3D point cloud classification segmentation method based on adaptive laplacian convolution network

Abstract

The application discloses a 3D point cloud classification and segmentation method based on an adaptive Laplacian convolution network, and comprises the following steps: using a public data set after preprocessing and sampling grouping to train the network, improving a traditional Laplacian convolution network, introducing a vector attention method in a feature extraction part to realize adaptive adjustment of weights, establishing a skip connection mode between modules in the feature extraction part to realize multi-scale fusion of features, and designing an enhancement module to enhance local feature information extracted. Finally, the trained network is used for classification and segmentation of the 3D point cloud to be tested. The application solves the problem of difficult extraction of local information caused by irregular characteristics of the 3D point cloud, and provides a new solution for the problem of performance drop of part of the network when facing sparse point clouds and noisy point clouds. Moreover, the application pursues lightweight design, has good real-time performance, and can accurately complete the classification and segmentation task of the 3D point cloud.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of 3D point cloud feature extraction and classification segmentation, in particular to a 3D point cloud classification segmentation method based on an adaptive Laplacian convolution network. BACKGROUND

[0002] It is an urgent need in robot, autonomous driving, augmented reality and other applications to extract features directly from 3D point clouds for classification and segmentation. Unlike 2D images, point clouds are unordered and unstructured, making it challenging to design neural networks to process them.

[0003] There are two solutions to this problem. One is to design a geometric structure extractor and a convolution operation to obtain the local geometric features of the point cloud. Since the ordinary convolution has poor applicability on 3D point clouds and is not sensitive to the connection between point cloud data, the extraction performance is low when this method is applied. The second solution is to use a multi-head attention mechanism to extract global features. However, this method will bring a large computational overhead, so that real-time tasks cannot achieve fast iteration. Moreover, both of these two solutions will have a sharp drop in performance when dealing with sparse and noisy point clouds, and have poor robustness. SUMMARY

[0004] The present application aims to overcome the shortcomings and deficiencies of the prior art and provides a 3D point cloud classification segmentation method based on an adaptive Laplacian convolution network, which can effectively extract local and global features of point cloud data, pursue lightweight in network structure design, and reduce unnecessary additional overhead. Moreover, this method does not have a sharp drop in accuracy when dealing with sparse point clouds and noisy point clouds, and has strong robustness.

[0005] To achieve the above-mentioned purpose, the technical solution provided by the present application is as follows: a 3D point cloud classification segmentation method based on an adaptive Laplacian convolution network, comprising the following steps:

[0006] 1) Use a public 3D point cloud dataset containing true value point cloud class labels and a plurality of point clouds, and preprocess the dataset to obtain point cloud data with uniform quantity and uniform distribution;

[0007] 2) Use the farthest point sampling algorithm and the ball query algorithm to sample and group the point cloud data obtained after preprocessing, divide the point cloud data into a plurality of ball domain groups, each ball domain group has a center point, and each ball domain group contains all point clouds within a specified radius of a sphere;

[0008] 3) Extract features from the plurality of ball domain groups obtained after sampling and grouping to extract corresponding feature vectors, including center point features and neighborhood features, wherein the neighborhood features refer to the features of the point clouds other than the center point in the ball domain group;

[0009] 4) Design a Laplacian feature convolution mechanism containing vector attention, which adaptively assigns attention weights to feature vectors using the Laplacian feature convolution mechanism, and respectively strengthens the center point features and neighborhood features of the point cloud in the ball group according to the attention weight size, to realize the convolution between the center point features and the neighborhood features;

[0010] 5) Design an enhanced module containing similarity matching calculation to further enhance the extracted feature vectors and improve the accuracy of network training;

[0011] 6) Train an adaptive Laplacian convolution network using feature vectors, use multi-class cross-entropy as a loss function, and output the corresponding point cloud category; wherein the adaptive Laplacian convolution network is an improved Laplacian convolution network, which is improved by adding vector attention to the original Laplacian convolution network, adaptively obtaining Laplacian weights, and adding a multi-scale fusion mode of skip connection, inputting point cloud, using grouping sampling to obtain ball groups for feature extraction, obtaining local and global features of point cloud data, and making it obtain more perfect overall features on the corresponding space;

[0012] 7) Apply the trained adaptive Laplacian convolution network to input the 3D point cloud data to be tested into the network to obtain the category of the point cloud and the part segmentation effect.

