Point cloud classification segmentation network for three-dimensional point cloud classification segmentation and classification segmentation method thereof

Abstract

The invention provides a point cloud classification segmentation network used for three-dimensional point cloud classification segmentation and a classification segmentation method thereof, and the method comprises the steps: firstly, constructing a Coordinate Transformation module, enabling the network to use the same feature extractor to effectively process all irregular point cloud data through the Coordinate Transformation module, and then employing a ResMLP module to learn complex nonlinear features, and meanwhile, static and dynamic context information is fused on the basis of a traditional self-attention mechanism through a Contextual Transform module, so that global features are effectively obtained, the global modeling capability of the model is supplemented, and the modeling efficiency is improved. And fusing the two features to obtain a point cloud classification and segmentation network for three-dimensional point cloud classification and segmentation, and carrying out a point cloud classification and segmentation task, thereby improving the precision and robustness of point cloud classification and segmentation.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and lidar point cloud processing, and more specifically to a point cloud classification and segmentation network and a classification and segmentation method thereof for three-dimensional point cloud classification and segmentation. Background Art

[0002] In recent years, the rapid development and widespread application of LiDAR technology has fueled a surge in research on 3D point cloud data processing. As a crucial information carrier in areas such as autonomous driving, 3D scene reconstruction, and smart cities, point cloud data has become a key component in numerous technological applications. With the continuous advancement of deep learning technology, researchers have gradually applied it to point cloud data processing, encompassing tasks such as data preprocessing, classification, and segmentation. However, the disordered, sparse, and unstructured nature of point cloud data poses numerous challenges to tasks such as classification and segmentation.

[0003] Currently, deep learning-based point cloud processing methods can be roughly divided into three categories: projection-based, voxel-based, and point-based. Projection-based methods project point cloud data in different directions to generate a two-dimensional image, which is then extracted using a two-dimensional convolutional neural network (CNN). Although simple to implement, this method may lose some spatial information during the projection process, resulting in reduced performance when processing complex three-dimensional structures. Voxel-based methods divide the point cloud into uniform three-dimensional grid cells (i.e., voxels) and extract features using a three-dimensional convolutional neural network (3D CNN). This method can directly process three-dimensional data and has strong expressive power, but the need to process a large number of three-dimensional voxels leads to significant computational overhead and memory consumption, limiting its application in large-scale datasets. Point-based methods are more direct. Instead of discretizing the point cloud, they operate directly on the raw point cloud and learn features using a network architecture designed to handle the disorder of the point cloud. This method can maximize the preservation of the original information of the point cloud data and has advantages when processing complex geometric shapes.

[0004] However, point cloud data is typically unevenly distributed, and point density can vary significantly due to factors such as sensor resolution, scanning angle, or occlusion. Existing network models have limited robustness when processing non-uniform point cloud data, and are prone to feature extraction imbalances. They also perform poorly when the actual collected point cloud data may contain a lot of noise or incomplete areas, failing to handle these situations well. Therefore, improving the performance of point cloud processing networks and maintaining their robustness in the presence of complex noise interference has become an important research direction. Summary of the Invention

[0005] In order to solve the problems of poor performance of existing point cloud processing networks and limited robustness of existing network models in processing non-uniform point cloud data, which easily leads to imbalanced feature extraction, the present invention proposes a point cloud classification and segmentation network and its classification and segmentation method for three-dimensional point cloud classification and segmentation.

[0006] In order to achieve the above technical effects, the technical solutions of the present invention are as follows:

[0007] A point cloud classification and segmentation network for three-dimensional point cloud classification and segmentation, comprising:

[0008] The first Contextual Transformer module, the second Contextual Transformer module, the third Contextual Transformer module, the fourth Contextual Transformer module and the first Concat module are connected in sequence;

[0009] A first Coordinate Transformation module, a first ResMLP module, a second Coordinate Transformation module, a second ResMLP module, a third Coordinate Transformation module, a third ResMLP module, a fourth Coordinate Transformation module, and a fourth ResMLP module connected in sequence;

[0010] The second Concat module and the first MaxPool module are connected in sequence, the output end of the first MaxPool module is connected to the input end of the first Classification module, and the output end of the first MaxPool module is also connected to the input end of the first Segmentation module;

[0011] The output end of the first input module is connected to the input end of the first Contextual Transformer module, and the output end of the first input module is also connected to the input end of the first Coordinate Transformation module; the input end of the second Concat module is connected to the output end of the first Concat module, and the input end of the second Concat module is also connected to the output end of the fourth ResMLP module.

