Point cloud classification segmentation method based on rapid sampling and attention mechanism

Abstract

The invention discloses a point cloud classification segmentation method based on rapid sampling and an attention mechanism. The method comprises the following steps: loading a point cloud classification data set Dc and a point cloud segmentation data set Ds; a point cloud classification segmentation network is constructed, the network is based on a Point Net + + network, a downsampling layer is constructed to replace an original sampling layer, a segmentation network NSt based on an attention mechanism is constructed to replace an original segmentation network, interpolation operation is improved in the NSt, and a self-attention encoder is additionally arranged during jump connection; and training the point cloud classification segmentation network by using Dc and Ds to obtain a point cloud classification segmentation model which is used in a classification or segmentation task of point cloud data. According to the method, the complexity in the downsampling process can be reduced, space and related information in the point cloud can be concerned and extracted, and the problem of inaccurate interpolation caused by dependence on the size of the selected neighborhood and the distribution condition of the points during interpolation can be solved. Therefore, the sampling efficiency, the feature extraction precision and the segmentation precision can be improved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a point cloud classification and segmentation method based on fast sampling and attention mechanism. Background Art

[0002] With the rapid development and widespread application of advanced depth sensing technologies such as LiDAR, RGB-D cameras, 3D structured light cameras, and ToF cameras, it has become easier than ever to obtain and use point cloud data. Compared with traditional two-dimensional images, three-dimensional point cloud data contains richer information on form, size, and geometric structure, providing smart devices with a deeper level of environmental perception. Taking autonomous driving as an example, the LiDAR carried by the vehicle can continuously scan the surrounding environment and generate a detailed three-dimensional point cloud model including road layout, traffic signs, surrounding vehicles, pedestrians, and even roadside buildings, which greatly promotes the understanding and interaction of the surrounding world. In recent years, three-dimensional point cloud data has shown great potential in many fields closely related to human life, such as driverless cars, robotics, and geospatial mapping, prompting efficient processing and in-depth analysis of such data to become a frontier hotspot in current research.

[0003] In the field of 3D point cloud data processing, classification and segmentation are two core challenges. The goal of point cloud classification is to accurately assign category labels to each point, such as identifying objects such as chairs, cars or buildings in the point cloud; while point cloud segmentation is committed to dividing the point cloud data set into multiple subsets, and the points in each subset belong to the same object or share similar feature attributes. These 3D visual tasks are crucial for path planning of autonomous driving systems, autonomous navigation of robots, urban management optimization, ecological environment monitoring and even remote sensing image analysis. However, due to the inherent irregularity, disorder and uneven density distribution of point cloud data, its processing and analysis process faces high computational complexity and model design difficulties. Early studies mostly relied on traditional image processing algorithms and machine learning strategies. Although they have achieved certain results in specific situations, these methods are often sensitive to noise, have limited generalization performance, and are unable to cope with complex and changeable actual scenes. Therefore, exploring new methods for 3D point cloud processing that are more robust, efficient and adaptable has become a key issue that needs to be broken through.

[0004] In recent years, deep learning technology has made remarkable achievements in the field of computer vision and has gradually become the mainstream method for 3D point cloud processing. Most of the 3D point cloud processing methods based on deep learning are neural network (NN) model architectures. Neural network models such as PointNet and PointNet++ use multilayer perceptrons (MLP) or convolutional neural networks (CNN) as the backbone network for feature extraction, and then combine the farthest point sampling (FPS) and nearest neighbor search for local feature extraction, which performs well in point cloud classification and segmentation tasks.

[0005] Pointnet++ mainly consists of three parts: Hierarchical point set feature learning module, segmentation network and classification network. Regarding the hierarchical point set feature learning module: multiple SA (Set Abstraction) layers are used for feature extraction. The difference is that the number of sampling points (K) and sampling radius (R) of each SA layer are different. As the level deepens, K and R will increase accordingly. The SA layer contains three key layers: Sampling Layer, Grouping Layer, and PointNetLayer. Among them, the Sampling Layer selects several center points from the point cloud to form a point set based on FPS sampling. The Grouping Layer divides the point set into several regions based on the ball query and K nearest neighbor grouping method. The PointNet Layer performs MLP feature extraction, maximum pooling and other operations on each region to obtain data features, and then splices them into a set of multi-scale features. However, the existing methods of Pointnet++ still have some problems in sampling efficiency and spatial information extraction in point cloud processing: (1) PointNet++ selects the center points of local areas through FPS sampling and constructs local areas around these center points. This process involves multiple nearest neighbor searches, which significantly increases the amount of computation, especially when processing large-scale point cloud data. Although FPS is more accurate, it still needs to traverse the entire dataset to find the farthest point, which will result in higher computational costs when the amount of data is large. And as the number of points increases, it will bring higher computational complexity.

