A Damage Detection Method for Tensioned Membrane Structures Based on 3D Depth Point Cloud Segmentation

Abstract

This invention discloses a damage detection method for tensile membrane structures based on 3D depth point cloud segmentation. The method includes: acquiring 3D point cloud data of the tensile membrane structure surface; creating a 3D point cloud dataset of the tensile membrane structure surface; applying an attention mechanism to the PointNet++ network architecture to improve the PointNet++ architecture; inputting the improved PointNet++ model into the 3D point cloud dataset of the membrane surface for training; and inputting the trained improved PointNet++ model into a test set of the 3D point cloud dataset of the membrane surface for model testing. This method can solve the problems of slow inspection speed and susceptibility to subjective judgment and work experience in the visual inspection of tensile membrane structures with large and complex shapes and high ground elevations, thus reducing the accuracy of damage detection. It belongs to the field of damage monitoring of cable-membrane structure building materials.

Description

Technical Field

[0001] This invention relates to the fields of damage identification of tensile membrane structures and artificial intelligence three-dimensional point cloud technology, specifically to a method for detecting damage in tensile membrane structures based on three-dimensional depth point cloud segmentation. Background Technology

[0002] Tensile membrane structures fully utilize the characteristics of tensile curved membranes, offering a rich variety of shapes. They feature flowing, smooth, and elegant spatial curves, as well as transparent, simple, and bright column-free large-span spaces. Combined with slender yet powerful tension cables, stabilizing cables, and edge cables that balance the internal forces of the membrane, they are perfectly coordinated with the membrane surface, thus finding wider application in public buildings. The membrane surface of a tensile membrane structure is the most important structural unit maintaining the system; once the membrane surface ruptures, the entire structural support system will lose balance and collapse. According to incomplete statistics, since 2000, my country has built over a hundred large-scale cable-membrane structures. The earliest membrane structures built in China have been in service for nearly 20 years, and some tensile membrane structures have reached or are nearing their full service life. With more projects put into operation, the problems of membrane structure inspection and repair have gradually become apparent and increasingly serious. However, existing research in areas such as damage detection and health monitoring of tensile membrane structures is relatively lagging behind, currently relying mainly on staff to conduct inspections and monitoring periodically and at fixed locations using manual experience methods. However, as large public buildings, cable-membrane structures generally have a large and complex shape, with the membrane surface high above the ground, making manual damage detection challenging. Furthermore, manual visual inspection is slow and susceptible to subjective biases and experience. Differences in the professional experience of different inspectors and engineers, coupled with fatigue from high-intensity work, reduce the accuracy of damage detection in cable-membrane structures, easily leading to missed detections. If damage to tensile membrane structures could be quickly detected and identified, and maintenance personnel could be promptly notified for repair, the efficiency of damage detection would undoubtedly be greatly improved. Therefore, using more objective technical means to detect and monitor the performance of existing cable-membrane structures has become increasingly important.

[0003] In recent years, deep learning, as a crucial foundation of artificial intelligence, has developed rapidly, and data processing and computer vision have been widely applied to damage detection in civil structures. The combination of computer vision and deep learning overcomes the shortcomings of manual visual methods, greatly improving detection efficiency. However, the detection effect of this method is easily affected by natural light and ignores the three-dimensional features of the target structure, making it less effective for detecting surface defects, especially small targets. Tensioned membrane structures, with their spatial curved surfaces formed by tensioning, are characterized by limited recognition content and difficulty in identifying early, minor damage. Using computer vision to inspect the membrane surface of tensioned membrane structures is difficult for accurate detection and localization. In contrast, three-dimensional vision, as a non-contact, zero-destructive detection method, can achieve information integration, automation, intelligence, and precise control. Three-dimensional point cloud data contains rich information, including the three-dimensional coordinates, color information, normal vectors, and intensity values ​​of the measurement points. Compared to traditional image data, three-dimensional point cloud data is simpler; by recording the geometric structure of objects in space, it can display the spatial morphology of objects in three dimensions. Furthermore, this method is less affected by natural light. For early-stage, minor damage defects on the membrane surface, 3D point cloud data can better describe the defect's location on the membrane surface, enabling rapid detection and localization of defects. Moreover, the membrane surface detection method based on 3D point cloud data not only possesses the accuracy and real-time performance of deep learning methods but also provides spatial location information of the target, thus achieving better defect localization and classification.

