A three-dimensional point cloud data recognition method for neural networks

Abstract

The application provides a three-dimensional point cloud data recognition method for a neural network, comprising: obtaining three-dimensional point cloud data; obtaining a sampling point set of the three-dimensional point cloud data according to the three-dimensional point cloud data; obtaining rotation invariant feature data in the sampling point set when the three-dimensional point cloud data rotates; constructing a plurality of RISurFormer modules for processing the rotation invariant feature data into high-dimensional rotation invariant feature data one by one; constructing a Transformer-Encoder module for optimizing the high-dimensional rotation invariant feature data to obtain optimized high-dimensional rotation invariant feature data; and constructing a classifier for classifying the three-dimensional point cloud data according to the optimized high-dimensional rotation invariant feature data; the method can process three-dimensional point cloud data with arbitrary rotation, is not affected by the posture of a target object, does not need complex data preprocessing, and greatly improves the efficiency, and can be widely applied to neural networks, especially convolutional neural networks with original data as initial input data.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional information technology, and more specifically, to a method for recognizing three-dimensional point cloud data using neural networks. Background Technology

[0002] 3D point clouds are geometric representations of a series of points in space, recording a set of 3D coordinates and various attribute information for each point on the surface of a scanned object, such as texture, material, color, occupancy map, normal vector, reflection intensity, and feature data. 3D point cloud data is a geometric description of real objects and a new 3D model data format. As a primary carrier of information in visual communication scenarios, it can effectively represent static objects and scenes in visual media services and can also render accurate 3D models in real time, realistically describing dynamic objects or scene information. Therefore, 3D point cloud data can bring users an immersive consumption experience that combines virtual and real elements and allows for real-time interaction. With the rapid development of information technology, 3D point clouds have become an important data modality complementing images, audio, and video. As the main input data for tasks such as autonomous navigation, map building, and scene understanding, 3D point clouds have enormous application prospects in many fields, including autonomous driving, mobile robots, virtual / mixed reality, geographic information systems, remote sensing and mapping, and cultural heritage protection, and have become one of the research hotspots of the next generation of artificial intelligence technology.

[0003] With the development of artificial intelligence, a series of deep learning algorithms related to 3D point clouds have emerged, such as PointNet++, DGCNN, PointCNN, and PointMLP. These algorithms have undergone various improvements, from network architecture to convolution methods, resulting in continuous improvements in accuracy and speed, and achieving excellent performance on 3D point cloud datasets for different tasks. However, these methods cannot effectively solve the problem of rotation resistance; that is, when the 3D object in the actual test data rotates, these methods cannot handle it well. To solve this problem, existing solutions often incorporate data augmentation during the training process of neural networks. Specifically, this involves randomly rotating the 3D point cloud data or rotating it along the z-axis to alleviate the accuracy degradation caused by rotation. However, in practical applications, 3D objects can rotate freely along any axis, resulting in an infinite number of rotation possibilities. Therefore, no matter how many random rotations are added during training, it is still impossible to cover all rotation possibilities in reality. Furthermore, data augmentation during training increases training time and reduces efficiency. Summary of the Invention

[0004] The problem solved by this invention is to provide a method for recognizing 3D point cloud data for neural networks, which can process arbitrarily rotated 3D point cloud data, is not affected by the pose of the target object, and does not require complex data preprocessing. It can be widely used in neural networks, especially convolutional neural networks where the initial input data is the original data.

[0005] To address the above problems, this invention provides a method for recognizing 3D point cloud data using neural networks, comprising:

[0006] S1. Obtain 3D point cloud data;

[0007] S2. Obtain the sampling point set of the 3D point cloud data based on the 3D point cloud data;

[0008] S3. When the 3D point cloud data is rotated, obtain the rotation-invariant feature data in the sampling point set;

[0009] S4. Construct several RISurFormer modules to process rotation-invariant feature data into high-dimensional rotation-invariant feature data step by step.

