A multi-scale point cloud classification method and system

Through the multi-scale point cloud classification method, the parallel classification network and deep Q network extract and fusion features are used to extract and integrate features, the problem of local relationship neglect in point cloud classification is solved, the classification accuracy and robustness are improved, the training process is simplified, and it is suitable for practical purposes. application.

CN114419372BActive Publication Date: 2024-11-01NANJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210038066.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-13
Publication Date
2024-11-01
Estimated Expiration
2042-01-13

AI Technical Summary

Technical Problem

The prior art ignores the local relationship between points and points in point cloud classification, resulting in insufficient accuracy and robustness of classification segmentation, and deep learning methods are complex in calculations and long training time, making it difficult to adapt to practical application scenarios.

Method used

The multi-scale point cloud classification method is adopted to extract and fuse features at different levels through pre-trained parallel classification networks and deep Q networks, including local relationships between points and points, and improve feature extraction accuracy and classification accuracy.

Benefits of technology

It improves the accuracy and robustness of point cloud classification, simplifies the training process, and is suitable for practical application scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114419372B_ABST
    Figure CN114419372B_ABST
Patent Text Reader

Abstract

The present invention discloses a multi-scale point cloud classification method, which comprises the following steps: obtaining 3D disordered point cloud data; using a pre-trained parallel classification network to extract features and classify the obtained point cloud data to obtain an output result; wherein the parallel classification network comprises a plurality of basic networks with the same structure; using a pre-trained deep Q network to fuse the output results of the parallel networks to obtain the final result of point cloud classification. The present invention can improve the accuracy and robustness of point cloud classification.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a multi-scale point cloud classification method and system, and belongs to the technical field of computer 3D vision. Background Art

[0002] Compared with 2D vision, 3D vision technology can obtain richer and more comprehensive environmental information, and has become a key technology in the fields of face recognition, unmanned driving, augmented / virtual reality, industrial inspection, etc. 3D vision technology can better overcome the problem that 2D images cannot reflect the real world, and the coordinates in the point cloud format can reflect its actual position information in the real world. Object recognition in three-dimensional space is a current research hotspot, and the key to object recognition is object classification and segmentation. With the development of data acquisition technology, from pictures to three-dimensional models, various types of data acquisition methods are becoming more and more abundant. Point cloud is the main expression mode of three-dimensional data, and the original point cloud data can be directly collected through devices such as laser radar and depth sensors. Point cloud can be widely used in various fields, such as virtual reality, three-dimensional printing, unmanned driving, object classification and segmentation, etc. Therefore, more and more scholars have invested a lot of work in the processing of point cloud data.

[0003] Point cloud classification is similar to image classification, and the point cloud set is correctly identified according to the corresponding label. Point cloud segmentation is to classify point cloud data according to rules, and points with the same features are usually marked into one category. Traditional methods generally classify and segment point cloud data by manually extracting features, but such methods rely on manual professional level and experience, the process is relatively complicated, and are only applicable to specific tasks. In order to improve the degree of automation of point cloud classification and segmentation, more and more methods use deep learning to replace traditional feature extraction methods. Deep learning methods can extract high-dimensional features of input data according to the target to be learned. In recent years, a large number of researchers have applied deep learning technology in the two-dimensional field to three-dimensional unstructured point cloud data, which can be directly applied to disordered and sparse point cloud data. However, most methods only extract the global features of point clouds, ignoring the local relationship between points, resulting in insufficient accuracy and robustness of classification and segmentation. In addition, the calculation of deep learning is relatively complex and requires a lot of training time, which is not conducive to transplanting the model into actual application scenarios. Therefore, improving the accuracy of point cloud classification and improving training efficiency have strong practical significance. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a multi-scale point cloud classification method and system that can improve the accuracy and robustness of point cloud classification. To achieve the above purpose, the present invention is implemented by adopting the following technical solutions:

[0005] In a first aspect, the present invention provides a multi-scale point cloud classification method, comprising:

[0006] Get 3D unordered point cloud data;

[0007] Using a pre-trained parallel classification network to extract features and classify the acquired point cloud data to obtain an output result; wherein the parallel classification network includes multiple basic networks with the same structure;

[0008] The output results of the parallel networks are fused using a pre-trained deep Q network to obtain the final result of point cloud classification.

