A large-scene point cloud semantic segmentation method for digital twin application

Abstract

A large-scene point cloud semantic segmentation method for digital twin application belongs to the field of point cloud semantic segmentation, comprising the following steps: step one, neighborhood feature construction; step two, density adaptive down-sampling; step three, global feature acquisition; step four, triple information encoding; step five, triple mixed attention pooling; step six, encoder structure design; step seven, network decoding and semantic segmentation result calculation. The application firstly adjusts the sampling weight to reserve key details, generates optimized point features for semantic segmentation; then captures the global dependence of the point cloud through an efficient attention mechanism, reduces the computational complexity and avoids gradient degradation; then fuses the point cloud space, color and semantic features through multi-dimensional measurement, strengthens the feature representation; finally, the local and global features are aggregated through attention weighting, the point features are further optimized to ensure the segmentation effect. The application improves the accuracy and efficiency of point cloud semantic segmentation.

Description

Technical Field

[0001] This invention belongs to the field of point cloud semantic segmentation, specifically involving a large-scale point cloud semantic segmentation method for digital twin applications. Background Technology

[0002] With the continuous advancement of smart city, smart transportation, and various industrial park construction, digital twin technology has developed into a core supporting technology bridging the real and virtual spaces. Digital twin technology can construct complete virtual simulation models based on real-world objects. Based on these virtual models, people can view on-site conditions, dynamically monitor operational status, and simulate operational scenarios, and formulate scientific and rational management decisions. The key to the effective application of digital twin technology lies in its ability to accurately reproduce the three-dimensional form of the real world and transform this three-dimensional information into directly applicable structured data. Therefore, accurately reproducing the three-dimensional state of the real world and organizing this three-dimensional information into directly applicable structured data has become a pressing technical problem that needs to be solved.

[0003] Laser point clouds can accurately measure distances and completely reconstruct the environment, unaffected by light intensity, making them the most commonly used data acquisition method for building large-scale digital twin models. For the construction of digital twins in large-scale scenarios such as cities and large parks, laser point clouds are often used as the basic data source. However, raw point clouds only contain spatial coordinates and reflection intensity information, belonging to an unclassified and unrelated discrete set of points, which cannot be directly used for digital twin modeling. Point cloud semantic segmentation can process large amounts of point cloud data, assigning each point a corresponding category. It can distinguish different types of objects such as buildings, roads, greenery, and ground from the chaotic raw point cloud, transforming meaningless discrete points into 3D data with semantic information. Point cloud semantic segmentation is a key technology connecting 3D acquired data and practical digital twin models; its processing effect directly determines the modeling accuracy and practical application value of the digital twin.

[0004] Large-scale point cloud datasets are massive, with significant variations in the size of internal features, widespread object occlusion, complex terrain types, and marked differences in point cloud density across different regions. This places higher demands on the processing efficiency, recognition accuracy, and coverage of point cloud semantic segmentation. Currently used point cloud semantic segmentation techniques are mostly designed for small-scale, single-scene applications, exhibiting significant limitations when dealing with large-scale point cloud data. Furthermore, conventional techniques are inefficient and time-consuming when processing hundreds of millions of point cloud datasets, failing to meet the demands of rapid modeling and high-frequency updates for large-scale digital twins. Existing technologies include SPVCNN, which uses sparse voxel convolution as its backbone and introduces point branches for voxel feature interaction to strike a balance between efficiency and detailed representation; Cylinder3D, which employs a cylindrical voxel structure combined with contextual modeling to improve the efficiency of large-scale inference in vehicular outdoor scenes; and Stratified Transformer, which reduces the computational complexity of self-attention through hierarchical partitioning and windowed attention to adapt to larger-scale scenes. Although the above technologies have been improved for large-scale processing, they still generally face problems such as high computational and memory overhead, insufficient global consistency, and limited cross-regional semantic alignment capabilities in complex city-level scenarios and with hundreds of millions of points, making it difficult to simultaneously meet the engineering requirements of rapid modeling and high-frequency updates. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention proposes a large-scale point cloud semantic segmentation method for digital twin applications. The aim is to improve the accuracy and efficiency of point cloud semantic segmentation by balancing sampling weights across different density regions of the point cloud, capturing global dependencies, fusing multi-dimensional information from the point cloud, and weighted aggregation of local details and global contextual features. The method first adjusts sampling weights to retain key details, generating optimized point features for semantic segmentation. Then, it captures global dependencies in the point cloud through an efficient attention mechanism, reducing computational complexity and avoiding gradient degradation. Next, it fuses spatial, color, and semantic features of the point cloud through multi-dimensional metrics to enhance feature representation. Finally, it further optimizes point features by weighted aggregation of local and global features through attention to ensure effective segmentation.

[0006] The technical solution adopted in this invention is: A semantic segmentation method for large-scale point clouds in digital twin applications includes the following steps: Step 1: Neighborhood Feature Construction: Construct a point cloud and feature set, and obtain point neighborhood features using the K-Nearest Neighbors (KNN) algorithm, which is commonly used in machine learning. Step 2, Density-adaptive downsampling: By combining local density estimation with weight fusion to adjust the sampling probability through density-adaptive downsampling (DAS), the attention of different density regions in the point cloud is balanced, thereby improving the ability to segment details in complex scenes. Step 3, Global Feature Acquisition: By introducing an external memory unit through the External-Attention Block (EAB), the computational complexity is reduced, residual connections are added to avoid gradient degradation, and a progressive single-multi-head design is adopted to accurately capture the global remote dependencies of the point cloud and balance feature learning, outputting global features of the point cloud that fuse the global context. Step 4, Triple Information Encoding: Using the Triple Information Encoding Block (TIEB), distance is measured from three dimensions: spatial geometry, color, and features. Multi-dimensional information is fused and the contextual features of neighboring points are aggregated to enhance the feature representation capability of point clouds. Step 5, Tri-Fusion Attention Pooling Block: Through Tri-Fusion Attention Pooling Block (TFAP), multi-dimensional fusion features and global features are concatenated, attention scores are calculated through non-linear transformation, neighboring point features are weighted and aggregated, and local details and global context are fused. Step Six: Encoder Structure Design: The execution steps from Step Three to Step Five are denoted as EAF-Net. EAF-Net is then executed repeatedly in sequence. Each execution is recorded as one layer of encoding module. Before each layer of encoding module is executed, point cloud dimensionality reduction and sampling optimization are completed through density adaptive downsampling (step two) to realize hierarchical progressive extraction of point cloud features, aggregate multi-scale feature information, provide high-quality feature input for the decoder, and support the subsequent improvement of semantic segmentation accuracy. Step 7: Calculation of network decoding and semantic segmentation results.