[0012] A three-dimensional point cloud classification and segmentation method includes the following steps:

[0013] Step 1: Input the point cloud feature vector into the first Contextual Transformer module for feature extraction;

[0014] Step 2: The feature vector obtained in step 1 is input into the second Contextual Transformer module for feature extraction. After passing through four Contextual Transformer modules in the same way, the output of each module is fused through the first Concat module to obtain the final global feature and output it;

[0015] Step 3: Input the point cloud feature vector into the first Coordinate Transformation module so that the irregular geometric structure of the local area of the point cloud can be processed by a unified feature extractor;

[0016] Step 4: Input the coordinate-transformed feature vector into the first ResMLP module to extract local features;

[0017] Step 5: Input the features extracted in step 4 into the second Coordinate Transformation module, and then input them into the second ResMLP module. After passing through four Coordinate Transformation modules and four ResMLP modules in the same way, the final local features are obtained and output;

[0018] Step 6: The outputs of step 5 and step 2 are fused through the second Concat module and the final extracted features are obtained through maximum pooling;

[0019] Step 7: Input the features finally extracted in step 6 into the first Classification module and the first Segmentation module respectively to obtain the final classification and segmentation results.

[0020] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0021] The present invention proposes a point cloud classification and segmentation network and a classification and segmentation method for three-dimensional point cloud classification and segmentation. By combining the advantages of the residual multi-layer perceptron (ResMLP) and the contextual transformer (Contextual Transformer), the accuracy and efficiency of three-dimensional point cloud data processing are improved. The residual structure used by ResMLP allows the network to learn more complex nonlinear features and can better learn local features. At the same time, the coordinate transformation of irregular point clouds is performed so that the same ResMLP can stably extract features. The Contextual Transformer module can effectively enhance the global modeling capability and solve the problem of the network ignoring global features, thereby improving the accuracy and robustness of point cloud classification and segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1This is a diagram showing the overall structure of a point cloud classification and segmentation network according to an embodiment of the present invention.

[0023] Figure 2 This is a flow chart of a three-dimensional point cloud classification and segmentation method according to an embodiment of the present invention.

[0024] Figure 3 This is a diagram of the Contextual Transformer module structure shown in an embodiment of the present invention.

[0025] Figure 4 This is a structural diagram of the Point-ResMLP module according to an embodiment of the present invention.

[0026] Figure 5 A comparison diagram of the point cloud component segmentation effects on the ShapeNet dataset shown in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, like numbers in different figures represent like or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all possible embodiments consistent with the present invention. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present invention, as detailed in the appended claims.

[0028] The terms used in this invention are for the purpose of describing specific embodiments only and are not intended to limit the invention. The singular forms "a," "the," and "the" used in this invention and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0029] It should be understood that although the terms "first," "second," "third," etc. may be used in the present invention to describe various information, such information should not be limited to these terms. These terms are merely used to distinguish information of the same type from one another. For example, first information may also be referred to as second information, and similarly, second information may also be referred to as first information, without departing from the scope of the present invention. Depending on the context, the term "if" as used herein may be interpreted as "when," "when," or "in response to determining."

[0030] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] Example 1

[0032] This embodiment proposes a point cloud classification and segmentation network for three-dimensional point cloud classification and segmentation, and its structural diagram is shown in FIG. Figure 1 As shown, including:

[0033] The first Contextual Transformer module, the second Contextual Transformer module, the third Contextual Transformer module, the fourth Contextual Transformer module and the first Concat module are connected in sequence;

[0034] A first Coordinate Transformation module, a first ResMLP module, a second Coordinate Transformation module, a second ResMLP module, a third Coordinate Transformation module, a third ResMLP module, a fourth Coordinate Transformation module, and a fourth ResMLP module connected in sequence;

[0035] The second Concat module and the first MaxPool module are connected in sequence, the output end of the first MaxPool module is connected to the input end of the first Classification module, and the output end of the first MaxPool module is also connected to the input end of the first Segmentation module;

[0036] The output end of the first input module is connected to the input end of the first Contextual Transformer module, and the output end of the first input module is also connected to the input end of the first Coordinate Transformation module; the input end of the second Concat module is connected to the output end of the first Concat module, and the input end of the second Concat module is also connected to the output end of the fourth ResMLP module.