[0010] S5. Construct the Transformer-Encoder module to optimize high-dimensional rotation-invariant feature data and obtain optimized high-dimensional rotation-invariant feature data;

[0011] S6. Construct a classifier to classify 3D point cloud data based on optimized high-dimensional rotation-invariant feature data.

[0012] This invention obtains a set of sampling points from 3D point cloud data. When the 3D point cloud data rotates, rotation-invariant feature data can be extracted from the sampling point set. This rotation-invariant feature data is then processed one by one through each RISurFormer module, and then optimized by the Transformer-Encoder module. Finally, the 3D point cloud data can be effectively classified. This method can be widely applied in neural networks, such as neural networks that require classification tasks, especially convolutional neural networks where the initial input data is the original data.

[0013] Furthermore, in step S2, the sampling point set is obtained using the farthest point sampling algorithm. The advantage of this design is that the farthest point sampling algorithm can cover all points in the entire 3D point cloud data.

[0014] Further, step S3 includes:

[0015] S31. Select any rotation point from the three-dimensional point cloud sampling point set;

[0016] S32. Using the currently selected rotation point as the center point, generate a neighborhood, which includes several neighboring points;

[0017] S33. Select any neighboring point;

[0018] S34. Find the first and second nearest neighbors to the currently selected neighbor;

[0019] S35. Generate a first plane based on the currently selected neighboring point, the first neighboring point, and the currently selected rotation point; generate a second plane based on the currently selected neighboring point, the second neighboring point, and the currently selected rotation point.

[0020] S36. Extract rotation-invariant feature data from the first and second planes;

[0021] S37. Determine if there are any unselected neighboring points in the neighborhood. If there are, select the next neighboring point and return to step S34. If all neighboring points in the neighborhood have been selected, proceed to step S38.

[0022] S38. Determine if there are any unselected rotation points in the sampling point set. If so, select the next rotation point and return to step S32. If all rotation points in the sampling point set have been selected, proceed to step S4.

[0023] The beneficial effect of this design is that, in the rotating 3D point cloud data, based on all rotating points, the neighborhood of each rotating point is constructed in sequence through the above steps, and a plane is established in sequence, which can effectively cover each rotating point and effectively perform mathematical modeling.

[0024] Furthermore, in step S36, the formula for extracting rotation-invariant feature data is as follows:

[0025] , , , , , , , , , , , , , , ;

[0026] Wherein, RISP means selecting the currently chosen neighboring nodes. The operation of mapping to rotation-invariant feature data, where P represents the center point. This indicates the currently selected neighboring point. Indicates the first neighboring point, Let N represent the second neighboring point, and N represent the normal vector. This represents the normal vector of the currently selected neighboring point. Let the normal vector of the first neighbor be denoted as . This represents the normal vector of the second neighboring point. It is a unique length quantity, representing the distance between the center point and the currently selected neighboring points; ∠ represents a vector; ∠(,) represents the angle between the two vectors or the product of the two vectors within the brackets.

[0027] The beneficial effect of this design is that, in the rotated 3D point cloud data, based on all the rotated points, Euclidean space features and tangent space features are extracted from the rotated 3D point cloud data in sequence using the above formula, which can effectively perform mathematical modeling.

[0028] Furthermore, the RISurFormer module in step S4 includes:

[0029] The first multilayer perceptron module is used to map rotation-invariant feature data to a high-dimensional space to obtain initial high-dimensional rotation-invariant feature data.

[0030] The first self-attention module, which is connected to the first multilayer perceptron module, is used to optimize the initial high-dimensional rotation-invariant feature data and obtain the optimized initial high-dimensional rotation-invariant feature data.

[0031] The feature data concatenation module is connected to the first self-attention module. It is used to determine whether there is a previous RISurFormer module. If there is a previous RISurFormer module, the optimized initial high-dimensional rotation-invariant feature data is concatenated with the high-dimensional rotation-invariant feature data output by the previous RISurFormer module to obtain the concatenated initial high-dimensional rotation-invariant feature data. If there is no previous RISurFormer module, the optimized initial high-dimensional rotation-invariant feature data is directly output.

