Point cloud feature extraction method and device based on graph convolutional neural network

Through the point cloud feature extraction method based on graph convolution neural network, combined with the fusion of local and global features, the problem of difficulty in learning local and global features in the existing technology is solved, and the accuracy and accuracy of three-dimensional semantic segmentation are improved , achieving fast and accurate three-dimensional point cloud semantic segmentation effect.

CN113902920BActive Publication Date: 2024-12-06CHINA UNIV OF GEOSCIENCES (WUHAN)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111136226.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-09-27
Publication Date
2024-12-06
Estimated Expiration
2041-09-27

AI Technical Summary

Technical Problem

It is difficult for the prior art to learn the local and global features of three-dimensional point clouds at the same time, resulting in low accuracy and accuracy of three-dimensional semantic segmentation.

Method used

Using a point cloud feature extraction method based on graph convolution neural network, local features are obtained by aggregating the spatial information and multi-dimensional features of point clouds, and graph convolution calculation is combined with random graphs to obtain global features, and finally local and global features are fused to improve Semantic segmentation effect.

Benefits of technology

It realizes better segmentation of three-dimensional point cloud data, improves the accuracy and accuracy of semantic segmentation, and can quickly and accurately perform three-dimensional point cloud semantic segmentation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113902920B_ABST
    Figure CN113902920B_ABST
Patent Text Reader

Abstract

The present invention discloses a point cloud feature extraction method and device based on a graph convolutional neural network, including: constructing a training data set and a test data set; building a point cloud feature extraction network based on a graph convolutional neural network; inputting the training data set into the point cloud feature extraction network, and using a K-nearest neighbor search algorithm to construct an adjacency graph for each point; calculating and obtaining the spatial geometric information and multidimensional features of each point, and fusing them to obtain the local features of each point; calculating and obtaining the global features of each point according to the adjacency graph and local features of each point; fusing the global features and local features to obtain the fused features; after iterative network training with the training data set and the fused features, a trained point cloud feature extraction network is obtained; inputting the test data set into the trained point cloud feature extraction network, and outputting the semantic prediction results of the point cloud features. The present invention can simultaneously learn local features and global features, thereby improving the accuracy of semantic prediction and classification of point clouds.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of three-dimensional point cloud processing, and more specifically, to a point cloud feature extraction method and device based on a graph convolutional neural network. Background Art

[0002] With the rapid development of 3D acquisition technology, 3D sensors are becoming more and more accessible and affordable, including various types of 3D scanners, LiDARs, and RGB-D cameras (such as Kinect, RealSense, and AppleDepth cameras). The 3D data collected by these sensors can provide rich geometry, shape, and scale information. 3D data complements 2D images and provides opportunities for better understanding of the environment around the machine. 3D data has many applications in different fields, including autonomous driving, robotics, remote sensing, medical, and design industries. 3D data can usually be represented in different formats, including depth images, point clouds, meshes, and volume grids. Point cloud representation, as a commonly used representation format, can maintain the original geometric information without any discretization in 3D space. Therefore, point cloud is the preferred representation for many scene understanding related applications (such as autonomous driving and robotics).

[0003] With the rise of deep learning, the research direction of 3D point cloud semantic segmentation has gradually moved towards deep learning. Due to its end-to-end characteristics, deep learning does not require manual feature screening and has become the most important research direction for 3D point cloud semantic segmentation. The applications of deep learning in 3D point cloud semantic segmentation mainly include: multi-view-based algorithms, voxel-based algorithms, and algorithms that directly process point clouds. Among them, the multi-view-based method represents the 3D data by a 2D image of multiple views, which can be processed based on a 2D convolutional neural network, but this will cause serious information loss. The voxel-based method can represent the point cloud as a regular 3D volume grid, so that the 3D convolutional neural network can be directly applied to the point cloud, but the voxel structure will take up a lot of storage space, which may lead to higher computing and memory requirements. The method of directly processing points generally establishes a feature extraction network, using a multi-layer perceptron and a simple maximum pooling layer to learn the features of each point, but since this can only learn the features of the point, the local features near the point and the global features of the point cloud as a whole are difficult to learn.

[0004] The above methods all have a common limitation, that is, they cannot learn both the local and global features of point clouds at the same time, so the precision and accuracy of semantic segmentation are not high. Summary of the invention

[0005] A major technical problem to be solved by the present invention is to provide a point cloud feature extraction network that can simultaneously learn the local features and global features of the point cloud, thereby improving the precision and accuracy of three-dimensional semantic segmentation.

