A universal three-dimensional point cloud model classification method and system

Abstract

The present invention relates to the field of deep learning and computer vision technology, and in particular to a universal three-dimensional point cloud model classification method and system, wherein the method includes S1, obtaining several key areas of the three-dimensional point cloud model; S2, generating a dynamic convolution kernel suitable for the three-dimensional point cloud model through a weight library calculation module, a score calculation module and a dynamic convolution kernel construction module; S3, performing feature extraction on several key areas of the three-dimensional point cloud model respectively through the dynamic convolution kernel, and obtaining local classification results through a SoftMax function; S4, fusing the local classification results into the final overall classification results of the three-dimensional point cloud model through a key area voting algorithm. The present invention simplifies the global classification problem into multiple local classification problems, can overcome rigid changes and non-rigid changes at the same time, and reduces the difficulty of subsequent neural network training. In addition, the local area classification results are reasonably utilized to make the classification effect more significant.

Description

Technical Field

[0001] The present invention relates to the field of deep learning and computer vision technology, and in particular to a universal three-dimensional point cloud model classification method and system. Background Art

[0002] With the development of artificial intelligence technology, the feature extraction technology of 3D point cloud models has made great progress. It is also a key technology in the fields of target classification, autonomous driving, robot navigation, augmented reality / virtual reality, etc. According to different composition forms, 3D point cloud models can be divided into models such as views, voxels, point clouds, grids, octrees, etc. At present, there are many deep learning neural networks based on different modalities, and they have shown excellent performance. For the problem of target recognition and classification of 3D point cloud models, due to the disorder and irregularity of point clouds (the basic expression of 3D data) that are difficult to handle, the early deep learning methods for 3D data analysis and understanding first converted point clouds into regular views, voxels and other modalities, and then used a more mature neural network framework to extract features to complete specific tasks, which led to the loss of geometric information during the conversion process and limited the performance of the network. In order to avoid this problem, some methods for directly processing point cloud data have been proposed, and have led to more and more subsequent research on raw point cloud data.

[0003] In addition, according to whether it is easy to deform, the 3D point cloud model can be divided into a rigid 3D point cloud model (not easy to deform) and a non-rigid 3D point cloud model (easy to deform). The non-rigid 3D point cloud model contains twelve non-rigid transformations, which makes the feature description, classification and retrieval of the 3D shape more complicated and faces more difficulties. In addition, the current advanced deep learning methods for rigid 3D point cloud models are not applicable to non-rigid 3D point cloud models. Summary of the invention

[0004] The first purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art and provide a universal three-dimensional point cloud model classification method. First, by using the farthest geodesic distance to extract several key areas of the three-dimensional point cloud model, the global classification problem is simplified into multiple local classification problems, which can overcome both rigid changes and non-rigid changes, and reduce the difficulty of subsequent neural network training. In addition, compared with the Euclidean distance, the use of geodesic distance also makes the division of these key areas more in line with the rules of manifolds. Secondly, by using a network based on a multi-layer perceptron MLP, feature extraction is made more flexible and more learnable. Finally, the key area voting algorithm is used to reasonably utilize the local area classification results, increase the gap between classes, and make the classification effect more significant.

[0005] The second object of the present invention is to provide a universal three-dimensional point cloud model classification system.

[0006] The first object of the present invention is achieved by the following technical solution: A universal three-dimensional point cloud model classification method comprises the following steps:

[0007] S1. Obtain several key areas of the three-dimensional point cloud model;

[0008] S2, generating a dynamic convolution kernel suitable for a three-dimensional point cloud model through a weight library calculation module, a score calculation module and a dynamic convolution kernel construction module;

[0009] S3, using the constructed dynamic convolution kernel to extract features from several key areas of the three-dimensional point cloud model, and obtaining local classification results through the SoftMax function;

[0010] S4. Through the key area voting algorithm, the local classification results are fused into the final overall classification results of the 3D point cloud model.

