An aviation laser point cloud semantic segmentation method and device based on a multi-level context feature fusion network

Abstract

The application relates to an aviation laser point cloud semantic segmentation method and device based on a multi-level context feature fusion network, which comprises the following steps: S1, data down-sampling preprocessing is performed on acquired target region point cloud; S2, the processed aviation laser point cloud data is input into a pre-established multi-level context feature fusion network coding layer, and each layer feature is input into a cross-layer attention module to obtain cross-layer connection features; S3, the coding features are up-sampled to the size of the point cloud to be segmented through a decoding layer, multi-level feature fusion modules are used for fusing the features of all levels, and then a Softmax function is used to obtain the final segmentation result of the aviation laser point cloud; and S4, a hybrid loss function is constructed to obtain the error between the segmentation result and the true value, and according to the error value, a random gradient descent method is used for back propagation training, model parameters are dynamically adjusted, and a stable segmentation result is obtained. The method and device can effectively enhance the association of point feature expression and spatial information, improve the information perception ability of multi-scale context features of the point cloud, and obtain a more fine semantic segmentation result.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology and relates to a method and apparatus for semantic segmentation of aerial laser point clouds based on a multi-level contextual feature fusion network. Background Technology

[0002] LiDAR (LiDAR) is a technology that acquires 3D information about objects by emitting laser pulses and measuring their return time. Compared to traditional remote sensing technologies, LiDAR offers advantages such as high resolution, illumination independence, and strong penetration capabilities. By mounting LiDAR equipment on manned or unmanned aircraft for aerial scanning, aerial laser point cloud data covering large areas can be obtained. This technology is unaffected by lighting and shadows, enabling accurate ground feature observation under varying times and weather conditions. In urban point clouds, aerial laser point cloud data contains rich information about human and natural targets, making it crucial for applications such as semantic segmentation, urban 3D modeling, and high-precision mapping. However, due to the massive volume of aerial laser point cloud data, the complexity of urban areas, and the diverse geometric attributes of ground features, existing algorithms still face challenges in the fine extraction of semantic features for semantic segmentation.

[0003] Traditional point cloud processing methods mainly include feature extraction-based methods, clustering-based methods, and segmentation-based methods. Feature extraction-based methods identify and classify ground features by calculating feature values ​​of geometric, color, or other attributes in the point cloud, such as surface normals and curvature. Clustering-based methods divide point cloud data into different clusters or groups, each representing a ground feature category; commonly used algorithms include K-means and DBSCAN. Segmentation-based methods divide the point cloud into smaller sub-regions and achieve semantic segmentation through algorithms such as region growing and watersheds.

[0004] However, traditional point cloud processing methods have several drawbacks. First, these methods are sensitive to complex variations in ground feature geometry and noise, leading to inaccurate and unstable results. Second, traditional methods typically require manual selection of parameters or features, which is impractical for large-scale datasets and is highly user-dependent. Furthermore, traditional methods are computationally intensive and slow for large-scale, high-density point cloud data. Additionally, traditional methods perform poorly when dealing with challenges such as occlusion, incomplete data, and complex backgrounds. Finally, these methods often lack advanced semantic understanding and contextual reasoning capabilities, making it difficult to achieve fine-grained classification and recognition of ground features.

[0005] To overcome these shortcomings, researchers are exploring new point cloud processing methods, such as the application of deep learning and machine learning techniques. These methods utilize neural network models and large-scale training data to automatically learn and extract features, exhibiting stronger adaptability and generalization capabilities. Currently, deep learning-based methods have achieved significant improvements in 3D point cloud semantic segmentation, demonstrating better performance compared to traditional point cloud segmentation methods. Based on the organization and management type of point cloud data, existing techniques can be summarized into three categories: projection-based methods, voxel-based methods, and point-based methods.

[0006] Projection-based methods utilize the projection information of point cloud data onto a two-dimensional plane for segmentation. Typically, the point cloud is projected onto the image plane or a specific geometric projection space, and then a two-dimensional convolutional neural network (CNN) is used for semantic segmentation. This approach leverages mature techniques in image segmentation and deep learning architectures to efficiently process and analyze point cloud data.