[0037] In an optional embodiment, any one of the first ResMLP module, the second ResMLP module, the third ResMLP module, and the fourth ResMLP module includes:

[0038] The first MLP connected sequentially pre module, the first maximum pooling layer module and the first MLP pos module.

[0039] In an alternative embodiment, the first MLP pre Modules include:

[0040] The first Point-ResMLP module, the second Point-ResMLP module and the third Point-ResMLP module are connected in sequence.

[0041] In an alternative embodiment, the first MLP pos Modules include:

[0042] The fourth Point-ResMLP module, the fifth Point-ResMLP module and the sixth Point-ResMLP module are connected in sequence.

[0043] In an optional embodiment, any one of the first Point-ResMLP module, the second Point-ResMLP module, the third Point-ResMLP module, the fourth Point-ResMLP module, the fifth Point-ResMLP module, and the sixth Point-ResMLP module includes:

[0044] The first Conv module, the first BactchNorm module, the first ReLU module, the second Conv module, the second BactchNorm module and the second ReLU module are connected in sequence.

[0045] In this embodiment, the classification and segmentation performance of the point cloud classification and segmentation network for three-dimensional point cloud classification and segmentation is superior to existing classical methods and models. It can accurately and efficiently process point cloud data and can also effectively cope with noise interference in complex real-world situations. It is suitable for application scenarios such as autonomous driving, virtual reality, and medical imaging that require high-precision point cloud data processing, and has broad application prospects and practical value. It first constructs a Coordinate Transformation module, which enables the network to effectively process all irregular point cloud data using the same feature extractor. It then uses ResMLP to learn complex nonlinear features to obtain better local features. At the same time, the ContextualTransformer module fuses static and dynamic context information on the basis of the traditional self-attention mechanism to effectively obtain global features and supplement the global modeling capabilities of the model. These two parts of features are then fused to obtain the point cloud classification and segmentation network for three-dimensional point cloud classification and segmentation in the present invention to perform point cloud classification and segmentation tasks, thereby improving the accuracy and robustness of point cloud classification and segmentation.

[0046] Example 2

[0047] Based on Example 1, a 3D point cloud classification and segmentation method is proposed, such as Figure 2 Shown is a flow chart of this embodiment.

[0048] Step 1: Input the point cloud feature vector into the first Contextual Transformer module for feature extraction;

[0049] Step 2: The feature vector obtained in step 1 is input into the second Contextual Transformer module for feature extraction. After passing through four Contextual Transformer modules in the same way, the output of each module is fused through the first Concat module to obtain the final global feature and output it;

[0050] Step 3: Input the point cloud feature vector into the first Coordinate Transformation module so that the irregular geometric structure of the local area of the point cloud can be processed by a unified feature extractor;

[0051] Step 4: Input the coordinate-transformed feature vector into the first ResMLP module to extract local features;

[0052] Step 5: Input the features extracted in step 4 into the second Coordinate Transformation module, and then input them into the second ResMLP module. After passing through four Coordinate Transformation modules and four ResMLP modules in the same way, the final local features are obtained and output;

[0053] Step 6: The outputs of step 5 and step 2 are fused through the second Concat module and the final extracted features are obtained through maximum pooling;

[0054] Step 7: Input the features finally extracted in step 6 into the first Classification module and the first Segmentation module respectively to obtain the final classification and segmentation results.

[0055] Furthermore, the point cloud feature vector of size n×f is input into a Contextual Transformer module for feature extraction. The structure of the Contextual Transformer module is shown in the figure below. Figure 3 shown.

[0056] The obtained feature vector is input into the next Contextual Transformer module for feature extraction. In the same way, after passing through four Contextual Transformer modules, the output of each module is fused through the concat operation to obtain the final global feature N c .

[0057] Furthermore, the output dimensions of the four Contextual Transformer modules used are [64, 64, 128, 256] respectively.

[0058] The Contextual Transformer module integrates contextual information mining and self-attention learning into the same architecture, which can effectively utilize the contextual information between adjacent keys and significantly improve the global feature expression ability of the model.