[0011] Further, the step S1 comprises the following steps:

[0012] S101. Randomly select a point of the 3D point cloud model As a starting point;

[0013] S102, As the center, take the geodesic distance from it as in the area point or the point with the largest curvature points, forming a key area of ​​the 3D point cloud model ;

[0014] S103, Calculation The geodesic distance to each point of the 3D point cloud model is obtained dimensional geodesic distance set , and take out the point with the maximum geodesic distance ,in is the number of points in the 3D point cloud model;

[0015] S104, As the center, take the geodesic distance from it as in the area point or the point with the largest curvature points, forming a key area of ​​the 3D point cloud model ;

[0016] S105, calculate all points in the 3D point cloud model The geodesic distance for Points , and If the geodesic distance is less than The first Bit , that is, point and The geodesic distance between the point and The geodesic distance of dimensional geodesic distance set L The Bit , will soon and The geodesic distance of is updated to and The geodesic distance of

[0017] S106, Select dimensional geodesic distance set The point corresponding to the maximum value is ;

[0018] S107 . Repeat steps S103 to S106 until a set number of key areas are obtained.

[0019] Further, the step S2 comprises the following steps:

[0020] S201. Define weight matrix library ,in Is the size of The weight matrix of It is a weight matrix library Medium weight matrix the number of

[0021] S202, obtaining several evenly distributed sampling points in the key area through average sampling ,in Indicates the number of sampling points;

[0022] S203, for each sampling point, select several discrete neighboring points with the smallest Euclidean distance to it ;

[0023] S204, for each group is expressed as Point pairs, use a multi-layer perceptron to learn the positional relationship between discrete neighborhood points;

[0024] S205: For the positional relationship output in step S204, the fractional coefficient is output through the SoftMax function. , where the SoftMax function is used to ensure the output fractional coefficient The range is , that is, each weight matrix is ​​associated with a discrete neighborhood point The relationship between is taken as probability. The larger the probability, the closer the weight matrix is ​​to the discrete neighborhood points. The closer the relationship is;

[0025] S206: The weight matrix library defined in step S201 is All weight matrices of and the fractional coefficients outputted in step S205 are obtained by:

[0026]

[0027] Combined to generate a dynamic convolution kernel used to calculate point cloud features; represents the output dynamic convolution kernel, represents the fractional coefficient, represents the weight matrix, Indicates its quantity.

[0028] Furthermore, in step S3, several key areas of the three-dimensional point cloud model are taken as input, and features are extracted respectively using the constructed dynamic convolution kernel, and the local classification results of the key areas are obtained through the SoftMax function.

[0029] Further, the step S4 comprises the following steps:

[0030] S401, Generate elements and all are List ,in Indicates the number of categories, list Indicates that the probability of each category in the initial prediction is the same;

[0031] S402: Count the number of times a certain category appears as the maximum classification probability in the local classification results of different key areas, and generate a list ;

[0032] S403, list With list Multiply bitwise, update list , as a result of voting;

[0033] S404, in order to ensure the probability of output In the list The SoftMax function is sent to obtain the overall classification probability of the 3D point cloud model, and the maximum probability result is used as the final overall classification result of the 3D point cloud model.

[0034] The second objective of the present invention is achieved by the following technical solution: A universal three-dimensional point cloud model classification system, comprising:

[0035] The key area acquisition unit is used to acquire several key areas of the three-dimensional point cloud model; wherein, firstly, a point of the three-dimensional point cloud model is randomly selected. As the starting point, then calculate the remaining points and The geodesic distance of Maximum in range point or the point with the greatest curvature point, as The key area neighborhood points are then calculated and the remaining points in the 3D point cloud model are The geodesic distance of , and by calculating the geodesic distances of the remaining points from this point, the geodesic distances in the range The maximum point or the point with the greatest curvature points; finally, repeat the above steps until the set number of key areas are obtained;

[0036] A dynamic convolution kernel construction unit, used to construct a dynamic convolution kernel for calculating point cloud features; wherein the dynamic convolution kernel construction unit includes a weight library calculation module, a score calculation module and a dynamic convolution kernel construction module;

[0037] The weight library calculation module is used to first define the weight matrix library ,in Is the size of The weight matrix of It is a weight matrix library Medium weight matrix The number of; through average sampling, obtain several evenly distributed sampling points in the key area ,in Represents the number of sampling points; for each sampling point, select several discrete neighboring points with the smallest Euclidean distance to it ; For each group, it is represented as Point pairs, use a multi-layer perceptron to learn the positional relationship between discrete neighborhood points;