[0007] Voxel-based methods convert point cloud data into voxel (3D pixel) representations and use them as input for semantic segmentation. By dividing the point cloud space into a regular voxel grid, the point cloud information in each voxel can be encoded as voxel features. Then, a 3D convolutional neural network (3D CNN) is used to segment the voxels to obtain the semantic information of the point cloud. This method can preserve more structural information of the point cloud, but it may be limited by computational and storage resource challenges when processing large-scale point clouds.

[0008] Point-based methods directly segment point cloud data, with each point serving as the network input. These methods can fully utilize the geometric and positional information of point cloud data. For example, Thomas et al. proposed the kernel-point continuous convolutional network KPConv. However, this network only contains one kernel-point convolution (KPConv) per downsampling layer, failing to achieve robust multi-level feature extraction for point clouds of varying densities. Huang et al. proposed a local differential attention and global perceptual attention network (GraNet) for semantic segmentation of aerospace laser point cloud data. This method first extracts features through attention weighting to learn representative geometric information. However, this method does not consider the global contextual information of unordered point clouds, thus failing to fully recognize fine-grained point cloud semantic features.

[0009] In summary, existing techniques for semantic segmentation of airborne laser point clouds do not adequately consider multi-level contextual feature learning and neglect the imbalance between semantic categories. Therefore, proposing a method capable of multi-level feature extraction and fusion to improve the robustness and accuracy of semantic segmentation is a significant research endeavor. Summary of the Invention

[0010] In view of this, the purpose of this invention is to provide a semantic segmentation method and apparatus for airborne laser point clouds based on a multi-level context feature fusion network. The squeezed extended multilayer perceptron module (SE-MLP) compresses and expands channel information using multiple MLPs, and enhances the local feature representation of the kernel point convolution operator (KPConv) by combining residual connections. For the downsampling path, to improve the effectiveness of feature information propagation in skip connections, a channel self-attention module (CSA) is used to extract contextual features of the point cloud data through weighted aggregation of feature information. The cross-layer attention fusion module (CAF) performs attention-weighted learning on features from different decoding stages, which can improve the discriminative power of fine-grained point cloud segmentation. This method and apparatus can better acquire and analyze airborne laser point clouds and has advantages over other methods in segmenting imbalanced point clouds.

[0011] To achieve the above objectives, the present invention provides the following technical solution:

[0012] A semantic segmentation method for airborne laser point clouds based on a multi-level context feature fusion network (MCFN) is characterized by the following steps: S1: preprocessing the acquired target region point cloud by downsampling; S2: inputting the processed airborne laser point cloud data into the pre-established multi-level context feature fusion network encoding layer, and inputting the features of each layer into a cross-layer attention module to obtain cross-layer connection features; S3: upsampling the encoded features to the scale of the point cloud to be segmented through the decoding layer, fusing the decoded features at each level through a multi-level feature fusion module, and then obtaining the final segmentation result of the airborne laser point cloud by the Softmax function; S4: constructing a hybrid loss function to obtain the error between the segmentation result and the true value, and using stochastic gradient descent for backpropagation training based on the error value to dynamically adjust the model parameters and obtain a robust segmentation result.

[0013] Furthermore, in step S1, the KD-tree algorithm is used to perform downsampling preprocessing on the point cloud data of the acquired target.

[0014] 7. Further, in step S2, the processed airborne laser point cloud data is input into the established multi-level context feature fusion network encoding layer, that is, the collected point cloud training data is input into the kernel convolution block (KPConv) to learn the point cloud features and obtain the features after kernel convolution; in downsampling, the compressed extended multilayer perceptron (SE-MLP) module is used to aggregate the local neighborhood features of the point cloud features after kernel convolution KPConv operation again to achieve effective representation of features in the downsampling stage. Specifically, the kernel convolution operation is first used to extract the point cloud features to obtain feature x. kp ,as follows:

[0015]

