Automatic selection method of 3D point cloud classification algorithm based on meta-learning

Through a meta-learning-based method, a meta-classifier is constructed, and the optimal classification algorithm is predicted, which solves the long training time and insufficient generalization ability of the three-dimensional point cloud classification algorithm in the existing technology when scene changes are long and the generalization ability is insufficient in the existing technology. Problem, achieving faster and more accurate algorithm selection.

CN113486902BActive Publication Date: 2024-11-12NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110724839.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-06-29
Publication Date
2024-11-12
Estimated Expiration
2041-06-29

AI Technical Summary

Technical Problem

The existing three-dimensional point cloud classification algorithm needs to be retrained when the scene changes, the generalization ability is insufficient, and the algorithm selection process is time-consuming and labor-intensive, lacks flexibility and scalability.

Method used

The automatic selection method of point cloud classification algorithm based on meta-learning is adopted. By extracting the global features of point cloud and the meta-element of data sets, metadata is constructed, and the metaclassifier is trained using a random forest and an improvement algorithm to predict the optimal classification algorithm to adapt to different scenarios.

Benefits of technology

This greatly saves model selection time, improves the accuracy and flexibility of algorithm selection, and makes it better applicable in a variety of complex environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113486902B_ABST
    Figure CN113486902B_ABST
Patent Text Reader

Abstract

The present invention provides a method for automatically selecting a three-dimensional point cloud classification algorithm based on meta-learning, which belongs to the field of point cloud data processing. The technical key points include: determining a set of point cloud classification algorithms to be selected; extracting global features of point clouds to complete point cloud classification; extracting meta-features of point cloud data sets to construct metadata; selecting a meta-classifier and using metadata to train a meta-classification model; predicting the optimal classification algorithm for a target point cloud data set in a new scene, so as to at least solve the problems of time-consuming and laborious point cloud algorithms and difficulty in selecting algorithms in the prior art, so as to quickly select the algorithm most suitable for the current scene from a large number of point cloud classification algorithms.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional point cloud processing, and in particular relates to an automatic selection method for a point cloud classification algorithm based on meta-learning. Background Art

[0002] 3D point cloud is a multi-dimensional complex data set consisting of XYZ position information, RGB color information and intensity information of spatial points, obtained by 3D scanning devices such as LiDAR and RGB-D cameras. Compared with 2D images, 3D point cloud has incomparable advantages. It can provide rich geometry, shape and scale information, and is not easily affected by changes in light intensity and occlusion by other objects. Through 3D point cloud, we can better understand the environment in which the machine is located and realize scene understanding in complex environments. These make 3D point cloud widely used in many fields such as autonomous driving, robotics and augmented reality (AR).

[0003] Object detection and recognition based on 3D point cloud data are the main technologies for solving scene understanding, and 3D point cloud object classification is the basis of these technologies. By assigning a specific meaning (label) to each object in space, we can accurately describe the type of objects in space, giving us a more detailed understanding of spatial objects, and the assigned labels are meaningful in the real world.

[0004] At present, 3D point cloud classification algorithms can be divided into two categories, namely, traditional feature-based machine learning methods and deep learning-based methods. Traditional machine learning methods first extract the global features of point clouds, and then apply machine learning methods such as KNN, SVM, NB, DT, RF, etc. for classification. They have short training time and strong real-time performance, but are not suitable for situations with a large number of point cloud types. Deep learning-based methods construct various deep learning models, directly use the original 3D point cloud as input, and train the model by continuously optimizing parameters. The more commonly used deep learning models include Pointnet, Pointnet++, PointCNN, DGCNN, etc. Compared with traditional machine learning methods, deep learning models are well suited for situations with a large number of point cloud types. The only drawback is that their model training time is much longer than traditional machine learning methods.

[0005] These methods can accomplish the task of point cloud classification very well, but they are all trained on a large amount of point cloud data for a specific scene. The generalization ability of the model is usually not ideal. If the scene changes, we need to retrain the model. How to improve the generalization ability of the model so that the model can be applicable in many different scenarios has always been a difficult point in research. The general method is to continuously optimize the model parameters to enable it to handle more different tasks, but this greatly increases the complexity of the model and the training time will increase dramatically. Instead of trying hard to adapt to the problem by constantly improving the algorithm, it may be better to actively let the problem choose the algorithm. This idea is called algorithm recommendation (selection).