[0038] The score calculation module is used to calculate the position relationship output by the weight library calculation module, and output the score coefficient through the SoftMax function. , where the SoftMax function is used to ensure the output fractional coefficient The range is , that is, each weight matrix is ​​associated with a discrete neighborhood point The relationship between is taken as probability. The larger the probability, the closer the weight matrix is ​​to the discrete neighborhood points. The closer the relationship is;

[0039] The dynamic convolution kernel construction module is used to All weight matrices and the score coefficients output by the score calculation module are obtained by:

[0040]

[0041] To combine and generate dynamic convolution kernels used to calculate point cloud features; among them, represents the output dynamic convolution kernel, represents the fractional coefficient, represents the weight matrix, Indicates its quantity;

[0042] The local classifier unit is used to process the key area features of the 3D point cloud model to obtain the key area category of the 3D point cloud model; wherein, the constructed dynamic convolution kernel is used to extract the features of the key areas of several 3D point cloud models respectively, and then the classification probability of the key areas of the 3D point cloud model is obtained through the SoftMax function, and finally the maximum probability result is used as the final local classification result of the key area;

[0043] The global classifier unit is used to process the local classification results of the key areas of the 3D point cloud model, merge them into the final overall classification results, and obtain the category of the 3D point cloud model; wherein, first generate elements and all are List ,in is the number of categories. Then, the number of times a certain category appears as the maximum classification probability in the local classification results of different key areas is counted to generate a list , then the list With list Multiply bitwise, update list , as the voting result, and sent to the SoftMax function to obtain the overall classification probability of the 3D point cloud model, and the maximum probability result is used as the final overall classification result of the 3D point cloud model.

[0044] One of the above technical solutions has the following beneficial effects:

[0045] (1) The present invention designs a universal 3D point cloud model classification method that uses the farthest geodesic distance to extract several key areas of the 3D point cloud model, simplifies the global classification problem into multiple local classification problems, and can simultaneously overcome rigid changes and non-rigid changes, thereby reducing the difficulty of subsequent neural network training. In addition, compared with the Euclidean distance, the use of the geodesic distance also makes the division of these key areas more in line with the rules of the manifold, which enhances the expression ability of the key areas to a certain extent and enhances the spatial information of the 3D point cloud.

[0046] (2) In the feature extraction network part and the local classifier part, the multi-layer perceptron (MLP) is used rationally, and the convolution kernel is designed through dynamic data-driven, so that it can adaptively learn the relational features from the positions of different neighborhood points and effectively overcome the differences between different key areas. It has greater flexibility for the irregularity and disorder of point clouds.

[0047] (3) In the global classifier part, the key area voting algorithm is used to make statistics on the classification of each area, and the initial probability is updated in the form of a product. The local area classification results are reasonably utilized to increase the gap between classes, which improves the difference between different categories and makes the classification effect more significant. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flowchart of a universal three-dimensional point cloud model classification method of the present invention;

[0049] Figure 2 A schematic diagram of a process for extracting key areas of a universal three-dimensional point cloud model classification method of the present invention;

[0050] Figure 3 A schematic diagram of a flow chart of a feature extraction network of a universal three-dimensional point cloud model classification method of the present invention;

[0051] Figure 4 A schematic diagram of a process of voting for key areas of a universal three-dimensional point cloud model classification method of the present invention;

[0052] Figure 5 It is a schematic diagram of the architecture of the universal three-dimensional point cloud model classification system of the present invention; DETAILED DESCRIPTION

[0053] The technical solution of the present invention is further described below with reference to the accompanying drawings and through specific implementation methods.

[0054] See also Figure 1 As shown, the universal three-dimensional point cloud model classification method provided in this embodiment includes the following steps:

[0055] S1. Obtain several key areas of the 3D point cloud model, see Figure 2 As shown, the specific process is as follows:

[0056] S101. Randomly select a point of the 3D point cloud model As a starting point;

[0057] S102, As the center, take the geodesic distance from it as in the area point or the point with the greatest curvature points, forming a key area of ​​the 3D point cloud model ;

[0058] S103, Calculation The geodesic distance to each point of the 3D point cloud model is obtained dimensional geodesic distance set , and take out the point with the maximum geodesic distance ,in is the number of points in the 3D point cloud model;

[0059] S104, As the center, take the geodesic distance from it as in the area point or the point with the greatest curvature points, forming a key area of ​​the 3D point cloud model ;

[0060] S105, calculate all points in the 3D point cloud model The geodesic distance for Points , and If the geodesic distance is less than dimensional geodesic distance set L The Bit , that is, point and The geodesic distance between the point and The geodesic distance of dimensional geodesic distance set L The Bit , will soon and The geodesic distance of is updated to and The geodesic distance of

[0061] S106, Select dimensional geodesic distance set The point corresponding to the maximum value is ;

[0062] S107 . Repeat steps S103 to S106 until a set number of key areas are obtained.