[0016] in, For coding layers of different scales, For input point cloud sample features, KPConv() represents the KPConv function, whose formula is as follows:

[0017]

[0018] in, It is a point cloud Internal fixed radius ( The neighborhood set of ), i.e. , It belongs to point clouds any subset, It is a subset of Dianyun Corresponding features, kernel function The formula is as follows:

[0019]

[0020] in, This indicates the position of the spherical core point in 3D space. Representing kernel function The number of core points, It is the weight matrix corresponding to the kernel points;

[0021] Features after kernel convolution operation The features are compressed and expanded using SE-MLP to obtain discriminative features. Simultaneously, batch normalization (BN) is introduced to improve the trainability of the proposed network. The features are then processed by a multilayer perceptron. The representation is as follows:

[0022]

[0023] In the formula, For coding layers of different scales, Here, ∑ represents the learnable parameters of different layers in SE-MLP, and ∑ is the summation operation. Dot product of matrices;

[0024] based on A SE-MLP module is designed using residual information to improve network depth and feature representation capabilities. This module effectively performs compression and expansion operations, aggregates local features, and enhances the diversity of point cloud feature representation. The output of SE-MLP can be obtained as follows:

[0025]

[0026] Here, ReLU() represents the activation function, which is ReLU(x) = max(0, x).

[0027] 8. Furthermore, to enhance information propagation between the downsampling and upsampling paths of MCFN, and to obtain multi-level point cloud context features within a channel from a single skip connection path, a channel self-attention module (CSA) suitable for KPConv operations was designed; given the input feature matrix... For a point set, Projecting onto three different feature subspaces, we form query, key, and value, respectively. The calculations for R, S, and V are as follows:

[0028]

[0029] Where i represents different downsampling layers. This represents a pairwise MLP operation, which compresses the input features by half and then restores them to their original scale. Finally, the output feature map with skip connections can be calculated as follows:

[0030]

[0031] In the formula, It is used to map tensors to probability distributions; The CSA module output features are matrix dot products. , where i represents different downsampling layers. This module highlights the low-level features of skip connections, which can help similar semantic features obtain better discriminative power.

[0032] 9. Further, after the five layers of the above encoding modules, the decoding operation is performed. The shallow features are concatenated with the high-level features after passing through the CSA module. The concatenated tensor is then subjected to a 1x1 convolution operation and upsampled. A cross-layer attention fusion module (CAF) is used to fuse the high-level feature information from different upsampled layers after the 1x1 convolution operation in the decoding stage. The formula is as follows:

[0033]

[0034] in, For learnable weights, For decoding layers of different scales, This is the normalized scaling factor in the BN layer. and These correspond to the output and input features of different upsampling layers in the CAF module, respectively. It is the sogmoid function, that is , This indicates the average pooling operation. The CAF module extracts different point cloud features at different upsampling layers, achieving multi-level information aggregation. The aggregation formula is as follows:

[0035]

[0036] in, This represents the aggregated features of the 3 upsampling layers. This is a tensor splicing operation.

[0037] 10. Furthermore, the semantic segmentation result obtained by passing the CAF tensor through two fully connected layers and the Softmax function is then processed. To address the data imbalance issue, a focus loss function is introduced, which is set as follows:

[0038]

[0039] Here, λ is typically set to 2. This represents the category weight parameter, where N represents the number of point clouds. This represents the probability that the j-th sample belongs to category c;

[0040] To enhance the robustness of MCFN to imbalanced data, a weighted cross-entropy loss function is introduced. This function can better handle imbalanced data by assigning different weights to each class. The weighted cross-entropy is expressed as:

[0041]

[0042] in, Indicates the number of samples. Indicates the first Does the _ sample belong to the _ ... Categories Indicates belonging to, This indicates that it does not belong to; The model represents the first The sample belongs to the first The natural logarithm of the predicted probabilities for each category; based on the above loss function, the overall loss function can be obtained as:

[0043]

[0044] The model parameters of the semantic segmentation framework are optimized based on the overall loss function and the backpropagation process. After training, a trained semantic segmentation framework is obtained. The trained semantic segmentation framework is used to discriminate input samples and output robust semantic segmentation results.