[0006] The most intuitive method of algorithm selection is to select the algorithm with the best performance through repeated experiments, which requires too much computing resources and is quite time-consuming; another method is to guide algorithm selection through expert experience, but there are also great defects. On the one hand, it is costly to obtain expert experience, and expert experience is not necessarily reliable; on the other hand, the developed expert system lacks scalability, which reduces the flexibility of the system. With the rapid development of machine learning, some scholars have realized that algorithm selection is actually a learning task, and algorithm selection has formed a meta-learning field in the field of machine learning.

[0007] Meta-learning refers to the process of learning meta-knowledge by applying machine learning methods (meta-algorithms) to seek the mapping between the characteristics of the problem (meta-features) and the relative performance measures of the algorithm. The algorithm selection task can be regarded as a learning task. In the initial stage, each candidate algorithm is applied to each data set, its performance is evaluated, and the morphological characteristics (meta-features) of these data sets are extracted to form metadata; then an appropriate meta-algorithm is used to meta-learn the metadata to obtain the mapping relationship between the characteristics of the data set and the performance of the algorithm, forming a meta-knowledge base. The selection of the algorithm needs to rely on the guidance of meta-knowledge. When a new data set is added, the feature extraction of the data set can be performed first, and then the optimal algorithm can be selected or the algorithms can be sorted according to the degree of feature matching with the meta-knowledge base.

[0008] Based on the above background conditions and technical problems, there is an urgent need for a point cloud classification algorithm selection method to at least solve the problems of the existing technology of time-consuming and laborious point cloud algorithms and difficulty in algorithm selection. Summary of the invention

[0009] The embodiment of the present application provides a method for automatically selecting a point cloud classification algorithm based on meta-learning, so as to at least solve the problems in the prior art of time-consuming and laborious point cloud algorithms and difficulty in selecting algorithms, so as to quickly select the algorithm most suitable for the current scenario from a large number of point cloud classification algorithms.

[0010] In one embodiment of the present application, a method for automatically selecting a point cloud classification algorithm based on meta-learning is proposed, comprising:

[0011] Determine a set of point cloud classification algorithms to be selected;

[0012] Extract global features of point cloud and complete point cloud classification;

[0013] Extract meta features of point cloud dataset and construct metadata;

[0014] Select a meta-classifier and use the metadata to train a meta-classification model;

[0015] Predict the optimal classification algorithm for the target point cloud dataset in the new scene.

[0016] Preferably, the step of determining a set of to-be-selected point cloud classification algorithms comprises:

[0017] The set of candidate point cloud classification algorithms is selected from a feature-based machine learning method and a deep learning-based method.

[0018] Preferably, the step of extracting global features of point clouds and completing point cloud classification includes:

[0019] According to the selected point cloud classification algorithm, point cloud classification algorithm models based on machine learning and deep learning are constructed respectively;

[0020] According to the constructed point cloud classification algorithm model, the global features of the point cloud are extracted to complete the point cloud classification.

[0021] Preferably, the step of extracting metadata features of the point cloud data set and constructing metadata includes:

[0022] The metadata features of the point cloud dataset include the number of point cloud classes, the number of point cloud samples, the number of point cloud points, and the entropy H(p) of the point cloud class.

[0023] Among them, n represents the number of point cloud classes, p i It represents the probability that the i-th type of point cloud samples accounts for the total number of samples.

[0024] Preferably, the metadata is composed of metadata features of the data set and an optimal classification algorithm for the data set.

[0025] Preferably, the process of obtaining the optimal classification algorithm includes:

[0026] Classify the point cloud dataset through the point cloud classification algorithm model;

[0027] obtaining at least one parameter of each classification algorithm;

[0028] The point cloud classification algorithm with the best performance is calculated and selected through the above parameters as the optimal classification algorithm.

[0029] Preferably, in the process of obtaining the optimal classification algorithm, the parameters include the accuracy of each classification algorithm, the running time of each classification algorithm, and the F1 score of each classification algorithm, which are calculated by the following formula:

[0030]

[0031]

[0032] Among them, TP stands for true positive examples, TN stands for true negative examples, FP stands for false positive examples, FN stands for false negative examples, F1 stands for the F1 score of each classification algorithm, which is a weighted average of the model precision and recall rate, and can better reflect the algorithm performance when the data set is unbalanced. ACC stands for the accuracy of each classification algorithm.

