A 3D Point Cloud Classification Method Based on a Small Amount of Training Data

Abstract

The present invention relates to the field of feature information recognition technology and discloses a 3D point cloud classification method based on a small amount of training data. The method comprises the following steps: establishing a 3D point cloud model of a target point, randomly acquiring multiple 3D point cloud sample data, and training the model using a Siamese network deep learning model; improving the PointNet++ network, utilizing the concept of a convolutional neural network for hierarchical feature learning, and adding the distance between two points for global feature extraction; using the improved PointNet++ network as a shared network of the Siamese network to establish a 3D point cloud classification model; collecting all 3D point cloud data within the target point, and classifying the obtained 3D point cloud data using the 3D point cloud classification model. The present invention has the beneficial effect of completing classification using a small amount of sample data and improving the accuracy of classification results in the case of limited point cloud data.

Description

Technical Field

[0001] The present invention relates to the technical field of feature information recognition, and in particular to a three-dimensional point cloud classification method based on a small amount of training data. Background Art

[0002] With the promotion and popularization of 3D point cloud technology, the technology of using image recognition technology combined with 3D point cloud technology to accurately classify target objects has gradually been widely used, especially for objects with very large scene restrictions, such as the identification and classification of industrial equipment. Because it is very difficult to manually collect equipment information and then accurately classify it, the current application rate of 3D point cloud classification technology in industry is relatively low.

[0003] On the other hand, the classification of three-dimensional point clouds also has many defects. For example, a large amount of training data is required as classification sample support to ensure the accuracy of the classification results. Currently, models based on traditional machine learning and deep learning require a large amount of data for model training, but in practical applications, a large number of data samples are often not available, especially in industry. Not only is the equipment difficult to move, but the equipment is also too large to collect full-view information of the equipment. In addition, some existing three-dimensional point cloud classification methods do not consider the integration of global features for sample training, resulting in the point cloud information features extracted by the model being not comprehensive enough, which leads to inaccurate classification results. Summary of the Invention

[0004] The present invention aims to provide a 3D point cloud classification method based on a small amount of training data to improve the accuracy of 3D point cloud classification results for industrial equipment.

[0005] To achieve the above object, the present invention adopts the following technical solution: a three-dimensional point cloud classification method based on a small amount of training data, comprising the following steps:

[0006] Step S1: Establish a 3D point cloud model of the target point, randomly obtain multiple 3D point cloud sample data, and use the Siamese network deep learning model for model training;

[0007] Step S2: Improve the PointNet++ network, use the idea of convolutional neural network to perform hierarchical feature learning, and add the distance between two points to perform global feature extraction;

[0008] Step S3: Use the improved PointNet++ network as a shared network of the Siamese network to establish a 3D point cloud classification model;

[0009] Step S4: Collect all three-dimensional point cloud data within the target point, and classify the obtained three-dimensional point cloud data using a three-dimensional point cloud classification model.

[0010] The principles and advantages of this solution are as follows: in practical applications, after collecting a small amount of data, the Siamese deep learning model is used for model training. This model measures the similarity between two inputs by learning the similarities and dissimilarities of features between samples, thereby achieving accurate classification with a small amount of sample data. On the other hand, this solution uses an improved PointNet++ network, using the Euclidean distance from the point cloud coordinates to the origin coordinates as the global feature of the point cloud data. By integrating local features extracted by the PointNet++ network, they serve as the basis for point cloud data classification, improving classification accuracy. Compared with existing technologies, the advantage of this solution is that, compared with other point cloud classification algorithms, this method effectively integrates local and global features under limited sample conditions, and combines the advantages of the Siamese network to further improve classification accuracy under limited point cloud data conditions. It can accurately train the model with a small amount of training. Combined with the improved PointNet++ network, it can accurately classify three-dimensional point cloud data, greatly reducing the amount of data calculation while effectively ensuring the accuracy of the classification results.

[0011] Preferably, as an improvement, in step S1, when performing model training, the similarity between two input sample data is measured by learning the similarity and dissimilarity of features between the sample data.

[0012] Beneficial effects: The Siamese network uses two different inputs and passes through two similar sub-networks with the same architecture, parameters, and weights. These two sub-networks mirror each other, so any changes to the architecture, parameters, and weights of any sub-network also apply to the other sub-networks, greatly improving the applicability and accuracy of the training model.