[0006] In order to achieve the above purpose, the present invention provides a point cloud feature extraction method and device based on a graph convolutional neural network, which obtains local features of point clouds by aggregating spatial information and multidimensional features of point clouds, and performs graph convolution in combination with random graphs to calculate global features of point clouds, and aggregates local features and global features of point clouds to obtain fused features. By effectively learning fused features, the network can better segment three-dimensional point cloud data.

[0007] According to one aspect of the present invention, the present invention provides a point cloud feature extraction method based on a graph convolutional neural network, the method comprising the following steps:

[0008] Construct training and testing datasets;

[0009] Build a point cloud feature extraction network based on graph convolutional neural network;

[0010] Input the training data set into the point cloud feature extraction network based on graph convolutional neural network, and use K nearest neighbor search algorithm to construct an adjacency graph for each point;

[0011] Calculate and obtain spatial geometric information of each point and multidimensional features of each point, and fuse the spatial geometric information and the multidimensional features to obtain local features of each point;

[0012] Calculate and obtain the global feature of each point according to the adjacency graph and the local feature of each point;

[0013] Fusing the global feature and the local feature of each point to obtain a fused feature of each point;

[0014] Iteratively training the point cloud feature extraction network based on the graph convolutional neural network through the training data set and the fusion features of each point, and obtaining a trained point cloud feature extraction network after the training is completed;

[0015] The test data set is input into the trained point cloud feature extraction network, and the semantic prediction results of the point cloud features are output.

[0016] Furthermore, the steps of constructing a training data set and a test data set specifically include:

[0017] Get point cloud data;

[0018] The point cloud data is downsampled to obtain a downsampled sparse point cloud data set;

[0019] The downsampled sparse point cloud dataset is divided into a training dataset and a test dataset.

[0020] Furthermore, the downsampling operation is performed on the point cloud data by using a random downsampling method.

[0021] Furthermore, the step of constructing an adjacency graph for each point using the K nearest neighbor search algorithm specifically includes:

[0022] For each point in the point cloud input into the point cloud feature extraction network based on graph convolutional neural network, a K-nearest neighbor search algorithm is used to search for the point closest to the point, and the point is added as an edge to the adjacency graph to construct an adjacency graph for each point.

[0023] Furthermore, the step of calculating and obtaining the spatial geometric features of each point specifically includes:

[0024] For each point p in the point cloud input to the point cloud feature extraction network based on the graph convolutional neural network i , use the K nearest neighbor search algorithm to search for the k points closest to it

[0025] Calculate p separately i arrive The Euclidean distance RD and absolute distance AD ​​are calculated as follows:

[0026]

[0027]

[0028] Use shared MLP to transform point p i Coordinates and points The coordinates, Euclidean distance RD and absolute distance AD ​​are aggregated to obtain the spatial geometric features of each point.

[0029] Furthermore, the step of calculating and obtaining the multi-dimensional features of each point specifically includes:

[0030] Each point p in the point cloud is input into the point cloud feature extraction network based on the graph convolutional neural network. i , all have the original features f i , search for the k features closest to the point through the K nearest neighbor search algorithm

[0031] Calculate the original features and The matrix difference is to make a difference in each feature dimension to get the difference feature of each point;

[0032] The gap feature and the original feature are aggregated using a shared MLP to obtain a multi-dimensional feature of each point.

[0033] Furthermore, after the step of inputting the test data set into the trained point cloud feature extraction network and outputting the semantic prediction result of the point cloud feature, the method further includes:

[0034] The semantic prediction results of the point cloud features are evaluated, and the evaluation indicators include: precision P, recall rate R, F1-score and overall classification accuracy OA, which are defined as follows:

[0035] P=(TP) / (TP+FP)

[0036] R=(TP) / (TP+FN)

[0037] F1-score = 2 × [(P × R) / (P + R)]

[0038] oA=(TP+TN) / (TP+FN+FP+TN)

[0039] Among them, TP represents the number of target semantics that are correctly predicted and not unrecognized items, FP represents the number of samples whose original semantics are unrecognized but predicted incorrectly, FN represents the number of samples whose semantics are predicted to be unrecognized but originally have semantics, OA represents the ratio of the number of correctly classified samples to the total number of samples, and F1-score is an indicator used to measure the accuracy of the binary classification model, taking into account both precision P and recall rate R.