[0009] In combination with the first aspect, further, there are 5 basic networks, and each basic network includes a feature alignment layer, a feature extraction layer and a fully connected layer in sequence.

[0010] In combination with the first aspect, further, each basic network is distinguished according to different feature dimensions of the input basic network point cloud, and the dimensions of output features of different layers in the same basic network are the same.

[0011] In combination with the first aspect, further, obtaining the output result includes:

[0012] Pass the n×3 point cloud data through the feature alignment layer to align the point cloud features;

[0013] After the aligned n×3 point cloud data passes through the feature extraction layer and undergoes three convolution operations, the output dimensions are n×128, n×256, n×512, n×1024, and n×2048 features;

[0014] After the maximum pooling operation, the maximum value is calculated in the second dimension and the output dimensions are 1×128, 1×256, 1×512, 1×1024, and 1×2048 global features;

[0015] After processing through a fully connected layer composed of linear layers, the scores of k categories of size N×k are obtained, where N is the batch size, as the output result of the basic network.

[0016] In combination with the first aspect, further, the feature extraction layer uses a feature pyramid network to extract features, including:

[0017] Features are extracted in order from low dimension to high dimension, and the extracted features are upsampled twice from low dimension to high dimension and added, and the features of the corresponding feature dimensions are output.

[0018] In combination with the first aspect, further, a loss function is used to train each basic network constituting the parallel network, and the loss function is expressed by the following formula:

[0019]

[0020] In formula (1), x i represents the sample, m represents the number of training samples, P G (x i ; θ) represents the distribution model of each sample data determined by parameter θ.

[0021] In combination with the first aspect, preferably, the data set used for training the basic network is the ModelNet40 training sample of Stanford University.

[0022] In combination with the first aspect, further, the data set used for training the deep Q network is the object category with the best classification effect of each basic model through manual annotation.

[0023] In a second aspect, the present invention provides a multi-scale point cloud classification system, comprising:

[0024] Acquisition module: used to obtain 3D disordered point cloud data;

[0025] Feature extraction module: used to extract features and classify the acquired point cloud data using a pre-trained parallel classification network to obtain output results; wherein the parallel classification network is composed of multiple basic networks with the same structure;

[0026] Feature fusion module: used to fuse the output results of the parallel network using the pre-trained deep Q network to obtain the final result of point cloud classification.

[0027] In a third aspect, the present invention provides a computing device, including a processor and a storage medium;

[0028] The storage medium is used to store instructions;

[0029] The processor is used to operate according to the instructions to execute the steps of the method described in the first aspect.

[0030] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0031] Compared with the prior art, the multi-scale point cloud classification method provided by the embodiment of the present invention has the following beneficial effects:

[0032] The present invention obtains 3D disordered point cloud data; uses a pre-trained parallel classification network to extract features and classify the obtained point cloud data to obtain an output result; wherein the parallel classification network is composed of multiple basic networks with the same structure. The present invention uses a parallel classification network to obtain features of different levels and fuse them, including the local relationship between points, which can improve the accuracy of feature extraction and the accuracy and robustness of point cloud classification;

[0033] The output results of the parallel networks are fused using a pre-trained deep Q network to obtain the final result of point cloud classification. The present invention fuses the output results of the parallel networks based on the deep Q network, which can improve the accuracy of the classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a network architecture of a parallel classification network of a multi-scale point cloud classification method provided in the first embodiment of the present invention;

[0035] Figure 2 It is a network structure of a basic network of a multi-scale point cloud classification method provided in the first embodiment of the present invention;

[0036] Figure 3 It is a schematic diagram of feature extraction in a basic network of a multi-scale point cloud classification method provided in the first embodiment of the present invention;

[0037] Figure 4 The 3D unordered point cloud target to be classified provided by the second embodiment of the present invention;

[0038] Figure 5 The point cloud target including 128, 256, 512, and 1024 points provided in the second embodiment of the present invention;