[0006] (2) PointNet++ designed a skip link module for the segmentation task. However, this skip link module makes less use of the spatial information between points, which limits the ability of feature extraction and leads to poor results in feature extraction accuracy.

[0007] (3) PointNet++ uses the nearest neighbor search interpolation method when upsampling segmentation tasks. This method depends on the selected neighborhood size and point distribution during interpolation, which can lead to inaccurate interpolation.

[0008] Glossary: ​​The JL (Johnson-Lindenstrauss) lemma is a lemma widely used in mathematics and computer science, mainly used to deal with the dimensionality reduction problem of high-dimensional data. The lemma states that for n points in a high-dimensional space, they can be mapped to a low-dimensional space through a random projection while keeping the distances between these points roughly unchanged. Summary of the invention

[0009] The purpose of the present invention is to provide a point cloud classification and segmentation method based on fast sampling and attention mechanism, which can solve the problems of slow sampling speed, poor local feature extraction accuracy, inaccurate interpolation, etc. of Pointnet++.

[0010] In order to achieve the above object, the technical solution adopted by the present invention is as follows: a point cloud classification and segmentation method based on fast sampling and attention mechanism, comprising the following steps; S1, load the point cloud classification dataset D consisting of point cloud classification samples c , a point cloud segmentation dataset D consisting of point cloud segmentation samples s ; S2, constructing a point cloud classification and segmentation network, including steps S21 to S24; S21, select a PointNet++ network, including a hierarchical point set feature learning module, a segmentation network, and a classification network, wherein the hierarchical point set feature learning module includes two first SA layers L connected in sequence SA1 , the second SA layer L SA2 , the two have the same structure, both of which include a sampling layer, a grouping layer and a point network layer connected in sequence; the hierarchical point set feature learning module inputs point cloud data and outputs corresponding multi-scale features, and the point cloud data is a point cloud classification sample or a point cloud segmentation sample; S22, constructing a downsampling layer to downsample the point cloud data based on the sparse matrix; S23, construct a segmentation network N based on the attention mechanism St , including a first interpolation layer, a first splicing layer, a second interpolation layer, a second splicing layer, two fully connected layers and a segmentation layer connected in sequence; S24, select two downsampling layers and label them as L Ds1 , L Ds2 , respectively replace L SA1 and L SA2 The sampling layer in the L SA1 , L SA2 Marked as the first improved SA layer , the second improved SA layer ; in L Ds1 The input end is connected to the input end of the first concatenation layer through the first self-attention encoder. Ds2 The input end is connected to the input end of the second concatenation layer through the second self-attention encoder, and N St Replace the segmentation network to obtain the point cloud classification segmentation network; One point cloud segmentation sample X S Send it to the point cloud classification and segmentation network. , The output features are obtained respectively , , Enter N St middle; The first interpolation layer is based on Perform interpolation to obtain the first feature F St1 , the first concatenation layer is used to convert X S With F St1 Splice and get the second feature F St2 , the second interpolation layer is used to St2 Interpolation operation is performed to obtain the third feature F St3 , the second stitching layer is used to With F St3 After splicing, we get the fourth feature F St4 , F St4 The two layers of fully connected layers are used to reduce the dimension, and then the segmentation layer is used to perform semantic segmentation point by point, and the segmentation result F is output. S-out ; S3, with D c and D s Train until the point cloud classification and segmentation network converges to obtain a point cloud classification and segmentation model; S4, obtaining the point cloud data to be processed, and classifying or segmenting it using a point cloud classification and segmentation model.

[0011] Preferably, in S1, the point cloud classification dataset is the Modelnet40 dataset, and the point cloud segmentation dataset is the ShapeNet or S3DIS dataset.

[0012] Preferably, the downsampling method of the downsampling layer comprises steps Sa1 to Sa3; Sa1, obtain a point cloud data X, which contains N data points, each of which contains m-dimensional features, and construct X into an N×m matrix; Sa2, determine the number of data points n after downsampling, from the normal distribution Randomly sample an N×n sparse matrix Φ, Φ satisfies the JL lemma, and n<<N; Sa3, according to the formula Y=XΦ, obtain an n×m downsampled point cloud data Y.