[0032] The second multilayer perceptron module is connected to the feature data concatenation module and is used to summarize the concatenated high-dimensional rotation-invariant feature data to obtain the summarized initial high-dimensional rotation-invariant feature data.

[0033] The pooling module, which is connected to the second multilayer perceptron module, is used to perform pooling operations on the aggregated rotation-invariant feature data to obtain the initial high-dimensional rotation-invariant feature data after pooling.

[0034] The second self-attention module is connected to the pooling module and is used to optimize the rotation-invariant feature data after pooling to obtain high-dimensional rotation-invariant feature data.

[0035] The beneficial effects of this design are as follows: the first multilayer perceptron module can map multiple input rotation-invariant feature datasets onto a single output dataset, relative to the initial aggregation; the first self-attention module can weight the features of the rotation-invariant feature data at different positions on the feature channels, making the features stand out; the feature data concatenation module can concatenate with the high-dimensional rotation-invariant feature data output by the previous-level RISurFormer module, so that the rotation-invariant feature data can be extracted by each RISurFormer module one by one, and the extracted features are all related and progressively simplified; the second multilayer perceptron module is used to aggregate again for subsequent output; the pooling module can compress and simplify the aggregated rotation-invariant feature data again, and highlight the features again through the second self-attention module, making the classification efficiency more efficient. Attached Figure Description

[0036] Figure 1 This is an overall flowchart of the present invention;

[0037] Figure 2 This is a detailed schematic diagram of the RISurFormer module in this invention;

[0038] Figure 3 This is a schematic diagram illustrating the generation of the first and second planes in this invention;

[0039] Figure 4 This is a schematic diagram of rotation-invariant feature data extraction.

[0040] Explanation of reference numerals in the attached figures:

[0041] 1-First multilayer perceptron module, 2-First self-attention module, 3-Feature data concatenation module, 4-Second multilayer perceptron module, 5-Pooling module, 6-Second self-attention module. Detailed Implementation

[0042] To make the above-mentioned objectives, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0043] As attached Figure 1 As shown, the present invention provides a method for recognizing 3D point cloud data for neural networks, comprising:

[0044] S1. Obtain 3D point cloud data;

[0045] S2. Obtain the sampling point set of the 3D point cloud data based on the 3D point cloud data;

[0046] S3. When the 3D point cloud data is rotated, obtain the rotation-invariant feature data in the sampling point set;

[0047] S4. Construct several RISurFormer modules to process rotation-invariant feature data into high-dimensional rotation-invariant feature data step by step.

[0048] S5. Construct the Transformer-Encoder module to optimize high-dimensional rotation-invariant feature data and obtain optimized high-dimensional rotation-invariant feature data;

[0049] S6. Construct a classifier to classify 3D point cloud data based on optimized high-dimensional rotation-invariant feature data.

[0050] Three-dimensional point cloud data can be obtained through three-dimensional scanning devices with three-dimensional sensors, such as TOF cameras, DGBD cameras, and LiDAR. In step S2, a set of sampling points of the three-dimensional point cloud data is obtained based on the input three-dimensional point cloud data, so that rotation-invariant feature data of each rotation point can be extracted when the three-dimensional point cloud data is rotated.

[0051] Step S3 includes:

[0052] S31. Select any rotation point from the three-dimensional point cloud sampling point set;

[0053] S32. Using the currently selected rotation point as the center point, generate a neighborhood, which includes several neighboring points;

[0054] S33. Select any neighboring point;

[0055] S34. Find the first and second nearest neighbors to the currently selected neighbor;

[0056] S35. Generate a first plane based on the currently selected neighboring point, the first neighboring point, and the currently selected rotation point; generate a second plane based on the currently selected neighboring point, the second neighboring point, and the currently selected rotation point.