[0013] Preferably, as an improvement, in step S2, when performing hierarchical feature learning, a local hierarchical module consisting of a sampling layer, a grouping layer and a feature extraction layer is constructed.

[0014] Beneficial effect: Through this setting, it is possible to utilize the synergy between multiple levels to quickly complete multi-level feature learning, thereby greatly improving the efficiency of model training while ensuring the effectiveness of model training.

[0015] Preferably, as an improvement, the sampling layer selects the local area centroid from the original points by the farthest point sampling method; the grouping layer groups the points around the centroid by the ball query method; the feature extraction layer extracts local features by the PointNet network, and then completes multi-level feature learning through the connected hierarchical modules.

[0016] Beneficial effects: By selecting the centroid of the local area during sampling, the most ideal target point can be collected quickly and accurately. The collected points are then grouped through the ball query method, which not only improves the efficiency of grouping but also ensures the accuracy of the grouping results. Finally, feature extraction is performed to quickly complete multi-level feature learning, making the model training progress more efficient.

[0017] Preferably, as an improvement, adding the distance between two points for global feature extraction includes the following steps: first, the original point cloud data is arranged in ascending order in the X-axis direction, and then a number of points are evenly selected after setting the step size, and the distance between each point and the origin is calculated by the Euclidean distance formula, and the distance feature is added to the PointNet++ network.

[0018] Beneficial effect: Through this setting, the Euclidean distance from the point cloud coordinates to the origin coordinates can be used as the global feature of the point cloud data, and then the local features extracted by the PointNet++ network can be integrated as the basis for point cloud data classification, effectively improving the accuracy of 3D point cloud classification.

[0019] Preferably, as an improvement, the step size is set to 8, and the number of points is selected to be 128.

[0020] Beneficial effect: This setting can not only effectively ensure the simplification of training data, thereby improving the training effect of the model, but also ensure the accuracy of the trained model in classifying three-dimensional point clouds.

[0021] Preferably, as an improvement, the input of the three-dimensional point cloud classification model is a pair of identical or different sample data, and the final output is the degree of similarity between the two sample data.

[0022] Beneficial effect: Through this setting, regardless of whether the input data is the same or different, the model can determine the similarity between the two samples, thereby providing the most direct and reliable basis for the subsequent classification of the three-dimensional point cloud and facilitating the classification and identification of target points.

[0023] Preferably, as an improvement, in step S4, when the obtained three-dimensional point cloud data is classified using the three-dimensional point cloud classification model, all three-dimensional point cloud data within the target point are classified according to the similarity of the sample data output by the model.

[0024] Beneficial effect: Through this setting, the similarity between three-dimensional point clouds can be associated with the similarity of the identified objects themselves, thereby facilitating the identification and classification of objects.

[0025] Preferably, as an improvement, when the Siamese network deep learning model is used for model training, an iterative farthest point sampling algorithm is used to select a subset from the given input points, and then the point farthest from other points is found from the subset.

[0026] Beneficial effect: Through this setting, the iterative farthest point sampling algorithm can be used to select and locate data, and the optimal distance point can be found from the located data, thereby quickly completing the training of the model.

[0027] Preferably, as an improvement, in step S4, after the obtained 3D point cloud data is classified using the 3D point cloud classification model, the classified 3D point cloud data is matched with the actual object one by one, and the matching results are uploaded to the cloud server.

[0028] Beneficial effect: Through this setting, after all the three-dimensional point cloud data within the target point are classified and identified, they can be matched according to the characteristics of different actual objects and the attributes of the three-dimensional point cloud data, so as to quickly and accurately convert the three-dimensional point cloud classification into the classification of the actual object. Finally, the classification results are uploaded to the cloud for storage, which not only ensures storage security but also facilitates access and viewing. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a flow chart of a first embodiment of a three-dimensional point cloud classification method based on a small amount of training data according to the present invention.

[0030] Figure 2 Schematic diagram of a Siamese network in accordance with an embodiment of a three-dimensional point cloud classification method based on a small amount of training data according to the present invention.

[0031] Figure 3 This is a schematic diagram of feature extraction in Example 1 of a three-dimensional point cloud classification method based on a small amount of training data according to the present invention.