[0013] Preferably, the first interpolation layer and the second interpolation layer both perform interpolation operations based on a bilinear interpolation method.

[0014] Preferably, the classification network comprises a dot-net layer, two fully connected layers, and a classification layer which are connected in sequence.

[0015] Compared with the prior art, the advantages of the present invention are: (1) A method for downsampling point cloud data based on a sparse matrix is ​​designed and executed by the downsampling layer of the present invention. This method can obtain sampling points through a single matrix multiplication, thereby improving the sampling speed and is particularly suitable for large-scale point cloud data.

[0016] (2) Improve the skip connection in the original PointNet++ network, and transform the point cloud data X S The first feature F output by the first interpolation layer St1 Splice and improve the first SA layer Output The third feature F output by the second interpolation layer St3 Splicing is performed to solve the problem that existing studies ignore the spatial information between each point, so that the overall model can extract spatial and related information in the sampled point cloud, improving the accuracy of segmentation.

[0017] (3) In the segmentation network, the first interpolation layer and the second interpolation layer are interpolated based on bilinear interpolation, which overcomes the problem of inaccurate interpolation caused by the selected neighborhood size and point distribution during interpolation. We invented a bilinear interpolation method. Experimental results show that the test results of this method on multiple data sets are significantly improved compared with existing methods.

[0018] In summary, the present invention designs a point cloud classification and segmentation network based on fast sampling and attention mechanism, and a point cloud classification and segmentation method based on the network, which can improve sampling efficiency, feature extraction accuracy, and segmentation accuracy. A large number of experiments have verified that the present invention has achieved better results than PointNet++ in classification and segmentation tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1This is a schematic diagram of the point cloud classification and segmentation network structure; Figure 2 Schematic diagram of the interpolation operation of the first interpolation layer. DETAILED DESCRIPTION

[0020] The present invention will be further described below in conjunction with the accompanying drawings.

[0021] Example 1: See Figure 1 and Figure 2 ,A point cloud classification and segmentation method based on fast sampling and attention mechanism, includes the following steps; S1, load the point cloud classification dataset D consisting of point cloud classification samples c , a point cloud segmentation dataset D consisting of point cloud segmentation samples s ; S2, constructing a point cloud classification and segmentation network, including steps S21 to S24; S21, select a PointNet++ network, including a hierarchical point set feature learning module, a segmentation network, and a classification network, wherein the hierarchical point set feature learning module includes two first SA layers L connected in sequence SA1 , the second SA layer L SA2 , the two have the same structure, both of which include a sampling layer, a grouping layer and a point network layer connected in sequence; the hierarchical point set feature learning module inputs point cloud data and outputs corresponding multi-scale features, and the point cloud data is a point cloud classification sample or a point cloud segmentation sample; S22, constructing a downsampling layer to downsample the point cloud data based on the sparse matrix; S23, construct a segmentation network N based on the attention mechanism St , including a first interpolation layer, a first splicing layer, a second interpolation layer, a second splicing layer, two fully connected layers and a segmentation layer connected in sequence; S24, select two downsampling layers and label them as L Ds1 , L Ds2 , respectively replace L SA1 and L SA2 The sampling layer in the L SA1 , L SA2 Marked as the first improved SA layer , the second improved SA layer ; in L Ds1 The input end is connected to the input end of the first concatenation layer through the first self-attention encoder. Ds2 The input end is connected to the input end of the second concatenation layer through the second self-attention encoder, and N St Replace the segmentation network to obtain the point cloud classification segmentation network; One point cloud segmentation sample X S Send it to the point cloud classification and segmentation network. , The output features are obtained respectively , , Enter N St middle; The first interpolation layer is based on Perform interpolation to obtain the first feature F St1 , the first concatenation layer is used to convert X S With F St1 Splice and get the second feature F St2 , the second interpolation layer is used to St2 Interpolation operation is performed to obtain the third feature F St3 , the second stitching layer is used to With F St3 After splicing, we get the fourth feature F St4 , F St4 The two layers of fully connected layers are used to reduce the dimension, and then the segmentation layer is used to perform semantic segmentation point by point, and the segmentation result F is output. S-out ; S3, with D c and D s Train until the point cloud classification and segmentation network converges to obtain a point cloud classification and segmentation model; S4, obtaining the point cloud data to be processed, and classifying or segmenting it using a point cloud classification and segmentation model.

