poisson disk downsampling method based on the adaptive sampling strategy to generate sparse point cloud data.

[0009] Transmit sparse point cloud data to the server via a low-latency communication protocol;

[0010] A pre-trained dynamic graph convolutional neural network is used to upsample the sparse point cloud data and reconstruct the original point cloud data.

[0011] Furthermore, the sparse processing of the original point cloud data using the Poisson disk downsampling method based on the adaptive sampling strategy specifically includes:

[0012] Calculate each point The local point cloud density;

[0013] Adaptively update the sampling radius based on the local density;

[0014] Initialize the sparse point cloud data set, select sampling points from the sparse point cloud data set, select new sampling points from the original point cloud data, and if the distance between each new sampling point and the already selected sampling point is not less than the sampling radius corresponding to the currently selected sampling point, add the new sampling point to the sparse point cloud data set, and generate a sparse point cloud data set through iteration.

[0015] Furthermore, the sampling radius is:

[0016]

[0017] in, Yes Optimized sampling radius; Indicate point The local point cloud density; is a constant.

[0018] Furthermore, the sparse point cloud data set is:

[0019]

[0020] in, , , N is the number of original point cloud data.

[0021] Furthermore, the dynamic graph convolutional neural network includes an input layer, an edge feature construction layer, a convolution operation layer and an upsampling module, wherein:

[0022] The input layer is used to input sparse point cloud data , and the three-dimensional coordinates of each point ;

[0023] The edge feature construction layer converts the sparse point cloud dataset Convert to graph structure ,in, is a vertex set, representing a sparse point cloud dataset The point in For edge sets, calculate the characteristics of the edges , the K nearest neighbor algorithm is used to construct the adjacency matrix of the graph;

[0024] The convolution operation layer performs convolution operation on the edge set through dynamic graph convolution operation to update the features of each point;

[0025] The upsampling module gradually converts the sparse point cloud data into a three-dimensional space by using an enhanced multi-scale graph convolutional network based on DenseGCN and node reordering and reconstruction. Upsampling to restore high-precision point cloud .

[0026] Furthermore, the edge features for:

[0027]

[0028] in, Indicate point The eigenvector of is initially the three-dimensional coordinates of the point, The nearest neighbor The characteristic vector of Represents the connection function.

[0029] Furthermore, the output of the convolution operation layer is:

[0030]

[0031] in, It is Node feature representation of the layer; is the adjacency matrix of the graph, including self-connected edges; is the node degree matrix, with diagonal elements , are the rows and columns of the adjacency matrix respectively; is a trainable weight matrix; is the activation function.

[0032] Furthermore, the dynamic graph convolutional neural network is optimized using a loss function, and the loss function is:

[0033]

[0034] in is a point in the original point cloud, is the reconstructed point.

[0035] Furthermore, the enhanced multi-scale graph convolution and node reordering based on DenseGCN and the reconstruction back to three-dimensional space operations specifically include:

[0036] The features of the point cloud are extracted through three serial DenseGCN-based enhanced multi-scale graph convolutional network modules:

[0037]

[0038] in, To enhance the multi-scale graph convolutional network module The input features of To enhance the multi-scale graph convolutional network module The output features of It is an enhanced multi-scale graph convolutional network module The convolution kernel size, It is an enhanced multi-scale graph convolutional network module Expansion rate, It is an enhanced multi-scale graph convolutional network module The number of output channels;

[0039] Through a layer of GCN, the node features are transformed from the original dimension Expand to , the expanded node features Rearrange The dimension characteristics become ,Right now:

[0040]

[0041]

[0042]

[0043] in, is the inverse of the downsampling ratio, is the number of nodes of sparse point cloud data, is the number of nodes of the original point cloud data, and is a learnable parameter, To reshape the function, is the convolution operation layer Node characteristics of the layer;

[0044] Using multi-layer perceptron Point Cloud Reconstruct into three-dimensional space.

[0045] A fast point cloud data transmission system, comprising:

[0046] The client collects raw point cloud data through sensors, uses the Poisson disk downsampling method based on the adaptive sampling strategy to perform sparse processing on the raw point cloud data, generates sparse point cloud data, and transmits the sparse point cloud data to the server through a low-latency communication protocol;

[0047] On the server side, a pre-trained dynamic graph convolutional neural network is used to upsample the sparse point cloud data and reconstruct the original point cloud data.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) The present invention proposes an adaptive radius Poisson disk downsampling technology to perform sparse processing on the original point cloud data. This method analyzes the local density information of each point and adaptively adjusts the sampling radius, thereby reducing the number of redundant points while ensuring the global structural integrity of the point cloud, thereby improving the efficiency of point cloud transmission and processing;

[0050] (2) The present invention transmits sparse point cloud data to the server, reducing transmission cost and time;

[0051] (3) The dynamic graph convolutional neural network (DGCNN) upsampling method of the present invention integrates the enhanced multi-scale graph convolutional network of DenseGCN, which can efficiently extract the local and global structural information of the point cloud, quickly and effectively restore the key information of the point cloud, and achieve high-precision point cloud upsampling reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of a fast point cloud data transmission method proposed by the present invention.