[0007] In step seven, during the decoding stage, multi-scale features of the encoder are restored by layer-by-layer interpolation upsampling and skip connections to restore the original point cloud resolution and enhance the semantic representation of features. Finally, the class probability distribution of each point is calculated and output by the MLP classification head and the Softmax activation function, thereby obtaining the final semantic segmentation result of the point cloud.

[0008] Furthermore, the process of step one is as follows; Step 1.1) Define a point cloud dataset: Point cloud data consists of a set of discrete three-dimensional spatial points. These points have a continuous spatial distribution and share similar geometric and semantic attributes with their neighbors. Given a point cloud dataset containing N points... ,in Let represent the three-dimensional spatial coordinates of the i-th point, which has 3 feature channels. Let represent the color information of the i-th point, with 3 feature channels. For each point, after input, it passes through a fully connected (FC) layer to obtain initial semantic features, given the feature set. ,in Let represent the semantic features of the i-th point, and d represent the number of semantic feature channels; Step 1.2) Obtain neighborhood features: Use the KNN algorithm based on Euclidean distance to collect local neighborhood features of the point cloud. This allows for the determination of the spatial geometric information of the local point cloud. Color information and semantic feature information Where K represents the number of nearest neighbors set in the KNN algorithm, and k represents the k-th neighbor.

[0009] Furthermore, the process of step two is as follows; Step 2.1) Obtaining Location Weights: According to the traditional random downsampling method, during the sampling process, a sampling probability is maintained for each point, initialized to a very small random value. In each iteration, the point with the lowest current sampling probability is selected as the sampling center point. Subsequently, for points in its local neighborhood, geometric distance weights are calculated and assigned. , representing the sampling weight of the i-th neighboring point; the closer the distance, the larger the weight. ; in This represents the Euclidean distance from a neighboring point i to the center point. This represents the maximum distance between all points in this sampling and the center point; Step 2.2) Local density characterization: For each point in the point cloud Calculate the Euclidean distances from the node to its K nearest neighbors, and take the median to characterize the local density. This method has better robustness to noise points and outliers: ; in Represents the median function. The set notation packages the distance values ​​of the K nearest neighbors specified in the KNN algorithm into a distance array. This indicates the calculation of Euclidean distance; Step 2.3) Adaptive Weight Fusion: Subtract the local density of the current point from the maximum local density of all points in the current sampling, and then divide by the difference between the maximum and minimum local density values ​​to obtain the normalized density weight. Adaptive weight fusion is then used, with geometric distance weights as the basis, and the gain intensity of the density weights is controlled by hyperparameters to achieve adaptive weight enhancement, resulting in the fused sampling weights. : ; in This is a hyperparameter used to control the contribution of density weights. and These represent the maximum and minimum values ​​of the local density index in this sampling, respectively. Finally, the sampling probabilities of all points in the neighborhood of each sampling point are accumulated and fused using sampling weights to obtain new sampling probabilities. : ; in Indicates the initial sampling probability. This strategy ensures that the model does not ignore regions with a small number of points but which may contain important structural information, which is particularly crucial for improving the ability to segment details in complex scenes.

[0010] Furthermore, the process of step three is as follows: Step 3.1) Introduction of External Attention Infrastructure: To capture global long-range dependencies in the point cloud and reduce the computational overhead of standard self-attention, an external attention mechanism based on external memory units is adopted. This technique utilizes two independent external memory units, namely the key memory unit. Value memory unit To achieve efficient feature extraction, where S is the size of the memory unit, much smaller than the number of input points N, and d represents the feature dimension, thus significantly improving processing efficiency, the following steps are taken: First, the point cloud dataset defined in step one is sampled using the downsampling method described in step two. Then, the features of the sampled points are input into a single linear layer with shared parameters (SharedMLP1) for linear transformation to generate input features. Then The similarity response is obtained by multiplying the key memory unit matrix by its transpose and then doubly normalized; subsequently, it is multiplied by the value memory unit matrix to obtain the final output feature. The formula is as follows: ; in This represents a double normalization operation, used to mitigate the sensitivity of the attention matrix to the scale of the input features and improve numerical stability. This represents the transpose of the key memory unit matrix; A dual normalization model is designed. First, similarity is exponentially normalized along the memory unit dimension. Then, the Softmax function, commonly used in deep learning, is applied to each column of the original similarity matrix to ensure that the sum of the weights of each memory unit across all input points is 1. Next, each row of the intermediate weight matrix obtained in the first step is summed and normalized to ensure that the sum of the weights of each input point across all memory units is 1, thus obtaining a stable attention weight distribution. The specific formula is as follows: ; ; ; in Represents the memory unit for the i-th point and key. The similarity between rows j in the middle, This represents an exponential function, where s represents the index of the memory cell; Step 3.2) Large-scale point cloud semantic segmentation application.

[0011] Preferably, the process of step 3.2) is as follows: Step 3.2.1) Residual Connection: To ensure the network learns global context using the attention mechanism without losing the most basic local geometric structure and detailed information of the point cloud, thereby improving segmentation accuracy, and to avoid the gradient vanishing and network degradation problems that are inevitably encountered when building deep networks for large-scale point cloud segmentation, a residual connection is added at the output, as shown in the following formula: ; Step 3.2.2) Multi-head external attention mechanism Since single-head external attention mechanisms generate only a set of attention weights through a single projection, their expressive power is limited, making it difficult to simultaneously characterize feature associations at multiple scales and of multiple types. Multi-head external attention, on the other hand, can focus on different information within an unknown representation subspace. Therefore, this module is used to extract global features from point clouds. First, the input features are divided according to the channel dimension. Each sub-feature is fed into an external attention function to calculate the output of each attention head. Then, the output features of all attention heads are concatenated along the channel dimension and linearly fused using a learnable weight matrix to achieve multi-head information aggregation. Finally, residual connections are introduced to add the fused result to the original input to enhance feature propagation and improve training stability. The multi-head external attention mechanism with residual connections can be represented as follows: ; ; in ExternalAttention Represents the external attention function. This indicates that the original input features are divided equally by channel and then fed into the input features of the i-th head. MultiHea d represents the multi-head attention function. Concat Indicates the feature concatenation operator, and H Let i and n represent the number of heads, respectively. This represents the learnable weight matrix.

[0012] Step 3.2.3) Progressive Attention: Although multi-head attention can focus on information from different subspaces, in the low-dimensional layers of the network, where point cloud features typically have limited representation capabilities, using a multi-head mechanism too early may lead to attentional distraction, which is detrimental to the stable learning of basic features. Therefore, a single-head external attention mechanism is used in the low-dimensional layers of the network to prioritize the construction of robust basic feature representations; a multi-head mechanism is then used in higher-dimensional layers to capture richer global contextual information. This progressive design aims to balance the feature learning needs at different network depths.