[0040] According to another aspect of the present invention, the present invention also provides a point cloud feature extraction device based on a graph convolutional neural network, comprising the following modules:

[0041] Dataset construction module, used to construct training dataset and test dataset;

[0042] Network building module, used to build a point cloud feature extraction network based on graph convolutional neural network;

[0043] An adjacency graph construction module, used to input the training data set into the point cloud feature extraction network based on the graph convolutional neural network, and use the K nearest neighbor search algorithm to construct an adjacency graph for each point;

[0044] A local feature extraction module is used to calculate and obtain the spatial geometric information of each point and the multidimensional features of each point, and fuse the spatial geometric information and the multidimensional features to obtain the local features of each point;

[0045] A global feature extraction module, used for calculating and obtaining the global feature of each point according to the adjacency graph and the local feature of each point;

[0046] A fusion feature calculation module, used to fuse the global feature and the local feature of each point to obtain a fusion feature of each point;

[0047] A network training module, used for iteratively training the point cloud feature extraction network based on the graph convolutional neural network through the training data set and the fusion features of each point, and obtaining a trained point cloud feature extraction network after the training is completed;

[0048] The semantic prediction module is used to input the test data set into the trained point cloud feature extraction network and output the semantic prediction results of the point cloud features.

[0049] Furthermore, the local feature extraction module specifically includes:

[0050] The spatial geometric information extraction module is used to calculate and obtain the spatial geometric information of each point;

[0051] Multidimensional feature extraction module, used to calculate and obtain the multidimensional features of each point;

[0052] The fusion module is used to fuse the spatial geometric information with the multi-dimensional features to obtain the local features of each point.

[0053] Furthermore, the data set construction module specifically includes:

[0054] Point cloud data acquisition module, used to acquire point cloud data;

[0055] A downsampling module, used for downsampling the point cloud data to obtain a sparse point cloud data set after downsampling;

[0056] The data set division module is used to divide the downsampled sparse point cloud data set into a training data set and a test data set.

[0057] Furthermore, the point cloud local feature extraction device also includes:

[0058] The accuracy evaluation module is used to evaluate the semantic prediction results of the point cloud features, and the evaluation indicators include: precision P, recall rate R, F1-score and overall classification accuracy OA.

[0059] The beneficial effects brought by the technical solution provided by the present invention are:

[0060] 1. The present invention proposes a point cloud feature extraction network based on graph convolutional neural network, which can quickly and accurately perform three-dimensional point cloud semantic segmentation;

[0061] 2. The present invention proposes a local feature extraction module, which uses the K nearest neighbor search algorithm to search for the neighboring points and the nearest features of each point, calculates the spatial information and multidimensional features of each point, and aggregates the spatial information and multidimensional features of each point to obtain the local features of each point;

[0062] 3. The present invention proposes a global feature extraction module and a fusion feature calculation module, which are used to calculate the global features of the point cloud and aggregate the global features and local features to obtain the fusion features of the point cloud. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0064] Figure 1 It is a technical flow chart of a point cloud feature extraction method based on a graph convolutional neural network in an embodiment of the present invention;

[0065] Figure 2 A technical flow chart of spatial information calculation in an embodiment of the present invention;

[0066] Figure 3 A technical flow chart of multi-dimensional feature calculation in an embodiment of the present invention;

[0067] Figure 4 This is a flow chart of a point cloud feature extraction model training based on a graph convolutional neural network in an embodiment of the present invention;

[0068] Figure 5 A schematic diagram of a data set in an embodiment of the present invention;

[0069] Figure 6 It is a structural diagram of a point cloud feature extraction device based on a graph convolutional neural network in an embodiment of the present invention. DETAILED DESCRIPTION

[0070] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0071] Please refer to Figure 1 , Figure 1 The technical flow chart of a point cloud feature extraction method based on a graph convolutional neural network in an embodiment of the present invention includes the following steps:

[0072] S1, build training data set and test data set;

[0073] S1 specifically includes:

[0074] Get point cloud data;

[0075] The point cloud data is downsampled to obtain a downsampled sparse point cloud data set;

[0076] The downsampled sparse point cloud dataset is divided into a training dataset and a test dataset.