[0032] Figure 4 Schematic diagram of a three-dimensional point cloud classification model according to an embodiment of a three-dimensional point cloud classification method based on a small amount of training data of the present invention.

[0033] Figure 5 Schematic diagram of machine tool types in Example 1 of a three-dimensional point cloud classification method based on a small amount of training data according to the present invention.

[0034] Figure 6 This is a schematic diagram of one point cloud enhancement according to an embodiment of a three-dimensional point cloud classification method based on a small amount of training data of the present invention.

[0035] Figure 7 Schematic diagram comparing experimental results of Example 1 of a three-dimensional point cloud classification method based on a small amount of training data of the present invention. DETAILED DESCRIPTION

[0036] The following is further described in detail through specific implementation methods:

[0037] Example 1:

[0038] This embodiment is basically as shown in the attached Figure 1 As shown: A 3D point cloud classification method based on a small amount of training data includes the following steps:

[0039] Step S1: Establish a 3D point cloud model of the target point, randomly obtain multiple 3D point cloud sample data, and use the Siamese network deep learning model for model training;

[0040] Step S2: Improve the PointNet++ network, use the idea of convolutional neural network to perform hierarchical feature learning, and add the distance between two points to perform global feature extraction;

[0041] Step S3: Use the improved PointNet++ network as a shared network of the Siamese network to establish a 3D point cloud classification model;

[0042] Step S4: Collect all three-dimensional point cloud data within the target point, and classify the obtained three-dimensional point cloud data using a three-dimensional point cloud classification model.

[0043] Specifically, in step S1, when the model is trained, the similarity and dissimilarity of features between the sample data are learned to measure the similarity between the two input sample data, thereby enabling accurate classification with a small amount of sample data.

[0044] As attached Figure 2 As shown in the figure, the Siamese network is generally composed of two symmetrical neural networks, which share weights and structures and are finally connected together by an energy function. The purpose of the Siamese network is to determine whether two input values are similar or dissimilar. The Siamese network is used in this scheme to ensure that data features can be learned from a small amount of sample data to the greatest extent, thereby facilitating subsequent 3D point cloud data feature recognition and classification.

[0045] On the other hand, two symmetrical neural networks are selected to form a Siamese network, and they share weights and structures. Finally, the two networks are connected together through an energy function to determine whether the two variable values ​​input at the beginning are similar, which facilitates the subsequent three-dimensional point cloud classification. During the training process of sample data, the processing of the sample first reaches the ideal value state, otherwise the sample data will need to be processed twice, thereby improving the accuracy of data processing and streamlining the subsequent process.

[0046] To address the PointNet network's inability to extract local contextual features, this solution improves the network to create the PointNet++ network. PointNet++ draws on concepts similar to convolutional neural networks for hierarchical feature learning. First, a local hierarchical module is constructed, consisting of a sampling layer, a grouping layer, and a feature extraction layer. The sampling layer selects the local neighborhood centroid from the original points using the farthest point sampling method. The grouping layer groups points around the centroid using a sphere query method. The feature extraction layer extracts local features using the PointNet network, thus completing multi-level feature learning through connected hierarchical modules.

[0047] As attached Figure 3 As shown, to better capture the global features of point cloud data, this solution, based on the PointNet++ network, uses the Euclidean distance between the point cloud coordinates and the origin coordinates as the global feature of the point cloud data. By integrating the local features extracted by the PointNet++ network, they serve as the basis for point cloud data classification. First, the original point cloud data is sorted in ascending order along the X-axis. Then, 128 points are evenly selected with a step size of 8. The distance between each point and the origin is calculated using the Euclidean distance formula. The distance feature is then added to the PointNet++ network for model training.

[0048] Specifically, the distance formula is as follows:

[0049]

[0050] Among them, (X1, X2, ..., X n ), (y1, y2, …y n ) are the coordinates of the point.

[0051] At the same time, the contrast loss function is used to train the network. The contrast loss function is:

[0052]

[0053] Where d = || a n -b n ||2 represents the Euclidean distance between the two sample features, y is the label indicating whether the two samples match, and margin is the set threshold. In this embodiment, margin is set to 2.0.

[0054] On the other hand, an adaptive moment estimation optimizer is used, and the initial learning rate is set to 0.005.