[0013] Further, in step 2), the point cloud data obtained after preprocessing is sampled and grouped, and the specific process is as follows:

[0014] Input point cloud data U = {u1, u2,..., un}, wherein un represents the nth point cloud; the selection of the center point set U = {u1, u2,..., um} uses the iterative farthest point sampling algorithm, wherein um represents the mth center point; U = FPS{U}, for the sampled center point set U, use the ball query algorithm to determine the neighborhood N = {u1, u2,..., uj} of each center point set U according to the distance size, wherein uj represents the jth adjacent point of u, and N represents the neighborhood of um. n}, wherein u n represents the nth point cloud; the selection of the center point set U cm = {u c1 , u c2 ,..., u cm} uses the iterative farthest point sampling algorithm, wherein u cm represents the mth center point; U cm = FPS{U}, for the sampled center point set U cm , use the ball query algorithm to determine the neighborhood N cm = {u cm |u cm , u m1 ,..., u m2} of each center point set U mj , wherein u mj represents the jth adjacent point of u cm , and N cm represents the neighborhood of um. cmN cm is the set of ball regions centered at U cm , N cm is obtained by the ball query algorithm. cm

[0015] Further, in step 3), the adaptive Laplacian backbone of the adaptive Laplacian convolution network is used to extract features from the ball region set, and the extracted feature vector includes center point features and neighborhood features; the adaptive Laplacian backbone is stacked by 6 convolution modules with different parameters, and multi-scale feature fusion is used between these convolution modules to ensure that more comprehensive local information is extracted; the convolution module not only includes linear layers, convolution operations and activation functions, but also combines vector attention mechanisms to improve the effect and accuracy of feature extraction.

[0016] Further, in step 4), the vector attention and Laplace feature convolution mechanism is implemented as follows:

[0017] Vector attention: a set of parameters op obeying the standard normal distribution is randomly generated, op is normalized and converted into a learnable parameter, and the weight W a is obtained by the dot product of op after the softmax function and N cm , and N cm is concatenated by all the center point set U cm ={u c1 ,u c2 ,...,u cm} through a fully connected layer, and finally the output F on is obtained, as shown in the following formula:

[0018] F on =Conv1d(concat(U cm ,sum(W a N cm )))

[0019] In the formula, sum is a summation operation, concatenate is a concatenation operation using the concatenate operation along the vector dimension, and the concatenated data is then subjected to Conv1d, which is a one-dimensional convolution operation, and the vector attention is calculated by the attention method;

[0020] Laplacian feature convolution mechanism: the input of the Laplacian convolution module is denoted as F in , the weighted sum of the Laplace features learned by the neural network is denoted as F a , and the relationship between F a and F in is represented by the following formula:

[0021] F a =∑softmax(θ·(N cm -u cm ))·(N cm -u cm )

[0022] F in =N cm -u cm

[0023] In the formula, θ=(θ1, θ2,.., θ m ) is the parameter learned by the neural network, θ m m represents the mth parameter learned; the Laplace feature convolution mechanism introduces an adaptive strategy in deep learning, which not only is equivalent to using a neural network layer to infer the weight, but also introduces higher flexibility and accuracy in the point cloud classification and segmentation task, and adopts a completely different discretization method from the prior art, abandoning the dependence on uniform discretization or specific geometric information, thereby realizing a deeper understanding of complex geometric shapes, so that the network can dynamically strengthen the relationship between the center point features and the neighborhood features in the ball domain group according to the characteristics of specific data.