[0013] The process of step four is as follows: Step 4.1) Triple distance metric: Based on the local neighborhood of the point cloud obtained from Step 1 The spatial geometric information, color information, and semantic features of point clouds are measured, and the process is as follows: Step 4.1.1) Spatial Information Quantization: In spatial coordinates, geometric information is the essence of point clouds. In order to simultaneously characterize directionality and scale information, this invention considers the relative distance between point coordinates. Euclidean distance , This indicates the calculation of the Euclidean distance between adjacent points and the center point; Step 4.1.2) Color Information Quantification: Color information can help distinguish areas that are geometrically similar but different in material or category. Therefore, this invention examines the relative distance between colors to construct the relative color difference. ; Step 4.1.3) Semantic Feature Information Quantification: In a high-dimensional semantic feature space, abnormal fluctuations or significant differences in individual dimensions can excessively amplify the overall distance, potentially masking consistency across many other dimensions. The mean absolute error of features (MAE), however, is insensitive to outliers or drastic changes in individual dimensions, providing a more robust similarity measure. Therefore, this invention calculates the mean absolute error between semantic features. This measures the similarity of high-dimensional semantic features to ensure sufficient collection of contextual information. This represents the mean function used to calculate the distance to the mean. Represents the L1 norm; Step 4.2) Design of the triple information coding unit, the process is as follows: Step 4.2.1) Triple Relationship Vector Concatenation: Concatenate the quantized results of the relationship between the spatial coordinate domain, color domain, and feature space obtained in Step 4.1 to obtain the triple relationship vector between the center point and its neighboring points: ; in This indicates a splicing operation.

[0014] Step 4.2.2) Information Encoding: The triple relation is encoded through a multilayer perceptron network, including batch normalization (BN) operation on the vector to improve training stability and convergence speed. Subsequently, information is encoded through two linear transformations with ReLU activation functions to achieve nonlinear mapping and fusion of relation features, outputting a triple-enhanced relation feature representation. The two-layer linear transformation is implemented using a shared parameter approach, that is, the same set of network parameters is reused for all center points and their neighboring point pairs to reduce the parameter size and improve generalization ability. The specific formula is as follows: ; in This represents a multilayer perceptron network.

[0015] The process of step five is as follows: Step 5.1) Feature splicing The features obtained through the external attention module are treated as new point features. It is controlled by points through an external attention module. Dynamically updated, click The new KNN feature representation is as follows ,make Features obtained with triple information coding unit By cascading and combining these features, a triple hybrid attention feature is formed. , means as follows:

[0016] in This indicates a splicing operation.

[0017] Step 5.2) Calculation of normalized attention weights For splicing features A shared multilayer perceptron (MLP) is applied for nonlinear transformation, and the normalized attention weights corresponding to each neighborhood feature are calculated using the Softmax function. ; Among them, attention weight It is positively correlated with the importance of the corresponding feature.

[0018] Step 5.3) Context-enhanced feature calculation Using the attention weights to fuse features Element-wise weighting is performed, and a weighted summation operation is executed over a K-neighborhood. Then, through batch normalization and two linear transformations with ReLU activation functions, the point is obtained. Contextual enhancement features : .

[0019] In step six, the execution steps from step three to step five are denoted as EAF-Net. EAF-Net consists of three major modules: EAB, TIEB, and TFAP. EAF-Net is executed repeatedly in sequence. Each execution is recorded as one layer of encoding module. Before each layer of encoding module is executed, point cloud dimensionality reduction and sampling optimization are completed through density adaptive downsampling (step two) to achieve hierarchical progressive extraction of point cloud features, aggregate multi-scale feature information, provide high-quality feature input for the decoder, and support the subsequent improvement of semantic segmentation accuracy.

[0020] In step seven, during the decoding stage, multi-scale features of the encoder are restored by layer-by-layer interpolation upsampling and skip connections to restore the original point cloud resolution and enhance the semantic representation of features. Finally, the class probability distribution of each point is calculated and output by the MLP classification head and the Softmax activation function, thereby obtaining the final semantic segmentation result of the point cloud.

[0021] Preferably, the process of step seven is as follows: Step 7.1) Layer-by-layer upsampling: First, calculate the Euclidean distance between each sampling point in two adjacent point cloud layers to establish a spatial correspondence between the low-resolution point set and the point set to be restored; then, for each target point in the layer to be restored, select its nearest neighbor in the low-resolution point set of the previous layer, and use nearest neighbor interpolation to propagate the features of the nearest neighbor in the previous layer to the target point, thereby realizing the layer-by-layer transfer of sparse features to high-resolution features; Step 7.2) Skip Connection Multi-Scale Feature Fusion: After single-layer upsampling, the interpolated features are spliced ​​and fused with the corresponding scale features in the encoder through skip connections in the channel dimension. This preserves both deep semantic information and shallow geometric details, thereby improving the ability to identify boundary regions and small-scale targets. Subsequently, the spliced ​​fused features are processed by a shared multilayer perceptron to reduce computational complexity by compressing redundant multi-channel feature dimensions, mining spatial structure and semantic associations, and achieving compact integration of multi-scale features to provide refined features for subsequent decoding layers. Step 7.3) Point-level semantic classification output: The above upsampling and feature fusion process is repeated layer by layer along the decoding path until the original point cloud resolution is restored, obtaining a point-level high-dimensional feature representation containing global contextual semantics and local spatial structure information. Finally, the point-level high-dimensional features are input into the multilayer perceptron classification module for nonlinear mapping and channel transformation, and the probability distribution of each point in the preset semantic category is calculated through the Softmax activation function. The semantic category label of each point is output according to the maximum probability correspondence, thus obtaining the final semantic segmentation result of the point cloud.

[0022] The beneficial effects of this invention are mainly reflected in the following aspects: 1. This invention proposes density adaptive downsampling, which adaptively adjusts the sampling probability of each point by jointly modeling the geometric distance distribution and local point cloud density information, effectively preserving key structural features in sparse regions while reducing redundant computation. This method overcomes the problem that traditional random or uniform sampling easily ignores important details, thereby significantly improving the precision and stability of point cloud semantic segmentation in large-scale digital twin scenarios. 2. This invention constructs a progressive multi-head external attention mechanism, which achieves efficient global information interaction by introducing external memory units, avoiding redundant modeling caused by direct correlation between points in traditional self-attention. At the same time, combined with residual connections and progressive attention design from single-head to multi-head, the network maintains stable training under deep stacking conditions and can aggregate complementary global context information in parallel in different representation subspaces, thereby more accurately capturing long-range semantic dependencies in point clouds, significantly enhancing global feature expression capabilities and semantic segmentation performance in complex scenarios, and improving the overall processing efficiency of large-scale point clouds. 3. This invention designs a triple information encoding and hybrid attention pooling fusion mechanism, which jointly measures neighborhood relationships from three dimensions: spatial geometry, color attributes, and semantic features. It also achieves adaptive fusion of local detail information and global context information through attention weighting. This structure breaks through the limitations of traditional single feature modeling and simple feature splicing methods, realizes multi-level semantic complementarity and discriminative enhancement, thereby comprehensively improving the point cloud feature representation quality and semantic segmentation accuracy, and exhibits stronger robustness and generalization ability in large-scale complex scenarios of digital twins. Attached Figure Description

[0023] Figure 1 This is a diagram of the overall network architecture.