[0055] m t =β1m t-1 +(1-β1)g t

[0056] v t =β2v t-1 +(1-β2)(g t ) 2 ,

[0057]

[0058] Among them, β1 is the first-order moment attenuation coefficient, β2 is the second-order moment attenuation coefficient, L is the loss function, η is the learning rate, g t is the loss function for θ t Find the partial derivative, m t is the first-order moment estimate of the gradient at time t in momentum form, v t is the second-order moment estimate of the gradient at time t in momentum form, is the bias-corrected first-order moment estimate, is the bias-corrected second moment estimate.

[0059] To further improve the accuracy of point cloud classification with limited sample data, this solution adds the distance between two points to perform global feature extraction on the basis of the original PointNet++ network. The improved PointNet++ network is used as a shared network of the Siamese network to establish a small sample learning model, and the final 3D point cloud classification model of this solution is obtained, which can ensure that the 3D point cloud classification task can be completed with limited sample data.

[0060] Specifically, as attached Figure 4 As shown, the input of the 3D point cloud classification model is a pair of identical or different sample data, and the final output is the similarity between the two sample data. This enables the recognition and classification of corresponding objects to be completed by comparing the differences between different objects and classifying the 3D point cloud.

[0061] Specifically, during the data training process, the model uses the iterative farthest point sampling algorithm to select a subset from the given input points, and then finds the point farthest from other points in the subset; at the same time, when using iterative farthest point sampling, the local area centroid of the data sample points is given priority.

[0062] After the model training is completed, the model is tested and the one-shot K-way and N-shot K-way test modes are constructed. Specifically, in the one-shot K-way test mode, given a test sample to be classified And each class has a sample reference support set S, which contains a total of K reference classification samples (y1…y k ),

[0063] S={(x1,y1),…,(xK ,y K )}

[0064] Then, the most similar sample result is output from the twin model according to the paired input in the support set as the classification result for the test sample:

[0065] Then, in the N-shot K-way test mode, given a support set S with N sample references for each class, (S1, ..., S N ), and then determine which support set class the test sample is most similar to to determine its category, as shown in the following formula: Simulate N-shot K-way testing by using N one-shot K-way tests.

[0066] For example, the Five-shot K-way method repeats the one-shot K-way test five times to simulate the test. Finally, five test similarity vectors (P1, P2, P3, P4, P5) are obtained, and the probability of the same reference category is accumulated. The maximum cumulative probability value is found to determine the most similar category as the classification result of the test sample.

[0067] Specifically, in step S4, when the 3D point cloud classification model is used to classify the obtained 3D point cloud data, all 3D point cloud data within the target point are classified based on the similarity of the sample data output by the model. The trained model can quickly and accurately identify the similarity between input samples, thereby completing the recognition and classification of the similarity between different 3D point cloud data. Subsequently, based on the distinctiveness of the 3D point clouds, the corresponding object characteristics are distinguished, completing the recognition and classification of the physical object.

[0068] As attached Figure 5 As shown, in this embodiment, the selected target point is the production workshop, and the objects that need to be identified and classified are the 10 common types of machine tools in the workshop, namely, tool grinders, gear shaping machines, bow grinders, vertical lathes, balancing cranes, ordinary lathes, grinding machines, digital milling machines, milling machines, and shapers.

[0069] Because each machine tool has a different size and shape, the point cloud data obtained for each machine tool is quite unevenly distributed. Therefore, the machine tool point cloud data must be resampled to meet the network's requirements. In point cloud, resampling refers to the process of removing and adding points from the existing point cloud according to a certain sampling rule to achieve a certain point cloud quantity requirement. This ensures that the overall geometric characteristics of the point cloud remain unchanged while changing the point cloud density to meet the network's requirements.

[0070] Point cloud resampling includes upsampling and downsampling, of which the three most common downsampling methods are voxel downsampling, uniform downsampling, and random downsampling. In order to avoid destroying the regularity of point cloud distribution and density uniformity, and to ensure minimal information loss during the downsampling process, this embodiment uses voxel downsampling to sample the original point cloud. Voxels divide the three-dimensional space into small three-dimensional grids. Voxel downsampling represents the points falling in each voxel with one point, and generally takes the average value of the coordinates of all points in the voxel. This method can obtain a fixed number of point clouds, thereby reducing the complexity of the calculation. In the dataset of this article, the number of point clouds for each type of machine tool is uniformly downsampled to 5000.