[0077] In this embodiment, please refer to Figure 5The data set is the point cloud data set S3DIS published by Stanford University, where: Figure 5 (a)- Figure 5 (f) are schematic diagrams of area_1 data, area_2 data, area_3 data, area_4 data, area_5 data and area_6 data in the point cloud dataset S3DIS; area_1 data, area_2 data, area_3 data, area_4 data and area_6 data are used as training data, and area_5 data is used as test data. The random downsampling method is used, and the sampling rate is set to 1%.

[0078] S2. Build a point cloud feature extraction network based on graph convolutional neural network;

[0079] S3, inputting the training data set into the point cloud feature extraction network based on graph convolutional neural network, and constructing an adjacency graph for each point using K nearest neighbor search algorithm;

[0080] In this embodiment, the number of neighbors k is set to 20.

[0081] S4, calculating and obtaining the spatial geometric information of each point and the multidimensional features of each point, and fusing the spatial geometric information and the multidimensional features to obtain the local features of each point;

[0082] S5. Calculate and obtain the global feature of each point according to the adjacency graph and the local feature of each point;

[0083] S6, fusing the global feature and the local feature of each point to obtain a fused feature of each point;

[0084] S7, iteratively training the point cloud feature extraction network based on the graph convolutional neural network through the training data set and the fusion features of each point, and obtaining a trained point cloud feature extraction network after the training is completed;

[0085] S8. Input the test data set into the trained point cloud feature extraction network, and output the semantic prediction results of the point cloud features.

[0086] In this embodiment, there are 13 categories of point cloud classification: ceiling, floor, wall, beam, column, window, door, table, chair, sofa, bookshelf, bed and sundries.

[0087] Please refer to Figure 2 , Figure 2 This is a technical flow chart of spatial information calculation in an embodiment of the present invention. The steps of calculating and obtaining the spatial geometric features of each point specifically include:

[0088] S41, for each point p in the point cloud input to the point cloud feature extraction network based on the graph convolutional neural network i , use the K nearest neighbor search algorithm to search for the k points closest to it

[0089] S42, calculate p respectively i arrive The Euclidean distance RD and absolute distance AD ​​are calculated as follows:

[0090]

[0091]

[0092] S43, use shared MLP to convert point p i Coordinates and points The coordinates, Euclidean distance RD and absolute distance AD ​​are aggregated to obtain the spatial geometric features of each point.

[0093] Please refer to Figure 3 , Figure 3 This is a technical flow chart of multi-dimensional feature calculation in an embodiment of the present invention. The steps of calculating and obtaining the multi-dimensional features of each point specifically include:

[0094] S51, input each point p of the point cloud in the point cloud feature extraction network based on the graph convolutional neural network i , all have the original features f i , search for the k features closest to the point through the K nearest neighbor search algorithm

[0095] S52, calculate the original features and The matrix difference is to make a difference in each feature dimension to get the difference feature of each point;

[0096] S53: Use the shared MLP to aggregate the gap feature and the original feature to obtain a multi-dimensional feature of each point.

[0097] refer to Figure 4 , Figure 4 The present invention is a flowchart of a point cloud feature extraction model training based on a graph convolutional neural network in an embodiment of the present invention; the flowchart includes the following steps:

[0098] S71, inputting the downsampled point cloud data set (training data set) as an input point cloud into a point cloud feature extraction model based on a graph convolutional neural network;

[0099] S72, obtaining an adjacency graph of the point cloud according to the method in S3, and obtaining multi-dimensional features and spatial geometric information of the point cloud according to the method in S4;

[0100] S73, inputting the local features fused with the multi-dimensional features of the point cloud and the spatial geometric information into the encoding layer composed of the graph convolution calculation module. In this embodiment, two graph convolution layers are used, and the convolution kernel sizes are: 4096×64×20 and 4096×256×20, respectively. The extracted global features are input into the 1×256×1 pooling layer;

[0101] S74, aggregate the global features and the local features to obtain fused features, and input them into the decoding layer composed of a multi-layer perceptron (MLP). In this embodiment, a three-layer multi-layer perceptron is used, and the sizes are 4096×1024×1, 4096×512×1, and 4096×256×1, respectively, to obtain the feature data after decoding.

[0102] S75. Input the decoded feature data into the 4096×13×1 softmax layer for classification, perform dimensionality reduction mapping, and output the classification result of the point cloud.