[0033] Preferably, the meta-classifier is configured to use random forest and boosting algorithms, take metadata as input, train the meta-classification model through cross-validation, and use accuracy to measure model performance.

[0034] Preferably, the predicting of the optimal classification algorithm for the target point cloud data set in the new scene includes: constructing target metadata, and predicting the optimal classification algorithm for the target data set through a trained meta-classifier.

[0035] Through the embodiments of the present application, since the algorithm has "preferences", a meta-classifier is established to seek the "preference" relationship between the algorithm performance and the meta-features of the data set, thereby converting the algorithm selection problem into a classification problem. The examples that meet a certain type of features are classified into one category and a proven "good" algorithm is recommended to them. By extracting the meta-features of the data set for the point cloud data set in the new scene, the meta-classifier is used to predict the optimal classification algorithm for the data set, and then the algorithm is used to classify each point cloud data in the data set. Compared with the traditional method of selecting the best performance algorithm based on repeated experiments and the method of guiding algorithm selection based on expert experience, the automatic selection method of the three-dimensional point cloud classification algorithm provided by the present invention learns the optimal classification algorithm in different scenarios through a meta-learning framework, is applicable to various complex environments, greatly saves the time of model selection and has more accurate selection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0037] Figure 1 is a schematic diagram of the internal structure of a point cloud classification algorithm model according to an embodiment of the present application;

[0038] Figure 2It is a schematic diagram of the structure of metadata according to this application;

[0039] Figure 3 is a schematic diagram of VFH features of different types of point clouds according to an embodiment of the present application;

[0040] Figure 4 It is the classification accuracy and F1 score of the algorithm according to the embodiment of the present application;

[0041] Figure 5 This is an optional flowchart of a point cloud classification algorithm selection method according to an embodiment of the present application. DETAILED DESCRIPTION

[0042] The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0043] It should be noted that the terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0044] The present application embodiment provides a method for automatically selecting a point cloud classification algorithm based on meta-learning. Figure 1 Schematic diagram of the corresponding modules of the method. The method disclosed in the present application includes:

[0045] First, a set of candidate point cloud classification algorithms is determined, wherein the candidate point cloud classification algorithms are selected from a feature-based machine learning method and a deep learning-based method;

[0046] Extract the global features of the point cloud and complete the point cloud classification; according to the selected point cloud classification algorithm, construct the point cloud classification algorithm models based on machine learning and deep learning respectively; according to the constructed point cloud classification algorithm model, extract the global features of the point cloud and complete the point cloud classification.

[0047] The feature-based machine learning method has a dedicated manually designed global feature descriptor. Its feature extraction process and point cloud classification process are isolated from each other. The global features of the point cloud are extracted first, and then classification is performed based on this global feature. It is an unsupervised learning process, and its feature extraction process and point cloud classification process only need to be executed once.

[0048] The point cloud classification method based on deep learning is a supervised learning process, which requires repeated feature extraction and point cloud classification processes, which is specifically manifested in the continuous optimization of model parameters through back propagation until the model reaches a fitting state. When the model is fitted, we believe that the model can complete the point cloud classification task well.

[0049] Extract the meta-features of the point cloud dataset and construct metadata. The meta-features of the point cloud dataset include the number of point cloud classes, the number of point cloud samples, the number of point cloud points, and the entropy H(p) of the point cloud class.

[0050]

[0051] Among them, n represents the number of point cloud classes, p i It represents the probability that the i-th type of point cloud sample accounts for the total number of samples. The constructed metadata is composed of the dataset meta-features and the optimal classification algorithm under the dataset.

[0052] Its structure is as Figure 2 As shown in Figure 1, metadata is composed of the metadata features of the dataset and the optimal classification algorithm for the dataset.

[0053] It is necessary to provide additional explanation on the process of obtaining the optimal classification algorithm. The specific process is as follows: Use the constructed 3D point cloud classification algorithm model to classify the 3D point cloud models in the data set, and obtain the accuracy (Accuracy), F1 score (F1-score), running time T and other parameters of each classification algorithm. Use these parameters to measure different classification algorithms and select the point cloud classification algorithm with the best performance. It should be noted that in the process of selecting the optimal classification algorithm, not all of the above parameters will be considered, and researchers can choose different parameters according to different needs.

