Urban mobile laser scanning point cloud semantic segmentation method based on double-domain attention

Abstract

The invention provides an urban mobile laser scanning point cloud semantic segmentation method based on double-domain attention, and the method comprises the following steps: firstly, constructing a local space attention block, extracting local semantic information through a space self-attention mechanism, and enhancing the perception capability of local features; secondly, designing a global channel attention block, focusing on relevance modeling between feature channels to capture global context information, and improving the expression ability of the model to a complex scene; further, an information flow control module is introduced, an information transmission path is regulated and controlled through a gating mechanism, key details are reserved, and redundant information interference is suppressed; and finally, a complete semantic segmentation network is constructed based on the modules, and efficient semantic analysis of the point cloud data is realized. Compared with the existing point cloud segmentation method, the method provided by the invention has higher semantic recognition precision under the urban scene mobile laser scanning point cloud, the extracted urban street structure is more complete and clearer, and the method shows stronger adaptability and robustness to a complex point cloud environment.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional point cloud information processing, and more specifically, to a semantic segmentation method for urban mobile laser scanning point clouds based on deep learning. Background Technology

[0002] Mobile laser scanning systems, with their flexibility and high precision, have become a key technology for acquiring urban 3D point cloud data, and are widely used in urban surveying, autonomous driving, infrastructure management, and building information modeling. Fine-grained semantic segmentation of point cloud data—that is, correctly classifying each point into categories such as buildings, vehicles, pedestrians, and vegetation—is the foundation and prerequisite for realizing these advanced applications. However, urban mobile laser scanning point clouds are characterized by their large scale, uneven density, complex structure, and significant occlusion, posing a severe challenge to semantic segmentation.

[0003] In recent years, deep learning techniques, especially those based on the attention (Transformer) architecture, have demonstrated enormous potential in point cloud processing due to their powerful long-range dependency modeling capabilities. Currently, point cloud segmentation methods based on the attention (Transformer) architecture are evolving primarily in two directions: first, constructing hierarchical attention mechanisms, combining downsampling and self-attention to capture multi-scale features; and second, improving local geometry awareness, such as combining graph convolution and offset attention to enhance feature robustness. These methods have achieved significant progress in point cloud classification and segmentation tasks.

[0004] However, directly applying traditional attention-based (Transformer) architectures to semantic segmentation of urban mobile laser scanning point clouds still has significant shortcomings. Firstly, it suffers from high computational complexity: the computational cost of global self-attention mechanisms increases quadratically with the size of the point cloud, making it difficult to efficiently handle massive urban point cloud data. Secondly, it suffers from loss of local details: when focusing on the global context, it easily overlooks the crucial local fine structures that affect the segmentation results, such as the geometric and semantic information of small targets like utility poles and signs. This leads to weak detail preservation capabilities and low accuracy in small target recognition in complex urban scenes. Therefore, how to effectively capture the semantics of the global scene while accurately preserving and enhancing local detail features is a core challenge that current technologies urgently need to overcome. Summary of the Invention

[0005] The technical problem to be solved by this invention is that, in response to the shortcomings of existing deep learning-based point cloud segmentation methods on urban mobile laser scanning data, such as easy loss of local details, poor segmentation effect of small-scale targets, and difficulty in effective coordination of global and local features, this invention provides a semantic segmentation method for urban mobile laser scanning point clouds based on dual-domain attention.

[0006] This invention addresses the aforementioned problems encountered in the semantic segmentation of urban moving laser scanning point clouds using current deep neural networks. It introduces a semantic segmentation network for urban moving laser scanning point clouds based on dual-domain attention, excelling in semantic segmentation of urban moving laser scanning point clouds. It employs a two-pronged strategy across spatial and channel dimensions to preserve scene details, such as small urban objects. First, in the local spatial attention block, spatial intelligent attention is used to observe local patterns, thereby obtaining enhanced self-attention containing extensive local semantic cues. This block also transmits local spatial information in parallel from different representation subspaces embedded in each input, enhancing its ability to learn local spatial semantics. Then, based on the feature pyramid framework of fusing and refining the local spatial attention module, we introduce a global channel attention block, which effectively captures the global context by focusing on the interrelationships of feature channels. On one hand, our method utilizes Transformers in both the spatial and channel domains to examine and fuse features from multiple layers, effectively summarizing the semantic context and enriching spatial details with multi-scale information. On the other hand, we enhance semantic features by integrating semantic features with aggregated context and use a gating mechanism for selective information delivery.