[0024] Figure 2 This is a diagram of the EAF module structure.

[0025] Figure 3 This is a schematic diagram of a triple information coding unit.

[0026] Figure 4 This is a schematic diagram of triple hybrid attention pooling. Detailed Implementation

[0027] The present invention will now be further described with reference to the accompanying drawings.

[0028] Reference Figures 1-4 A semantic segmentation method for large-scale point clouds in digital twin applications, characterized by the following steps: Step 1: Neighborhood feature construction, the process is as follows: Step 1.1) Define a point cloud dataset: Point cloud data consists of a set of discrete three-dimensional spatial points. These points have a continuous spatial distribution and share similar geometric and semantic attributes with their neighbors. Given a point cloud dataset containing N points... ,in Let represent the three-dimensional spatial coordinates of the i-th point, which has 3 feature channels. Let represent the color information of the i-th point, with 3 feature channels. For each point, after input, it passes through a fully connected (FC) layer to obtain initial semantic features, given the feature set. in Let represent the semantic features of the i-th point, and d represent the number of semantic feature channels; Step 1.2) Obtain neighborhood features: Use the KNN algorithm based on Euclidean distance to collect local neighborhood features of the point cloud. This allows for the determination of the spatial geometric information of the local point cloud. Color information and semantic feature information Where K represents the number of nearest neighbors set in the KNN algorithm, and k represents the k-th neighbor.

[0029] Step 2, density adaptive downsampling, the process is as follows: Step 2.1) Obtaining Location Weights: According to the traditional random downsampling method, during the sampling process, a sampling probability is maintained for each point, initialized to a very small random value. In each iteration, the point with the lowest current sampling probability is selected as the sampling center point. Subsequently, for points in its local neighborhood, geometric distance weights are calculated and assigned. , representing the sampling weight of the i-th neighboring point; the closer the distance, the larger the weight. ; in This represents the Euclidean distance from a neighboring point i to the center point. This represents the maximum distance between all points in this sampling and the center point; Step 2.2) Local density characterization: For each point in the point cloud Calculate the Euclidean distances from the node to its K nearest neighbors, and take the median to characterize the local density. This method has better robustness to noise points and outliers: ; in Represents the median function. The set notation packages the distance values ​​of the K nearest neighbors specified in the KNN algorithm into a distance array. This indicates the calculation of Euclidean distance; Step 2.3) Adaptive Weight Fusion: Subtract the local density of the current point from the maximum local density of all points in the current sampling, and then divide by the difference between the maximum and minimum local density values ​​to obtain the normalized density weight. Adaptive weight fusion is then used, with geometric distance weights as the basis, and the gain intensity of the density weights is controlled by hyperparameters to achieve adaptive weight enhancement, resulting in the fused sampling weights. : ; in This is a hyperparameter used to control the contribution of density weights. and These represent the maximum and minimum values ​​of the local density index in this sampling, respectively. Finally, the sampling probabilities of all points in the neighborhood of each sampling point are accumulated and fused using sampling weights to obtain new sampling probabilities. : ; in Indicates the initial sampling probability. This strategy ensures that the model does not ignore regions with a small number of points but which may contain important structural information, which is particularly crucial for improving the ability to segment details in complex scenes.

[0030] Step 3: Global feature acquisition, the process is as follows: Step 3.1) Introduction of External Attention Infrastructure: To capture global long-range dependencies in the point cloud and reduce the computational overhead of standard self-attention, this invention employs an external attention mechanism based on external memory units. This technology utilizes two independent external memory units, namely the key memory unit. Value memory unit To achieve efficient feature extraction, where S is the size of the memory unit, much smaller than the number of input points N, and d represents the feature dimension, thus significantly improving processing efficiency, the following steps are taken: First, the point cloud dataset defined in step one is sampled using the downsampling method described in step two. Then, the features of the sampled points are input into a single linear layer with shared parameters (SharedMLP1) for linear transformation to generate input features. Then The similarity response is obtained by multiplying the key memory unit matrix by its transpose and then doubly normalized; subsequently, it is multiplied by the value memory unit matrix to obtain the final output feature. The formula is as follows: ; in This represents a double normalization operation, used to mitigate the sensitivity of the attention matrix to the scale of the input features and improve numerical stability. This represents the transpose of the key memory unit matrix; A dual normalization model is designed. First, the similarity is exponentially normalized along the memory unit dimension. Then, the Softmax function, commonly used in deep learning, is applied to each column of the original similarity matrix to ensure that the sum of the weights of each memory unit across all input points is 1. Next, each row of the intermediate weight matrix obtained in the first step is summed and normalized to ensure that the sum of the weights of each input point across all memory units is 1, thus obtaining a stable attention weight distribution. The formula is as follows: ; ; ; in Represents the memory unit for the i-th point and key. The similarity between rows j in the middle, This represents an exponential function, where s represents the index of the memory cell; Step 3.2) Large-scale point cloud semantic segmentation application, the process is as follows: Step 3.2.1) Residual Connection: To ensure the network learns global context using the attention mechanism without losing the most basic local geometric structure and detailed information of the point cloud, thereby improving segmentation accuracy, and to avoid the gradient vanishing and network degradation problems that are inevitably encountered when building deep networks for large-scale point cloud segmentation, this invention adds a residual connection at the output, as shown in the following formula: ; Step 3.2.2) Multi-head external attention mechanism: Since the single-head external attention mechanism generates only a set of attention weights through a single projection, its expressive power is limited and it is difficult to simultaneously characterize feature associations of multiple scales and types. Multi-head external attention, on the other hand, can focus on different information in unknown representation subspaces. Therefore, this invention uses this module to extract global features of point clouds. First, the input features are divided according to the channel dimension. Each sub-feature is fed into an external attention function to calculate the output of each attention head. Then, the output features of all attention heads are concatenated along the channel dimension and linearly fused using a learnable weight matrix to achieve multi-head information aggregation. Finally, residual connections are introduced to add the fused result to the original input to enhance feature propagation and improve training stability. The multi-head external attention mechanism with residual connections can be represented as follows: ; ; in ExternalAttention Represents the external attention function. This indicates that the original input features are divided equally by channel and then fed into the input features of the i-th head; MultiHead This represents a multi-head attention function. Concat Indicates the feature concatenation operator, H and H represent the i-th head and the number of heads, respectively. Represents the learnable weight matrix; Step 3.2.3) Progressive Attention: Although multi-head attention can focus on information from different subspaces, in the low-dimensional layers of the network, where point cloud features typically have limited representation capabilities, using a multi-head mechanism too early may lead to attentional distraction, which is detrimental to the stable learning of basic features. Therefore, this invention employs a single-head external attention mechanism in the low-dimensional layers of the network to prioritize the construction of robust basic feature representations; a multi-head mechanism is then used in higher-dimensional layers to capture richer global contextual information. This progressive design aims to balance the feature learning needs at different network depths.