[0004] 3D point cloud deep learning methods can achieve end-to-end feature learning of point cloud data through convolutional neural networks. Currently, PointNet and PointNet++ network architectures are commonly used. These models have the significant advantage of being lightweight, but their weakness lies in the need for improvement in capturing local features of the detected target. Therefore, to improve the local detection performance of 3D point clouds on the PointNet++ network architecture, it is necessary to strengthen the network architecture by applying point-by-point information from local regions to enhance the connections between features for network model training, thereby achieving better detection results. Currently, attention mechanisms, as a high-performance advanced feature learning module, have attracted extensive research in the field of machine vision. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the present invention aims to provide a method for detecting damage in tensile membrane structures based on three-dimensional deep point cloud segmentation. This method proposes applying an attention mechanism to the PointNet++ network architecture. By introducing attention weights to dynamically adjust the importance of different sampling points in the point cloud data, adaptive integration of local point features is achieved, resulting in guaranteed interpretability and generalization ability. This improves the accuracy and precision of three-dimensional point cloud semantic segmentation, enabling accurate detection of early-stage, minute local damage features on the surface of tensile membrane structures.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] This invention provides a method for damage detection of tensile membrane structures based on three-dimensional depth point cloud segmentation, specifically:

[0008] Step 1: Acquire 3D point cloud data of the tensile membrane structure surface. Use a 3D point cloud acquisition device to scan the membrane surface of the tensile membrane structure under different damage conditions, and collect point cloud data information of the membrane surface under different damage conditions.

[0009] Step 2: Create a 3D point cloud dataset of the tensile membrane structure surface;

[0010] Step 3: Apply the attention mechanism to the PointNet++ network architecture to improve the PointNet++ network architecture;

[0011] Step 4: Input the improved PointNet++ model with the membrane surface 3D point cloud dataset for training. Divide the processed tensile membrane structure point cloud dataset under different damage conditions into training and testing sets according to a certain ratio, and input them into the improved PointNet++ model for model training;

[0012] Step 5: Input the trained improved PointNet++ model into the membrane surface 3D point cloud test set for model testing. Input the membrane surface point cloud data of the test set into the trained improved PointNet++ model, infer the semantic category prediction values ​​of the test point cloud, and evaluate the prediction accuracy of the improved 3D point cloud model.

[0013] Preferably, in step 2, the collected point cloud information of the tensile membrane structure under different damage conditions is preprocessed. First, normal vector information is added to the point cloud information. Second, point cloud processing software is used to segment the point cloud data under different damage conditions. Then, labels are added to different damage types of the membrane surface. Finally, the preprocessed point cloud data of the membrane surface under different damage conditions is saved as follows: it contains the three-dimensional coordinate position information (x, y, z) and the three-dimensional normal vector information (N). x N y N z The text is a TXT file containing 7 columns of data, including data for labels and tags.

[0014] Preferably, in step 3, the attention mechanism method is applied to the PointNet++ network architecture. By introducing the AMS attention mechanism module based on the spatial dimension features of local point clouds and the AMC attention mechanism module based on the channel dimension features of local point clouds, the attention weights dynamically adjust the importance of different sampling points of the spatial dimension features and channel dimension features of local point clouds in the point cloud data. This achieves adaptive integration of local point features, thereby improving the local detection performance of 3D point clouds on the PointNet++ network architecture. It also strengthens the network architecture's ability to apply point-by-point information from local regions to enhance the connection between features, improves the semantic segmentation accuracy and precision of 3D point clouds, and enables accurate detection of early minor local damage features on the surface of tensile membrane structures.

[0015] Preferably, in step 3, the point cloud processed by the SA1 module, SA2 module and SA3 module is mapped to a high-dimensional feature (N2, D+C2), where D represents the point cloud data dimension, C2 represents the point cloud spatial feature dimension and N2 represents the point cloud channel feature dimension.