[0063] Steps S101-S107 split the three-dimensional point cloud model into several parts, while overcoming rigid changes and non-rigid changes, reducing the difficulty of subsequent network training, and making the method universal.

[0064] S2. Generate a dynamic convolution kernel suitable for the three-dimensional point cloud model through the weight library calculation module, the score calculation module and the dynamic convolution kernel construction module; see Figure 3 As shown, the specific process is as follows:

[0065] S201. Define weight matrix library ,in Is the size of The weight matrix of It is a weight matrix library Medium weight matrix the number of

[0066] S202, obtaining several evenly distributed sampling points in the key area through average sampling ,in Indicates the number of sampling points;

[0067] S203, for each sampling point, select several discrete neighboring points with the smallest Euclidean distance to it ;

[0068] S204, for each group is expressed as Point pairs, use a multi-layer perceptron to learn the positional relationship between discrete neighborhood points;

[0069] S205: For the positional relationship output in step S204, the fractional coefficient is output through the SoftMax function. , where the SoftMax function is used to ensure the output fractional coefficient The range is , that is, each weight matrix is ​​associated with a discrete neighborhood point The relationship between is taken as probability. The larger the probability, the closer the weight matrix is ​​to the discrete neighborhood points. The closer the relationship is;

[0070] S206: The weight matrix library defined in step S201 is All weight matrices of and the fractional coefficients outputted in step S205 are obtained by:

[0071]

[0072] Combined with the dynamic convolution kernel used to calculate the three-dimensional point cloud model; represents the output dynamic convolution kernel, represents the fractional coefficient, represents the weight matrix, Indicates its quantity.

[0073] S3. Use the constructed dynamic convolution kernel to extract features from several key areas of the 3D point cloud model, and obtain the local classification results of the key areas through the SoftMax function. The specific process is as follows:

[0074] Taking several key areas of the extracted 3D point cloud model as input, the constructed dynamic convolution kernel is used to extract features respectively, and the classification probability of the key areas of the 3D point cloud model is obtained through the SoftMax function. Finally, the maximum probability result is taken as the final local classification result of the key area.

[0075] S4, through the key area voting algorithm, the local classification results are integrated into the final overall classification result of the 3D point cloud model, see Figure 4 As shown, the specific process is as follows:

[0076] S401, Generate elements and all are List ,in Indicates the number of categories, list Indicates that the probability of each category in the initial prediction is the same;

[0077] S402: Count the number of times a certain category appears as the maximum classification probability in the local classification results of different key areas, and generate a list ;

[0078] S403, list With list Multiply bitwise, update list , as a result of voting;

[0079] S404, in order to ensure the probability of output In the list The SoftMax function is input to obtain the overall classification probability of the 3D point cloud model, and the maximum probability result is used as the final overall classification result of the 3D point cloud model.

[0080] See also Figure 5 As shown, this embodiment also provides a universal three-dimensional point cloud model classification system, including:

[0081] The key area acquisition unit is used to acquire several key areas of the three-dimensional point cloud model; wherein, firstly, a point of the three-dimensional point cloud model is randomly selected. As the starting point, then calculate the remaining points and The geodesic distance of Maximum in range point or the point with the largest curvature point, as The key area neighborhood points are then calculated and the remaining points in the 3D point cloud model are The geodesic distance of , and by calculating the geodesic distances of the remaining points from this point, the geodesic distances in the range The maximum point or the point with the greatest curvature points; finally, repeat the above steps until the set number of key areas are obtained;

[0082] A dynamic convolution kernel construction unit, used to construct a dynamic convolution kernel for calculating point cloud features; wherein, the dynamic convolution kernel construction unit includes a weight library calculation module, a score calculation module and a dynamic convolution kernel construction module;