[0071] After the collection is completed, considering that the amount of original point cloud data collected is too small, it may cause problems such as inability to train or poor training results for data-driven deep learning. Currently, data enhancement methods such as rotation, flipping, scaling, translation, and adding noise can be used to expand the data from the existing data.

[0072] As attached Figure 6 As shown, in this embodiment, a point cloud enhancement method using rotation around the gravity axis z-axis and translation perpendicular to the z-axis was selected. Data augmentation was performed on 10 types of machine tools, expanding each type to 240 data points. Of these, 100 were obtained through rotation and 140 through random translation. This ultimately resulted in a balanced machine tool point cloud dataset, with 190 data points from each type used as the training set and 50 data points used as the test set. In the figure, a shows the original machine tool point cloud, and b shows the rotated and translated machine tool point cloud.

[0073] In order to speed up the convergence of the model and eliminate the impact of unit and scale differences between data, it is necessary to normalize the data. In this embodiment, the Max-Min normalization method is used to normalize the data:

[0074]

[0075] Among them, x * represents the normalized data, x is the original data, min(x) and max(x) represent the minimum and maximum values in the original data respectively.

[0076] After model training and verification, a specific experimental comparison was conducted on the models of this solution. Three models were designed, including the original PointNet++ model, the small sample model based on the original PointNet++, and the small sample model based on the improved PointNet++ network in this solution. Comparative experimental verification was carried out under limited sample conditions. 40, 70, 100, 130, 160, and 190 training samples were randomly selected from each category, and a series of comparative experiments were conducted.

[0077] As attached Figure 7 As shown in the figure, when the number of samples per class is 40 and 70, the accuracy of the small-sample model based on the improved PointNet++ is 69.4% and 77.0%, respectively, which is nearly 10 percentage points higher than the original PointNet++ model. The accuracy of the small-sample model based on the original PointNet++ is 63.8% and 73.9%, respectively, which is also higher than the original PointNet++ model. This proves the effectiveness of the small-sample model with limited data. This is due to the fact that the Siamese twin network is a similarity measurement method that can effectively extract sample features when the sample size is small.

[0078] On the other hand, the accuracy of all three models increases with the number of samples. When the number of samples per class is greater than 100, which means that the number of samples is sufficient, the accuracy of the three models is not much different. Therefore, the 3D point cloud classification model in this solution has better performance on limited datasets, which makes it more advantageous in scenarios where sufficient training samples are not available.

[0079] When the number of training samples is 190, the recognition accuracy of the model of this scheme for all categories on the machine tool equipment dataset is compared with PointNet++ as shown in Table 1.

[0080] Table 1: Comparison of category recognition accuracy

[0081] Object PointNet++ 3D point cloud classification model Tool grinder 0.88 0.84 gear shaping machine 0.82 0.84 Bow grinder 0.74 0.70 Vertical rotary bed 0.90 0.96 Balance crane 0.92 0.98 Ordinary lathe 0.92 0.94 grinding machine 0.88 0.96 Digital milling machine 0.90 0.88 milling machine 0.84 0.84 Shaper 0.86 0.80

[0082] Therefore, the model in this scheme has improved the recognition accuracy of balancing cranes, vertical turning machines, ordinary lathes, and grinding machines. This shows that by adding distance as a global feature in this scheme, machine tool information with obvious contour features can be better collected, thereby improving the recognition ability of the model.

[0083] The specific implementation process of this embodiment is as follows:

[0084] The first step is to build a 3D point cloud model of the production workshop. Then, 3D point cloud sample data of multiple devices in the workshop is randomly obtained. The Siamese network deep learning model is used for model training. By learning the similarities and dissimilarities of features between sample data, the similarity between two input sample data is measured, thereby achieving accurate classification with a small amount of sample data.

[0085] The second step is to improve the PointNet++ network and construct a local hierarchical module consisting of a sampling layer, a grouping layer, and a feature extraction layer. The sampling layer selects the local area centroid from the original points through the farthest point sampling method, the grouping layer groups the points around the centroid through the ball query method, and the feature extraction layer extracts local features through the PointNet network, thereby completing multi-level feature learning through the connected hierarchical modules.