[0057] S36. Extract rotation-invariant feature data from the first and second planes;

[0058] S37. Determine if there are any unselected neighboring points in the neighborhood. If there are, select the next neighboring point and return to step S34. If all neighboring points in the neighborhood have been selected, proceed to step S38.

[0059] S38. Determine if there are any unselected rotation points in the sampling point set. If so, select the next rotation point and return to step S32. If all rotation points in the sampling point set have been selected, proceed to step S4.

[0060] In this embodiment, the RISurFormer module is as follows: Figure 1As shown, there are a total of 5 modules, allowing the rotation-invariant feature data to pass through 5 RISurFormer modules sequentially, being processed step-by-step into high-dimensional rotation-invariant feature data.

[0061] In step S36, the formula for extracting rotation-invariant feature data is as follows:

[0062] , , , , , , , , , , , , , , ;

[0063] Wherein, RISP means selecting the currently chosen neighboring nodes. The operation of mapping to rotation-invariant feature data, where P represents the center point. This indicates the currently selected neighboring point. Indicates the first neighboring point, Let N represent the second neighboring point, and N represent the normal vector. This represents the normal vector of the currently selected neighboring point. Let the normal vector of the first neighbor be denoted as . This represents the normal vector of the second neighboring point. It is a unique length quantity, representing the distance between the center point and the currently selected neighboring points; ∠ represents a vector; ∠(,) represents the angle between the two vectors or the product of the two vectors within the brackets.

[0064] For details, see attached. Figure 3 As shown, the input 3D point cloud data is the 3D point cloud data of an aircraft. A sampling point set is obtained based on this aircraft's 3D point cloud data. When the aircraft's 3D point cloud data rotates, the points in the sampling point set will also rotate. Taking point P as the rotation point as an example, a neighborhood is generated centered on point P. The neighborhood contains 9 points: one is point P, and the other 8 are the neighbors of point P. An arbitrary neighbor is selected, and the rotation is performed using point P as the pivot point. For example, find the distance point. The two nearest neighbors are the first neighbor and the second nearest neighbor. Second neighbor At this time, the first neighboring point ,point The first plane is formed by the three points P, and the second neighboring point... ,point The second plane is formed by the three points, including point P.

[0065] As attached Figure 4 As shown, K×14 rotation-invariant feature data points are extracted from the first and second planes as points. The rotation-invariant feature data at a given location, where K represents the number of neighboring points, is as described in the formula above. The 14 rotation-invariant feature vectors include one length value and 13 angle values. The only length quantity, referring to point P and point [missing information]. The distance; φ1 and φ3 represent the distance between triangle P. The two angles, φ2 and φ4, represent triangle P. The two angles; φ5 represents triangle P. With triangle P The dihedral angle between points; αi, i=1,2, represents the normal vector at point P; βi, i=1,2, represents the normal vector at point Xi; θi, i=1,2, represents the normal vector at point P. The normal vector on the point; γi, i=1,2, represents the point The normal vector on.

[0066] The Transformer-Encoder module is the encoder in the Transformer network model. Since this invention only needs to perform recognition and classification, a decoder is not needed; the output is directly sent to the classifier through the encoder.

[0067] As attached Figure 2 As shown, the RISurFormer module in step S4 includes:

[0068] The first multilayer perceptron module 1 is used to map rotation-invariant feature data to a high-dimensional space to obtain initial high-dimensional rotation-invariant feature data;

[0069] The first self-attention module 2 is connected to the first multilayer perceptron module 1 and is used to optimize the initial high-dimensional rotation-invariant feature data to obtain the optimized initial high-dimensional rotation-invariant feature data.

[0070] Feature data concatenation module 3 is connected to the first self-attention module 2. It is used to determine whether there is a previous RISurFormer module. If there is a previous RISurFormer module, the optimized initial high-dimensional rotation-invariant feature data is concatenated with the high-dimensional rotation-invariant feature data output by the previous RISurFormer module to obtain concatenated initial high-dimensional rotation-invariant feature data. If there is no previous RISurFormer module, the optimized initial high-dimensional rotation-invariant feature data is directly output.