[0007] The technical solution adopted by this invention to solve its technical problem is: a semantic segmentation method for urban mobile laser scanning point clouds based on dual-domain attention, comprising the following steps:

[0008] S1. Point cloud data preprocessing is performed on the large-scale point cloud data generated by urban mobile laser scanning to obtain training dataset and test dataset;

[0009] S2. Introduce local spatial attention blocks to improve the self-attention mechanism and enhance local semantic expression. At the same time, design global channel attention blocks in the feature channel dimension to model inter-channel dependencies and capture global context.

[0010] S3. Perform multi-scale feature fusion on the local spatial features and global channel features obtained in step S2 to obtain dual-domain attention features based on spatial and channel dimensions. These fused features will be passed to the classifier module for further optimization to improve classification accuracy and complete the semantic classification task of point cloud.

[0011] S4. Use the training dataset obtained in step S1 to train the dual-domain attention network composed of S2-S3. During the training process, the loss function of the network is the multi-class cross-entropy loss function. Train until the loss function converges to obtain the final generative network model.

[0012] S5. Use the generative network model trained in step S4 to test the test dataset obtained in S1 to obtain the point cloud prediction results.

[0013] Furthermore, the processing of local spatial attention blocks consists of two parts: local patch embedding and spatial attention; When performing local patch embedding, random subsampling is used to process the original point cloud to reduce point density differences and decrease the computational cost of network training. Then, for each sampled point... K-nearest neighbors are used to obtain the set of its nearest neighbors. Then, the relative positions of the center point and its neighboring points are calculated through edge convolution. : ; And connect it with the absolute coordinates to obtain a local patch. Then, a multilayer perceptron (MLP) is used to aggregate them into a higher-dimensional feature space to generate local block embeddings. ; Then, the learned local block embeddings are... Compared with the feature vectors previously captured by the perceptron MLP Connect and obtain adaptive weights through MLP processing Then through SoftMax ( The function is used to normalize and aggregate the input feature embeddings within the neighborhood, forming enhanced local attention features. ; Based on this, an improved spatial self-attention mechanism is adopted, according to the features of the input. Generate query, key, and value vectors:

[0014] In the formula, the query vector, key vector and value vector are all generated by the same input feature through MLP linear mapping, wherein each linear mapping module is identical in structure and independent of each other in parameters;

[0015] Furthermore, a nonlinear maximally symmetric function incorporating relative positions is used to enhance the robustness of unordered point cloud feature aggregation, thereby strengthening local semantic learning. The function is defined as follows:

[0016]

[0017] In the formula: max represents the symmetric maximum aggregation function, and position embedding. It is by using relative positions The query vector and key vector are passed to the MLP; next, trainable positional embeddings are used to enhance the query vector and key vector to effectively integrate positional relationships into the model to obtain the attention map. :

[0018]

[0019] In the formula: This represents the dot product operation; subsequently, local semantic data is processed by a self-attention mechanism and connected to the attention map. Merge to generate local attention features Finally, fused features are generated through an MLP layer, and the final local features of the local spatial attention block are represented as follows: As shown below:

[0020]

[0021]

[0022] In the formula: SoftMax( The function is used for normalization. Represents a parameter that can be learned Shared multilayer perceptron.

[0023] Furthermore, generate local block embeddings. The process is shown in the following formula:

[0024]

[0025]

[0026] In the formula: Indicates a connection operation. Represents a parameter that can be learned Shared MLP.

[0027] Furthermore, the adaptive weights obtained through MLP processing The calculation formula is:

[0028] ;

[0029] Local attention characteristics The calculation formula is:

[0030] .

[0031] Furthermore, based on the U-Next pyramid structure, a global channel attention block is constructed, specifically including:

[0032] Features from different scales and propagation paths of the U-Next pyramid, including lateral propagation features, top-down propagation features, and bottom-up propagation features, are first concatenated along the channel dimension, and then feature encoding is performed through channel mapping operations to obtain multi-scale fused features. Based on this, the multi-scale fusion features are used to obtain a query matrix through three sets of structurally similar and parameter-independent channel mappings. Key matrix Sum matrix These matrices are used to construct channel relevance weights and perform cross-channel information aggregation, generating channel-level attention features, denoted as... Its definition is as follows:

[0033]

[0034]

[0035] In the formula: It is the number of channels used for feature fusion. For the final local features, Represents a parameter that can be learned Shared multilayer perceptron, SoftMax ( The ) function is used for normalization.