[0022] Example 2: See Figure 1 and Figure 2 , based on Example 1, in S1, the point cloud classification dataset is the Modelnet40 dataset, and the point cloud segmentation dataset is the ShapeNet or S3DIS dataset.

[0023] The downsampling method of the downsampling layer comprises steps Sa1 to Sa3; Sa1, obtain a point cloud data X, which contains N data points in total. Each data point contains m-dimensional features, where m is generally 3 or 6. The data points are numbered from 1 to N in sequence, so the point cloud data X can be represented as an N×m matrix; , In this matrix, each row represents a point cloud, x 11 ~x 1m Respectively represent the eigenvalues ​​of the 1st dimension feature to the mth dimension feature of the 1st data point, x N1 、x Nm Respectively represent the feature values ​​of the 1st dimension feature to the mth dimension feature of the Nth data point; Sa2, assuming that the number of data points after downsampling is n, then from the normal distribution Randomly sample an N×n sparse matrix Φ from it. Φ satisfies the JL lemma, and n << N. Usually, constructing an N×n matrix has a time complexity of O(N×n). However, since n << N, in this step, the time complexity of constructing the sparse matrix Φ can be approximated as O(N). For the FPS algorithm, the time complexity is usually O( ), where N is the number of points before sampling. From the perspective of the order of magnitude of time complexity, our sampling algorithm is far superior to FPS. Moreover, subsequently, we only need one matrix multiplication to complete the selection of all points, while FPS needs to calculate the Euclidean distance from the selected points to all the remaining points in each point selection iteration; Sa3, according to the formula Y = XΦ, obtain a downsampled point cloud data Y of n×m. Through this step, it is possible to randomly map high-dimensional data to a low-dimensional space.

[0024] Both the first interpolation layer and the second interpolation layer perform interpolation operations based on the bilinear interpolation method.

[0025] The classification network includes a point network layer, two fully connected layers, and a classification layer connected in sequence.

[0026] In addition, more specifically, regarding the grouping layer and the point network layer: Group the results obtained by the sampling layer according to the ball query method, that is, set three sets of radii (0.1, 0.2, and 0.4) with the sampled points as the center points, and group the number of points within the radii (16, 32, and 128 corresponding to the radii 0.1, 0.2, and 0.4 respectively). Each center point forms a cluster based on the set of points it composes, obtaining multi-scale point cloud data with three different clusters; then send it to the point network layer. First, extract the point cloud data features through the MLP for different clusters, then obtain three sets of data features through max pooling for each cluster, and then splice the three sets of features into one set of features to obtain the output of the sampling module.

[0027] Regarding the classification network, it is the original classification network of the PointNet++ network. In this classification network, the point network layer is also called the Pointnet layer, which is used to perform max pooling on the results input to the classification network to obtain a set of one-dimensional global features. Then, through two fully connected layers in sequence, the one-dimensional global features are dimensionally reduced to effectively extract and retain the key features in the point cloud data. Finally, through the classification layer, the scores of the corresponding point cloud data in different classifications are obtained, and the one with the highest score is selected as the classification result.

[0028] Regarding the segmentation network N based on the attention mechanism St , the principle of its segmentation layer is the same as that of the segmentation layer自带 in the PointNet++ network. N St In, when the fourth feature F St4After the dimension reduction by two layers of fully connected layers, it is sent to the segmentation layer. Assuming that there are C segmentation categories in the segmentation task, for each point input to the segmentation layer, it will output the score of each segmentation category. For point cloud data X, there are N points, and the segmentation layer will output N×C scores to form the segmentation result F. S-out .

[0029] About the self-attention encoder: It includes the first self-attention encoder and the second self-attention encoder, and the connection method is as follows Figure 1 As shown in . The self-attention mechanism is a special attention mechanism that allows the model to consider the relationship between each element and all other elements in the sequence when processing the sequence. In the present invention, the first and second self-attention encoders first perform self-attention encoding on their inputs, extract the spatial relationship of each point, and then each is connected to the segmentation network N based on the attention mechanism through a jump link. St middle.