[0016] Step 1: Along the C2 local point cloud spatial feature dimension, use parallel max pooling and average pooling to fuse the spatial features of the point cloud input, thereby generating two spatial feature representations;

[0017] Step 2: Use convolution to train the weights, and use a shared-parameter MLP to train the spatial feature dimension of the aggregated features to generate attention weights;

[0018] Step 3: Activate the weights using the activation function S, and perform a dot product with the original input point cloud spatial feature dimension to generate new point cloud spatial features.

[0019] Step 4: Along the N2 local point cloud channel feature dimension, use parallel max pooling and average pooling to perform feature fusion on the point cloud input channel features, thereby generating two channel feature representations;

[0020] Step 5: Use convolution to train the weights, and use a shared-parameter MLP to train the channel feature dimension of the aggregated features to generate attention weights.

[0021] Step 6: Activate the weights using the activation function S, and perform a dot product with the original input point cloud channel feature dimensions to generate new point cloud channel features.

[0022] Preferably, in step 3, the point cloud processed by modules SA1, SA2, and SA3 is mapped to high-dimensional features (N2, D+C2), where D represents the point cloud data dimension, C2 represents the point cloud spatial feature dimension, and N2 represents the point cloud channel feature dimension. First, along the N2 local point cloud channel feature dimension, parallel max pooling and average pooling are used to fuse the point cloud input channel features, generating two channel feature representations. Second, convolution is used for weight training, and a shared-parameter MLP is used to train the channel feature dimension of the aggregated features to generate attention weights. Finally, an activation function S is used to activate the weights, and a dot product is performed with the original input point cloud channel feature dimension to generate new point cloud channel features.

[0023] Preferably, in step 3, the AMS attention mechanism module based on the spatial dimension features of local point clouds and the AMC attention mechanism module based on the channel dimension features of local point clouds are combined and applied to the PointNet++ network architecture. This enables the network architecture to first perform feature fusion on the spatial feature dimension of local point clouds and then perform feature fusion on the channel feature dimension of local point clouds, thereby improving the local detection performance of the PointNet++ network architecture for 3D point clouds and strengthening the network architecture's ability to apply point-by-point information from local regions to enhance the connection between features.

[0024] Preferably, in step 3, the three SA (set abstraction) modules in the PointNet++ network model architecture function as follows: These modules first sample a certain number of key points, then select points within a spherical region of a multi-scale (MSG) radius (R) surrounding each key point as a group. PointNet is then applied to these groups for point cloud feature extraction, and through continuous iteration, the global point cloud features are finally obtained. Compared to PointNet, the advantage of the PointNet++ network architecture lies in its multi-layered feature extraction structure, effectively extracting point cloud features.

[0025] The beneficial effects of this invention are as follows: The method proposed in this invention uses membrane surface damage of tensile membrane structures as the segmentation and detection object, and successively embeds the AMS attention mechanism based on local point cloud spatial dimension features and the AMC attention mechanism based on local point cloud channel dimension features into the existing PointNet++ network architecture, which can achieve good recognition and detection functions for local point cloud spatial dimension features and local point cloud channel dimension features. This method and system are characterized by high efficiency, convenience and low cost, and have good applicability to tensile membrane structures, and can be widely used in engineering inspection practice. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 The main flowchart of a method for detecting damage to tensile membrane structures based on three-dimensional depth point cloud segmentation is provided for an embodiment.

[0028] Figure 2 The flowchart for preprocessing point cloud data of tensile membrane structure under different damage conditions is provided for the embodiments.

[0029] Figure 3 This example provides a network model architecture based on the PointNet++ model architecture with hybrid AMS and AMC attention mechanisms.

[0030] Figure 4 The model architecture of the AMS attention mechanism based on local point cloud spatial dimension features provided for the embodiment;

[0031] Figure 5 The model architecture of the AMC attention mechanism based on local point cloud channel dimension features provided for the embodiment;

[0032] Figure 6 The example provides a network model architecture for a hybrid local point cloud spatial dimension feature (AMS) and local point cloud channel dimension feature (AMC) attention mechanism.

[0033] Figure 7 This is a schematic diagram illustrating point cloud sampling on the membrane surface and extraction of global features from point cloud groups at different scales, provided for an embodiment.