[0036] Furthermore, step S2 also includes processing the output of the global channel attention block using an information control flow module. The processing procedure is as follows:

[0037]

[0038]

[0039] In the formula: Gating represents the gating mechanism unit, For GELU nonlinear activation layer, Indicates element-wise multiplication. This is a global channel feature.

[0040] Furthermore, the standard cross-entropy loss function L used during training is:

[0041]

[0042] In the formula: L represents cross-entropy; N is the total number of samples; K is the number of sample categories; is the one-hot encoding of the sample target value, which is 1 if the true class of sample i belongs to c, and 0 otherwise; is the probability that the model predicts that sample i belongs to class c, and its value is in (0,1).

[0043] Furthermore, after step S5, the extraction results obtained in step S5 are evaluated using overall accuracy (OA) and mean intersection-over-union ratio (mIoU) as evaluation metrics. These metrics are then compared and analyzed with other model algorithms from both numerical and visual perspectives. The definitions of OA and mIoU are as follows:

[0044]

[0045]

[0046] In the formula: K represents the number of categories, TP i This represents the number of correctly predicted points in the i-th class, where N represents the total number of points. mIoU measures the overlap between the predictions and the true labels; the IoU is calculated first for each class, then averaged. FP i Representing the The number of points in a class that are incorrectly predicted as belonging to that class, FN i This indicates the number of points that are actually of that class but are predicted as other classes.

[0047] The present invention also provides a semantic segmentation system for urban mobile laser scanning point clouds based on dual-domain attention, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute the semantic segmentation method for urban mobile laser scanning point clouds based on dual-domain attention as described in the above technical solution.

[0048] The present invention also provides a computer-readable storage medium, including a readable storage medium on which a computer program is stored, wherein when the computer program is executed, it implements the urban mobile laser scanning point cloud semantic segmentation method based on dual-domain attention as described in the above technical solution.

[0049] The semantic segmentation method for urban mobile laser scanning point clouds based on dual-domain attention, which implements the present invention, has the following advantages:

[0050] 1. This invention significantly improves the network's accuracy in perceiving and segmenting the geometric features of small-scale, slender urban elements such as utility poles, signs, and railings by using improved self-attention and relative position encoding in local spatial attention blocks.

[0051] 2. This invention performs efficient self-attention calculation in the channel dimension through global channel attention blocks, integrates multi-scale features with low computational cost, and enables the network to obtain powerful scene-level semantic understanding capabilities.

[0052] 3. This invention achieves adaptive and selective fusion of local detailed features and global context features through an information flow control gating mechanism, avoiding feature dilution or conflict, and enabling the network output to have both semantic consistency and rich details.

[0053] 4. The overall network structure constructed in this invention is highly efficient, achieving leading segmentation accuracy on publicly available large-scale urban point cloud benchmark datasets, while maintaining reasonable model complexity and inference speed, and has practical application potential. Attached Figure Description

[0054] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0055] Figure 1This is a flowchart illustrating the technical process of the method of the present invention.

[0056] Figure 2 This is a schematic diagram of the local spatial attention block in the method of the present invention;

[0057] Figure 3 This is a schematic diagram of the global channel attention block in the method of the present invention;

[0058] Figure 4 This is a schematic diagram of the information flow control module in the method of the present invention;

[0059] Figure 5 The results show the comparison of the evaluation index values ​​between the method of this invention and other model algorithms;

[0060] Figure 6 The results show the comparison of the visualization effects of the method of this invention with other model algorithms, where (A)-(F) are the areas that need to be magnified for display.

[0061] Figure 7 The results show the comparison of the visualization effects of the method of this invention with other model algorithms, where (A)-(E) are the areas that need to be magnified for display. Detailed Implementation

[0062] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0063] Referring to Figure 1, which is a flowchart of the method of the present invention, the semantic segmentation method for urban mobile laser scanning point clouds based on dual-domain attention proposed in this invention includes the following steps:

[0064] S1. Publicly available urban mobile laser scanning point cloud datasets are used, covering complex scenes such as urban roads, buildings, vegetation, and vehicles. The input of the datasets only uses the three-dimensional coordinates (x, y, z) of the point cloud to avoid interference from redundant information.