[0031] Step four, triple information encoding, the process is as follows: Step 4.1) Triple distance metric: such as Figure 3 The point cloud containing (d+3+3) dimensional features for each point is used as input to the triple information encoding unit, based on the local neighborhood of the point cloud obtained in step one. The spatial geometric information, color information, and semantic features (TriFusion Information) of the point cloud are measured, where d is the semantic feature dimension, and 3 and 3 represent spatial information and color information, respectively. The process is as follows: Step 4.1.1) Spatial Information Quantization: In spatial coordinates, geometric information is the essence of point clouds. In order to simultaneously characterize both directionality and scale information, this invention considers the relative distance between point coordinates. Reativate distance and Euclidean distance , This indicates the calculation of the Euclidean distance between adjacent points and the center point; Step 4.1.2) Color Information Quantification: Color information can help distinguish areas that are geometrically similar but different in material or category. Therefore, this invention examines the relative distance between colors to construct the relative color difference. ; Step 4.1.3) Semantic feature information quantification: In a high-dimensional semantic feature space, abnormal fluctuations or significant differences in individual dimensions can overemphasize the overall distance, which may mask the consistency in many other dimensions. However, the mean absolute error of features is not sensitive to outliers or drastic changes in individual dimensions, providing a more robust similarity measure. Therefore, this invention calculates the mean absolute error between semantic features. This measures the similarity of high-dimensional semantic features to ensure sufficient collection of contextual information. This represents the mean function used to calculate the distance to the mean. Represents the L1 norm; Step 4.2) Design of the triple information coding unit, the process is as follows: Step 4.2.1) Triple Relationship Vector Concatenation: Concatenate the quantized results of the relationship between the spatial coordinate domain, color domain, and feature space obtained in Step 4.1 to obtain the triple relationship vector between the center point and its neighboring points: ; in Indicates a splicing operation; Step 4.2.2) Information Encoding: The triple relation is encoded through a multilayer perceptron network, including batch normalization (BN) of the vector settings to improve training stability and convergence speed. Subsequently, information is encoded through two linear transformations (SharedMLP2) with ReLU activation functions to achieve nonlinear mapping and fusion of relation features, outputting a triple-enhanced relation feature representation. (Tri-Enhanced information) The two-layer linear transformation is implemented using a shared parameter approach, that is, the same set of network parameters is reused for all center points and their neighboring point pairs to reduce the parameter size and improve generalization ability, as shown in the following formula: ; in This represents a multilayer perceptron network.

[0032] Step 5: Triple Hybrid Attention Pooling, the process is as follows: Step 5.1) Feature splicing: such as Figure 4 The features obtained through the external attention module are treated as new point features. (External-Attention Feature), which uses an external attention module to focus on points. Dynamically updated, click The new KNN feature representation is as follows ,make Features obtained with triple information coding unit By cascading and combining these features, a triple hybrid attention feature is formed. , means as follows: ; in Indicates a splicing operation; Step 5.2) Normalized attention weight calculation: For concatenated features A shared multilayer perceptron (MLP) is applied for nonlinear transformation, and the normalized attention weights corresponding to each neighborhood feature are calculated using the Softmax function. ; Among them, attention weight It is positively correlated with the importance of the corresponding feature.

[0033] Step 5.3) Context-enhanced feature calculation: Utilize the attention weights to calculate the fused features. Element-wise weighting is performed, and a weighted summation operation is executed over a K-neighborhood. Then, through batch normalization (BN) and two linear transformations with ReLU activation functions (SharedMLP2), the point is obtained. Contextual enhancement features (Enhanced External-Attention Feature): .

[0034] Step Six: Encoder Structure Design: In the encoder stage, such as... Figure 1 First, the input point clouds pass through a fully-connected layer (FC) for feature dimension mapping, transforming the feature dimensions of each point to obtain an initial feature representation; for example... Figure 2 The execution steps from step three to step five are denoted as EAF-Net. EAF-Net consists of three main modules: EAB, TIEB, and TFAP. EAF-Net is executed repeatedly in sequence. Each execution is recorded as one layer of encoding module. Before each layer of encoding module is executed, point cloud dimensionality reduction and sampling optimization are completed through density adaptive downsampling (DAS) as described in step two. This enables hierarchical progressive extraction of point cloud features, aggregating multi-scale feature information, providing high-quality feature input for the decoder, and supporting the subsequent improvement of semantic segmentation accuracy.

[0035] Step 7: Network Decoding and Semantic Segmentation Result Calculation: In the decoder stage, a combination of layer-by-layer feature upsampling (US) and skip connections is used to progressively restore the low-resolution point cloud features obtained through continuous downsampling in the encoding stage, in order to reconstruct the original point cloud resolution and enhance the point-level semantic expression capability. The process is as follows: Step 7.1) Layer-by-layer upsampling: First, calculate the Euclidean distance between each sampling point in two adjacent point cloud layers to establish a spatial correspondence between the low-resolution point set and the point set to be restored; then, for each target point in the layer to be restored, select its nearest neighbor in the low-resolution point set of the previous layer, and use nearest neighbor interpolation to propagate the features of the nearest neighbor in the previous layer to the target point, thereby realizing the layer-by-layer transfer of sparse features to high-resolution features; Step 7.2) Skip Connection Multi-Scale Feature Fusion: After single-layer upsampling, the interpolated features are spliced ​​and fused with the corresponding scale features in the encoder through skip connections in the channel dimension. This preserves both deep semantic information and shallow geometric details, thereby improving the ability to identify boundary regions and small-scale targets. Subsequently, the spliced ​​fused features are processed by a shared multilayer perceptron to reduce computational complexity by compressing redundant multi-channel feature dimensions, mining spatial structure and semantic associations, and achieving compact integration of multi-scale features to provide refined features for subsequent decoding layers. Step 7.3) Point-level semantic classification output: The above upsampling and feature fusion process is repeated layer by layer along the decoding path until the original point cloud resolution is restored, obtaining a point-level high-dimensional feature representation containing global contextual semantics and local spatial structure information. Finally, the point-level high-dimensional features are input into the classification module, and pass through a fully connected layer, a dropout layer (DP), and another fully connected layer in sequence for nonlinear mapping and feature transformation. The probability distribution of each point in the preset semantic category is calculated using the Softmax activation function, and the semantic category label of each point is output according to the maximum probability correspondence, thus obtaining the final semantic segmentation result of the point cloud.