[0039] Figure 6 is the overall classification accuracy of the point cloud containing 1024, 512, 256, and 128 points provided by the second embodiment of the present invention;

[0040] Figure 7 It is the average classification accuracy of the point cloud containing 1024, 512, 256, and 128 points provided by the second embodiment of the present invention. DETAILED DESCRIPTION

[0041] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0042] Embodiment 1:

[0043] An embodiment of the present invention provides a multi-scale point cloud classification method, including:

[0044] Get 3D unordered point cloud data;

[0045] Using a pre-trained parallel classification network to perform feature extraction and classification on the acquired point cloud data to obtain an output result; wherein the parallel classification network is composed of multiple basic networks with the same structure;

[0046] The output results of the parallel networks are fused using a pre-trained deep Q network to obtain the final result of point cloud classification.

[0047] The specific steps are as follows:

[0048] Step 1: Obtain the 3D unordered point cloud data to be classified.

[0049] Use {P i |i=1,...,n} represents the 3D unordered point cloud to be classified, where each point P i It is composed of three channels (x, y, z) and an additional feature channel. i |i=1,...,n} consists of n points in no particular order. Although the points in the point cloud are disordered, they come from a space with distance measurement, and there are still local structural relationships and combination relationships of local structures between the points.

[0050] After preprocessing the 3D unordered point cloud to be classified by rotation and normalization, 1024 points are uniformly taken for each point cloud target for subsequent operations.

[0051] Step 2: Use the pre-trained parallel classification network to extract features and classify the acquired point cloud data to obtain the output results.

[0052] like Figure 1 As shown, the parallel classification network includes an input layer, five basic networks with the same structure, and an output layer. According to the different feature dimensions of the input point cloud of the basic network, each basic network is distinguished, and the dimensions of the output features of different levels in the same basic network are the same.

[0053] like Figure 2 As shown in Figure 1, each basic network includes a feature alignment layer, a feature extraction layer, and a fully connected layer. The feature alignment layer is used to learn point cloud position features, the feature extraction layer is used to extract global features, and the fully connected layer is used to output results.

[0054] The input layer of the parallel classification network loads the preprocessed dataset, that is, {P i |i=1,...,n}.

[0055] like Figure 1 , Figure 2 As shown, the output results of the basic network are obtained, including:

[0056] Pass the n×3 point cloud data through the feature alignment layer to align the point cloud features;

[0057] After the aligned n×3 point cloud data passes through the feature extraction layer and undergoes three convolution operations, the output dimensions are n×128, n×256, n×512, n×1024, and n×2048 features;

[0058] After the maximum pooling operation, the maximum value is calculated in the second dimension and the output dimensions are 1×128, 1×256, 1×512, 1×1024, and 1×2048 global features;

[0059] After processing through a fully connected layer composed of linear layers, the scores of k categories of size N×k are obtained, where N is the batch size, as the output result of the basic network.

[0060] The results of the five parallel classification networks are fused to obtain the output of the pre-trained parallel classification network.

[0061] The structure of the feature pyramid network is a pyramid structure with high-level semantic features, which is composed of semantics from low to high levels. The feature pyramid network is constructed from top to bottom, sampling lower-dimensional features to match the dimensional features of the upper layer, and then merging them by element addition. After repeated sampling, addition and merging, the final feature map is generated.

[0062] The feature extraction layer uses a feature pyramid network for feature extraction, including:

[0063] Features are extracted in order from low dimension to high dimension, and the extracted features are upsampled twice from low dimension to high dimension and added, and the features of the corresponding feature dimensions are output.

[0064] Specifically, Figure 3 As shown in the figure, features from low dimension to high dimension are extracted in sequence, with sizes of n×256, n×512, n×1024, and n×2048 respectively. The extracted features are upsampled twice from low dimension to high dimension and added in sequence to construct a feature pyramid network composed of low dimension to high dimension features. This method adopts a top-down path, through the gradual synthesis of low-dimensional and high-dimensional features, and repeats this fusion process until a global feature of size n×2048 is generated.