[0059] In an optional embodiment, for the input point cloud feature map X∈R B×D×N , the key, query and value are defined as K = X, Q = X and V = XW respectively V Unlike the classic self-attention mechanism that encodes each key through a 1×1 convolution, the Contextual Transformer module first uses a k×k group of convolutions on all adjacent keys in a k×k grid in space to encode the context of each key representation, which is expressed as:

[0060] K 1 =Conv k×k (K)

[0061] The learned context key K 1 ∈R B×D×N Reflects the static context information between local adjacent keys, and K 1 As a static context representation of the input X, it is then 1 and Q splicing, through two consecutive 1×1 convolutions (Conv 1×1 ) to get the attention matrix:

[0062] W θ =W δ =Conv 1×1

[0063] A=[K 1 ,Q]W θ W δ

[0064] Among them, K 1 represents the static context, Q represents the query, and A represents the context attention matrix;

[0065] According to the context attention matrix A, the attention feature map K is calculated by aggregating all V 2 :

[0066] K 2 =V*A

[0067] Among them, K 2 Represents dynamic context;

[0068] Since the attention feature map K 2 It captures the dynamic feature interactions between inputs and converts K 2 is named as the dynamic context representation of the input. Therefore, the final output of the Contextual Transformer module is the static context K 1 and dynamic context K 2 Obtained by fusion.

[0069] The point cloud feature vector of size n×f is input into the coordinate transformation module so that the irregular geometric structure of the local area of the point cloud can be processed by a unified feature extractor.

[0070] In an optional embodiment, the coordinate transformation expression of the Coordinate Transformation module is:

[0071]

[0072] Among them, f i,j Represents the point f obtained by the K-nearest neighbor algorithm i The K local neighboring points of the local area are represented by ε, which is a very small value used to keep the denominator non-zero. α and β represent learnable parameters. σ represents the standard deviation of the point features in the local area, which is used to measure the degree of variation or dispersion between the features of each point in the local area and the features of the local center point.

[0073] Furthermore, the coordinate transformation module uses σ to scale the features of points in the local neighborhood and maps the point features in the local neighborhood to a space with a uniform scale.

[0074] The feature vector after coordinate transformation is input into ResMLP to extract local features.

[0075] The extracted local features are input into the next coordinate transformation module, and then into the second ResMLP. After four coordinate transformation modules and four ResMLPs in the same way, the final local feature output N is obtained. r .

[0076] In an optional embodiment, the ResMLP module is represented as a whole as follows:

[0077] p i =MLP pos (MaxPool(MLP pre (f i,j ),|j=1,...,K))

[0078] Among them, pi ResMLP module, MaxPool represents the maximum pooling operation, f i,j Represents the point f obtained by the K-nearest neighbor algorithm i The K local neighbors of .

[0079] Furthermore, ResMLP consists of six point-residual multi-layer perceptrons (Point-ResMLP) and a maximum pooling layer, where the three Point-ResMLP layers before the maximum pooling layer are used to learn shared weight parameters in local areas, denoted as MLP. pre , the three Point-ResMLPs after the maximum pooling are used to extract deep aggregation features, denoted as MLP pos , the Point-ResMLP module structure diagram is as follows Figure 4 shown.

[0080] Furthermore, each Point-ResMLP consists of multiple convolutional layers (Conv), regularization layers (BN), activation function layers (LeakyReLU) and residual connections, and its structure can be expressed as the following formula:

[0081]

[0082] The definitions of the regularization layer and activation function layer are as follows:

[0083] LeakyReLu(X)=max(0,X)+leaky×min(0,X)

[0084]

[0085] Among them, X represents input data, leaky represents a constant, γ and β represent learnable parameters, E represents mean, and Var represents variance.

[0086] In this embodiment, ResMLP is used as the main feature extractor to obtain local features, and the point cloud is adjusted through the coordinate transformation module before extraction, so that the network can use the same feature extractor to effectively process all irregular point cloud data; Contextual Transformer is used to supplement the missing global feature information, so that the model can improve accuracy while maintaining high robustness.

[0087] The global feature N c With local features N r The output of is fused through concat and the final extracted feature N is obtained through maximum pooling.

[0088] In an optional embodiment, the calculation expression of step six is:

[0089] N=MaxPool(concat(N r ,N c )

[0090] Among them, N represents the final extracted features, N c represents the global feature, N r Represents the fourth local feature.