[0034] Figure 8 The network model architecture based on the PointNet++ model architecture provided in the embodiment, with the addition of a hybrid AMS and AMC attention mechanism, improves the training accuracy of the membrane surface damage;

[0035] Figure 9 The network model architecture based on the PointNet++ model architecture provided in the embodiment, with the addition of hybrid AMS and AMC attention mechanisms, demonstrates the membrane damage detection performance. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] This invention discloses a method for damage detection of tensile membrane structures based on three-dimensional depth point cloud segmentation, the steps of which include:

[0038] Step 1: Collect 3D point cloud data of the tensile membrane structure surface;

[0039] Step 2: Create a 3D point cloud dataset of the tensile membrane structure surface;

[0040] Step 3: Apply the attention mechanism to the PointNet++ network architecture to improve the PointNet++ network architecture;

[0041] Step 4: Input the improved PointNet++ model into the 3D point cloud dataset of the membrane surface and train it;

[0042] Step 5: Input the trained, improved PointNet++ model into the 3D point cloud test set of the membrane surface for model testing. Compared with machine vision methods, this invention is less affected by natural light, and for early, minor damage defects on the membrane surface, the 3D point cloud data can better describe the location information of the defect on the membrane surface, enabling rapid detection and localization of membrane surface defects.

[0043] Therefore, the membrane surface detection method based on 3D point cloud data not only possesses the accuracy and real-time performance of deep learning methods but also provides spatial location information of the detection target, thereby better enabling defect localization and classification. Addressing the challenges of the spatial curved surface formed by tensioning of tensile membrane structures, such as limited identification content and difficulty in identifying early, minute damages, as well as the lack of dedicated datasets and intelligent models in existing AI-based 3D point cloud methods, this invention proposes targeted improvement strategies. This results in a 3D point cloud semantic segmentation damage detection method and system for tensile membrane structures that is efficient, convenient, and low-cost, demonstrating good applicability to tensile membrane structures and making it widely applicable in engineering inspection practices.

[0044] like Figure 1 The diagram shown is the main flowchart of a method for detecting damage to tensile membrane structures based on three-dimensional depth point cloud segmentation, provided by an embodiment of the present invention.

[0045] The specific implementation process for collecting the three-dimensional point cloud data of the tensile membrane structure surface is as follows:

[0046] The membrane surface of the tensile membrane structure under different damage conditions was scanned using a 3D point cloud acquisition device. Considering the types of membrane surface damage in actual use and the influence of obstructions during daily inspections, the membrane surface point cloud damage dataset built in this patent includes four types: stains, scratches, holes, and obstructions (leaves and branches). A total of 20 3D point cloud files of damaged membrane surfaces were collected under different damage conditions of the tensile membrane structure.

[0047] The specific implementation process for creating the three-dimensional point cloud dataset of the tensile membrane structure surface is as follows:

[0048] like Figure 2 As shown, the point cloud information of the tensile membrane structure under four damage conditions is preprocessed. First, a normal vector (N) is added to the point cloud information of the damaged membrane surface. x N y N z The data is processed by first identifying the membrane surface point cloud data under different damage conditions, then segmenting the data using point cloud processing software, and finally adding labels to the membrane surface for different damage types. The preprocessed point cloud data under different damage conditions is then saved as separate data sets containing the three-dimensional coordinates (x, y, z) and three-dimensional normal vectors (N). x N y N z The dataset consists of 7 columns of TXT text containing data including labels and other information. After processing, there are 20 TXT texts of 3D point cloud datasets for different damaged membrane surfaces, each containing approximately 920,000 rows of data. The PointNet++ network model was trained using 80% of the sample data from the constructed 3D point cloud datasets for different damaged membrane surfaces of the tensile membrane structure. After training, the network model was tested using the remaining 20% ​​of the dataset.

[0049] The application of the attention mechanism to the PointNet++ network architecture improves the PointNet++ network architecture. The specific implementation process is as follows:

[0050] like Figure 3As shown, the attention mechanism method is applied to the PointNet++ network architecture. By introducing the AMS attention mechanism module based on the spatial dimension features of local point clouds and the AMC attention mechanism module based on the channel dimension features of local point clouds, the attention weights dynamically adjust the importance of different sampling points of the spatial dimension features and channel dimension features of local point clouds in the point cloud data. This achieves adaptive integration of local point features, thereby improving the local detection performance of 3D point clouds on the PointNet++ network architecture. It also strengthens the network architecture by applying point-by-point information from local regions to enhance the connection between features, improves the semantic segmentation accuracy and precision of 3D point clouds, and enables accurate detection of early minor local damage features on the surface of tensile membrane structures.