[0054] Select the meta-classifier. The meta-classifier is configured to use random forest and boosting algorithms, take metadata as input, train the meta-classification model through cross-validation, and use accuracy to measure model performance. Cross-validation is a common method used in machine learning to build and verify models, and is generally used to evaluate the performance of a machine learning model. Cross-validation, as the name suggests, is to repeatedly use data, split the obtained sample data, and combine it into different training sets and test sets. The training set is used to train the model, and the test set is used to evaluate the quality of the model prediction. On this basis, multiple different training sets and test sets can be obtained. A sample in a training set may become a sample in the test set next time, which is the so-called "crossover". This method can effectively prevent model overfitting.

[0055] The process of obtaining the optimal classification algorithm includes:

[0056] Classify the point cloud dataset through the point cloud classification algorithm model;

[0057] obtaining at least one parameter of each classification algorithm;

[0058] The process of obtaining the optimal classification algorithm, the parameters include the accuracy of each classification algorithm, the running time of each classification algorithm, and the F1 score of each classification algorithm, which are calculated by the following formula:

[0059]

[0060]

[0061] Among them, TP stands for true positive examples, TN stands for true negative examples, FP stands for false positive examples, FN stands for false negative examples, F1 stands for the F1 score of each classification algorithm, which is a weighted average of the model precision and recall rate, and can better reflect the algorithm performance when the data set is unbalanced. ACC stands for the accuracy of each classification algorithm.

[0062] The point cloud classification algorithm with the best performance is calculated and selected through the above parameters as the optimal classification algorithm.

[0063] Finally, the optimal classification algorithm is predicted for the target point cloud dataset in the new scene. This includes constructing target metadata based on the new scene and predicting the optimal classification algorithm for the target dataset through the trained meta-classifier. The trained meta-classifier predicts the optimal classification algorithm for the dataset, and then uses the algorithm to classify each point cloud data in the dataset. Compared with the traditional method of selecting the best performance algorithm based on repeated experiments and the method of guiding algorithm selection based on expert experience, the automatic selection method of 3D point cloud classification algorithm based on meta-learning saves a lot of time, and the selected algorithm has better classification effect.

[0064] according to Figure 3-Figure 5 , take the following application examples as an example:

[0065] Step 1: constructing a 3D point cloud data cluster, which belongs to the data set preparation stage. This embodiment assumes that the data set has been prepared before model training.

[0066] Step 2: Determine the set of 3D point cloud classification algorithms to be selected

[0067] In this embodiment, from the perspective of real-time and simplicity, three machine learning methods, namely KNN, SVM and NB, are selected as candidate 3D point cloud classification algorithms, aiming to illustrate the effectiveness of the automatic selection method of the point cloud classification algorithm.

[0068] K-Nearest Neighbor (KNN) is a commonly used supervised learning method. Its working mechanism is very simple: given a test sample, find the K training samples closest to it in the training set based on a certain distance metric, and then make predictions based on the information of these K "neighbors". Usually, the "voting method" can be used in classification tasks, that is, the category label with the most occurrences in these K samples is selected as the prediction result.

[0069] Support Vector Machine (SVM) is a typical machine learning classification method. Its basic model is a linear classifier with the largest margin defined in the feature space. The largest margin makes it more suitable for classification tasks than the perceptron. For a given training set, SVM uses the maximum margin to find the optimal separation hyperplane, and then classifies the input test data based on this optimal hyperplane.

[0070] The Naive Bayes method is a classification method based on the Bayesian theorem and the feature conditional independence hypothesis. For a given training data set, the joint probability distribution of input and output is first learned based on the feature conditional independence hypothesis; then based on this model, for a given input x, the Bayesian theorem is used to find the output y with the maximum posterior probability.

[0071] Step 3: Build a point cloud classification algorithm model;

[0072] In step 2, simple KNN, SVM, and NB were selected as candidate 3D point cloud classification algorithms. They belong to traditional feature-based machine learning methods, which need to first extract the global features of the point cloud and then classify the point cloud based on this feature.