[0024] Further, in step 5), the specific implementation of the enhancement module is as follows:

[0025] The relationship between the neighborhood features and the true value point cloud category label is calculated by similarity matching, and then the vector features obtained by the adaptive Laplace backbone part are linearly interpolated to realize the enhancement of the features, and the similarity matching calculation formula is as follows:

[0026] L pre =exp(-γ·(1-∑F s ·L t T ))·L t T

[0027] In the formula, F s is the feature for similarity matching calculation, L t is the true value point cloud category label, L t T is the matrix transpose of L t , and γ is a matching coefficient measured to control the sensitivity of the matching calculation process, and the linear interpolation formula is as follows:

[0028] L cls =λ q ·L pre +λ n ·F b

[0029] wherein L pre is the result of the similarity matching calculation, F b is the sum of the features extracted by the adaptive Laplace backbone, and the two are linearly interpolated to obtain the final prediction value L cls , λ q and λ n are measured parameters used to control the degree of interpolation of the two, thereby achieving the regulation effect of feature enhancement; by introducing the enhancement module of similarity matching calculation, the real label information and the extracted features can be effectively integrated, so that the network can distinguish the importance of the features during the training process.

[0030] Further, in step 6), the adaptive Laplace convolution network processes the point cloud segmentation and classification, and the loss function used is the multi-class cross-entropy to minimize the distance between the predicted value and the true value point cloud class label; the output of the network is a probability value indicating the probability of belonging to N categories; for the multi-class problem, the true value point cloud class label is a one-hot encoded vector indicating which category the sample belongs to; the output of the entire network is a vector indicating the probability of each category, which is processed by the softmax function; assuming that there are N points, each point has C categories, the probability distribution output by the network is P, and the true value point cloud class label is Y; the loss function for point cloud classification is represented as:

[0031]

[0032] wherein N is the number of points, C is the number of categories, Y ij is the true label of point i belonging to category j, and the probability of the network predicting that the point belongs to category j is P ij , and category j is the output of the network. The trained adaptive Laplace convolution network can be used to complete the classification and segmentation of point clouds.

[0033] Further, in step 7), the specific steps of applying the trained adaptive Laplace convolution network to complete the 3D point cloud classification and segmentation are as follows:

[0034] 71) Scan the object that needs to be classified and segmented by the sensor to obtain the complete 3D point cloud data of the object to be measured, and send the preprocessed 3D point cloud data into the trained adaptive Laplace convolution network;

[0035] 72) After step 71), sample and group the obtained point cloud data to obtain different spherical domain groups, extract the corresponding features according to different spherical domain groups, use the Laplace feature convolution mechanism to extract deep local features, and finally use the feature enhancement module to enhance the extracted global and local features;

[0036] 73) After the feature extraction of the above-mentioned step 72), the category of the object to be measured is output for the classification task; for the segmentation task, the segmentation effect of each component of the object to be measured is output.

[0037] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0038] 1、The present application applies Laplacian convolution to 3D point cloud data, has strong local detail capture capability for point cloud, and can effectively extract features.

[0039] 2、The present application proposes a vector attention method, which uses a neural network to adaptively learn Laplacian weights, so that the Laplacian convolution is no longer dependent on a uniform discretization method, but can be adaptively adjusted according to different tasks.

[0040] 3、The neural network training of the present application adopts a lightweight design concept, which reduces redundant calculation while ensuring high classification and segmentation accuracy.

[0041] 4、The present application can process relatively sparse and noisy point clouds, and has strong robustness.

[0042] 5、The neural network structure of the present application introduces a point cloud feature enhancement branch, which can provide more feature information to make the classification and segmentation more accurate. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The present application is a schematic diagram of the logical flow.

[0044] Figure 2 The present application is a schematic diagram of the neural network structure.

[0045] Figure 3 The present application is a schematic diagram of the vector attention structure.

[0046] Figure 4 The present application is a classification and segmentation example effect diagram of the neural network. DETAILED DESCRIPTION

[0047] The present application will be further described in detail below in combination with the embodiments and drawings, but the embodiments of the present application are not limited thereto.