[0071] The second multilayer perceptron module 4 is connected to the feature data concatenation module 3 and is used to summarize the concatenated high-dimensional rotation-invariant feature data to obtain the summarized initial high-dimensional rotation-invariant feature data.

[0072] Pooling module 5, which is connected to the second multilayer perceptron module 4, is used to perform pooling operations on the summarized rotation-invariant feature data to obtain the initial high-dimensional rotation-invariant feature data after pooling.

[0073] The second self-attention module 6 is connected to the pooling module 5 to optimize the rotation-invariant feature data after pooling, thereby obtaining high-dimensional rotation-invariant feature data.

[0074] The first multilayer perceptron module 1 and the second multilayer perceptron module 4 are MLPs; the feature data concatenation module 3 concatenates the dimensions of the high-dimensional rotation-invariant feature data output from the previous-level RISurFormer module. For example, if the high-dimensional rotation-invariant feature data output from the first-level RISurFormer module has 32 dimensions, and the rotation-invariant feature data in this-level RISurFormer module also has 32 dimensions, then the rotation-invariant feature data output by the feature data concatenation module in this level will be 64 dimensions. In the above embodiment, the final rotation-invariant feature data output after passing through 5 RISurFormer modules is 512 dimensions. Figure 2 middle, This indicates the dimension of the rotation-invariant feature data output by the previous RISurFormer module. C represents the dimension of the rotation-invariant feature data in this level of the RISurFormer module. + The first self-attention module 2 and the second self-attention module 6 are the self-attention mechanism; the pooling module 5 performs the max pooling operation.

[0075] In a preferred embodiment of the present invention, the sampling point set is obtained in step S2 using the farthest point sampling algorithm. The farthest point sampling algorithm can cover all points of the entire 3D point cloud data.

[0076] This invention can be constructed as a neural network model, or the intermediate layer consisting of the RISurFormer modules, Transformer-Encoder modules, and classifiers of this invention can be connected to other neural networks for more complex processing of 3D point cloud data.

[0077] This invention achieves high-precision classification of 3D point cloud data, obtaining high-precision results on three different types of 3D point cloud data recognition datasets, while also exhibiting rotation invariance. Experiments were conducted on three of the most widely used 3D point cloud data recognition datasets: ModelNet40, ScanObjectNN, and FG3D. Three test modes were designed: z / z, SO3 / SO3, and z / SO3, where z represents random rotation of the 3D point cloud data along the z-axis, and SO3 represents rotation in any direction. z / z indicates that z-axis rotation was included during training, and random z-axis rotation was also performed on the 3D point cloud data during testing. SO3 / SO3 indicates that the model was arbitrarily rotated during training, and arbitrary angle rotation was performed on the 3D point cloud data during testing. z / SO3 indicates that z-axis rotation was included during training, and arbitrary angle rotation was performed on the 3D point cloud data during testing. The difficulty gradually increases from z / z to z / SO3.

[0078] First, the ModelNet40 dataset is a dataset for classifying 3D point cloud data of synthetic objects. It consists of 9843 training models and 2468 test models, divided into 40 classes, and is one of the most commonly used 3D point cloud data classification evaluation datasets. The input point cloud data size is 1024, and each point (x, y, z, nx, ny, nz) represents the 3D coordinates and 3D normal vector in Euclidean space, respectively. The results are shown in Table 1. We use two metrics for evaluation: mean accuracy and standard deviation of accuracy (Std.). Mean accuracy is a commonly used indicator to measure the performance of classification tasks, while standard deviation of accuracy measures the consistency across three different rotation modes (z / z, SO3 / SO3, z / SO3). Meanwhile, the rotation-invariant method is not affected by rotations present in the training and testing data, maintaining stability in accuracy. The experimental results show that the method of this invention has achieved a significant improvement in accuracy, currently ranking first in the world, surpassing not only current rotation-invariant methods but also non-rotation-invariant methods. It is worth noting that the accuracy of this method is very consistent across all test modes, while some existing methods only perform well in z / z mode or SO3 / SO3 mode, but perform poorly in z / SO3 mode, indicating that they are not truly rotation-invariant methods.