[0103] As shown in Table 1, the data in the table are the results obtained after the model is cross-validated on the S3DIS dataset. It can be seen that the overall classification accuracy OA (Overall Accuracy) and mean intersection over Union (mIoU) of this embodiment are significantly better than the pointnet method.

[0104] Table 1 Accuracy comparison between the proposed method and the pointnet method

[0105] method OA (%) mIoU(%) PointNet 78.6 47.6 Method of the present invention 88.1 63.7

[0106] Please refer to Figure 6 , Figure 6 : is a structural diagram of a point cloud feature extraction device based on a graph convolutional neural network in an embodiment of the present invention, the device includes:

[0107] Dataset construction module 1, used to construct training data set and test data set;

[0108] Network building module 2, used to build a point cloud feature extraction network based on graph convolutional neural network;

[0109] The adjacency graph construction module 3 is used to input the training data set into the point cloud feature extraction network based on the graph convolutional neural network, and use the K nearest neighbor search algorithm to construct an adjacency graph for each point;

[0110] The local feature extraction module 4 is used to calculate and obtain the spatial geometric information of each point and the multi-dimensional features of each point, and fuse the spatial geometric information and the multi-dimensional features to obtain the local features of each point;

[0111] A global feature extraction module 5, used for calculating and obtaining the global feature of each point according to the adjacency graph and the local feature of each point;

[0112] A fusion feature calculation module 6 is used to fuse the global feature and the local feature of each point to obtain a fusion feature of each point;

[0113] A network training module 7 is used to iteratively train the point cloud feature extraction network based on the graph convolutional neural network through the training data set and the fusion feature of each point, and after the training is completed, a trained point cloud feature extraction network is obtained;

[0114] The semantic prediction module 8 is used to input the test data set into the trained point cloud feature extraction network and output the semantic prediction results of the point cloud features.

[0115] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0116] The serial numbers of the embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In a unit claim that lists several means, several of these means may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order and these words may be interpreted as identifiers.

[0117] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A point cloud feature extraction method based on graph convolutional neural network, characterized in that: The following steps are involved: Construct training and testing datasets; Build a point cloud feature extraction network based on graph convolutional neural network; Input the training data set into the point cloud feature extraction network based on graph convolutional neural network, and use K nearest neighbor search algorithm to construct an adjacency graph for each point; Calculate and obtain spatial geometric information of each point and multidimensional features of each point, and fuse the spatial geometric information and the multidimensional features to obtain local features of each point; Calculate and obtain the global feature of each point according to the adjacency graph and the local feature of each point; Fusing the global feature and the local feature of each point to obtain a fused feature of each point; Iteratively training the point cloud feature extraction network based on the graph convolutional neural network through the training data set and the fusion features of each point, and obtaining a trained point cloud feature extraction network after the training is completed; Inputting the test data set into the trained point cloud feature extraction network, and outputting semantic prediction results of point cloud features; The step of calculating and obtaining the spatial geometric features of each point specifically includes: For each point p in the point cloud input to the point cloud feature extraction network based on the graph convolutional neural network i , use the K nearest neighbor search algorithm to search for the k points closest to it Calculate the point p separately i To point The Euclidean distance RD and absolute distance AD ​​are calculated as follows: Use shared MLP to transform point p i Coordinates and points The coordinates, Euclidean distance RD and absolute distance AD ​​are aggregated to obtain the spatial geometric features of each point; The step of calculating and obtaining the multi-dimensional features of each point specifically includes: Each point p in the point cloud is input into the point cloud feature extraction network based on the graph convolutional neural network. i , all have the original features f i , search for the k features closest to the point through the K nearest neighbor search algorithm Calculate the original features and The matrix difference is to make a difference in each feature dimension to get the difference feature of each point; The gap feature and the original feature are aggregated using a shared MLP to obtain a multi-dimensional feature of each point.

2. The point cloud feature extraction method based on graph convolutional neural network according to claim 1, characterized in that: The steps of constructing the training data set and the test data set specifically include: Get point cloud data; The point cloud data is downsampled to obtain a downsampled sparse point cloud data set; The downsampled sparse point cloud dataset is divided into a training dataset and a test dataset.

3. The point cloud feature extraction method based on graph convolutional neural network according to claim 1, characterized in that: The step of using the K nearest neighbor search algorithm to construct an adjacency graph for each point specifically includes: For each point in the point cloud input into the point cloud feature extraction network based on graph convolutional neural network, a K nearest neighbor search algorithm is used to search for the point closest to the point, and the point is added as an edge to the adjacency graph to construct an adjacency graph for each point.