[0053] Figure 2 This is a network model structure framework diagram in a fast point cloud data transmission method of the present invention. DETAILED DESCRIPTION

[0054] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings to help those skilled in the art better understand the present invention. It should be particularly noted that in the following description, in order to highlight the core content of the present invention, the details of the known functions and designs will not be described in detail if they do not directly affect the understanding of the present invention.

[0055] Combination Figure 1The present application provides a fast point cloud data transmission method integrating Poisson disk downsampling and dynamic graph convolutional neural network upsampling, which specifically includes the steps of:

[0056] Step 1: Provide an adaptive radius Poisson disk downsampling technology to perform sparse processing on the original point cloud data. The sparse processing of the optimization algorithm is performed on the sensor end to reduce the burden of data transmission. On the sensor end, the Poisson disk sampling technology is first used to perform sparse processing on the original point cloud data to reduce the volume of the point cloud data. Poisson disk sampling ensures that there is a sufficient minimum distance between the sampling points to ensure that the data is sparse while retaining the geometric features. Specifically, it includes:

[0057] Step 1-1 Capture of original point cloud data during autonomous driving

[0058] Get the raw point cloud data captured by the sensor , where each point Represents a position in three-dimensional space, and the number of points is .

[0059] Step 1-2 Optimization of Poisson disk sampling radius

[0060] According to the traditional Poisson disk sampling, a sampling radius is set , defines the distance threshold for Poisson disk sampling. Iteratively select sampling points from the point cloud, ensuring that the distance between each new sampling point and the already selected sampling point is no less than In order to further improve the accuracy and efficiency of Poisson disk sampling, this paper proposes an adaptive sampling strategy based on local geometric information. This method analyzes the local density information of each point and adaptively adjusts the sampling radius, thereby reducing the number of redundant points while ensuring the global structural integrity of the point cloud, and improving the efficiency of point cloud transmission and processing.

[0061] The present invention calculates each point The local density around it, based on the local density information, calculates the optimal sampling radius of the Poisson disk , the constraint formula for designing the sampling radius is:

[0062]

[0063] Among them, it is point The sampling radius, is the local point cloud density of the point, is a user-defined constant that controls sparsity.

[0064] This optimization strategy uses a smaller sampling radius in high-density areas to capture detailed information, and a larger sampling radius in low-density areas to reduce redundant sampling. In this way, Poisson disk sampling can more effectively compress point cloud data while preserving its geometric structure and detailed features. This adaptive sampling radius optimization method greatly improves the flexibility and efficiency of sampling.

[0065] Step 1-3 Recursive Poisson disk sampling to generate a sparse point cloud dataset ,in, .

[0066] Apply the Poisson disk sampling algorithm. First, select an initial point in the point cloud data, use it as a seed point, and mark it as selected. Then, according to the sampling radius Constraints are used to filter out other adjacent distances less than The process is repeated recursively until the entire point cloud is processed and sparse point cloud data is generated. ,in, ;

[0067] The mathematical expression is:

[0068]

[0069] Through this process, a sparse point cloud is obtained. , the number of points is significantly reduced, but the geometric structure information of the original point cloud can still be retained.

[0070] Step 1-4 Sampling completion judgment

[0071] Determine whether the density of the sparse point cloud meets the preset conditions (such as error threshold, resolution requirements, etc.). If the conditions are met, end the downsampling process and output the sparse point cloud; otherwise, readjust the sampling radius or other parameters as required, repeat steps S102 and S103, and the final downsampling ratio is.

[0072]

[0073] Step 1-5 Downsampling and compressing point cloud data transmission

[0074] The downsampled sparse point cloud data is transmitted to the server through a low-latency communication protocol to reduce transmission costs and time.

[0075] Step 2: Provide a dynamic graph convolutional neural network to upsample the coefficient point cloud. On the server side, use a pre-trained dynamic graph convolutional neural network (DGCNN) to upsample the sparse point cloud data and reconstruct the original point cloud.