[0065] S2. The semantic segmentation method for urban mobile laser scanning point clouds based on dual-domain attention includes the following core technical processes: First, a local spatial attention block is constructed, focusing on spatial transformation attention to observe local patterns, effectively integrating local information from the input data into the feature vector to provide richer information representation. (Reference) Figure 2 The local spatial attention block consists of two parts: local patch embedding and spatial attention. During local patch embedding, random subsampling of the original point cloud is used to reduce point density differences and decrease network training computation. Then, for each sampled point... The nearest neighbor set of points is obtained using K-nearest neighbors (K=16). Then, the relative positions of the center point and its neighboring points are calculated through edge convolution. :

[0066] ,

[0067] And connect it with the absolute coordinates to obtain a local patch. Then, a multilayer perceptron (MLP) is used to aggregate them into a higher-dimensional feature space to generate local block embeddings. The process is shown in the following formula:

[0068]

[0069]

[0070] In the formula: Indicates a connection operation. Represents a parameter that can be learned A shared MLP. Subsequently, the learned local block embeddings are... Compared with the feature vectors previously captured by the perceptron MLP Connect and obtain adaptive weights through MLP processing :

[0071] ,

[0072] Then through SoftMax( The function is used to normalize and aggregate the input feature embeddings within the neighborhood, forming enhanced local attention features. :

[0073]

[0074] Based on this, an improved spatial self-attention mechanism is adopted, according to the features of the input. Generate query, key, and value vectors:

[0075]

[0076] In the formula, the query vector, key vector, and value vector are all generated from the same input feature through MLP linear mapping, where each linear mapping module is structurally identical and parameter-independent. Furthermore, a nonlinear maximally symmetric function incorporating relative positions is used to enhance the robustness of unordered point cloud feature aggregation, thereby strengthening local semantic learning. This function is defined as follows:

[0077]

[0078] In the formula: max represents the symmetric maximum aggregation function, and position embedding. It is by using relative positions The query vector and key vector are then passed to the MLP. Next, trainable positional embeddings are used to augment the query and key vectors, effectively integrating positional relationships into the model to obtain the attention map. .

[0079]

[0080] In the formula: This represents the dot product operation. Subsequently, the local semantic data is processed through a self-attention mechanism and connected to the attention map. Merge to generate local attention features Finally, fused features are generated through an MLP layer, and the final attention features of the local spatial attention block are represented as follows: As shown below:

[0081]

[0082]

[0083] In the formula: SoftMax( The function is used for normalization. Represents a parameter that can be learned Shared multilayer perceptron.

[0084] Secondly, based on the structure of the U-Next pyramid, this method constructs a global channel attention block, which performs global feature fusion by focusing on the relationships between channels. (Reference) Figure 3 Global feature fusion is achieved by modeling the correlation between different feature channels. Specifically, features from different scales and propagation paths, including lateral propagation features, top-down propagation features, and bottom-up propagation features, are first concatenated along the channel dimension, and then feature encoding is performed through channel mapping operations to obtain multi-scale fused features. Based on this, the multi-scale fusion features are used to obtain a query matrix through three sets of structurally similar and parameter-independent channel mappings. Key matrix Sum matrix These matrices are used to construct channel relevance weights and perform cross-channel information aggregation, generating channel-level attention features, denoted as... Its definition is as follows:

[0085]

[0086]

[0087] In the formula: It is the number of channels used to fuse features.

[0088] Finally, to regulate information flow, an information flow control module was designed, referencing... Figure 4 Selective feature propagation is achieved through gating mechanisms:

[0089]

[0090] Wherein: Gating represents the gating mechanism unit. For GELU nonlinear activation layer, This represents element-wise multiplication. The final output, enhanced global channel features, is achieved through residual connections. :

[0091]

[0092] This dual-domain attention feature extraction network structure collaboratively enhances local spatial details and global channel context, providing multi-level feature representations for subsequent semantic segmentation.

[0093] S3. Perform multi-scale feature fusion on the local spatial features and global channel features obtained in step S2 to obtain dual-domain attention features based on spatial and channel dimensions. These fused features will be passed to the classifier module for further optimization to improve classification accuracy and complete the semantic classification task of point clouds.