[0065] like Figure 1 As shown in the figure, all dimensions of features use a shared fully connected layer, and a fixed feature dimension is used in all feature maps. In this method, matching fully connected layers are used for features of different dimensions, matching 128, 256, 512, 1024, and 2048 dimensional features respectively, and outputting the corresponding classification results.

[0066] It should be noted that the present invention uses Stanford University's ModelNet40 training samples as the data set and uses the loss function for guidance to train the basic network.

[0067] Specifically, the dataset contains 12,311 point cloud objects, including 9,843 training set objects and 2,468 test set objects, with a total of 40 object categories. Each point cloud object used in training and testing consists of 1,024 points.

[0068] Specifically, the adopted loss function is introduced from the negative log-likelihood loss.

[0069] Assume that there is a data distribution P data (x), define a parameter θ and a data distribution model P determined by θ G (x; θ), where θ makes P as much as possible G (x;θ) and P data (x) is similar. data (x) Collect m samples x1, x2, ... x3, and calculate the data distribution model of each sample as P G (x; θ), so the likelihood function is defined as:

[0070]

[0071] Taking the logarithm of (1) gives formula (2):

[0072]

[0073] Since the log-likelihood is the logarithm of the probability distribution, the value of the probability p(x) is in the interval [0,1], and after taking the logarithm, it is in the interval (-∞,0). Then add a minus sign in front, and it becomes the interval [0,+∞), thus we can get formula (3):

[0074] Therefore, the negative log-likelihood loss is defined as:

[0075]

[0076] In formulas (1)-(3), x i represents the sample, m represents the number of training samples, P G (x i ; θ) represents the distribution model of each sample data determined by parameter θ.

[0077] When using negative log-likelihood loss, two tensors, i.e. two multidimensional arrays, are required: one is the prediction vector, i.e. the output result, and the other is the label. The prediction vector is a one-dimensional tensor of size N×k, where N is the number of point cloud objects in a batch and k is the classification probability of each class. The size of the label is also N×k, indicating the index number of each class in the k classes of N point cloud objects.

[0078] Step 3: Use the pre-trained deep Q network to fuse the output results of the parallel network to obtain the final result of point cloud classification.

[0079] The mainstream feature fusion method uses convolutional neural networks to achieve result selection. However, there is a bottleneck in the accuracy of result selection by convolutional neural networks. In order to solve this defect, the present invention uses a deep Q network to fuse the output results of the parallel network. The deep Q network uses deep learning to solve the problem of the best strategy in Q learning. Given an environmental state, a corresponding behavior is selected according to a certain strategy, and after executing this behavior, the environmental state will change according to the behavior and turn to a new state. Each time a new behavior is executed according to the strategy, an incentive value will be obtained. The corresponding program adjusts the execution strategy according to the size of the incentive obtained, so that when the final state and the terminal state are reached, the accumulated incentive value accumulated by executing the current strategy is the largest.

[0080] In the application scenario of the present invention, what the deep Q network needs to learn is the strategy of selecting the best results in the parallel models. By manually labeling the item categories that are best classified by each model and using them as data sets to train the deep Q network, a better fusion effect can be obtained than the convolutional neural network method.

[0081] For each state, the deep Q network only needs to perform one forward calculation, instead of a number of forward calculations proportional to the number of behaviors as in traditional neural networks. This solves the problem of insufficient attention to local objects and difficulty in classifying complex and small objects, while improving the robustness of the system and solving the problem of low overall classification accuracy due to insufficient attention to local details of point clouds.

[0082] Use function fitting Q learning, that is, Q(s,a;λ)≈Q * (s, a), where λ represents the model parameter, and the Q value is obtained by end-to-end fitting of a deep neural network.

[0083] Embodiment 2:

[0084] This embodiment uses the ModelNet40 data set to verify the accuracy of a multi-scale point cloud classification method provided by the present invention.

[0085] like Figure 4 Shown is a 3D unordered point cloud target to be classified, consisting of 1024 points.