[0086] In the third step, the Euclidean distance from the point cloud coordinates to the origin coordinates is used as the global feature of the point cloud data. The local features extracted by the PointNet++ network are integrated and used as the basis for point cloud data classification. The original point cloud data is arranged in ascending order in the X-axis direction, and then 128 points are evenly selected with a step size of 8. The distance between each point and the origin is calculated using the Euclidean distance formula, and the distance feature is added to the PointNet++ network for model training.

[0087] In the fourth step, the improved PointNet++ network is used as a shared network for the Siamese network to establish a small sample learning model to obtain the final 3D point cloud classification model. Then, the 3D point cloud data of all equipment in the workshop is collected and the 3D point cloud classification model is used to compare the similarities between the 3D point cloud data. All 3D point cloud data is identified and classified, and the classification of all equipment in the workshop is completed.

[0088] With the promotion and popularization of 3D point cloud technology, the technology of using image recognition technology combined with 3D point cloud technology to accurately classify target objects has gradually been widely used, especially for objects with very large scene restrictions, such as the identification and classification of industrial equipment. Because it is very difficult to manually collect equipment information and then accurately classify it, the current application rate of 3D point cloud classification technology in industry is relatively low.

[0089] In the industrial field, especially in the classification of equipment in large factories, due to the large size of the equipment and the wide variety of equipment, the conventional classification methods for these equipment cannot be completed quickly. They can only rely on manual investigation and classification one by one. Not only is the process cumbersome but the efficiency is very low. The classification method using three-dimensional point clouds can achieve fast and accurate classification. Using the model to train sample data from the beginning can not only greatly simplify the samples, but also compare the differences between samples, and then use them to measure the similarity of the input. For example, two devices with similar appearances can compare their difference points from the very beginning of the input. When sampling, the local area centroid is selected from the original point by the farthest point sampling method, and then the points around the centroid are grouped by the grouping layer through the ball query method, so that the collected sample points have certain unified rules, which is convenient for subsequent three-dimensional point cloud classification.

[0090] This solution uses a Siamese deep learning model for model training. This model measures the similarity between two inputs by learning the similarities and dissimilarities of features between samples, enabling accurate classification of 3D point clouds of industrial objects with a small number of samples. Furthermore, by improving the PointNet++ network, the Euclidean distance between point cloud coordinates and the origin coordinates is used as a point cloud data feature. By integrating local features extracted by the PointNet++ network, this method serves as the basis for point cloud data classification. This effectively integrates local and global features, and combined with the advantages of the Siamese network, further improves classification accuracy with limited point cloud data. Testing has shown that this solution can achieve an accuracy of 92.18% for 3D point cloud classification with a small number of samples.

[0091] Example 2:

[0092] This embodiment is basically the same as the first embodiment, except that: in step S4, after the obtained three-dimensional point cloud data is classified using the three-dimensional point cloud classification model, the classified three-dimensional point cloud data is matched one by one with the actual object, and the matching results are uploaded to the cloud server.

[0093] Specifically, the cloud server used in this embodiment is Alibaba Cloud Server.

[0094] The specific implementation process of this embodiment is the same as that of the first embodiment, except that:

[0095] In the fourth step, the improved PointNet++ network is used as a shared network for the Siamese network to establish a small sample learning model and obtain the final 3D point cloud classification model. 3D point cloud data of all equipment in the workshop is then collected and used to compare the similarities between the 3D point cloud data using the 3D point cloud classification model. This allows for identification and classification of all 3D point cloud data, ultimately completing the classification of all equipment in the workshop. After classification, the 3D point cloud data is matched one by one with the actual equipment, and the matching results are uploaded to the Alibaba Cloud server.

[0096] This setup enables mapping and matching of 3D point clouds with actual workshop equipment and uploads them to Alibaba Cloud servers for storage. This not only ensures storage security but also facilitates access and viewing. As the amount of identified and classified data continues to increase, big data will gradually form, providing a reference for subsequent 3D point cloud classification and identification of similar objects.

[0097] Example 3:

[0098] This embodiment is essentially the same as the first embodiment, differing in that a 3D point cloud conditional random field classifier is constructed during model training. Specifically, the 3D point cloud of each individual object is classified, taking into account its individual, neighborhood, and local characteristics. The point model, edge model, and high-order clique model of the conditional random field are designed using independent points, connecting edges, and separating segments to complete the construction of the 3D point cloud conditional random field classifier.