[0094] S4. Train the dual-domain attention network model constructed in step S2 using the training dataset obtained in step S1. During training, the cross-entropy loss function is used for optimization. The network continuously adjusts its parameters to minimize the loss function until the loss function converges, resulting in the final generative network model. After training, the network will be able to generate category predictions based on the input point cloud data. The definition of the loss function L is as follows:

[0095]

[0096] In the formula: L represents cross-entropy; N is the total number of samples; K is the number of sample categories; is the one-hot encoding (0 or 1) of the sample target value, which is 1 if the true category of sample i belongs to c, and 0 otherwise; is the probability that the model predicts that sample i belongs to category c, and its value is within (0,1).

[0097] S5. Using the network model obtained in step S4, extract features from the 3D point cloud of the test dataset obtained in step S1, and input it into the classifier to obtain the classification prediction for each point. The final test results are then visualized. Figure 6 and Figure 7 .

[0098] S6. (Optional) Evaluate the extraction results obtained in step S5, using the overall accuracy (OA) and mean intersection-union ratio (mIoU) as evaluation metrics. Their definitions are as follows:

[0099]

[0100]

[0101]

[0102] In the above formula: Represented as the number of categories, This represents the number of correctly predicted points in the i-th class, where N represents the total number of points. mIoU measures the overlap between the predictions and the true labels; the IoU is calculated first for each class, and then averaged. This represents the number of points in the i-th class that were incorrectly predicted as belonging to that class. This indicates the number of points that are actually of that class but are predicted as other classes.

[0103] The method of this invention is compared and analyzed with other model algorithms from two aspects: numerical evaluation indicators and visualization results. For example... Figure 5 As shown, the method of this invention outperforms the comparative algorithm in both overall accuracy (OA) and mean intersection-union ratio (mIoU), indicating that this invention can effectively improve the semantic discrimination ability after multi-scale feature fusion. Combined with... Figure 6 and Figure 7 The visualization results show that the method of this invention has a better ability to distinguish small targets and adjacent semantic regions in complex point cloud scenes, with more continuous and complete semantic boundaries and a significant reduction in misclassified points. This indicates that by introducing spatial and channel dual-domain feature modeling and a controlled flow information filtering mechanism, this invention can effectively preserve local detail information while ensuring global semantic consistency.

[0104] This invention also provides a semantic segmentation system for urban mobile laser scanning point clouds based on dual-domain attention, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute the semantic segmentation method for urban mobile laser scanning point clouds based on dual-domain attention as described in the above technical solution.

[0105] This invention also provides a computer-readable storage medium, including a computer program stored on the storage medium, wherein when the computer program is executed, it implements the urban mobile laser scanning point cloud semantic segmentation method based on dual-domain attention as described in the above technical solution.

[0106] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A semantic segmentation method for urban mobile laser scanning point clouds based on dual-domain attention, comprising the following steps: S1, Point cloud data preprocessing is performed on the large-scale point cloud data generated by urban mobile laser scanning to obtain training dataset and test dataset; S2 introduces local spatial attention blocks to improve the self-attention mechanism and enhance local features, while designing global channel attention blocks in the feature channel dimension to model inter-channel dependencies and capture global features; S3. Multi-scale feature fusion is performed on the local and global features obtained in step S2 to obtain dual-domain attention features based on spatial and channel dimensions. These features are then passed to the classifier module for further optimization to complete the semantic classification of the point cloud. S4. Use the training dataset obtained in step S1 to train the semantic segmentation network composed of S2-S3 to obtain the final generative network model. S5. Use the generative network model trained in step S4 to test the test dataset obtained in S1 to obtain the point cloud prediction results.