[0045] The beneficial effects of this invention are as follows:

[0046] The compressed extended multilayer perceptron module proposed in this invention can effectively aggregate local neighborhood features of point clouds after kernel convolution operations. This invention proposes a channel self-attention module suitable for kernel convolution to mine multi-level contextual information and generate discriminative semantic features. The cross-layer attention fusion module proposed in this invention fuses attention features from different decoding stages to achieve inter-layer information interaction and improve the fine segmentation effect of point clouds. This invention proposes a semantic segmentation method for aerospace laser point clouds based on a multi-level contextual feature fusion network (MCFN). This network focuses on multi-level feature representation learning and has strong representation capabilities for aerospace laser point clouds. Experimental results on two aerospace laser scanning point cloud datasets show that the proposed MCFN outperforms the current state-of-the-art aerospace laser point cloud semantic segmentation methods.

[0047] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0048] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0049] Figure 1 This is a flowchart of the method of the present invention;

[0050] Figure 2 A schematic diagram of point cloud data collected by an airborne platform;

[0051] Figure 3 The graph consists of a Multi-Level Context Feature Fusion Network (MCFN), a Compressed Extended Multilayer Perceptron Module (SE-MLP), a Channel Self-Attention Module (CSA), and a Cross-Layer Attention Fusion Module (CAF) for semantic segmentation of airborne laser point clouds.

[0052] Figure 4 The figure shows the experimental comparison results of the MCFN network of this invention. Detailed Implementation

[0053] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0054] Figure 1This is a flowchart of the method of the present invention. The present invention provides a method and apparatus for semantic segmentation of airborne laser point clouds based on a multi-level contextual feature fusion network. As shown in the figure, in the point cloud acquisition stage, a LiDAR-equipped UAV or manned aircraft can be used to quickly acquire spatial remote sensing information, including geographical location, reflection intensity, etc., to complete the acquisition of point cloud data of the target area. The deep learning network used for semantic segmentation is as follows: Figure 3 As shown, it can learn detailed and representative ground feature features from airborne laser point clouds. The network consists of a kernel-point convolutional module (KPConv), a compressed extended multilayer perceptron module (SE-MLP), a channel self-attention module (CSA), and a cross-layer attention fusion module (CAF). First, the point cloud data obtains local features through a five-layer kernel-point convolutional module. Then, the SE-MLP module extracts and aggregates semantic features with multi-level detailed representation capabilities from these local features. Next, the CSA module is embedded between the concatenation of low-level and high-level features to perform attention-weighted learning of contextual semantic information. Finally, in the decoding stage, the CAF module is used to mine high-level semantics at different decoding layers and fuse them with the network output. This invention proposes a multi-level contextual feature fusion network focusing on multi-level fusion learning of key information to mine the intrinsic information of airborne laser point clouds with unbalanced category distribution and generate discriminative high-level semantic features.

[0055] Specifically, the technical solution of the present invention includes the following:

[0056] 1. Data preprocessing and downsampling: The KD-tree algorithm is used to perform downsampling preprocessing on the acquired target area point cloud data.

[0057] 2. Input the processed point cloud training data into a kernel-based convolutional block (KPConv), such as... Figure 3 As shown, the collected point cloud training data is input into a kernel-convolutional block (KPConv) to learn the point cloud features, obtaining the features after kernel-convolution. During downsampling, a compressed extended multilayer perceptron (SE-MLP) module is used to further aggregate local neighborhood features of the point cloud features after the kernel-convolutional KPConv operation, achieving effective feature representation in the downsampling stage. Specifically, kernel-convolutional operations are first used to extract the point cloud features, obtaining feature x. kp ,as follows:

[0058]

[0059] in, For coding layers of different scales, For input point cloud sample features, KPConv() represents the KPConv function, whose formula is as follows:

[0060]

[0061] in, It is a point cloud Internal fixed radius ( The neighborhood set of ), i.e. , It belongs to point clouds any subset, It is a subset of Dianyun Corresponding features, kernel function The formula is as follows:

[0062]

[0063] in, This indicates the position of the spherical core point in 3D space. Representing kernel function The number of core points, It is the weight matrix corresponding to the kernel points;

[0064] Features after kernel convolution operation The features are compressed and expanded using SE-MLP to obtain discriminative features. Simultaneously, batch normalization (BN) is introduced to improve the trainability of the proposed network. The features are then processed by a multilayer perceptron. The representation is as follows:

[0065]

[0066] In the formula, For coding layers of different scales, Here, ∑ represents the learnable parameters of different layers in SE-MLP, and ∑ is the summation operation. Dot product of matrices;

[0067] based on A SE-MLP module is designed using residual information to improve network depth and feature representation capabilities. This module effectively performs compression and expansion operations, aggregates local features, and enhances the diversity of point cloud feature representation. The output of SE-MLP can be obtained as follows:

[0068]

[0069] Here, ReLU() represents the activation function, which is ReLU(x) = max(0, x).

[0070] 3. To enhance information propagation between the downsampling and upsampling paths of MCFN, a Channel Self-Attention (CSA) module suitable for KPConv operations was designed to obtain multi-level point cloud context features within a channel from a single skip connection path; given the input feature matrix... For a point set, Projecting onto three different feature subspaces, we form query, key, and value, respectively. The calculations for R, S, and V are as follows:

[0071]

[0072] Where i represents different downsampling layers. This represents a pairwise MLP operation, which compresses the input features by half and then restores them to their original scale. Finally, the output feature map with skip connections can be calculated as follows:

[0073]

[0074] In the formula, It is used to map tensors to probability distributions; The CSA module output features are matrix dot products. , where i represents different downsampling layers. This module highlights the low-level features of skip connections, which can help similar semantic features obtain better discriminative power.

[0075] 4. After five layers of the above encoding modules, the decoding operation is performed. The shallow features are concatenated with the high-level features after passing through the CSA module. The concatenated tensor is then subjected to a 1x1 convolution operation and upsampled. A cross-layer attention fusion module (CAF) is used to fuse the high-level feature information from different upsampled layers after the 1x1 convolution operation in the decoding stage. The formula is as follows:

[0076]

[0077] in, For learnable weights, For decoding layers of different scales, This is the normalized scaling factor in the BN layer. and These correspond to the output and input features of different upsampling layers in the CAF module, respectively. It is the sogmoid function, that is , This indicates the average pooling operation. The CAF module extracts different point cloud features at different upsampling layers, achieving multi-level information aggregation. The aggregation formula is as follows:

[0078]

[0079] in, This represents the aggregated features of the 3 upsampling layers. This is a tensor splicing operation.

[0080] 5. The semantic segmentation result obtained by passing the CAF tensor through two fully connected layers and the Softmax function is then processed. To address the data imbalance issue, a focus loss function is introduced, and the loss function is set as follows:

[0081]

[0082] Here, λ is typically set to 2. This represents the category weight parameter, where N represents the number of point clouds. This represents the probability that the j-th sample belongs to category c;

[0083] To enhance the robustness of MCFN to imbalanced data, a weighted cross-entropy loss function is introduced. This function can better handle imbalanced data by assigning different weights to each class. The weighted cross-entropy can be expressed as:

[0084]

[0085] in, Indicates the number of samples. Indicates the first Does the _ sample belong to the _ ... Categories Indicates belonging to, This indicates that it does not belong to; The model represents the first The sample belongs to the first The natural logarithm of the predicted probabilities for each category; based on the above loss function, the overall loss function can be obtained as:

[0086]

[0087] The model parameters of the semantic segmentation framework are optimized based on the overall loss function and the backpropagation process. After training, a trained semantic segmentation framework is obtained. The trained semantic segmentation framework is used to discriminate input samples and output robust semantic segmentation results.