[0086] This embodiment designs hyperparameters based on Point Net and adjusts them based on actual conditions. When the batch size is set larger, the convergence speed is faster, but too large a batch size will cause generalization problems and consume video memory; if the batch size is set too small, it will be difficult to converge in time, so in this embodiment, the size of each batch is set to 64. The video memory of the device 1080ti is 11GB. When the size of each batch is set to 64, the video memory consumption is 10609MB.

[0087] The parallel classification network is trained using the negative log-likelihood loss function, and the learning rate is set to 0.001, which is optimized based on the callback function. The specific monitoring indicator is the test set accuracy, the decay waiting iteration number is set to 7, and the decay rate is set to 0.7. When the tracking accuracy does not improve after 7 consecutive iterations, the learning rate will be changed. This embodiment intends to iterate the training 120 times, and the reference indicator for model preservation is the test set accuracy. The model will be saved only when the test set accuracy is higher than the best test set accuracy.

[0088] The results are divided into overall accuracy and classification accuracy. The overall accuracy is the ratio of the number of correctly predicted samples to the total number of samples, and the average category accuracy specifically refers to the average classification accuracy of each of the 40 object categories.

[0089] The results of training the parallel classification network using the negative log-likelihood loss function and the PointNet and PointNet++ codes are shown in Table 1.

[0090] Table 1 Average classification accuracy and overall classification accuracy

[0091] method Average classification accuracy Overall classification accuracy PointNet 86.0 89.2 PointNet++ 87.30 90.7 The present invention 87.34 90.8

[0092] At the same time, the object classification accuracy of each category in the ModelNet40 classification was tested and compared with the object classification accuracy of PointNet and PointNet++. The object classification results are shown in Table 2.

[0093] Table 2 Object classification accuracy

[0094] airplane tub Cone Dressing Table Glass Box Guitar desk lamp piano sofa PointNet 0.99 0.9101 0.8906 0.8055 0.9508 0.9665 0.8906 0.9040 0.9732 PointNet++ 1.0 0.9023 0.8906 0.7170 0.9486 0.9799 0.8906 0.9330 0.9732 The present invention 1.0 0.9218 0.9453 0.8263 0.9575 1.0 0.9453 0.9397 0.9821

[0095] Among them, cone, guitar, lamp, piano have been significantly improved, while airplane, bathtub, dressing table, glass box, sofa have been improved less. From the classification results, we can see that the network has achieved good performance and the results have been significantly improved.

[0096] The network designed by the present invention is compared with PointNet and PointNet++ in sparsity experiments to compare their robustness. Figure 5 As shown in the figure, the number of points contained in a single point cloud target is 1024, 512, 256, and 128 respectively. The recognition of the point cloud set composed of different numbers of sampling points decreases as the number of sampling points decreases.

[0097] The overall classification accuracy of the multi-scale point cloud classification method proposed in this invention under the above different point numbers is as follows: Figure 6 As shown in Figure 7As shown. When the number of points constituting the point cloud target is 1024, the overall classification accuracy of the multi-scale point cloud classification method proposed in the present invention and PointNet and PointNet++ are 0.90834, 0.89191, and 0.90513, respectively, and the average classification accuracy is 0.87132, 0.84801, and 0.86758, respectively. When the number of selected points is reduced to 128 points, the overall classification accuracy of the multi-scale point cloud classification method proposed in the present invention and PointNet and PointNet++ are 0.90233, 0.87669, and 0.78663, respectively, and the average classification accuracy is 0.86130, 0.82923, and 0.73670, respectively. The overall accuracy and average accuracy of the multi-scale point cloud classification method proposed in the present invention decreased by 0.006 and 0.01 respectively when the number of points increased from 1024 to 128, while PointNet decreased by 0.0153 and 0.019 respectively, and PointNet++ decreased by 0.1185 and 0.13088 respectively. Due to the local sampling mechanism adopted in PointNet++, the network has high requirements for the number of points in the point cloud set. When the number of points decreases rapidly, the accuracy will decrease rapidly. Compared with PointNet, the network designed by the present invention can select results for objects of different sizes, so it also performs better than PointNet in robustness tests.