[0079] ;

[0080] Table 1: Comparison of model results on the ModelNet40 dataset for synthetic 3D object recognition task

[0081] scanobjectNN is a dataset for recognizing 3D point cloud data captured by real-world RGB-D cameras. It consists of 2902 point clouds sampled from real indoor scenes, divided into 15 categories. For our evaluation, we used processed files and selected OBJ_ONLY, the simplest variable; OBJ_BG contains only objects with / without background points, without rotation, translation, or scaling; and PB_T50_RS, the most difficult variable used in this dataset, which has 50% bounding box translation, rotation around the gravity axis, and random scaling, resulting in rotation and partial data. The results are shown in Table 2. The evaluation metrics for this dataset are consistent with those for the ModelNet40 dataset, using instance mean accuracy. Our invention significantly outperforms all comparable methods on the OBJ_ONLY and PB_T50_RS variants, and also significantly outperforms state-of-the-art rotation-invariant methods on these two variables.

[0082] ;

[0083] Table 2: Comparison of model results on the real-world 3D object classification task dataset scanobjectNN

[0084] FG3D is a fine-grained 3D object recognition dataset. It includes three main categories: airplanes, chairs, and cars. These three main categories are further divided into 66 fine-grained subclasses, totaling 25,552 objects. Due to the small differences between subclasses, higher classification accuracy is required from the network. Compared to ModelNet40 and ScanObjectNN, which use instance-average accuracy, FG3D uses class-average accuracy as the evaluation metric. This invention still maintains excellent performance and has promising prospects for practical applications.

[0085] ;

[0086] Table 3: Comparison of model results on the 3D object fine-grained classification dataset FG3D

[0087] This invention has a wide range of applications, such as controlling industrial robotic arms based on visual positioning. Robotic arms are widely used in modern industrial production lines. During the disordered grasping process of the robotic arm, visual positioning is required, i.e., the identification and pose estimation of the target object. In real-world scenarios, the object's posture is not fixed, necessitating the use of sensors on the robotic arm to capture information about the target object. The acquired point cloud data is then used to identify the object, thereby automatically controlling the robotic arm to perform operations on the production line. Another example is in the field of autonomous driving, which involves perception, decision-making, and control. In the perception part, onboard sensors are needed to acquire information about the surrounding environment, such as vehicles, roads, pedestrians, and road signs. Currently, commonly used sensors include LiDAR, cameras, and millimeter-wave radar. These three types of sensors each have their advantages and disadvantages, complementing each other. LiDAR and millimeter-wave radar both capture point cloud data. We need to use the captured point cloud datasets for target detection and segmentation, and then identify and determine the category of the segmented object to help the vehicle decide on the next operation, such as obstacle identification and avoidance. In practical applications, the pose of the target is random, which poses a significant challenge to current algorithms. This invention can effectively solve this problem. For example, when building large scenes, relying on traditional camera modeling to acquire panoramic images is time-consuming and labor-intensive, insensitive to distance information, and prone to occlusion. Currently, LiDAR is widely used in these areas, such as topographic mapping, surveying and designing mine earthwork projects, drawing urban public road models, and modeling digital factories or cities. These areas all require the use of various LiDAR-captured 3D point cloud data, and then various point cloud processing algorithms to achieve tasks related to 3D scene reconstruction, point cloud segmentation, and recognition. Since the angle of the acquired data is not fixed, there are relatively high requirements for the anti-rotational properties of the network model, which this invention can also effectively solve.

[0088] While the disclosure is as stated above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of this disclosure, and all such changes and modifications will fall within the protection scope of this invention.