[0036] This embodiment compares and analyzes other point cloud semantic segmentation methods with the method of this invention, including the following steps: Step 1, Introduction to point cloud semantic segmentation methods, is as follows: SCF-Net: SCF-Net proposes a learnable SCF module for extracting spatial context features from large-scale point clouds. This module is based on local neighborhoods and constructs attention coefficients through a joint approach of geometric and feature distances. It can be flexibly embedded into various backbone networks to improve semantic segmentation accuracy. This feature effectively captures spatial relationships between points, alleviating the challenges of processing large-scale scenes.

[0037] BAAF-Net: BAAF-Net employs a bilateral enhancement module to deeply fuse geometric and semantic features to strengthen local context. It also introduces an adaptive fusion strategy to integrate inter-point differences across multiple resolutions, achieving high-precision semantic segmentation. This mechanism balances geometric relationships and semantic associations, significantly improving the ability to represent local information.

[0038] RandLA-Net: RandLA-Net employs a random sampling strategy to significantly reduce computational overhead, while preserving key geometric details through a local feature aggregation module, thus achieving a balance between accuracy and efficiency. Random sampling is an efficient downsampling method, and local feature aggregation enhances contextual information through extended neighborhoods and attention pooling.

[0039] LACV-Net: LACV-Net constructs a local adaptive feature enhancement module that strengthens the context by learning the centroid and the similarity between neighborhood points. It then uses a comprehensive VLAD to fuse multi-scale features and introduces an aggregation loss to optimize the boundary. The local adaptive features dynamically adjust weights to capture geometric details, while the comprehensive VLAD constructs a more discriminative global description.

[0040] LEARD-Net: Designed specifically for colored point clouds, LEARD-Net mines color information through local augmentation modules, learns robust semantic features using adaptive residual modules, and employs a dynamic fusion strategy to integrate multi-scale information to improve accuracy. Adaptive residuals can flexibly adjust connection strength, and dynamic fusion achieves efficient aggregation based on attention.

[0041] HFA-Net: HFA-Net comprehensively extracts and enhances multimodal information, including geometric, color, and semantic information, through hybrid feature extraction and bilateral augmentation components. It then dynamically fuses these features using a hybrid attention mechanism to achieve robust semantic segmentation. The hybrid attention mechanism combines self-attention and channel attention, significantly expanding the receptive field.

[0042] LSBE-Net: LSBE-Net first employs a local boundary feature extraction module to accurately capture the edge details of the point cloud, and then fuses global and local context through a spatial attention aggregation mechanism to achieve refined semantic segmentation. In this method, local boundary features focus on edge characterization, while spatial attention further enhances the spatial relationships between points.

[0043] PointNet++: PointNet++ proposes a hierarchical ensemble abstraction module for extracting multi-scale local features from large-scale point clouds. This module constructs local neighborhoods based on farthest point sampling and ball queries, leverages PointNet aggregated features, and can be flexibly embedded into various backbone networks to improve semantic segmentation accuracy. This module effectively captures both local and global context, alleviating the challenges of uneven point cloud density and large-scale operations.

[0044] PointConv: PointConv proposes a density-reweighted convolution operation for constructing deep convolutional networks for point clouds. This operation achieves a continuous convolution approximation through efficient weight function calculation and can be flexibly embedded into the backbone network to improve segmentation performance. This convolution effectively handles density differences, alleviating the problems of point cloud disorder and non-uniformity.

[0045] DCNet proposes a discriminative and efficient feature aggregation module for large-scale point cloud semantic segmentation. This module employs a dense-sparse sampling strategy combined with a similarity metric and utilizes a fast random sampling encoding / decoding structure, which can be flexibly embedded into the network to enhance feature discriminative power. This aggregation effectively reduces the negative impact of sampling and alleviates the challenges of processing large-scale scenes.

[0046] EAF-Net: This invention proposes a semantic segmentation method for large-scale point cloud applications in digital twins.

[0047] Step 2, Introduction to the point cloud semantic segmentation dataset, is as follows: Toronto-3D is a large-scale outdoor mobile LiDAR dataset for semantic segmentation of urban road scenes. This dataset was collected by a vehicle-mounted mobile LiDAR scanning system along a 1-kilometer stretch of street in Toronto, Canada, and contains approximately 78.3 million 3D points, each finely labeled with eight semantic categories. During training, 3D coordinates and color features can be acquired simultaneously. Following the original authors' recommended partitioning scheme, this invention designates region 2 as the test set, and the remaining three regions are used to train the model.

[0048] Semantic3D is a large-scale, high-quality outdoor point cloud semantic segmentation dataset. Acquired via a terrestrial laser scanner, it provides over 4 billion points, including 15 training scenarios and 15 online test scenarios. The point cloud covers various urban scenes such as churches, streets, and railways, with a real-world coverage area of ​​up to 160 x 240 x 30 meters. The original points in the dataset belong to 8 semantic categories, and each point contains seven-dimensional features (such as 3D coordinates, RGB color, and intensity). To ensure a fair comparison with mainstream methods, this invention uses only 3D coordinates and color information to train and test its network model.

[0049] Step 3: Define the evaluation indicators as follows: This invention uses Intersection over Union (IoU) per class, mean Intersection over Union (mIoU), and overall accuracy (OA) as evaluation metrics. To ensure the model can effectively learn and output prediction results with good probabilistic interpretability, this invention uses cross-entropy loss as the loss function of the proposed network, calculated as follows: ; ; ; ; Where TP represents true positive samples, FP represents false positive samples, FN represents false negative samples, n represents the number of semantic categories, N represents the total number of samples, and C represents the number of categories. and represents the true label vector and predicted label vector for category Cth, respectively. Larger values ​​for both mIou and OA indicate better model performance.

[0050] Step 4: Evaluate and compare the results as follows: Table 1 presents the quantitative evaluation results of the proposed method and other comparative models on the Toronto-3D region binary dataset. The proposed method and the state-of-the-art RandLA-Net improve overall accuracy and mIoU by 2.24% and 1.16%, respectively, validating the effectiveness of the proposed model in large-scale point cloud semantic segmentation tasks. Furthermore, the results in Table 1 show that the proposed method outperforms all comparative methods in segmentation IoU for artificial scene categories such as road markings, buildings, utility lines, and fences, fully demonstrating the model's advantage in fine segmentation of artificial structural targets in complex outdoor point cloud scenes.