[0076] Combination Figure 2The dynamic graph convolutional neural network (DGCNN) upsampling module of the present invention introduces a multi-scale feature extractor and a node rearrangement module (NodeShuffle) to efficiently upsample sparse point cloud data and enhance the reconstruction accuracy of point cloud details. This structure draws on the idea of ​​PU-GCN (Point Cloud Upsampling Graph Convolutional Network), but is optimized and improved to improve the reconstruction efficiency of point cloud data.

[0077] The model structure of the dynamic graph convolutional neural network (DGCNN) includes: input layer, edge feature construction module, convolution operation layer and upsampling module, among which,

[0078] The input layer inputs the point cloud ,Include points, each with three-dimensional coordinates ;

[0079] The edge feature construction module uses the K nearest neighbor algorithm to establish an adjacency relationship for each point, construct a dynamic edge set, and calculate the edge features. ,in Indicate point and The relationship between them is defined as:

[0080]

[0081] in Indicate point The eigenvector of , initially the three-dimensional coordinates of the point.

[0082] The convolution operation layer performs convolution operations on edge sets through dynamic graph convolution operations to update the features of each point:

[0083]

[0084] in is the weight matrix, is the bias term, Indicate point 's neighborhood.

[0085] The upsampling module gradually converts the sparse point cloud into a three-dimensional space by enhancing multi-scale graph convolution based on DenseGCN, reordering nodes, and reconstructing them into three-dimensional space. Upsampling to restore high-precision point cloud .

[0086] The dynamic graph convolutional neural network ensures the accuracy of upsampling by minimizing the loss function of point cloud reconstruction. The loss function is defined as:

[0087]

[0088] in is a point in the original point cloud, is the reconstructed point.

[0089] The sparse point cloud data is upsampled using the pre-trained dynamic graph convolutional neural network (DGCNN), specifically including the following steps:

[0090] Step 2-1: sparse point cloud data Convert to graph representation

[0091] Sparse point cloud Convert to graph structure , where the vertex set represents a point in the point cloud, The adjacency matrix of the graph is constructed based on the spatial relationship between points. For example, the K-nearest neighbor (K-NN) algorithm is used to obtain Establish a certain number of adjacency relationships for each point, and finally obtain the adjacency matrix of the graph.

[0092] Step 2-2, process the point cloud features through dynamic graph convolution operation, the process is as follows:

[0093]

[0094] in: It is Node feature representation of the layer; is the adjacency matrix of the graph, including self-connected edges; is the node degree matrix, whose diagonal elements , here are the rows and columns of the adjacency matrix respectively; is a trainable weight matrix; is the activation function (such as ReLU);

[0095] The last layer That is, a simple graph representation of the received transmission ; The features extracted through these operations can characterize the local geometric structure in the point cloud and provide effective feature representation for the upsampling process.

[0096] Step 2-3, use the dense graph convolutional network module (DenseGCN) as the basic structure of the enhanced multi-scale feature extractor to extract the features of the point cloud through the dense graph convolutional network block. The dense graph convolutional network module consists of parameters Definition, where: is the number of neighbors (convolution kernel size), that is, the number of neighboring nodes selected in the convolution operation; is the dilation rate, which is used to control the receptive field of the convolution operation; is the number of output channels, which specifies the dimension of features generated by each dense graph convolutional network block.

[0097] The enhanced multi-scale feature extractor structure uses two parallel and three serial dense graph convolutional network modules (DenseGCN) blocks, and the expansion rate of each block is Different, in order to achieve multi-scale feature extraction. The formula for feature extraction is:

[0098]

[0099] in For modules The input features of For modules The final output feature vector is obtained by feature-level connection:

[0100]

[0101] Step 2-4, adopt the node reordering strategy to reorder the nodes in the graph to enhance the robustness of the model. The node reordering process aims to enhance the robustness and feature expression ability of the model. The process can be divided into two steps, as follows:

[0102] Step 2-4-1 Channel expansion

[0103] Through a layer of GCN, the node features are transformed from the original dimension Expand to ,in The inverse of the expansion ratio set for the previous downsampling.

[0104]

[0105] The channel expansion formula is:

[0106]

[0107] in, and are still learnable parameters, is the expanded node feature, where That is .

[0108] Step 2-4-2 Periodic Shuffle

[0109] Expanded node features be rearranged The dimension characteristics become ,in is the number of nodes. The reordering process is implemented by the following steps:

[0110]

[0111] Step 2-5: Point cloud reconstruction after upsampling

[0112] The server uses the previously trained dynamic graph convolutional neural network model to upsample and reconstruct the transmitted sparse point cloud. Specifically, the server loads the weights learned in the training phase. and bias ,These weights and biases can accurately capture the local and global structural features of the point cloud during the upsampling process, and realize the restoration of sparse point cloud to high-density point cloud.