[0048] The present embodiment discloses a 3D point cloud classification and segmentation method based on an adaptive Laplacian convolution network, which uses a deep learning algorithm and uses a public 3D point cloud dataset for training, the logical flow is as shown in Figure 1 , the neural network structure is as shown in Figure 2 , the vector attention structure is as shown in Figure 3 , and the classification and segmentation effect is as shown in Figure 4 , which comprises the following steps:

[0049] 1) Use a public 3D point cloud dataset that contains ground truth point cloud category labels and several point clouds, and preprocess the dataset to obtain uniform and evenly distributed point cloud data;

[0050] The public 3D point cloud dataset contains ground truth point cloud category labels and several point clouds. Classification training is performed on the ModelNet40 dataset and the ScanObjNN dataset. The ModelNet40 dataset contains 40 categories of CAD mesh models, of which 9,843 are trained and 2,468 are tested; the ScanObjNN

[41] dataset contains 15,000 objects, which are divided into 15 categories and 2,902 unique object instances in the real world. The preprocessing step is to read the point cloud dataset, distinguish between training data and test data, and then downsample the point clouds in the dataset to ensure that the number of point clouds fed into the network is uniform and evenly distributed.

[0051] 2) Using the farthest point sampling algorithm and the sphere query algorithm, the pre-processed point cloud data is sampled and grouped into several sphere groups. Each sphere group has a center point. A sphere group refers to all point clouds contained in a sphere with a certain radius.

[0052] Input point cloud data U={u1,u2,...,u n}, where u n Represents the nth point cloud; the center point set U cm ={u c1 ,u c2 ,...,u cm} is selected using the iterative farthest point sampling algorithm, where u cm represents the mth center point; U cm = FPS{U}, for the sampled center point set U cm , use the ball query algorithm to determine each center point set U according to the distance cm Neighborhood N cm ={u cm |u m1 ,u m2 ,...,u mj}, where u mj Indicates u cm The adjacent points of the jth point as the center, N cm Indicates u cm The spherical region group with the center point N cm =queryBall{U cm}, N cm By U cm Obtained through the ball query algorithm;

[0053] The basic principle of the farthest point sampling is to select a starting point from the data set, and then iteratively select the point farthest from the current selected point set. This method can ensure that the sampling points are relatively uniformly distributed in space, thereby effectively preserving the geometric structure characteristics of the data set. The algorithm steps are as follows:

[0054] S1: Assume that the input set P contains n points, select a point p0 from the set P as the starting point, and obtain the sampling point set S = {p0};

[0055] S2: Calculate the distance of all points to p0 to form an array L, and select the point corresponding to the maximum value from L as p1, and update the sampling point set S = {p0, p1};

[0056] S3: Calculate the distance of all points to p1, for each point p i , if the distance between p i and p1 is less than L[i], update L[i] = d(p a , p1), and L always stores the nearest distance of each point to the sampling point set S;

[0057] S4: Select the point corresponding to the maximum value in L as p2, and update the sampling point set S = {p0, p1, p2};

[0058] S5: Repeat steps S2-S4 until the sampling is completed.

[0059] 3) Feature extraction is performed on the multiple spherical domain groups obtained through sampling grouping, and corresponding feature vectors are extracted, including center point features and neighborhood features, wherein the neighborhood features refer to the features of the point clouds other than the center point in the spherical domain group;

[0060] As shown in the figure, the feature extraction part is the adaptive Laplacian backbone part of the adaptive Laplacian convolution network, which is stacked by 6 convolution modules with different parameters, and multi-scale feature fusion is performed between the convolution modules to extract more comprehensive local information. The convolution module not only includes linear layers, convolution operations and activation functions, but also combines a vector attention mechanism.

[0061] 4) A Laplacian feature convolution mechanism containing vector attention is designed, which adaptively allocates attention weights to the feature vectors using the Laplacian feature convolution mechanism, and respectively strengthens the center point features and neighborhood features of the point clouds in the spherical domain group according to the attention weight size, and realizes the convolution between the center point features and the neighborhood features.

[0062] As shown in the figure, the feature extraction part is the adaptive Laplacian backbone part of the adaptive Laplacian convolution network, which is stacked by 6 convolution modules with different parameters, and multi-scale feature fusion is performed between the convolution modules to extract more comprehensive local information. The convolution module not only includes linear layers, convolution operations and activation functions, but also combines a vector attention mechanism. Figure 3As shown in the figure, it is a specific flow chart of vector attention. Vector attention enables the network to adaptively learn the features of different local areas. The weights will continue to learn and change according to training instead of being fixed to the default value. This helps to capture the complex local structure information in the 3D point cloud, thereby improving the performance of the network.