[0083] The weight library calculation module is used to first define the weight matrix library ,in Is the size of The weight matrix of It is a weight matrix library Medium weight matrix The number of; through average sampling, obtain several evenly distributed sampling points in the key area ,in Represents the number of sampling points; for each sampling point, select several discrete neighboring points with the smallest Euclidean distance to it ; For each group, it is represented as Point pairs, use a multi-layer perceptron to learn the positional relationship between discrete neighborhood points;

[0084] The score calculation module is used to calculate the position relationship output by the weight library calculation module, and output the score coefficient through the SoftMax function. , where the SoftMax function is used to ensure the output fractional coefficient The range is , that is, each weight matrix is ​​associated with a discrete neighborhood point The relationship between is taken as probability. The larger the probability, the closer the weight matrix is ​​to the discrete neighborhood points. The closer the relationship is;

[0085] The dynamic convolution kernel construction module is used to All weight matrices and the score coefficients output by the score calculation module are obtained by:

[0086]

[0087] To combine and generate dynamic convolution kernels used to calculate point cloud features; among them, represents the output dynamic convolution kernel, represents the fractional coefficient, represents the weight matrix, Indicates its number; the local classifier unit is used to process the key area features of the 3D point cloud model to obtain the key area category of the 3D point cloud model; wherein, the constructed dynamic convolution kernel is used to extract features of the key areas of several 3D point cloud models respectively, and then the classification probability of the key area of ​​the 3D point cloud model is obtained through the SoftMax function, and finally the maximum probability result is used as the final local classification result of the key area;

[0088] The global classifier unit is used to process the local classification results of the key areas of the 3D point cloud model, merge them into the final overall classification results, and obtain the category of the 3D point cloud model; wherein, first generate elements and all are List ,in is the number of categories. Then, the number of times a certain category appears as the maximum classification probability in the local classification results of different key areas is counted to generate a list , then the list With list Multiply bitwise, update list , as the voting result, and sent to the SoftMax function to obtain the overall classification probability of the 3D point cloud model, and the maximum probability result is used as the final overall classification result of the 3D point cloud model.

[0089] The technical principle of the present invention is described above in conjunction with specific embodiments. These descriptions are only for explaining the principle of the present invention and cannot be interpreted as limiting the scope of protection of the present invention in any way. Based on the explanations herein, those skilled in the art can think of other specific embodiments of the present invention without creative work, and these equivalent variations or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A universal three-dimensional point cloud model classification method, characterized in that: The following steps are involved: S1. Obtain several key areas of the three-dimensional point cloud model; S2, generating a dynamic convolution kernel suitable for a three-dimensional point cloud model through a weight library calculation module, a score calculation module and a dynamic convolution kernel construction module; S3, extract features from several key areas of the 3D point cloud model through dynamic convolution kernels, and obtain local classification results through the SoftMax function; S4, through the key area voting algorithm, the local classification results are integrated into the final overall classification results of the 3D point cloud model; The step S1 comprises the following steps: S101. Randomly select a point of the 3D point cloud model As a starting point; S102, As the center, take the geodesic distance from it as in the area point or the point with the largest curvature points, forming a key area of ​​the 3D point cloud model ; S103, Calculation The geodesic distance to each point of the 3D point cloud model is obtained dimensional geodesic distance set , and take out the point with the maximum geodesic distance ,in is the number of points in the 3D point cloud model; S104, As the center, take the geodesic distance from it as in the area point or the point with the largest curvature points, forming a key area of ​​the 3D point cloud model ; S105, calculate all points in the 3D point cloud model The geodesic distance for Points , and If the geodesic distance is less than The first Bit , that is, point and The geodesic distance between the point and The geodesic distance of dimensional geodesic distance set L The Bit , will soon and The geodesic distance of is updated to and The geodesic distance of S106, Select dimensional geodesic distance set The point corresponding to the maximum value is ; S107, repeating steps S103 to S106 until a set number of key areas are obtained; The step S2 comprises the following steps: S201. Define weight matrix library ,in Is the size of The weight matrix of It is a weight matrix library Medium weight matrix the number of S202, obtaining several evenly distributed sampling points in the key area through average sampling ,in Indicates the number of sampling points; S203, for each sampling point, select several discrete neighboring points with the smallest Euclidean distance to it ; S204, for each group is expressed as Point pairs, use a multi-layer perceptron to learn the positional relationship between discrete neighborhood points; S205: For the positional relationship output in step S204, the fractional coefficient is output through the SoftMax function. , where the SoftMax function is used to ensure the output fractional coefficient The range is , that is, each weight matrix is ​​associated with a discrete neighborhood point The relationship between is taken as probability. The larger the probability, the closer the weight matrix is ​​to the discrete neighborhood points. The closer the relationship is; S206: The weight matrix library defined in step S201 is All weight matrices of and the fractional coefficients outputted in step S205 are obtained by: Combined with the dynamic convolution kernel used to calculate the three-dimensional point cloud model; represents the output dynamic convolution kernel, represents the fractional coefficient, represents the weight matrix, Indicates its quantity.