[0091] The obtained features N are input into the classification and segmentation task head respectively to obtain the final classification and segmentation results.

[0092] Furthermore, the classification task head classificate consists of two linear layers, and its structure is shown in the following formula:

[0093]

[0094] Among them, the output dimensions of Linear1, Linear2, and Linear3 are [512, 256, 40], and the dimensions of BN1 and BN2 are [512, 256] respectively.

[0095] The segmentation task header structure is as follows:

[0096]

[0097] Among them, x i is the input vector, and the output dimensions of Conv1 and Conv2 are [128,50] respectively.

[0098] In this embodiment, a point cloud classification and segmentation method based on a residual multilayer perceptron and a contextual transformer network (RCT-Net) is described. The residual structure used by the residual multilayer perceptron (ResMLP) allows the network to learn more complex nonlinear features, which can better learn local features. At the same time, the coordinate transformation of irregular point clouds allows them to be stably extracted by the same residual multilayer perceptron. The contextual transformer module can effectively enhance the global modeling capability and solve the problem of the network ignoring global features, thereby improving the accuracy and robustness of point cloud classification and segmentation.

[0099] Example 3

[0100] This example compares the classification, segmentation performance, and robustness of this method with some existing networks. The comparison results are shown in Tables 1-5 below:

[0101] Table 1: Performance comparison of different methods on the point cloud object classification test dataset ModelNet40

[0102]

[0103]

[0104] Table 2: Performance comparison of different methods on different subsets of the point cloud classification dataset ScanObjectNN

[0105]

[0106]

[0107] Table 3: Performance comparison of different methods on the point cloud object part segmentation test dataset ShapeNet

[0108] Network Model Average intersection-merge ratio (%) Pointnet 83.7 Pointnet++ 85.1 DGCNN 85.2 DTNet 85.6 3DGCN 85.1 GCN3D 85.2 GateNet 85.6 RCT-Net 86.0

[0109] Table 4: Comparison of model robustness under different number of point clouds

[0110]

[0111] Table 5: Comparison of model robustness under noise influence

[0112]

[0113]

[0114] As can be seen from the above three tables, this method can effectively improve the results of point cloud classification and segmentation tasks. Its performance on ModelNet40, ScanObjectNN classification task dataset, and ShapeNet segmentation task dataset is better than other compared models, including the commonly used classic models pointnet, pointnet++, and DGCNN. In the robustness test, this method can maintain the highest robustness under the influence of reduced point cloud number and noise. The above experiments prove the effectiveness of this method. Figure 5 Shown is a comparison of the point cloud component segmentation effects on the ShapeNet dataset.

[0115] Each embodiment of the present invention is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple. For the relevant parts, refer to the partial description of the method embodiment. The device embodiment described above is merely exemplary. The modules described as separate components may or may not be physically separated. When implementing the scheme of the present invention, the functions of each module can be implemented in the same one or more software and / or hardware. It is also possible to select some or all of the modules according to actual needs to achieve the purpose of the scheme of this embodiment.

[0116] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A point cloud classification and segmentation network for three-dimensional point cloud classification and segmentation, characterized in that: include: The first Contextual Transformer module, the second Contextual Transformer module, the third Contextual Transformer module, the fourth Contextual Transformer module and the first Concat module are connected in sequence; A first Coordinate Transformation module, a first ResMLP module, a second Coordinate Transformation module, a second ResMLP module, a third Coordinate Transformation module, a third ResMLP module, a fourth Coordinate Transformation module, and a fourth ResMLP module connected in sequence; The second Concat module and the first MaxPool module are connected in sequence, the output end of the first MaxPool module is connected to the input end of the first Classification module, and the output end of the first MaxPool module is also connected to the input end of the first Segmentation module; The output end of the first input module is connected to the input end of the first Contextual Transformer module, and the output end of the first input module is also connected to the input end of the first Coordinate Transformation module; the input end of the second Concat module is connected to the output end of the first Concat module, and the input end of the second Concat module is also connected to the output end of the fourth ResMLP module.

2. A point cloud classification and segmentation network for three-dimensional point cloud classification and segmentation according to claim 1, characterized in that: Any one of the first ResMLP module, the second ResMLP module, the third ResMLP module, and the fourth ResMLP module includes: The first MLP connected sequentially pre module, the first maximum pooling layer module and the first MLP pos module.