[0063] The specific steps of vector attention are: randomly generate a set of parameters op that obey the standard normal distribution, normalize op and convert them into learnable parameters. a By the op and N after the softmax function cm The dot product is obtained. And N cm With all center point sets U cm ={u c1 ,u c2 ,...,u cm}, and spliced ​​together through the fully connected layer. Finally, the output F on , as shown in the formula:

[0064] F on =Conv1d(concat(U cm ,sum(W a N cm )))

[0065] Among them, sum is a summation operation, concat is a concatenate operation, which is used to splice along the vector dimension. The spliced ​​data is then passed through Conv1d, which is a conventional one-dimensional convolution operation. Vector attention uses the attention method to calculate the weight;

[0066] The specific implementation of the Laplace feature convolution mechanism: the input of the Laplace convolution module is denoted as F in , the weighted sum of the Laplace features learned by training the neural network is expressed as F a , F a and F in The relationship between can be expressed by the following formula:

[0067] F a =∑softmax(θ·(N cm -u cm ))·(N cm -u cm )

[0068] F in =N cm -u cm

[0069] where N cm is the neighborhood, u cm is the center point, θ=(θ1,θ2,..,θm ) is the parameter learned by the neural network, and θ m represents the mth learned parameter. The Laplace feature convolution mechanism is equivalent to using a neural network layer to infer the weight, and is suitable for the corresponding point cloud classification and segmentation task. Compared with the traditional Laplace convolution method, its discretization method is more flexible and no longer depends on uniform discretization or specific geometric information. The Laplace feature convolution mechanism introduces an adaptive strategy in deep learning, which not only is equivalent to using a traditional neural network layer to infer the weight, but also introduces higher flexibility and accuracy in the point cloud classification and segmentation task. The innovation of this method lies in that it adopts a completely different discretization method from the prior art, abandoning the dependence on uniform discretization or specific geometric information, thereby realizing a deeper understanding of complex geometric shapes, so that the model can dynamically strengthen the relationship between the center point features and the neighborhood features in the ball domain group according to the characteristics of specific data. Compared with the traditional method, it shows stronger robustness and generalization ability.

[0070] 5) Design an enhanced module containing similarity matching calculation to further enhance the extracted feature vector and improve the accuracy of network training;

[0071] The relationship between the neighborhood features and the true value point cloud category label is calculated by similarity matching, and then the vector features obtained by the adaptive Laplace backbone part are linearly interpolated to realize the enhancement of the features. The similarity matching calculation formula is as follows:

[0072] L pre =exp(-γ·(1-∑F s ·L t T ))·L t T

[0073] In the formula, F s is the feature for similarity matching calculation, L t is the true value point cloud category label, L t T is the matrix transpose of L t , and γ is the matching coefficient measured to control the sensitivity of the matching calculation process. The linear interpolation formula is as follows:

[0074] L cls =λ q ·L pre +λ n ·F b

[0075] In the formula, L pre is the result of similarity matching calculation, and F bis the sum of the features extracted from the adaptive Laplace backbone, and the two are linearly interpolated to obtain the final prediction value L cls , λ q and λ n are parameters determined to control the degree of interpolation of the two, so as to achieve the regulation effect of feature enhancement; by introducing the enhancement module of similarity matching calculation, the real label information and the extracted features can be effectively integrated, so that the network can distinguish the importance of the features in the training process. This strategy not only improves the accuracy of feature representation, but also enhances the sensitivity of the network to complex data information.