[0113] In this embodiment, a multi-layer perceptron is used according to the obtained features. Reconstruct the point cloud into three-dimensional space and predict and restore the reconstructed point cloud of the upsampled dense point cloud .

[0114]

[0115] This embodiment also provides a fast point cloud data transmission system, including:

[0116] The client collects raw point cloud data through sensors, uses the Poisson disk downsampling method based on the adaptive sampling strategy to perform sparse processing on the raw point cloud data, generates sparse point cloud data, and transmits the sparse point cloud data to the server through a low-latency communication protocol;

[0117] On the server side, a pre-trained dynamic graph convolutional neural network is used to upsample the sparse point cloud data and reconstruct the original point cloud data.

[0118] The method of the present invention performs sparse processing on the point cloud by adopting a Poisson disk downsampling algorithm with an adaptive radius on the sensor side to reduce the burden of data transmission; a deep learning model based on a dynamic graph convolutional neural network is used on the server side, and the enhanced multi-scale graph convolution module of DenseGCN is combined to perform point cloud upsampling to restore high-precision three-dimensional point cloud data; this method can efficiently compress data while ensuring the recovery of high-precision point cloud data after transmission, and is suitable for application scenarios such as autonomous driving and drone collaborative perception.

[0119] Example

[0120] In order to verify the performance of the point cloud data fast transmission method based on Poisson disk downsampling and dynamic graph convolutional neural network (DGCNN) upsampling of the present invention, this embodiment uses the Semantic KITTI dataset for experimental comparison in the field of autonomous driving. The Semantic KITTI dataset is a large-scale point cloud dataset extracted from the LiDAR SLAM system. The point cloud data is dense and complex, and is suitable for evaluating point cloud compression, transmission, and reconstruction tasks. Due to the large amount of point cloud data, the efficiency of compression and transmission is crucial for real-time applications, so this dataset helps to accurately evaluate the actual performance of the point cloud transmission method.

[0121] In this embodiment, on the client side, i.e., the driving car side, the present invention is compared and tested by combining Poisson disk downsampling with several common point cloud upsampling methods. The same simulation transmission channel and edge computing end and service computing receiving end are used to achieve the goal of the same transmission quality as much as possible. The following combinations are compared:

[0122] Poisson disk downsampling and PUNet upsampling: Poisson disk sampling is combined to reduce the amount of point cloud data, and PUNet is used for upsampling and reconstruction.

[0123] Poisson disk downsampling and PUGAN upsampling: After Poisson disk downsampling, PUGAN is used to reconstruct the point cloud.

[0124] The present invention (Poisson disk downsampling and DGCNN upsampling): The method of the present invention combines Poisson disk downsampling with a dynamic graph convolutional neural network for point cloud reconstruction.

[0125] This example compares and evaluates different methods from multiple dimensions, including:

[0126] (1) Compression ratio: measures the ratio of the compressed volume of the point cloud to its original volume.

[0127] (2) Chamfer Distance: Use the common point cloud evaluation indicator chamfer distance to measure the point cloud quality. CD (Chamfer Distance) is a measurement method used to measure the similarity between the reconstructed point cloud and the real point cloud. Here, the chamfer distance between the point clouds after upsampling and before downsampling is measured.

[0128] (3) Transmission speed (ms / frame): Calculate the average transmission time of each frame of point cloud.

[0129] (4) Total encoding and decoding time (ms / frame): Calculate the total time required for data compression and decompression.

[0130] Table 1 Comparison of results between various transmission methods and the fast point cloud data transmission method in the present invention

[0131]

[0132] Table 1 shows the experimental results of different methods, and each indicator is based on the average results of multiple frames of point cloud in the Semantic KITTI dataset. Combined with Table 1, the experimental results show that the present invention significantly improves the transmission efficiency while maintaining the quality of the point cloud. Compared with other comparative methods, the compression ratio of the present invention method is the lowest (2.20), indicating that the details can be better preserved while reducing the amount of data. The chamfer distance (2.95) is also low, slightly better than PUNet and PUGAN upsampling, showing the advantage of the present invention in reconstruction accuracy. In terms of transmission speed (3.44 ms / frame), the present invention is significantly faster than other methods, indicating that it has significant advantages in point cloud transmission efficiency and is conducive to meeting real-time requirements. At the same time, the total encoding and decoding time (3.12ms / frame) is shorter, lower than other methods, further proving the advantages of the present invention method in data processing efficiency and its applicability to the field of autonomous driving.