[0088] like Figure 4The experimental results of the MCFN semantic segmentation network described in this invention are shown on an open-source airborne city point cloud dataset, LASDU. Figure (a) shows the ground truth, Figure (b) shows the result of the KPConv algorithm, and Figure (c) shows the result of our proposed algorithm, MCFN. Compared with the ground truth, it can be seen that the categories in the elliptical regions of the MCFN algorithm are well segmented. The segmentation effect of this invention can be further illustrated by comparative experiments. The method of this invention is compared with other existing methods, PointNet++, PointCNN, PointConv, DGCNN, ShellNet, PosPool, and KPConv, on the LASDU dataset, as shown in Table 1. The overall accuracy (OA) and average F1 index (Avg. F1) are calculated respectively. A higher overall accuracy (OA) indicates a higher proportion of correct predictions among all correctly predicted results; a higher average F1 index (Avg. F1) indicates a better overall evaluation of the results. Table 1 shows the values ​​of various indicators for the detection results of different methods:

[0089] Table 1 Comparison of MCFN and various methods on the LASDU dataset

[0090] Methods ground building tree low_veg artifact OA Avg.F1 PointNet++ 87.7 90.6 82.0 63.2 31.3 82.8 71.0 PointCNN 89.3 92.8 84.1 62.8 31.7 85.0 72.1 PointConv 89.6 94.3 84.6 67.5 36.4 85.9 74.5 DGCNN 90.5 93.2 81.6 63.3 37.1 85.5 73.1 ShellNet 89.5 92.8 84.6 68.5 35.7 85.6 74.2 PosPool 88.3 93.7 83.9 61.0 38.3 83.5 73.0 KPConv 89.1 93.4 83.2 59.7 31.9 83.7 71.5 ours (MCFN) 91.6 96.7 85.9 67.1 43.8 88.0 77.0

[0091] As can be seen, the method of this invention achieves the best OA and Avg. F1 on this dataset. This also demonstrates that the performance of the method described in this invention is superior to other aerial laser point cloud semantic segmentation methods. The method proposed in this invention can achieve better results and has advantages over other methods in terms of categories such as ground, buildings, trees, and man-made objects.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications should be covered within the scope of the claims of the present invention.

Claims

1. A semantic segmentation method for airborne laser point clouds based on a multi-level contextual feature fusion network, characterized in that: The method includes the following steps: S1: Perform data downsampling preprocessing on the acquired point cloud of the target area; S2: Input the processed airborne laser point cloud data into the pre-established multi-level context feature fusion network encoding layer, and input the features of each layer into the cross-layer attention module to obtain cross-layer connection features; S3: The encoded features are upsampled to the scale of the point cloud to be segmented through the decoding layer, the decoded features at each level are fused through a multi-level feature fusion module, and then the final segmentation result of the aerospace laser point cloud is obtained by the Softmax function; S4: Construct a hybrid loss function to obtain the error between the segmentation result and the true value. Based on the error value, use stochastic gradient descent to perform backpropagation training and dynamically adjust the model parameters to obtain a robust segmentation result. In step S2, the processed airborne laser point cloud data is input into the established multi-level context feature fusion network encoding layer. This involves inputting the collected point cloud training data into a kernel convolutional block (KPConv) to learn the point cloud features and obtain the kernel-convolutioned features. During downsampling, the compressed extended multilayer perceptron (SE-MLP) module is used to further aggregate local neighborhood features of the point cloud features after the KPConv operation, achieving effective feature representation in the downsampling stage. Specifically, kernel convolution is first used to extract the point cloud features, obtaining feature x. kp ,as follows: in, For coding layers of different scales, For input point cloud sample features, KPConv(·) represents the KPConv function, and its formula is as follows: in, It is a point cloud Internal fixed radius The neighborhood set, ,Right now , It belongs to point clouds any subset, It is a subset of Dianyun Corresponding features, kernel function The formula is as follows: in, This indicates the position of the spherical core point in 3D space. Representing kernel function The number of core points, These are the weight matrices corresponding to the kernel points; the features after kernel point convolution operation. The features are compressed and expanded using SE-MLP to obtain discriminative features. Simultaneously, batch normalization (BN) is introduced to improve the trainability of the proposed network. The features are then processed by a multilayer perceptron. The representation is as follows: In the formula, For coding layers of different scales, Here, ∑ represents the learnable parameters of different layers in SE-MLP, and ∑ is the summation operation. Dot product of matrices; based on A SE-MLP module is designed using residual information to improve network depth and feature representation capabilities. This module effectively performs compression and expansion operations, aggregates local features, and enhances the diversity of point cloud feature representation. The output of SE-MLP can be obtained as follows: Here, ReLU(·) represents the activation function, which is ReLU(x) = max(0, x).