[0073] This example extracts point cloud VFH features as input to KNN, SVM, and NB classifiers, and uses VFH features to classify point clouds. Figure 3 The VFH features of different types of point cloud targets are shown. It can be clearly seen that the VFH features of different types of point cloud targets are very different and can be used for point cloud classification tasks. Viewpoint Feature Histogram VFH (Viewpoint Feature Histogram) is a feature descriptor widely used in point cloud classification and clustering. It is evolved on the basis of Fast Point Feature Histogram FPFH (Fast Point Feature Histogram), and well inherits the reliable identification ability and fast processing ability of FPFH. VFH extends the estimation object of FPFH to the entire point cloud, takes the point pair composed of the center point of the point cloud and all other points on the surface of the point cloud as the basic calculation unit, adds additional information between the viewpoint direction and the estimated normal of each point, and calculates the VFH feature component of the point cloud by counting the histogram of the angle between the viewpoint direction and each normal. The VFH feature is a 308-dimensional array that describes the global features of the point cloud. The VFH features of different types of point clouds are quite different. Through feature differences, the classification and recognition of point clouds can be effectively completed. FPFH describes the local features of point clouds and is widely used in semantic segmentation and part segmentation of point clouds.

[0074] Step 4: Data cluster format conversion;

[0075] In step 3, building a point cloud classification algorithm model requires extracting point cloud VFH features. This process generally requires converting the point cloud into PCD format, and then completing feature extraction based on the C++PCL library. The PCD format is the official format specified by the PCL library. Any point cloud processing based on the PCL library uses the PCD format as its standard input. Each PCD file consists of a file header and a data field. The file header is used to determine and declare certain characteristics of the point cloud data stored in the file, and the data field is used to store the XYZ coordinates of the point cloud space points.

[0076] There are many storage formats for 3D point clouds, the most commonly used ones are LAS, TXT, PLY, PCD, etc. Each file in these storage formats can only store one point cloud. Considering the problems of data compression and batch processing, many researchers store a single 3D point cloud as a 2D array, and store a 3D array composed of multiple 3D point clouds as a .h5 file.

[0077] The three-dimensional point cloud data set used in this embodiment is taken from Modelnet40, which stores the three-dimensional point cloud and its class labels in a .h5 file, which needs to be converted into PCD format first. The .h5 file is a container containing two objects: a dataset and a group. A dataset is a data set similar to an array (numpy), and a group is a container similar to a folder, which is used to store datasets and other groups. The three-dimensional point cloud data of Modelnet40 is stored in the data dataset in the form of a three-dimensional array, and the class label data is stored in the label dataset. To convert the three-dimensional point cloud data in array form into PCD format, it is necessary to add a PCD file header in front of each two-dimensional array representing the point cloud, and use the array as the data field of the PCD file. The two are combined to form a PCD file to complete the format conversion. One of the contributions of this embodiment is to complete the data conversion between array point cloud and PCD format point cloud.

[0078] Step 5: Extract point cloud VFH features.

[0079] The PCD format point cloud converted in step 4 is used for VFH feature extraction based on the PCL library. PCL is an open source C++ library that integrates many simple point cloud processing algorithms and related data structures. The VFH feature extraction of point clouds can be completed through simple combination. The specific steps are as follows:

[0080] The PCD format point cloud converted in step 4 only contains the XYZ coordinates of the spatial data points. Extracting the VFH features of the point cloud requires that the point cloud contain both the XYZ coordinates and its normal direction. This method performs a neighbor search for each point in the space based on Kd-tree, and estimates the normal direction of the point cloud through the queried neighborhood points. After obtaining the normal direction of the point cloud, the VFH feature calculation function of the PCL library is called in combination with the XYZ coordinates of the space to extract the VFH features of the point cloud. It should be noted that when performing a neighbor search for a point in the point cloud, the search radius needs to be set. The fixed search radius selected in this embodiment is 0.04.

[0081] Step 6: Select the optimal classification algorithm for the 3D point cloud dataset.

[0082] For each 3D point cloud data set in the data cluster, this embodiment extracts the VFH features of the training set and the test set based on the method of step 5, and constructs the 3D point cloud classification algorithm determined in step 2 based on the Python sklearn library. For the VFH features of the training set, each classification algorithm is trained in turn, and the VFH features of the test set are classified based on the obtained training model, and multiple performance parameters (such as accuracy, F1 score, time consumption, etc.) of each classification algorithm are obtained. Different classification algorithms are comprehensively measured through these performance parameters, and the algorithm with the highest comprehensive score is selected as the algorithm we want to select. This embodiment only uses the F1 score as a measurement indicator for algorithm selection to select the optimal classification algorithm.