2. The method as described in claim 1, characterized in that: The processing of local spatial attention blocks consists of two parts: local patch embedding and spatial attention; When performing local patch embedding, random subsampling is used to process the original point cloud to reduce point density differences and decrease the computational cost of network training. Then, for each sampled point... K-nearest neighbors are used to obtain the set of its nearest neighbors. Then, the relative positions of the center point and its neighboring points are calculated through edge convolution. : ； And connect it with the absolute coordinates to obtain a local patch. Then, a multilayer perceptron (MLP) is used to aggregate them into a higher-dimensional feature space to generate local block embeddings. ; Then, the learned local block embeddings are... Compared with the feature vectors previously captured by the perceptron MLP Connect and obtain adaptive weights through MLP processing Then through SoftMax ( The function is used to normalize and aggregate the input feature embeddings within the neighborhood, forming enhanced local attention features. ; Based on this, an improved spatial self-attention mechanism is adopted, according to the features of the input. Generate query, key, and value vectors: ； In the formula, the query vector, key vector and value vector are all generated by the same input feature through MLP linear mapping, wherein each linear mapping module is identical in structure and independent of each other in parameters; Furthermore, a nonlinear maximally symmetric function incorporating relative positions is used to enhance the robustness of unordered point cloud feature aggregation, thereby strengthening local semantic learning. The function is defined as follows: ； In the formula: max represents the symmetric maximum aggregation function, and position embedding. It is by using relative positions The query vector and key vector are passed to the MLP; next, trainable positional embeddings are used to enhance the query vector and key vector to effectively integrate positional relationships into the model to obtain the attention map. : ； In the formula: This represents the dot product operation; subsequently, local semantic data is processed by a self-attention mechanism and connected to the attention map. Merge to generate local attention features Finally, fused features are generated through an MLP layer, and the final local features of the local spatial attention block are represented as follows: As shown below: ； ； In the formula: SoftMax( The function is used for normalization. Represents a parameter that can be learned. Shared multilayer perceptron.

3. The method as described in claim 2, characterized in that: Generate local block embeddings The process is shown in the following formula: ； ； In the formula: Indicates a connection operation. Represents a parameter that can be learned. Shared MLP.

4. The method as described in claim 2, characterized in that: Adaptive weights The calculation formula is: ； Local attention characteristics The calculation formula is: 。 5. The method as described in claim 2, characterized in that: Based on the U-Next pyramid structure, a global channel attention block is constructed, which specifically includes: Features from different scales and propagation paths of the U-Next pyramid, including lateral propagation features, top-down propagation features, and bottom-up propagation features, are first concatenated along the channel dimension, and then feature encoding is performed through channel mapping operations to obtain multi-scale fused features. Based on this, the multi-scale fusion features are used to obtain a query matrix through three sets of structurally similar and parameter-independent channel mappings. Key matrix Sum matrix These matrices are used to construct channel relevance weights and perform cross-channel information aggregation, generating channel-level attention features, denoted as... Its definition is as follows: ； ； In the formula: It is the number of channels used for feature fusion. For the final local features, Represents a parameter that can be learned. Shared multilayer perceptron, SoftMax ( The ) function is used for normalization.

6. The method as described in claim 1, characterized in that: Step S2 also includes processing the output of the global channel attention block using the information control flow module. The processing procedure is as follows: ； ； In the formula: Gating represents the gating mechanism unit, For GELU nonlinear activation layer, Indicates element-wise multiplication. This is a global channel feature.

7. The method as described in claim 1, characterized in that: The standard cross-entropy loss function L used during training: ； In the formula: L represents cross-entropy; N is the total number of samples; K is the number of sample categories; is the one-hot encoding of the sample target value, which is 1 if the true class of sample i belongs to c, and 0 otherwise; is the probability that the model predicts that sample i belongs to class c, and its value is in (0,1).

8. The method as described in claim 1, characterized in that: After step S5, the extraction results obtained in step S5 are evaluated using overall accuracy (OA) and mean intersection-union ratio (mIoU) as evaluation metrics. The results are compared and analyzed with other model algorithms from both numerical and visual perspectives. The definitions of OA and mIoU are as follows: ； ； In the formula: K represents the number of categories, TP i This represents the number of correctly predicted points in the i-th class, where N represents the total number of points. mIoU measures the overlap between the predictions and the true labels. For each class, the IoU is calculated first, then averaged. FP i Representing the The number of points in a class that are incorrectly predicted as belonging to that class, FN i This indicates the number of points that are actually of that class but are predicted as other classes.

9. A semantic segmentation system for urban mobile laser scanning point clouds based on dual-domain attention, characterized in that: It includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the program instructions in the memory to execute the urban mobile laser scanning point cloud semantic segmentation method based on dual-domain attention as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It includes a readable storage medium on which a computer program is stored, and when the computer program is executed, it implements the urban mobile laser scanning point cloud semantic segmentation method based on dual-domain attention as described in any one of claims 1-8.

Images