[0133] In summary, the present invention achieves efficient transmission and reconstruction of point cloud data through the combination of low compression ratio and high transmission speed, and has excellent real-time performance and data compression capabilities.

[0134] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0135] Obviously, those skilled in the art can make various changes and modifications to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, if these modifications and variations of the embodiments of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A fast point cloud data transmission method, characterized in that: include: Collect raw point cloud data through sensors on the client side; The original point cloud data is sparsely processed by using the Poisson disk downsampling method based on the adaptive sampling strategy to generate sparse point cloud data. Transmit sparse point cloud data to the server via a low-latency communication protocol; Use a pre-trained dynamic graph convolutional neural network to upsample the sparse point cloud data and reconstruct the original point cloud data; The dynamic graph convolutional neural network includes an input layer, an edge feature construction layer, a convolution operation layer and an upsampling module, wherein: The input layer is used to input the sparse point cloud dataset , and the three-dimensional coordinates of each point ; The edge feature construction layer converts the sparse point cloud dataset Convert to graph structure ,in, is a vertex set, representing a sparse point cloud dataset The point in For edge sets, calculate the characteristics of the edges , the K nearest neighbor algorithm is used to construct the adjacency matrix of the graph; The convolution operation layer performs convolution operation on the edge set through dynamic graph convolution operation to update the features of each point; The upsampling module gradually converts the sparse point cloud data into a three-dimensional space by using an enhanced multi-scale graph convolutional network based on DenseGCN and node reordering and reconstruction. Upsampling to restore high-precision point cloud ; Specifically include: The features of the point cloud are extracted through three serial DenseGCN-based enhanced multi-scale graph convolutional network modules: ； in, To enhance the multi-scale graph convolutional network module The input features of To enhance the multi-scale graph convolutional network module The output features of It is an enhanced multi-scale graph convolutional network module The convolution kernel size, It is an enhanced multi-scale graph convolutional network module Expansion rate, It is an enhanced multi-scale graph convolutional network module The number of output channels; Transform node features from the original dimension Expand to , the expanded node features Rearrange The dimension characteristics become ,Right now: ； ； ； in, is the inverse of the downsampling ratio, is the number of nodes of sparse point cloud data, is the number of nodes of the original point cloud data, and is a learnable parameter, To reshape the function, is the convolution operation layer Node characteristics of the layer; Using multi-layer perceptron Point Cloud Reconstruct into three-dimensional space.

2. A fast point cloud data transmission method according to claim 1, characterized in that: The method of using the Poisson disk downsampling method based on the adaptive sampling strategy to perform sparse processing on the original point cloud data specifically includes: Calculate each point The local point cloud density; Adaptively update the sampling radius based on the local density; Initialize the sparse point cloud data set, select sampling points from the sparse point cloud data set, select new sampling points from the original point cloud data, and if the distance between each new sampling point and the already selected sampling point is not less than the sampling radius corresponding to the currently selected sampling point, add the new sampling point to the sparse point cloud data set, and generate a sparse point cloud data set through iteration.

3. A fast point cloud data transmission method according to claim 2, characterized in that: The sampling radius is: ； in, Yes Optimized sampling radius; Indicate point The local point cloud density; is a constant.

4. A fast point cloud data transmission method according to claim 2, characterized in that: The sparse point cloud dataset is: ； in, , , N is the total number of original point cloud data, is the total number of sparse point cloud data.

5. A fast point cloud data transmission method according to claim 1, characterized in that: The characteristics of the edge for: ； in, Indicate point The eigenvector of is initially the three-dimensional coordinates of the point, The nearest neighbor The characteristic vector of Represents the connection function.

6. A fast point cloud data transmission method according to claim 1, characterized in that: The output of the convolution operation layer is: ； in, It is Node feature representation of the layer; is the adjacency matrix of the graph, including self-connected edges; is the node degree matrix, with diagonal elements , are the rows and columns of the adjacency matrix respectively; is a trainable weight matrix; is the activation function.

7. A fast point cloud data transmission method according to claim 1, characterized in that: The dynamic graph convolutional neural network is optimized using a loss function, and the loss function is: ； in is a point in the original point cloud, is the reconstructed point.

8. A fast point cloud data transmission system for implementing the method described in any one of claims 1 to 7, characterized in that: include: The client collects raw point cloud data through sensors, uses the Poisson disk downsampling method based on the adaptive sampling strategy to perform sparse processing on the raw point cloud data, generates sparse point cloud data, and transmits the sparse point cloud data to the server through a low-latency communication protocol; On the server side, a pre-trained dynamic graph convolutional neural network is used to upsample the sparse point cloud data and reconstruct the original point cloud data.

Images