An airborne LiDAR urban point cloud semantic segmentation method and system of a recursive residual double-attention kernel point convolution network

Abstract

The application relates to an airborne LiDAR urban point cloud semantic segmentation method and system of a recursive residual double-attention kernel point convolution network, and belongs to the technical field of computer vision. The method comprises the following steps: S1: acquiring target region point clouds; S2: pre-processing the acquired target region point clouds to obtain training sample data and test sample data; S3: inputting the labeled training sample into the recursive residual double-attention kernel point convolution network for training; and S4: after the training is completed, performing semantic segmentation on the test sample and obtaining a result. The performance of the method and system is superior to that of other airborne LiDAR urban point cloud semantic segmentation methods, the method and system can better acquire and analyze airborne LiDAR urban point clouds, and have an advantage over other methods in terms of segmenting non-balanced point clouds.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology and relates to an airborne LiDAR urban point cloud semantic segmentation method and system using a recursive residual dual attention kernel point convolutional network. Background Technology

[0002] LiDAR (Light Detection and Ranging) is an emerging 3D remote sensing technology that provides point cloud data reflecting the 3D structure of ground features. Unaffected by lighting conditions or shadows, it offers significant advantages in feature identification. By mounting LiDAR equipment on manned or unmanned aircraft for aerial scanning, airborne LiDAR point cloud data covering large areas can be obtained. Furthermore, airborne LiDAR urban point clouds contain a wealth of high-value information related to human and natural elements; semantic segmentation of this data is fundamental and crucial for subsequent applications such as urban 3D modeling and high-precision mapping. However, due to the large volume of LiDAR data, the complexity of urban areas, and the variability in the geometric attributes of ground features, existing algorithms struggle to achieve precise semantic segmentation of ground features.

[0003] Traditional point cloud processing methods mostly obtain features from point cloud data through specific, manually defined methods, and then combine them with classifiers such as random forests, support vector machines, and Gaussian mixture models to train a set of feature classifiers to complete the semantic segmentation task of scattered point cloud data. Clearly, the performance of these methods largely depends on expert experience and classification algorithms. Although many scholars at home and abroad have proposed methods for generating 3D global descriptors for various application scenarios, such as spin images, fast point feature histograms (FPFH), hot kernel signatures (HKS), and histogram localization signatures (SHOT), these methods independently estimate the local features of each point and predict each label without considering the consistency between neighboring points. Therefore, the segmentation results are often affected by noise and label inconsistencies, and cannot be fully applied to all semantic segmentation scenarios.

[0004] Currently, thanks to the rapid development of deep learning, applying deep learning methods to 3D point cloud semantic segmentation often yields better results than traditional point cloud segmentation methods. Based on the type of convolution operator, existing techniques can be summarized into two categories: 3D point cloud semantic segmentation based on discrete convolution operators and 3D point cloud semantic segmentation based on continuous convolution operators. In methods primarily based on discrete convolution operations, some researchers have shifted to projecting the original 3D point cloud from multiple angles onto a 2D plane, converting it into an image segmentation task, a typical example being SnapNet. However, this projection dimensionality reduction method inevitably loses spatial information to some extent, especially under complex spatial distributions, making it easier to obtain poor results. Alternatively, the concept of pixels can be extended to 3D space, using voxelization to organize the topological relationships between spatial points before inputting into a deep learning model, such as SEGCloud. However, this method comes at the cost of excessive memory consumption and struggles to capture high-resolution and fine-grained features. To avoid the cumbersome process of multi-directional projection and voxel construction, PointNet pioneered a deep learning architecture directly applied to points. It uses a shared multilayer perceptron and transform network to independently learn the features of each point and extracts a global representation with a simple aggregation operation. PointNet++ considers the local structure of the point cloud, processing a set of points in a hierarchical manner and then aggregating the obtained local features to generate higher-level features. However, the discrete convolution-based methods described above have not achieved significant results on airborne LiDAR urban point clouds.