[0076] 6) Using the feature vector to train the adaptive Laplace convolution network, using multi-class cross-entropy as the loss function, and outputting the corresponding point cloud category, wherein the adaptive Laplace convolution network is an improved Laplace convolution network, which is improved on the basis of the original Laplace convolution network by adding vector attention, adaptively obtaining Laplace weight, and increasing the multi-scale fusion mode of skip connection; input the point cloud, and use the spherical domain group obtained by grouping sampling to extract features respectively, to obtain local and global features of the point cloud data, so that it obtains more perfect overall features in the corresponding space;

[0077] The adaptive Laplace convolution network mainly processes point cloud segmentation and classification, and the loss function adopted is multi-class cross-entropy to minimize the distance between the predicted value and the true value point cloud category label; the output of the network is a probability value indicating the probability of belonging to N categories; for multi-classification problems, the true value point cloud category label is a one-hot encoded vector indicating which category the sample belongs to; the output of the entire network is a vector indicating the probability of each category, which is processed by the softmax function; assuming that there are N points, each point has C categories, the probability distribution output by the network is P, and the true value point cloud category label is Y; the loss function for point cloud classification is represented as:

[0078]

[0079] In the formula, N is the number of points, C is the number of categories, Y ij is the true label of point i belonging to category j, and the probability of the network predicting that the point belongs to category j is P ij , and category j is the output of the network. The trained adaptive Laplace convolution network can be used to complete the classification and segmentation of point clouds.

[0080] 7) Apply the trained adaptive Laplace convolution network to input the 3D point cloud data to be tested into the network to obtain the category to which the point cloud belongs and the part segmentation effect, and the specific steps are as follows:

[0081] 71) Scan the object that needs to be classified and segmented by the sensor, obtain the complete 3D point cloud data of the object to be measured, and send the obtained 3D point cloud data to the trained adaptive Laplacian convolution network after preprocessing;

[0082] 72) After the processing of step 71), the obtained point cloud data is sampled and grouped to obtain different spherical domain groups, corresponding features are extracted according to different spherical domain groups, deep local feature extraction is performed using a Laplacian feature convolution mechanism, and finally a feature enhancement module is used to complete the enhancement of the extracted global and local features;

[0083] 73) After the feature extraction of the above step 72), the category of the object to be measured is output for the classification task; for the segmentation task, the segmentation effect of each component of the object to be measured is output.

[0084] The experimental results of the experiment are described in detail as follows:

[0085] According to the final results of the network, the accuracy and speed of the improved network are evaluated from the overall accuracy (OA), the average class accuracy (mAcc), the average class intersection over union (mIoU), the parameter quantity (Params) and the floating point operation size (FLOPs) indicators.

[0086] The model size of the adaptive Laplacian convolution network and other networks is shown in Table 1.

[0087] Table 1

[0088] Method Params. (M) FLOPs (G) PointNet 3.5 0.9 PointConv 18.6 - KpConv 15.2 - PointMLP 13.2 31.3 Adaptive Laplacian Convolutional Network 4.5 10.5

[0089] The accuracy comparison results of the adaptive Laplacian convolution network and other algorithms are shown in Tables 2 and 3.

[0090] Table 2

[0091] Method mAcc (%) OA (%) PointNet 63.4 68.2 SpiderCNN 69.8 73.7 PointNet++ 75.4 77.9 PointCNN 75.1 78.5 Dgcnn 73.6 78.1 DRNet 78.0 80.3 GBNet 77.8 80.5 PRANet 79.1 82.1 Adaptive Laplacian Convolutional Network 82.1 84.2

[0092] Table 3

[0093]

[0094]

[0095] Wherein the input represents the dimensionality contained in the input to the network, xyz represents the information contained in the three-dimensional coordinates (x, y, z), and nr represents the information contained in the normal vector coordinates (n x ,n y ,n z ). The input point number represents the number of points sampled after preprocessing.

[0096] The results of the above table show that the adaptive Laplacian convolution network has obvious advantages in precision and speed compared with the traditional point cloud feature extraction classification segmentation network PointNet, PointNet++, and Dgcnn; in addition, in terms of model size, the adaptive Laplacian convolution network is more lightweight; in general, the classification and segmentation accuracy of the present application is higher, and the real-time performance is also better.

[0097] The adaptive Laplacian convolution network segmentation instance effect and true value comparison results are as shown in the figure Figure 4 As can be seen from the figure, the components are clearly hierarchical, the predicted category is correct, and it is basically close to the fact.

[0098] The above embodiments are the preferred embodiments of the present application, but the embodiments of the present application are not limited to the above embodiments, and any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present application should be equivalent replacement methods, and are all included in the protection scope of the present application.