[0051] Table 1 shows the quantitative evaluation results of the method of the present invention and other comparative models on the Toronto-3D region binary dataset;

[0052] Table 2 presents the segmentation accuracy comparison results of this invention with several classic or state-of-the-art methods, including RandLA-Net, DCNet, and LEARD-Net, on the Semantic3D dataset. It can be seen that EAF-Net achieves an overall mIoU of 75.9% and an OA of 95.1%, performing best among all methods. Compared to RandLA-Net, this method improves mIoU by 4.1 percentage points and OA by 0.9 percentage points, demonstrating more stable and superior semantic segmentation capabilities. Furthermore, its segmentation IoU index is superior to all compared methods in artificial terrain, tall vegetation, natural terrain, and hardscape categories, and it also performs well in other scenes, indicating that this method can better maintain structural consistency and category coherence in complex, large-scale outdoor scenes.

[0053] Table 2 shows the segmentation accuracy comparison results of the present invention with various classic or state-of-the-art methods, including RandLA-Net, DCNet, and LEARD-Net, on the Semantic3D dataset;

[0054] This embodiment uses specific point cloud data analysis as an example to further illustrate the effectiveness of the network module of the present invention, as follows: Step 1) Introduction to the network module; The EAB of this invention is used to capture the long-range dependencies of all points in a point cloud and overcomes the high computational complexity of standard self-attention mechanisms. The TIEB of this invention utilizes three types of information simultaneously to enhance the feature representation capability of each point. The TFAP of this invention uses an attention mechanism to assign different weights to neighboring points, achieving weighted aggregation of key neighborhood information, thereby generating enhanced features containing rich local details and global context. Simultaneously, this invention designs a DAS downsampling method to ensure that the model does not ignore regions with a small number of points but potentially containing important structural information, thereby improving the detail segmentation capability in complex scenes.

[0055] Step 2), Introduction to point cloud datasets; The S3DIS dataset, proposed by Armeni et al. in 2016, is a widely used benchmark for semantic segmentation of point clouds in indoor scenes. Collected using a Matterport scanner, the dataset covers six large-scale indoor regions from three different buildings, containing a total of 272 rooms. Each point cloud in the dataset is labeled with 13 semantic categories. In this study, each point is represented by a six-dimensional vector containing three-dimensional spatial coordinates (XYZ) and color information (RGB). To comprehensively evaluate the model's performance, the proposed network was evaluated on the widely used Region 5.

[0056] Step 3) Implement the design and process as follows: Table 3 presents the ablation quantitative analysis results of the proposed method on the S3DIS region 5 dataset for different modules. It can be seen that the complete model A5, including EAB, TIEB, TFAP, and DAS, achieves the highest mIoU of 66.7%. When EAB is replaced with local self-attention (A1), which only performs self-attention on adjacent local regions, the mIoU drops to 65.4%, indicating that the external attention mechanism is more effective in modeling long-range dependencies and global context of point clouds. After removing TIEB and using the relative position encoding module of RandLA-Net (A2), the mIoU decreases to 65.0%, indicating that the triple information encoding in TIEB provides a more discriminative feature representation. Removing the triple information encoding in TFAP and using only the global features generated by external attention for pooling (A3) causes the mIoU to drop to 64.3%, showing the most significant performance degradation, indicating that feature aggregation guided by multiple information is crucial for improving segmentation accuracy. Replacing density-adaptive downsampling with the random downsampling (A4) proposed by RandLA-Net also resulted in a performance drop of 66.1%, indicating that considering point density during downsampling helps preserve key structural information. Overall, removing or replacing any module caused varying degrees of performance degradation, demonstrating that EAB, TIEB, TFAP, and DAS all contribute positively to overall performance and are complementary to each other.

[0057] Table 3 shows the ablation quantitative analysis results of different modules using the method of the present invention on the S3DIS region 5 dataset;

[0058] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A semantic segmentation method for large-scale point clouds in digital twin applications, characterized in that, The method includes the following steps: Step 1: Construct a point cloud and feature set, and obtain the neighborhood features of the points using the K-nearest neighbor search algorithm; Step 2: By using density adaptive downsampling, combined with local density estimation and weight fusion, the sampling probability is adjusted to balance the attention of different density regions in the point cloud; Step 3: Through the external attention module, an external memory unit is introduced, residual links are added, and a progressive single-multi-head design is adopted to accurately capture the global remote dependencies of the point cloud and balance feature learning, outputting the global features of the point cloud that fuse the global context. Step 4: Using a triple information encoding unit, distance is measured from three dimensions: spatial geometry, color, and features. Multi-dimensional information is fused and the contextual features of neighboring points are aggregated. Step 5: Through triple hybrid attention pooling, multi-dimensional fusion features and global features are concatenated, attention scores are calculated through nonlinear transformation, neighboring point features are weighted and aggregated, and local details and global context are fused. Step Six: Repeat Steps Three through Five in sequence. Each execution is recorded as one layer of encoding module. Before each layer of encoding module is executed, point cloud dimensionality reduction and sampling optimization are completed through density adaptive downsampling to achieve hierarchical progressive extraction of point cloud features, aggregate multi-scale feature information, and provide high-quality feature input for the decoder. Step 7: Calculation of network decoding and semantic segmentation results.

2. The large-scene point cloud semantic segmentation method for digital twin applications as described in claim 1, characterized in that, The process of step one is as follows; Step 1.1) Define a point cloud dataset: Point cloud data consists of a set of discrete three-dimensional spatial points. Each point has a continuous spatial distribution relationship and shares similar geometric and semantic attributes with its neighboring points. Given a point cloud dataset containing N points; Step 1.2) Obtain neighborhood features: Use the KNN algorithm based on Euclidean distance to collect the local neighborhood of the point cloud, thereby determining the spatial geometric information, color information and semantic feature information of the local point cloud.

3. A large-scene point cloud semantic segmentation method for digital twin applications as described in claim 1 or 2, characterized in that, The process of step two is as follows: Step 2.1) Obtaining Location Weights: According to the random downsampling method, during the sampling process, a sampling probability is maintained for each point, initialized to a very small random value. In each iteration, the point with the lowest current sampling probability is selected as the sampling center point. Subsequently, for points in its local neighborhood, geometric distance weights are calculated and assigned. , represents the sampling weight of the i-th neighboring point; the closer the distance, the greater the weight. Step 2.2) Local density characterization: For each point in the point cloud Calculate the Euclidean distances from the node to its K nearest neighbors, and take the median to characterize the local density. ; Step 2.3) Adaptive Weight Fusion: Subtract the local density of the current point from the maximum local density of all points in the current sampling, and then divide by the difference between the maximum and minimum local density values ​​to obtain the normalized density weight. Adaptive weight fusion is then used, with geometric distance weights as the basis, and the gain intensity of the density weights is controlled by hyperparameters to achieve adaptive weight enhancement, resulting in the fused sampling weights. Finally, the sampling probabilities of all points in the neighborhood of each sampling point are accumulated and fused to obtain new sampling probabilities.