[0098] Embodiment three:

[0099] An embodiment of the present invention provides a multi-scale point cloud classification system, including:

[0100] Acquisition module: obtain 3D disordered point cloud data;

[0101] Feature extraction module: Use a pre-trained parallel classification network to extract features and classify the acquired point cloud data to obtain output results; wherein the parallel classification network includes multiple basic networks with the same structure;

[0102] Feature fusion module: Use the pre-trained deep Q network to fuse the output results of the parallel network to obtain the final result of point cloud classification.

[0103] Embodiment 4:

[0104] An embodiment of the present invention provides a computing device, including a processor and a storage medium;

[0105] The storage medium is used to store instructions;

[0106] The processor is used to operate according to the instructions to execute the steps of the method described in embodiment 1.

[0107] Embodiment five:

[0108] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the method described in the first embodiment are implemented.

[0109] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0110] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0111] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0112] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0113] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A multi-scale point cloud classification method, characterized in that: include: Get 3D unordered point cloud data; Use a pre-trained parallel classification network to extract features and classify the acquired point cloud data to obtain an output result; wherein the parallel classification network includes multiple basic networks with the same structure; wherein there are 5 basic networks, each of which includes a feature alignment layer, a feature extraction layer and a fully connected layer in sequence; according to the different feature dimensions of the point cloud input to the basic network, the basic networks are distinguished, and the dimensions of the output features of different layers in the same basic network are the same; Wherein, the output result is obtained, including: Pass the n×3 point cloud data through the feature alignment layer to align the point cloud features; After the aligned n×3 point cloud data passes through the feature extraction layer and undergoes three convolution operations, the output dimensions are n×128, n×256, n×512, n×1024, and n×2048 features; After the maximum pooling operation, the maximum value is calculated in the second dimension and the output dimensions are 1×128, 1×256, 1×512, 1×1024, and 1×2048 global features; After processing through a fully connected layer consisting of a linear layer, the scores of k categories of size N×k are obtained, where N is the batch size, as the output result of the basic network; The output results of the parallel networks are fused using a pre-trained deep Q network to obtain the final result of point cloud classification.

2. The multi-scale point cloud classification method according to claim 1, characterized in that: The feature extraction layer uses a feature pyramid network to extract features, including: Features are extracted in order from low dimension to high dimension, and the extracted features are upsampled twice from low dimension to high dimension and added, and the features of the corresponding feature dimensions are output.

3. The multi-scale point cloud classification method according to claim 1, characterized in that: The loss function is used to train each basic network constituting the parallel network. The loss function is expressed by the following formula: In formula (1), x i represents the sample, m represents the number of training samples, P G (x i ; θ) represents the distribution model of each sample data determined by parameter θ.

4. The multi-scale point cloud classification method according to claim 1, characterized in that: The dataset used for training the deep Q network is the item category with the best classification effect of each basic model through manual annotation.

5. A multi-scale point cloud classification system, characterized in that: include: Acquisition module: used to obtain 3D disordered point cloud data; Feature extraction module: used to use a pre-trained parallel classification network to extract features and classify the acquired point cloud data to obtain output results; wherein the parallel classification network is composed of multiple basic networks with the same structure; wherein there are 5 basic networks, each of which includes a feature alignment layer, a feature extraction layer and a fully connected layer in sequence; according to the different feature dimensions of the point cloud input to the basic network, the basic networks are distinguished, and the dimensions of the output features of different layers in the same basic network are the same; Wherein, the output result is obtained, including: Pass the n×3 point cloud data through the feature alignment layer to align the point cloud features; After the aligned n×3 point cloud data passes through the feature extraction layer and undergoes three convolution operations, the output dimensions are n×128, n×256, n×512, n×1024, and n×2048 features; After the maximum pooling operation, the maximum value is calculated in the second dimension and the output dimensions are 1×128, 1×256, 1×512, 1×1024, and 1×2048 global features; After processing through a fully connected layer consisting of a linear layer, the scores of k categories of size N×k are obtained, where N is the batch size, as the output result of the basic network; Feature fusion module: used to fuse the output results of the parallel network using the pre-trained deep Q network to obtain the final result of point cloud classification.

6. A computing device, characterized in that including processor and storage medium; The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.