2. The semantic segmentation method for airborne laser point clouds based on a multi-level contextual feature fusion network according to claim 1, characterized in that: In step S2, in order to enhance the information propagation between the downsampling path and the upsampling path of MCFN, and to obtain multi-level point cloud context features within the channel from a single skip connection path, a channel self-attention module (CSA) suitable for KPConv operation was designed. MCFN is a multi-level context feature fusion network. Given the input feature matrix For a point set, Projecting onto three different feature subspaces, we form query, key, and value, respectively. The calculations for R, S, and V are as follows: Where i represents different downsampling layers. This represents a pairwise MLP operation, which compresses the input features by half and then restores them to their original scale. Finally, the output feature map with skip connections can be calculated as follows: In the formula, It is used to map tensors to probability distributions; The CSA module output features are matrix dot products. , where i represents different downsampling layers. This module highlights the low-level features of skip connections, which can help similar semantic features obtain better discriminative power.

3. The semantic segmentation method for airborne laser point clouds based on a multi-level contextual feature fusion network according to claim 2, characterized in that: In step S3, after five layers of the above encoding modules, the decoding operation is performed. The shallow features are concatenated with the high-level features after passing through the CSA module. The concatenated tensor is then subjected to a 1x1 convolution operation and upsampled. The cross-layer attention fusion module CAF is used to fuse the high-level feature information from different upsampled layers after the 1x1 convolution operation in the decoding stage. The formula is as follows: in, For learnable weights, For decoding layers of different scales, This is the normalized scaling factor in the BN layer. and These correspond to the output and input features of different upsampling layers in the CAF module, respectively. It is the sogmoid function, that is , This indicates the average pooling operation. The CAF module extracts different point cloud features at different upsampling layers, achieving multi-level information aggregation. The aggregation formula is as follows: in, This represents the aggregated features of the 3 upsampling layers. This is a tensor splicing operation.

4. The semantic segmentation method for airborne laser point clouds based on a multi-level contextual feature fusion network according to claim 3, characterized in that: In step S4, the tensor after the CAF operation is passed through two fully connected layers, and the semantic segmentation result is obtained by the Softmax function. To address the data imbalance problem, a focus loss function is introduced, and the loss function is set as follows: Where λ is set to 2, This represents the category weight parameter, where N represents the number of point clouds. This represents the probability that the j-th sample belongs to category c; To enhance the robustness of MCFN to imbalanced data, a weighted cross-entropy loss function is introduced. This function can better handle imbalanced data by assigning different weights to each class. The weighted cross-entropy is expressed as: in, Indicates the number of samples. Indicates the first Does the _ sample belong to the _ ... Categories Indicates belonging to, This indicates that it does not belong to; The model represents the first The sample belongs to the first The natural logarithm of the predicted probabilities for each category; based on the above loss function, the overall loss function can be obtained as: The model parameters of the semantic segmentation framework are optimized based on the overall loss function and the backpropagation process. After training, a trained semantic segmentation framework is obtained. The trained semantic segmentation framework is used to discriminate input samples and output robust semantic segmentation results.

5. A semantic segmentation device for airborne laser point clouds using a multi-level contextual feature fusion network, characterized in that: The apparatus includes a processor, a memory, and an airborne laser point cloud semantic segmentation program based on a multi-level context feature fusion network, stored in the memory and executable by the processor, wherein the airborne laser point cloud semantic segmentation program based on the multi-level context feature fusion network, when executed, implements the steps of the method as described in any one of claims 1 to 4.

Images