4. The point cloud feature extraction method based on graph convolutional neural network according to claim 1, characterized in that: After the step of inputting the test data set into the trained point cloud feature extraction network and outputting the semantic prediction result of the point cloud feature, the method further includes: The semantic prediction results of the point cloud features are evaluated, and the evaluation indicators include: precision P, recall rate R, F1-score and overall classification accuracy OA, which are defined as follows: P=(TP) / (TP+FP) R=(TP) / (TP+FN) F1-score = 2 × [(P × R) / (P + R)] OA=(TP+TN) / (TP+FN+FP+TN) Among them, TP represents the number of target semantics that are correctly predicted and not unrecognized items, FP represents the number of samples whose original semantics are unrecognized but predicted incorrectly, FN represents the number of samples whose semantics are predicted to be unrecognized but originally have semantics, OA represents the ratio of the number of correctly classified samples to the total number of samples, and F1-score is an indicator used to measure the accuracy of the binary classification model, taking into account both precision P and recall rate R.

5. A point cloud feature extraction device based on graph convolutional neural network, characterized in that: Includes the following modules: Dataset construction module, used to construct training dataset and test dataset; Network building module, used to build a point cloud feature extraction network based on graph convolutional neural network; An adjacency graph construction module, used to input the training data set into the point cloud feature extraction network based on the graph convolutional neural network, and use the K nearest neighbor search algorithm to construct an adjacency graph for each point; A local feature extraction module is used to calculate and obtain the spatial geometric information of each point and the multidimensional features of each point, and fuse the spatial geometric information and the multidimensional features to obtain the local features of each point; A global feature extraction module, used for calculating and obtaining the global feature of each point according to the adjacency graph and the local feature of each point; A fusion feature calculation module, used to fuse the global feature and the local feature of each point to obtain a fusion feature of each point; A network training module, used for iteratively training the point cloud feature extraction network based on the graph convolutional neural network through the training data set and the fusion features of each point, and obtaining a trained point cloud feature extraction network after the training is completed; A semantic prediction module, used for inputting the test data set into the trained point cloud feature extraction network and outputting semantic prediction results of point cloud features; The step of calculating and obtaining the spatial geometric features of each point specifically includes: For each point p in the point cloud input to the point cloud feature extraction network based on the graph convolutional neural network i , use the K nearest neighbor search algorithm to search for the k points closest to it Calculate the point p separately I To point The Euclidean distance RD and absolute distance AD ​​are calculated as follows: Use shared MLP to transform point p i Coordinates and points The coordinates, Euclidean distance RD and absolute distance AD ​​are aggregated to obtain the spatial geometric features of each point; The step of calculating and obtaining the multi-dimensional features of each point specifically includes: Each point p in the point cloud is input into the point cloud feature extraction network based on the graph convolutional neural network. I , all have the original features f i , search for the k features closest to the point through the K nearest neighbor search algorithm Calculate the original features and The matrix difference is to make a difference in each feature dimension to get the difference feature of each point; The gap feature and the original feature are aggregated using a shared MLP to obtain a multi-dimensional feature of each point.

6. The point cloud feature extraction device based on graph convolutional neural network according to claim 5, characterized in that: The local feature extraction module specifically includes: The spatial geometric information extraction module is used to calculate and obtain the spatial geometric information of each point; Multidimensional feature extraction module, used to calculate and obtain the multidimensional features of each point; The fusion module is used to fuse the spatial geometric information with the multi-dimensional features to obtain the local features of each point.

7. The point cloud feature extraction device based on graph convolutional neural network according to claim 5, characterized in that: The data set construction module specifically includes: Point cloud data acquisition module, used to acquire point cloud data; A downsampling module, used for downsampling the point cloud data to obtain a sparse point cloud data set after downsampling; The data set division module is used to divide the downsampled sparse point cloud data set into a training data set and a test data set.

8. The point cloud feature extraction device based on graph convolutional neural network as claimed in claim 5, characterized in that: The point cloud local feature extraction device also includes: The accuracy evaluation module is used to evaluate the semantic prediction results of the point cloud features, and the evaluation indicators include: precision P, recall rate R, F1-score and overall classification accuracy OA.