[0030] Regarding the first interpolation layer and the second interpolation layer, the two have the same structure. For their interpolation methods, see Figure 2 :First, cut the three-dimensional point cloud data into two-dimensional form, and use f(x, y) to represent a point. In the figure, f(x 0 ,y 0 )、f(x 1 ,y 0 )、f(x 1 ,y 1 )、f(x 0 ,y 1 ) represent four points in space respectively; then perform three bilinear interpolation operations: the first one is for f(x 0 ,y 0 )、f(x 0 ,y 1 ) Perform a bilinear interpolation on these two points to get point z 0 , the second time for f(x 1 ,y 0 )、f(x 1 ,y 1 ) Perform a bilinear interpolation on these two points to get point z 1 , the third time to z 0 、z 1 Perform a bilinear interpolation on these two points to obtain point z, which is the final interpolation result.

[0031] About using D c and D s Train until the point cloud classification and segmentation network converges to obtain the point cloud classification and segmentation model. A simple training process is given below: Step 1: Data preparation and division Dataset selection: Point cloud classification dataset D cSelect Modelnet40 dataset, point cloud segmentation dataset D s Select ShapeNet.

[0032] Data preprocessing: 1) Preload points: Randomly sample a fixed number of points for each point cloud object to form point cloud data. In this embodiment, 1024 points are sampled; 2) Divide the data set: Divide the data set into a training set and a test set in proportion. The proportions of the present invention are 80% and 20%.

[0033] Step 2: Model architecture setup Network selection: 1) Improve the SA layer first , the second improved SA layer and classification network to form the classification task network structure; 2) Improve the first SA layer , the second improved SA layer , first self-attention encoder, second self-attention encoder, segmentation network based on attention mechanism N St Construct the network structure for segmentation tasks.

[0034] Layer grouping: such as the first improved SA layer Sampling 512 points, 128 points, etc., setting the grouping radius and number of neighbors. The present invention: the first improved SA layer Sampling 512 points, second improved SA layer The sampling points are 256. The grouping radius is 0.1, 0.2 and 0.4, corresponding to the number of neighbors 16, 32 and 128.

[0035] Step 3: Set hyperparameters Learning rate: 0.0001, using Adam optimizer, using Step Decay to adjust the learning rate.

[0036] Batch size: 16.

[0037] Loss function: cross entropy loss function.

[0038] Number of training rounds: 200.

[0039] Step 4: Training process First, the predefined model architecture of the present invention is loaded, the weight parameters are initialized, and then the point cloud data and its labels are loaded from the training set in batches and accelerated by GPU. Then the forward propagation process is: for the classification task, after , The multi-scale features are obtained and sent to the classification network. The classification result is output through the classification layer. The classification result is the prediction result of the classification task. , , first self-attention encoder, second self-attention encoder, N St After that, the output segmentation result is the score of each point belonging to each segmentation category, which is the prediction result of the segmentation task. Then the loss value between the prediction result and the label is calculated, the gradient is calculated by the back propagation algorithm, and the network parameters are updated. Then, the model is verified every several training rounds, the performance is evaluated on the verification set, and the model parameter file with the best performance is saved. Finally, the model is tested and evaluated.

[0040] Example 3: Based on Examples 1 and 2, the present invention is used to conduct comparative experiments with two typical point cloud task models. The experimental results are shown in Tables 1 and 2: Table 1 Comparison of classification tasks of different models Model Instance Acc (%) Class Acc(%) PointNet 89.2 86.2 PointNet++ 91.1 90.7 The present invention 92.1 91.0 Table 2 Comparison of segmentation tasks of different models Model MIoU (%) PointNet 83.7 PointNet++ 85.1 The present invention 85.2 In the table, Instance Acc: Instance Accuracy in English and instance accuracy in Chinese is a commonly used evaluation indicator in point cloud classification tasks. It refers to the total number of correct predictions made by the model in all predictions divided by the total number of samples involved in the prediction, regardless of the category.

[0041] Class Acc: Class Accuracy in English and classification accuracy in Chinese. It is also a commonly used evaluation indicator in point cloud classification tasks. It refers to the average accuracy (or recall) of each category. It first calculates the accuracy of each category and then averages these accuracy rates. This accuracy rate can better reflect the impact of class imbalance.

[0042] MIoU: Mean Intersection over Union in English, is a commonly used evaluation indicator in point cloud segmentation tasks. It measures model performance by calculating the ratio between the intersection and union of the predicted results and the true labels.