Claims

1. A 3D point cloud classification segmentation method based on an adaptive Laplacian convolution network, characterized in that, The method comprises the following steps: 1) using a public 3D point cloud dataset containing true value point cloud category labels and a plurality of point clouds, and preprocessing the dataset to obtain point cloud data with uniform quantity and uniform distribution; 2) using the farthest point sampling algorithm and the ball query algorithm, the point cloud data obtained after preprocessing is sampled and grouped, and the point cloud data is divided into a plurality of ball domain groups, each ball domain group has a center point, and each ball domain group contains all point clouds in a ball formed with a specified radius; 3) feature extraction is performed on the plurality of ball domain groups obtained after sampling and grouping, and corresponding feature vectors are extracted, including center point features and neighborhood features, wherein the neighborhood features refer to the features of the point clouds other than the center point in the ball domain group; 4) a Laplace feature convolution mechanism containing vector attention is designed, the feature vectors are adaptively assigned attention weights using the Laplace feature convolution mechanism, and the center point features and neighborhood features of the point clouds in the ball domain group are respectively strengthened according to the attention weight size, realizing convolution between the center point features and the neighborhood features; 5) an enhanced module containing similarity matching calculation is designed to further enhance the extracted feature vectors to improve the accuracy of network training; 6) an adaptive Laplace convolution network is trained using the feature vectors, a multi-class cross-entropy is used as a loss function, and a corresponding point cloud category is output; wherein the adaptive Laplace convolution network is an improved Laplace convolution network, which is improved by adding vector attention to the original Laplace convolution network, adaptively obtaining Laplace weight, increasing a multi-scale fusion mode of skip connection, inputting point clouds, using the ball domain groups obtained by grouping sampling to perform feature extraction, obtaining local and global features of the point cloud data, and making the point cloud data obtain more perfect overall features in the corresponding space; 7) the trained adaptive Laplace convolution network is applied, 3D point cloud data to be tested is input into the network, and a category to which the point cloud belongs and a component segmentation effect are obtained.

2. The 3D point cloud classification segmentation method based on the adaptive Laplacian convolution network according to claim 1, characterized in that: In step 2), the point cloud data obtained after preprocessing is sampled and grouped, and the specific process is as follows: Input point cloud data U={u1,u2,...,u n }, where u n Represents the nth point cloud; the center point set U cm ={u c1 ,u c2 ,...,u cm } is selected using the iterative farthest point sampling algorithm, where u cm represents the mth center point; U cm = FPS{U}, for the sampled center point set U cm , use the ball query algorithm to determine each center point set U according to the distance cm Neighborhood N cm ={u cm |u m1 ,u m2 ,...,u mj }, where u mj Indicates u cm The adjacent points of the jth point as the center, N cm Indicates u cm The spherical region group with the center point N cm =queryBall{U cm }, N cm By U cm Obtained through the ball query algorithm.

3. The 3D point cloud classification segmentation method based on the adaptive Laplacian convolution network according to claim 2, characterized in that: In step 3), the adaptive Laplace backbone part of the adaptive Laplace convolution network is used to extract features of the ball domain group, and the extracted feature vectors include center point features and neighborhood features; The adaptive Laplace backbone part is stacked by six convolution modules with different parameters, and multi-scale feature fusion is used between these convolution modules to ensure that more comprehensive local information is extracted, and the convolution module not only includes a linear layer, a convolution operation and an activation function, but also combines a vector attention mechanism to improve the effect and precision of feature extraction.