[0051] The AMS attention mechanism module based on local point cloud spatial dimension features is applied to the PointNet++ network architecture to improve the PointNet++ network architecture. The specific implementation process is as follows:

[0052] like Figure 4 As shown, the point cloud processed by modules SA1, SA2, and SA3 is mapped to high-dimensional features (N2, D+C2), where D represents the point cloud data dimension, C2 represents the point cloud spatial feature dimension, and N2 represents the point cloud channel feature dimension. First, along the local point cloud spatial feature dimension C2, parallel max pooling and average pooling are used to fuse the point cloud input spatial features, generating two spatial feature representations. Second, convolution is used for weight training, and a shared-parameter MLP is used to train the spatial feature dimension of the aggregated features to generate attention weights. Finally, the ReLU activation function is used to activate the weights, and a dot product is performed with the original input point cloud spatial feature dimension to generate new point cloud spatial features.

[0053] The AMC attention mechanism module based on local point cloud channel dimension features is applied to the PointNet++ network architecture to improve the PointNet++ network architecture. The specific implementation process is as follows:

[0054] like Figure 5As shown, the point cloud processed by modules SA1, SA2, and SA3 is mapped to high-dimensional features (N2, D+C2), where D represents the point cloud data dimension, C2 represents the point cloud spatial feature dimension, and N2 represents the point cloud channel feature dimension. First, along the N2 local point cloud channel feature dimension, parallel max pooling and average pooling are used to fuse the point cloud input channel features, generating two channel feature representations. Next, convolution is used for weight training, and a shared-parameter MLP is used to train the channel feature dimension of the aggregated features to generate attention weights. Finally, the ReLU activation function is used to activate the weights, and a dot product is performed with the original input point cloud channel feature dimension to generate new point cloud channel features.

[0055] The aforementioned application of the hybrid AMS attention mechanism module based on local point cloud spatial dimension features and the AMC attention mechanism module based on local point cloud channel dimension features to the PointNet++ network architecture improves the PointNet++ network architecture. The specific implementation process is as follows:

[0056] like Figure 6 As shown, the AMS attention mechanism module based on the spatial dimension features of local point clouds and the AMC attention mechanism module based on the channel dimension features of local point clouds are combined and applied to the PointNet++ network architecture. This enables the network architecture to first perform feature fusion on the spatial feature dimension of local point clouds and then perform feature fusion on the channel feature dimension of local point clouds. This improves the local detection performance of the PointNet++ network architecture for 3D point clouds and strengthens the network architecture's ability to use point-by-point information from local regions to enhance the connection between features.

[0057] The point cloud is mapped to high-dimensional features after processing by the SA1, SA2, and SA3 modules. The specific implementation process is as follows:

[0058] like Figure 7 As shown, the three SA (set abstraction) modules in the PointNet++ network model architecture function as follows: First, these modules sample a certain number of key points. Then, based on these key points, they select points within a spherical region of a multi-scale (MSG) radius (R) surrounding each key point as a group. Next, PointNet is applied to these groups for point cloud feature extraction. After continuous iteration, the global point cloud features are finally obtained. Compared to PointNet, the advantage of the PointNet++ network architecture lies in its multi-level feature extraction structure, effectively extracting point cloud features. The two multi-scale grouping radii used in this invention are (0.1, 0.2, 0.3) and (0.4, 0.8).

[0059] The specific implementation process for training the improved PointNet++ model on the input membrane surface 3D point cloud dataset is as follows:

[0060] The processed tensile membrane structure point cloud dataset under different damage conditions was divided into training set and test set according to a certain ratio, and then input into the improved PointNet++ model for model training.