[0043] In addition, based on the model of the present invention, we replaced some modules for comparative experiments. Among them, the corresponding positions of the first interpolation layer and the second interpolation layer in the segmentation network need to be interpolated, and the moving least squares interpolation method and the interpolation operation based on the bilinear interpolation method described in the present invention are respectively selected. The first self-attention encoder and the second self-attention encoder are combined into a spatial self-attention module, and the spatial self-attention module is set to 0, 1, and 2 for comparative experiments. The experimental results are shown in Table 3: Table 3 Comparison of different interpolation methods and the number of spatial self-attention modules Modules result Moving Least Squares Interpolation Number of spatial self-attention modules Bilinear interpolation Instance Acc (%) MIoU (%) √ - - 91.1 80.0 - - √ 91.0 81.2 - 1 - 94.3 85.2 - 2 √ 94.2 85.1

[0044] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A point cloud classification and segmentation method based on fast sampling and attention mechanism, characterized by: The steps include: S1, load the point cloud classification dataset D consisting of point cloud classification samples c , a point cloud segmentation dataset D consisting of point cloud segmentation samples s ; S2, constructing a point cloud classification and segmentation network, including steps S21 to S24; S21, select a PointNet++ network, including a hierarchical point set feature learning module, a segmentation network, and a classification network, wherein the hierarchical point set feature learning module includes two first SA layers L connected in sequence SA1 , the second SA layer L SA2 , the two have the same structure, both of which include a sampling layer, a grouping layer and a point network layer connected in sequence; the hierarchical point set feature learning module inputs point cloud data and outputs corresponding multi-scale features, and the point cloud data is a point cloud classification sample or a point cloud segmentation sample; S22, constructing a downsampling layer to downsample the point cloud data based on the sparse matrix; S23, construct a segmentation network N based on the attention mechanism St , including a first interpolation layer, a first splicing layer, a second interpolation layer, a second splicing layer, two fully connected layers and a segmentation layer connected in sequence; S24, select two downsampling layers and label them as L Ds1 , L Ds2 , respectively replace L SA1 and L SA2 The sampling layer in the L SA1 , L SA2 Marked as the first improved SA layer , the second improved SA layer ; in L Ds1 The input end is connected to the input end of the first concatenation layer through the first self-attention encoder. Ds2 The input end is connected to the input end of the second concatenation layer through the second self-attention encoder, and N St Replace the segmentation network to obtain the point cloud classification segmentation network; One point cloud segmentation sample X S Send it to the point cloud classification and segmentation network. , The output features are obtained respectively , , Enter N St middle; The first interpolation layer is based on Perform interpolation to obtain the first feature F St1 , the first concatenation layer is used to convert X S With F St1 Splice and get the second feature F St2 , the second interpolation layer is used to St2 Interpolation operation is performed to obtain the third feature F St3 , the second stitching layer is used to With F St3 After splicing, we get the fourth feature F St4 , F St4 The two layers of fully connected layers are used to reduce the dimension, and then the segmentation layer is used to perform semantic segmentation point by point, and the segmentation result F is output. S-out ; S3, with D c and D s Train until the point cloud classification and segmentation network converges to obtain a point cloud classification and segmentation model; S4, obtaining the point cloud data to be processed, and classifying or segmenting it using a point cloud classification and segmentation model.

2. According to claim 1, a point cloud classification and segmentation method based on fast sampling and attention mechanism is characterized in that: In S1, the point cloud classification dataset is the Modelnet40 dataset, and the point cloud segmentation dataset is the ShapeNet or S3DIS dataset.

3. The point cloud classification and segmentation method based on fast sampling and attention mechanism according to claim 1, characterized in that: The downsampling method of the downsampling layer includes steps Sa1 to Sa3; Sa1, obtain a point cloud data X, which contains N data points, each of which contains m-dimensional features, and construct X into an N×m matrix; Sa2, determine the number of data points n after downsampling, from the normal distribution Randomly sample an N×n sparse matrix Φ, Φ satisfies the JL lemma, and n<<N; Sa3, according to the formula Y=XΦ, obtain an n×m downsampled point cloud data Y.

4. The point cloud classification and segmentation method based on fast sampling and attention mechanism according to claim 1, characterized in that: The first interpolation layer and the second interpolation layer both perform interpolation operations based on bilinear interpolation.

5. The point cloud classification and segmentation method based on fast sampling and attention mechanism according to claim 1, characterized in that: The classification network includes a dot network layer, two fully connected layers, and a classification layer which are connected in sequence.

Images