[0083] Step 7: Extract metadata features of the 3D point cloud dataset.

[0084] Extract the meta-features of the data set constructed in step 1. In this implementation, the number of point cloud classes, the number of point cloud samples, the number of points in the point cloud, and the entropy of the point cloud class are selected as the meta-features of the data set.

[0085] Step 8: Construct metadata based on the dataset meta-features and the selected optimal algorithm.

[0086] The metadata is formed by combining the dataset metadata extracted in step 7 with the optimal algorithm selected in step 6.

[0087] Step 9: Select a meta-classifier and train its model using the metadata.

[0088] Select an appropriate meta-classifier and use the metadata to train the algorithm classification model. In this embodiment, Random Forest RF (Random Forest) is selected as the meta-classifier. Random Forest RF is a typical ensemble learning scheme that completes the learning task by building and combining multiple decision trees. It is worth noting that using RF as a meta-classifier requires manual setting of many hyperparameters, which will greatly affect the performance of the algorithm. In this embodiment, RandomizedSearchCV is used to enable the machine to automatically select hyperparameters, greatly reducing the negative effects caused by human factors.

[0089] Step 10: Predict the optimal classification algorithm for the point cloud dataset in the new scenario.

[0090] For the point cloud dataset in the new scene, the meta-features of the dataset are extracted by the method in step 7 and sent to the meta-classifier trained in step 9 to predict the optimal classification algorithm suitable for the scene.

[0091] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for automatically selecting a 3D point cloud classification algorithm based on meta-learning, characterized in that: include: Determine a set of point cloud classification algorithms to be selected; The set of point cloud classification algorithms to be selected is selected from a feature-based machine learning method and a deep learning-based method; Extracting global features of the point cloud and completing point cloud classification; wherein, according to the selected point cloud classification algorithm, constructing point cloud classification algorithm models based on machine learning and deep learning respectively; extracting global features of the point cloud and completing point cloud classification according to the constructed point cloud classification algorithm model; Extracting meta features of the point cloud data set, extracting the optimal classification algorithm of the point cloud data set, and constructing metadata; the metadata is composed of the meta features of the data set and the optimal classification algorithm under the data set; Select a meta-classifier and use the metadata to train a meta-classification model; Predict the optimal classification algorithm for the target point cloud dataset in the new scene; The process of obtaining the optimal classification algorithm includes: Classify the point cloud dataset through the point cloud classification algorithm model; obtaining at least one parameter of each classification algorithm; The point cloud classification algorithm with the best performance is calculated and selected through the above parameters as the optimal classification algorithm.

2. The method for automatically selecting a three-dimensional point cloud classification algorithm based on meta-learning according to claim 1, characterized in that: The step of extracting the metadata features of the point cloud dataset and constructing metadata includes: The metadata features of the point cloud dataset include the number of point cloud classes, the number of point cloud samples, the number of point cloud points, and the entropy H(p) of the point cloud class. Among them, n represents the number of point cloud classes, p i It represents the probability that the i-th type of point cloud samples accounts for the total number of samples.

3. The method for automatically selecting a three-dimensional point cloud classification algorithm based on meta-learning according to claim 2, characterized in that: The process of obtaining the optimal classification algorithm, the parameters include the accuracy of each classification algorithm, the running time of each classification algorithm, and the F1 score of each classification algorithm, which are calculated by the following formula: Among them, TP stands for true positive examples, TN stands for true negative examples, FP stands for false positive examples, FN stands for false negative examples, F1 stands for the F1 score of each classification algorithm, which is a weighted average of the model's precision and recall rate, and can better reflect the algorithm performance when the data set is unbalanced. ACC stands for the accuracy of each classification algorithm.

4. The method for automatically selecting a three-dimensional point cloud classification algorithm based on meta-learning according to claim 1, characterized in that: The meta-classifier is configured to use random forest and boosting algorithms, take metadata as input, train the meta-classification model through cross-validation, and use accuracy to measure model performance.

5. The method for automatically selecting a three-dimensional point cloud classification algorithm based on meta-learning according to claim 1, characterized in that: The optimal classification algorithm prediction for the target point cloud data set in the new scene includes: constructing target metadata, and predicting the optimal classification algorithm under the target data set through a trained meta-classifier.