4. A large-scene point cloud semantic segmentation method for digital twin applications as described in claim 1 or 2, characterized in that, The process of step three is as follows: Step 3.1) Introduction of External Attention Infrastructure: An external attention mechanism based on external memory units is employed, utilizing two independent external memory units, namely key memory units. Value memory unit To achieve efficient feature extraction, where S is the size of the memory unit, which is much smaller than the number of input points N, and d represents the feature dimension; First, the point cloud dataset defined in step one is used to obtain sampling points according to the downsampling method described in step two. Then, the features of the sampling points are input into a single linear layer with shared parameters for linear transformation to generate input features. Then The similarity response is obtained by multiplying the key memory unit matrix by its transpose and then doubly normalized; subsequently, it is multiplied by the value memory unit matrix to obtain the final output feature. ; The design employs a dual normalization model. The similarity is exponentially normalized according to the memory unit dimension. The Softmax function from deep learning is applied to each column of the original similarity matrix to make the sum of the weights of each memory unit on all input points equal to 1. Then, each row of the intermediate weight matrix obtained in the first step is summed and normalized to make the sum of the weights of each input point on all memory units equal to 1, thereby obtaining a stable attention weight distribution. Step 3.2) Large-scale point cloud semantic segmentation application.

5. A large-scene point cloud semantic segmentation method for digital twin applications as described in claim 4, characterized in that, The process of step 3.2) is as follows: Step 3.2.1) Residual connection: In order to ensure that the network does not lose the most basic local geometric structure and detailed information of the point cloud while learning global context using the attention mechanism, a residual connection is added at the output. Step 3.2.2) Multi-head external attention mechanism: First, the input features are divided according to the channel dimension. Each feature is individually fed into an external attention function to calculate the output of each attention head; then, the output features of all attention heads are concatenated along the channel dimension and linearly fused using a learnable weight matrix to complete multi-head information aggregation; finally, a residual connection is introduced to add the fusion result to the original input. Step 3.2.3) Progressive attention: A single-head external attention mechanism is used in the low-dimensional layers of the network to prioritize the construction of robust basic feature representations; The higher-dimensional layer then employs a multi-head mechanism to capture richer global context information.

6. A large-scene point cloud semantic segmentation method for digital twin applications as described in claim 1 or 2, characterized in that, The process of step four is as follows: Step 4.1) Triple distance metric: Based on the local neighborhood of the point cloud obtained from Step 1 The spatial geometric information, color information, and semantic features of point clouds are measured, and the process is as follows: Step 4.1.1) Spatial information quantification: In spatial coordinates, consider the relative distance and Euclidean distance between point coordinates; Step 4.1.2) Color information quantification: Examine the relative distance between colors to construct the relative color difference; Step 4.1.3) Semantic feature information quantification: Calculate the mean absolute error between semantic features to measure the similarity of high-dimensional semantic features, so as to ensure that sufficient contextual information is collected; Step 4.2) Design of the triple information coding unit, the process is as follows: Step 4.2.1) Triple relation vector concatenation: The quantization results of the relationship between the spatial coordinate domain, color domain and feature space obtained in Step 4.1 are concatenated to obtain the triple relation vector between the center point and the neighboring points; Step 4.2.2) Information Encoding: The triple relation is encoded through a multilayer perceptron network, including vector batch normalization, followed by two linear transformations with ReLU activation functions, outputting a triple-enhanced relation feature representation. The two-layer linear transformation is implemented using a shared parameter approach, which reuses the same set of network parameters for all center points and their neighboring point pairs to reduce the parameter size and improve generalization ability.

7. A large-scene point cloud semantic segmentation method for digital twin applications as described in claim 1 or 2, characterized in that, The process of step five is as follows: Step 5.1) Feature concatenation: Treat the features obtained through the external attention module as new point features. It is controlled by points through an external attention module. Dynamically updated, click The new KNN feature representation is as follows ,make Features obtained with triple information coding unit By cascading and combining these features, a triple hybrid attention feature is formed. ; Step 5.2) Normalized attention weight calculation: For concatenated features A shared multilayer perceptron is applied for nonlinear transformation, and the normalized attention weights corresponding to each neighborhood feature are calculated through the Softmax function. Step 5.3) Context-enhanced feature calculation: Utilize the attention weights to calculate the fused features. Perform element-wise weighted summation and then perform a weighted summation operation within a K-neighborhood to obtain the point. Contextual enhancement features .

8. A large-scene point cloud semantic segmentation method for digital twin applications as described in claim 1 or 2, characterized in that, In step six, the execution steps from step three to step five are denoted as EAF-Net. EAF-Net consists of three major modules: EAB, TIEB, and TFAP. EAF-Net is executed repeatedly in sequence. Each execution is recorded as one layer of encoding module. Before each layer of encoding module is executed, point cloud dimensionality reduction and sampling optimization are completed through density adaptive downsampling. This enables hierarchical progressive extraction of point cloud features, aggregating multi-scale feature information, providing high-quality feature input for the decoder, and supporting the subsequent improvement of semantic segmentation accuracy.

9. A semantic segmentation method for large-scale point clouds for digital twin applications as described in claim 1 or 2, characterized in that, In step seven, during the decoding stage, multi-scale features of the encoder are restored by layer-by-layer interpolation upsampling and skip connections to restore the original point cloud resolution and enhance the semantic representation of features. Finally, the class probability distribution of each point is calculated and output by the MLP classification head and the Softmax activation function, thereby obtaining the final semantic segmentation result of the point cloud.

10. A large-scene point cloud semantic segmentation method for digital twin applications as described in claim 1 or 2, characterized in that, The process of step seven is as follows: Step 7.1) Layer-by-layer upsampling: First, calculate the Euclidean distance between each sampling point in two adjacent point cloud layers to establish a spatial correspondence between the low-resolution point set and the point set to be restored; then, for each target point in the layer to be restored, select its nearest neighbor in the low-resolution point set of the previous layer, and use nearest neighbor interpolation to propagate the features of the nearest neighbor in the previous layer to the target point. Step 7.2) Skip connection multi-scale feature fusion: After completing single-layer upsampling, the interpolated features and the corresponding scale features in the encoder are spliced ​​and fused in the channel dimension through skip connections to simultaneously retain deep semantic information and shallow geometric details. Then, the spliced ​​fused features are processed by a shared multilayer perceptron. Step 7.3) Point-level semantic classification output: The above upsampling and feature fusion process is repeated layer by layer along the decoding path until the original point cloud resolution is restored, and a point-level high-dimensional feature representation containing global context semantics and local spatial structure information is obtained; finally, the point-level high-dimensional features are input into the multilayer perceptron classification module for nonlinear mapping and channel transformation, and the probability distribution of each point in the preset semantic category is calculated by the Softmax activation function. The semantic category label of each point is output according to the maximum probability correspondence, thereby obtaining the final semantic segmentation result of the point cloud.

Images