[0061] The process of testing the trained, improved PointNet++ model against a 3D point cloud dataset on a membrane surface is as follows:

[0062] like Figure 8 and Figure 9 As shown, the membrane point cloud data of the test set is input into the trained improved PointNet++ model to infer the semantic category prediction value of the test point cloud, evaluate the prediction accuracy of the improved 3D point cloud model, and perform 3D visualization of its test segmentation results.

[0063] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for detecting damage of a tensile membrane structure based on three-dimensional depth point cloud segmentation, characterized in that, The method comprises the following steps: Step 1: scanning the membrane surface of the tensile membrane structure under different damage conditions to collect the point cloud data information of the membrane surface of the tensile membrane structure under different damage conditions; Step 2: preparing a three-dimensional point cloud data set of the membrane surface of the tensile membrane structure; Step 3: applying the attention mechanism method to the PointNet++ network architecture and improving the PointNet++ network architecture; Step 4: inputting the three-dimensional point cloud data set of the membrane surface into the improved PointNet++ model for training; Step 5: inputting the point cloud data of the membrane surface of the test set into the trained improved PointNet++ model to infer the semantic category prediction value of the test point cloud and evaluate the prediction accuracy of the improved three-dimensional point cloud model; In step 2, the point cloud information of the membrane surface of the tensile membrane structure under different damage conditions is preprocessed. Firstly, the normal vector information is added to the point cloud information of the membrane surface. Secondly, the point cloud processing software is used to segment the point cloud data of the membrane surface under different damage conditions. Thirdly, the labels are added to the different damage types of the membrane surface in turn. Finally, the point cloud data of the membrane surface under different damage conditions after data preprocessing is saved as a TXT text containing point cloud three-dimensional coordinate position information, three-dimensional normal vector information and label information, a total of 7 columns of data. In step 3, the attention mechanism method is applied to the PointNet++ network architecture. By mixing the AMS attention mechanism module based on the local point cloud spatial dimension feature and the AMC attention mechanism module based on the local point cloud channel dimension feature, the importance of different sampling points of the local point cloud spatial dimension feature and the local point cloud channel dimension feature is dynamically adjusted.

2. The tensioned membrane structure damage detection method based on three-dimensional depth point cloud segmentation of claim 1, wherein, The point cloud mapping high-dimensional features (N2, D+C2) obtained after the SA1 module, the SA2 module and the SA3 module are processed are subjected to attention mechanism processing from the spatial feature dimension and the channel feature dimension, wherein D represents the point cloud data dimension, C2 represents the point cloud spatial feature dimension, and N2 represents the point cloud channel feature dimension; the method comprises the following steps: Step 1: along the C2 local point cloud spatial feature dimension, using parallel maximum pooling and average pooling to fuse the point cloud input spatial features to generate two kinds of spatial feature expressions; Step 2: using convolution to train the weight and using the MLP with shared parameters to train the spatial feature dimension of the aggregated features to generate the attention weight; Step 3: using the activation function S to activate the weight and performing dot product with the original input point cloud spatial feature dimension to generate new point cloud spatial features; Step 4: along the N2 local point cloud channel feature dimension, using parallel maximum pooling and average pooling to fuse the point cloud input channel features to generate two kinds of channel feature expressions; Step 5: using convolution to train the weight and using the MLP with shared parameters to train the channel feature dimension of the aggregated features to generate the attention weight; Step 6: using the activation function S to activate the weight and performing dot product with the original input point cloud channel feature dimension to generate new point cloud channel features.

3. The tensioned membrane structure damage detection method based on three-dimensional depth point cloud segmentation of claim 2, wherein, The AMS attention mechanism module based on local point cloud space dimension features and the AMC attention mechanism module based on local point cloud channel dimension features are combined before and after the PointNet++ network architecture.

4. The tensioned membrane structure damage detection method based on three-dimensional depth point cloud segmentation of claim 3, wherein, The SA1 module, the SA2 module and the SA3 module process the mapped high-dimensional features of the point cloud, and specifically, the three SA modules in the PointNet++ network model architecture have the following functions: the SA1 module, the SA2 module and the SA3 module first sample and select a configurable number of key points, then select points in a spherical region within a multi-scale radius around each point as a group, and then apply PointNet to the points in the group to extract point cloud features, and finally obtain the global features of the point cloud through continuous iteration.

Images