[0099] The specific implementation process of this embodiment is the same as that of the first embodiment, except that:

[0100] The first step is to establish a 3D point cloud model of the production workshop, and then randomly obtain 3D point cloud sample data from multiple devices in the workshop. The Siamese network deep learning model is used for model training. By learning the similarities and dissimilarities of the features between the sample data, the similarity between the two input sample data is measured, so that accurate classification can be achieved with a small amount of sample data; at the same time, the 3D point clouds of the individual classification objects are classified, and their individual characteristics, neighborhood characteristics and local characteristics are comprehensively considered. The point model, edge model and high-order clique model of the conditional random field are designed using independent points, connecting edges and separating fragments to complete the construction of the 3D point cloud conditional random field classifier.

[0101] Through this setting, the constructed three-dimensional point cloud conditional random field classifier can be used to improve the subsequent recognition and classification efficiency and accuracy of three-dimensional point clouds, thereby improving the classification accuracy of all equipment in the workshop.

[0102] The above is only an embodiment of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme is not described in detail here. Ordinary technicians in the field are aware of all common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the field can improve and implement this scheme in combination with their own abilities under the inspiration given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A 3D point cloud classification method based on a small amount of training data, characterized by: The following steps are involved: Step S1: Establish a 3D point cloud model of the target point, randomly obtain multiple 3D point cloud sample data, and use the Siamese network deep learning model for model training; Step S2: Improve the PointNet++ network, use the idea of convolutional neural network to perform hierarchical feature learning, and add the distance between two points to perform global feature extraction; The method of adding the distance between two points to extract global features includes the following steps: first, the original point cloud data is arranged in ascending order in the X-axis direction, then a number of points are evenly selected after setting the step size, and the distance between each point and the origin is calculated using the Euclidean distance formula, and the distance feature is added to the PointNet++ network; Step S3: Use the improved PointNet++ network as a shared network of the Siamese network to establish a 3D point cloud classification model; Step S4: Collect all three-dimensional point cloud data within the target point, and classify the obtained three-dimensional point cloud data using a three-dimensional point cloud classification model.

2. The three-dimensional point cloud classification method based on a small amount of training data according to claim 1, characterized in that: In step S1, when the model is trained, the similarity between two input sample data is measured by learning the similarity and dissimilarity of features between the sample data.

3. The three-dimensional point cloud classification method based on a small amount of training data according to claim 1, characterized in that: In step S2, when performing hierarchical feature learning, a local hierarchical module consisting of a sampling layer, a grouping layer and a feature extraction layer is constructed.

4. The three-dimensional point cloud classification method based on a small amount of training data according to claim 3, characterized in that: The sampling layer selects the local area centroid from the original points by the farthest point sampling method; The grouping layer groups points around the centroid using a ball query method; the feature extraction layer extracts local features using a PointNet network, and then completes multi-level feature learning through connected hierarchical modules.

5. The three-dimensional point cloud classification method based on a small amount of training data according to claim 1, characterized in that: The step size is set to 8, and the number of points is selected to be 128.

6. The three-dimensional point cloud classification method based on a small amount of training data according to claim 1, characterized in that: The input of the three-dimensional point cloud classification model is a pair of identical or different sample data, and the final output is the similarity between the two sample data.

7. The three-dimensional point cloud classification method based on a small amount of training data according to claim 6, characterized in that: In step S4, when the obtained three-dimensional point cloud data is classified using the three-dimensional point cloud classification model, all three-dimensional point cloud data within the target point are classified according to the similarity of the sample data output by the model.

8. The three-dimensional point cloud classification method based on a small amount of training data according to claim 1, characterized in that: When the Siamese network deep learning model is used for model training, an iterative farthest point sampling algorithm is used to select a subset from a given input point, and then the point farthest from other points is found from the subset.

9. The three-dimensional point cloud classification method based on a small amount of training data according to claim 1, characterized in that: In step S4, after the obtained three-dimensional point cloud data is classified using the three-dimensional point cloud classification model, the classified three-dimensional point cloud data is matched with the actual object one by one, and the matching results are uploaded to the cloud server.

Images