Claims

1. A method for three-dimensional point cloud data recognition of a neural network, characterized in that, The method comprises the following steps: S1, acquiring three-dimensional point cloud data; S2, obtaining a sampling point set of the three-dimensional point cloud data according to the three-dimensional point cloud data; S3, when the three-dimensional point cloud data is rotated, obtaining rotation-invariant feature data in the sampling point set; S4, constructing a plurality of RISurFormer modules for processing the rotation-invariant feature data into high-dimensional rotation-invariant feature data step by step; The RISurFormer module comprises: A first multi-layer perception module for mapping the rotation-invariant feature data to a high-dimensional space to obtain initial high-dimensional rotation-invariant feature data; A first self-attention module connected with the first multi-layer perception module for optimizing the initial high-dimensional rotation-invariant feature data to obtain optimized initial high-dimensional rotation-invariant feature data; A feature data phase concatenation module connected with the first self-attention module for judging whether there is an upper-level RISurFormer module, if there is the upper-level RISurFormer module, concatenating the optimized initial high-dimensional rotation-invariant feature data with high-dimensional rotation-invariant feature data output by the upper-level RISurFormer module to obtain concatenated initial high-dimensional rotation-invariant feature data, and if there is no upper-level RISurFormer module, directly outputting the optimized initial high-dimensional rotation-invariant feature data; A second multi-layer perception module connected with the feature data phase concatenation module for summarizing the concatenated high-dimensional rotation-invariant feature data to obtain summarized initial high-dimensional rotation-invariant feature data; A pooling module connected with the second multi-layer perception module for performing a pooling operation on the summarized rotation-invariant feature data to obtain pooled initial high-dimensional rotation-invariant feature data; A second self-attention module connected with the pooling module for optimizing the pooled rotation-invariant feature data to obtain the high-dimensional rotation-invariant feature data; S5, constructing a Transformer-Encoder module for optimizing the high-dimensional rotation-invariant feature data to obtain optimized high-dimensional rotation-invariant feature data; S6, constructing a classifier for classifying the three-dimensional point cloud data according to the optimized high-dimensional rotation-invariant feature data. 2.The method for 3D point cloud data recognition of neural network according to claim 1, characterized in that, The sampling point set is obtained by a farthest point sampling algorithm in the step S2.

3. The method of claim 1 or 2, wherein, The step S3 comprises: S31, selecting an arbitrary rotation point in the sampling point set; S32, generating a neighborhood with the currently selected rotation point as a center point, the neighborhood comprising a plurality of neighbor points; S33, selecting an arbitrary neighbor point; S34, finding a first neighbor point and a second neighbor point closest to the currently selected neighbor point; S35, generating a first plane according to the currently selected neighbor point, the first neighbor point and the currently selected rotation point, and generating a second plane according to the currently selected neighbor point, the second neighbor point and the currently selected rotation point; S36, extracting rotation-invariant feature data in the first plane and the second plane; S37, judging whether there is a neighbor point not selected in the neighborhood, if there is, selecting a next neighbor point and returning to step S34, if all neighbor points in the neighborhood have been selected, turning to step S38; S38, judging whether there is a rotation point not selected in the sampling point set, if there is, selecting a next rotation point and returning to step S32, if all rotation points in the sampling point set have been selected, turning to step S4.

4. The method of claim 3, wherein, In the step S36, a formula for extracting the rotation-invariant feature data is as follows: ； ， ， ， ， ， ， ， ， ， ， ， ， ， ； wherein RISP denotes the current selected neighbor point mapped to rotation invariant feature data, P denotes the center point, denotes the current selected neighbor point, denotes the first neighbor point, denotes the second neighbor point, N denotes the normal vector, denotes the normal vector of the current selected neighbor point, denotes the normal vector of the first neighbor point, denotes the normal vector of the second neighbor point, is a unique length quantity denoting the distance of the center point to the current selected neighbor point; denotes the vector; ∠(, ) denotes the angle between two vectors or two vector products in the parentheses.

Images