4. The 3D point cloud classification segmentation method based on the adaptive Laplacian convolution network according to claim 3, characterized in that: In step 4), the specific implementation of the vector attention and the Laplace feature convolution mechanism is as follows: Vector attention: randomly generate a set of parameters op subject to standard normal distribution, normalize op and convert it into learnable parameters, weight W a is obtained by the dot product of op and N cm , and N cm is the normalized vector of the center point set U cm ={u c1 ,u c2 ,...,u cm} through a fully connected layer, and finally the output F on is obtained, as shown in the following formula: F on = Conv1d(concat(U cm , sum(W a N cm ))) In the formula, sum is a summation operation, concatenate is a concatenate operation, and the concatenated data is subjected to Conv1d, which is a one-dimensional convolution operation, and the vector attention is calculated by an attention method; Laplacian feature convolution mechanism: the input of the Laplacian convolution module is denoted as F in , the weighted sum of the Laplacian features learned by the neural network is denoted as F a , the relationship between F a and F in is expressed by the following formula: F a =∑ softmax(θ·(N cm -u cm ))·(N cm -u cm ) F in = N cm - u cm In the formula, θ=(θ1, θ2,.., θ m ) is a parameter learned by the neural network, and θ m represents the mth parameter learned; the Laplace feature convolution mechanism introduces an adaptive strategy in deep learning, so that it is not only equivalent to using a neural network layer to infer the weight, but also introduces higher flexibility and accuracy in the point cloud classification and segmentation task, and it adopts a completely different discretization method from the prior art, abandoning the dependence on uniform discretization or specific geometric information, thereby achieving a deeper understanding of complex geometric shapes, so that the network can dynamically strengthen the relationship between the center point features and the neighborhood features in the ball domain group according to the characteristics of the specific data.

5. The 3D point cloud classification segmentation method based on adaptive Laplacian convolution network according to claim 4, characterized in that: In step 5), the specific implementation of the enhanced module is as follows: The relationship between the neighborhood features and the true value point cloud category label is calculated by similarity matching, and then linear interpolation is performed with the vector features obtained by the adaptive Laplace backbone part to enhance the features. The similarity matching calculation formula is as follows: L pre = exp(-γ · (1 - ∑F s · L t T )) · L t T where F s is the feature for similarity matching computation, L t is the ground truth point cloud class label, L t T is the matrix transpose of L t , and γ is the matching coefficient determined to control the sensitivity of the matching computation process, and the linear interpolation formula is as follows: L cls = λ q · L pre + λ n · F b In the formula, L pre is the result of similarity matching calculation, F b is the total feature extracted by adaptive Laplace backbone, and the two are linearly interpolated to obtain the final prediction value L cls , λ q and λ n are measured parameters, respectively used to control the degree of interpolation, so as to realize the regulation effect of feature enhancement; by introducing the enhancement module of similarity matching calculation, the real label information and the extracted features can be effectively integrated, so that the network can distinguish the importance of the features in the training process.

6. The 3D point cloud classification segmentation method based on an adaptive Laplacian convolution network according to claim 5, characterized in that: In step 6), the adaptive Laplace convolution network processes point cloud segmentation and classification, and a multi-class cross-entropy loss function is used to minimize the distance between the predicted value and the true value point cloud category label; the output of the network is a probability value indicating the probability of belonging to N categories; for a multi-class problem, the true value point cloud category label is a one-hot encoded vector indicating which category the sample belongs to; the output of the entire network is a vector indicating the probability of each category, which is processed by the softmax function; assuming there are N points, each point has C categories, the probability distribution output by the network is P, and the true value point cloud category label is Y; the loss function for point cloud classification is represented as: In the formula, N is the number of points, C is the number of categories, Y ij is the true label of the point i belonging to the category j, and the probability of the network predicting that the point belongs to the category j is P ij , the category j is taken as the output of the network, and the trained adaptive Laplacian convolution network can be used to complete the classification and segmentation task of the point cloud.

7. The 3D point cloud classification segmentation method based on the adaptive Laplacian convolution network according to claim 6, characterized in that: In step 7), the specific steps of applying the trained adaptive Laplace convolution network to complete 3D point cloud classification and segmentation are as follows: 71) Scan the object to be classified and segmented by the sensor to obtain the complete 3D point cloud data of the object to be measured, and then pre-process the obtained 3D point cloud data and input it into the trained adaptive Laplace convolution network; 72) After step 71), sample and group the obtained point cloud data to obtain different spherical domain groups, extract the corresponding features according to different spherical domain groups, use the Laplace feature convolution mechanism to extract deep local features, and finally use the feature enhancement module to enhance the extracted global and local features; 73) After the feature extraction in step 72) above, output the category of the object to be measured for the classification task; for the segmentation task, output the segmentation effect of each component of the object to be measured.

Images