3. The point cloud classification and segmentation network for three-dimensional point cloud classification and segmentation according to claim 2, characterized in that: First MLP pre Modules include: The first Point-ResMLP module, the second Point-ResMLP module and the third Point-ResMLP module are connected in sequence.

4. The point cloud classification and segmentation network for three-dimensional point cloud classification and segmentation according to claim 2, characterized in that: First MLP pos Modules include: The fourth Point-ResMLP module, the fifth Point-ResMLP module and the sixth Point-ResMLP module are connected in sequence.

5. A point cloud classification and segmentation network for three-dimensional point cloud classification and segmentation according to any one of claims 1 to 4, characterized in that: Any one of the first Point-ResMLP module, the second Point-ResMLP module, the third Point-ResMLP module, the fourth Point-ResMLP module, the fifth Point-ResMLP module, and the sixth Point-ResMLP module includes: The first Conv module, the first BactchNorm module, the first ReLU module, the second Conv module, the second BactchNorm module and the second ReLU module are connected in sequence.

6. A three-dimensional point cloud classification and segmentation method based on the point cloud classification and segmentation network for three-dimensional point cloud classification and segmentation according to any one of claims 1 to 5, characterized in that: The following steps are involved: Step 1: Input the point cloud feature vector into the first Contextual Transformer module for feature extraction; Step 2: The feature vector obtained in step 1 is input into the second Contextual Transformer module for feature extraction. After passing through four Contextual Transformer modules in the same way, the output of each module is fused through the first Concat module to obtain the final global feature and output it; Step 3: Input the point cloud feature vector into the first Coordinate Transformation module so that the irregular geometric structure of the local area of the point cloud can be processed by a unified feature extractor; Step 4: Input the coordinate-transformed feature vector into the first ResMLP module to extract local features; Step 5: Input the features extracted in step 4 into the second Coordinate Transformation module, and then input them into the second ResMLP module. After passing through four Coordinate Transformation modules and four ResMLP modules in the same way, the final local features are obtained and output; Step 6: The outputs of step 5 and step 2 are fused through the second Concat module and the final extracted features are obtained through maximum pooling; Step 7: Input the features finally extracted in step 6 into the first Classification module and the first Segmentation module respectively to obtain the final classification and segmentation results.

7. A three-dimensional point cloud classification and segmentation method according to claim 6, characterized in that: For the input point cloud feature map X∈R B×D×N , the key, query and value are defined as K = X, Q = X and V = XW respectively V The ContextualTransformer module first uses k×k groups of convolutions on all adjacent keys in a k×k grid in space to encode the context of each key representation, which is expressed as: K 1 =Conv k×k (K) The learned context key K 1 ∈R B×D×N It reflects the static context information between local adjacent keys. Based on the static context and query splicing, the attention matrix is obtained by two consecutive 1×1 convolutions: W θ =W δ =Conv 1×1 A=[K 1 ,Q]W θ W δ Among them, K 1 represents the static context, Q represents the query, and A represents the context attention matrix; Calculate the attention feature map based on the context attention matrix: K 2 =V*A Among them, K 2 Represents dynamic context; The static context K 1 and dynamic context K 2 The global features are obtained by fusion.

8. The three-dimensional point cloud classification and segmentation method according to claim 6, characterized in that: The coordinate transformation expression of the CoordinateTransformation module is: Among them, f i,j Represents the point f obtained by the K-nearest neighbor algorithm i The K local neighboring points of the local area are represented by ε, which is a very small value used to keep the denominator non-zero. α and β represent learnable parameters. σ represents the standard deviation of the point features in the local area, which is used to measure the degree of variation or dispersion between the features of each point in the local area and the features of the local center point.

9. The three-dimensional point cloud classification and segmentation method according to claim 6, characterized in that: The overall ResMLP module is represented as follows: p i =MLP pos (MaxPool(MLP pre (f i,j ),|j=1,...,K)) Among them, p i ResMLP module, MaxPool represents the maximum pooling operation, f i,j Represents the point f obtained by the K-nearest neighbor algorithm i The K local neighbors of .

10. The three-dimensional point cloud classification and segmentation method according to claim 6, characterized in that: The calculation expression of step six is: N=MaxPool(concat(N r ,N c ) Among them, N represents the final extracted features, N c represents the global feature, N r Represents the fourth local feature.

Images