2. A universal three-dimensional point cloud model classification method according to claim 1, characterized in that: Step S3 takes several key areas of the three-dimensional point cloud model as input, uses the constructed dynamic convolution kernel to extract features respectively, and obtains the local classification results of the key areas through the SoftMax function.

3. The universal three-dimensional point cloud model classification method according to claim 1, characterized in that: The step S4 comprises the following steps: S401, Generate elements and all are List ,in Indicates the number of categories, list Indicates that the probability of each category in the initial prediction is the same; S402: Count the number of times a certain category appears as the maximum classification probability in the local classification results of different key areas, and generate a list ; S403, list With list Multiply bitwise, update list , as a result of voting; S404, in order to ensure the probability of output In the list The SoftMax function is sent to obtain the overall classification probability of the 3D point cloud model, and the maximum probability result is used as the final overall classification result of the 3D point cloud model.

4. A universal 3D point cloud model classification system, characterized in that: include: The key area acquisition unit is used to acquire several key areas of the three-dimensional point cloud model; wherein, firstly, a point of the three-dimensional point cloud model is randomly selected. As the starting point, then calculate the remaining points and The geodesic distance of Maximum in range point or the point with the largest curvature point, as The key area neighborhood points are then calculated and the remaining points in the 3D point cloud model are The geodesic distance of , and by calculating the geodesic distances of the remaining points from this point, the geodesic distances in the range The maximum point or the point with the largest curvature points; finally, repeat the above steps until the set number of key areas are obtained; A dynamic convolution kernel construction unit, used to construct a dynamic convolution kernel for calculating point cloud features; wherein the dynamic convolution kernel construction unit includes a weight library calculation module, a score calculation module and a dynamic convolution kernel construction module; The weight library calculation module is used to first define the weight matrix library ,in Is the size of The weight matrix of It is a weight matrix library Medium weight matrix The number of; through average sampling, obtain several evenly distributed sampling points in the key area ,in Represents the number of sampling points; for each sampling point, select several discrete neighboring points with the smallest Euclidean distance to it ; For each group, it is represented as Point pairs, use a multi-layer perceptron to learn the positional relationship between discrete neighborhood points; The score calculation module is used to calculate the position relationship output by the weight library calculation module, and output the score coefficient through the SoftMax function. , where the SoftMax function is used to ensure the output fractional coefficient The range is , that is, each weight matrix is ​​associated with a discrete neighborhood point The relationship between is taken as probability. The larger the probability, the closer the weight matrix is ​​to the discrete neighborhood points. The closer the relationship is; The dynamic convolution kernel construction module is used to All weight matrices and the score coefficients output by the score calculation module are obtained by: To combine and generate dynamic convolution kernels used to calculate point cloud features; among them, represents the output dynamic convolution kernel, represents the fractional coefficient, represents the weight matrix, Indicates its quantity; The local classifier unit is used to process the key area features of the 3D point cloud model to obtain the key area category of the 3D point cloud model; wherein, the constructed dynamic convolution kernel is used to extract the features of the key areas of several 3D point cloud models respectively, and then the classification probability of the key areas of the 3D point cloud model is obtained through the SoftMax function, and finally the maximum probability result is used as the final local classification result of the key area; The global classifier unit is used to process the local classification results of the key areas of the 3D point cloud model, merge them into the final overall classification results, and obtain the category of the 3D point cloud model; wherein, first generate elements and all are List ,in is the number of categories. Then, the number of times a certain category appears as the maximum classification probability in the local classification results of different key areas is counted to generate a list , then the list With list Multiply bitwise, update list , as the voting result, and sent to the SoftMax function to obtain the overall classification probability of the 3D point cloud model, and the maximum probability result is used as the final overall classification result of the 3D point cloud model.

Images