[0005] In methods for semantic segmentation of 3D point clouds based on continuous convolution operators, a convolution operation in a continuous space is defined, where the weights of nearest neighbors are related to the spatial distribution of each center point. Researchers proposed the kernel-point continuous convolutional network KPConv; however, this network contains only one kernel-point convolution (KPConv) for each downsampling layer, failing to achieve robust multi-level feature extraction for point clouds of varying densities. Further, a local and global encoder network (LGENet) was proposed for semantic segmentation of airborne LiDAR urban point cloud data. This method first extracts features through 2D and 3D kernel-point continuous convolutions to learn representative geometric information. However, this method does not consider the global contextual information of unordered point clouds. Furthermore, the aforementioned continuous kernel-point convolution method fails to consider the imbalance and sparsity of airborne LiDAR point cloud data from both local and global perspectives, and how to perform multi-level feature learning, thus failing to fully identify fine-grained point cloud semantic features.

[0006] In summary, existing technologies for semantic segmentation of urban point clouds using airborne LiDAR do not fully consider multi-level feature learning at both the local and global levels, and neglect the imbalance between semantic categories. Therefore, proposing a deep learning method for semantic segmentation of urban point clouds using airborne LiDAR that can represent global and local features at multiple levels, thereby improving the accuracy of semantic segmentation, is a significant research objective. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a semantic segmentation method and system for airborne LiDAR urban point clouds based on a recursive residual dual-attention kernel point convolutional network. Firstly, this method and system proposes an attention kernel point convolution (AKPConv) module, which uses a batch normalization scaling factor to weight channel information. This factor utilizes the standard deviation to characterize the importance of channel weights, enhancing the local feature representation of the point cloud. Based on the AKPConv module, a recursive residual kernel point attention module (RRKA) is proposed, which aggregates diverse features of neighboring points through iterative cumulative learning. A five-layer encoding module is established based on the RRKA module, and then a corresponding five-layer decoding module is established based on the encoding module. Finally, to fully integrate the cross-layer features of the encoding and decoding layers, a global-local channel attention interaction module (GLCA) is proposed to fuse global and local information, thereby improving the discriminative power of fine-grained point cloud segmentation. This method and system can better acquire and analyze airborne LiDAR urban point clouds and has advantages over other methods in segmenting class-imbalanced point clouds.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A semantic segmentation method for airborne LiDAR urban point clouds using a recursive residual dual-attention kernel point convolutional network is proposed. The method includes the following steps: S1: acquiring point clouds of the target region; S2: preprocessing the acquired point clouds of the target region to obtain training sample data and test sample data; S3: inputting some labeled samples into the recursive residual dual-attention kernel point convolutional network for training; S4: performing semantic segmentation on the test samples after training and obtaining the results.

[0010] Furthermore, in step S1, a UAV or manned aircraft platform equipped with LiDAR is used to collect urban point cloud data. Specifically, this includes: S11, selecting the target area to be segmented based on the task requirements of urban point cloud semantic segmentation, setting the flight parameters of the airborne platform, including but not limited to flight altitude and speed, planning the flight route, and adopting a "Z" shaped flight route; S12, setting the scanning parameters of LiDAR, considering the overlap rate of point clouds to ensure the accuracy of point cloud scanning, selecting a non-repeating scanning-based LiDAR with an FOV of 70.4°×77.2°, a ranging accuracy of 3cm~1000cm, and a maximum multi-echo rate of 480,000 points / second; S13, during the actual data collection, the airborne platform equipped with LiDAR flies along the set scanning route to acquire point cloud data of the target area.

[0011] Furthermore, in step S2, the acquired target region point cloud is preprocessed, mainly including point cloud registration, noise removal, and radiometric correction. Then, training and test samples are extracted from the point cloud data in blocks, and semantic labels are assigned to the training samples. Specifically, to reduce the influence of the long-tailed distribution of the point cloud echo intensity, a gamma transform is used to correct the point cloud echo intensity data to a normal distribution to improve the network's robustness. The formula is as follows:

[0012]

[0013] Where I is the collected echo intensity; I γ The echo intensity after gamma correction ranges from 0 to 255; γ is a parameter with a value range of 0 ≤ γ ≤ 1; this formula can be used to map the original echo intensity to the image space.

[0014] Further, in step S3, the gamma-corrected point cloud training data is input into the attention kernel point convolutional block (AKPConv) to learn the point cloud features and obtain the attention kernel point convolutional features; in the downsampling, the AKPConv module can be used to downsample the number of points in the point cloud; specifically, the scaling factor in the sample normalization layer is first used to obtain the attention weights, and the channel attention features f are obtained. c ,as follows:

[0015]

[0016] in, Input point cloud sample features; for The mean, for The variance; To avoid a denominator of 0, set it to a small constant, 1 × 10. -5 ; Features of normalized point cloud samples; It is a product of elements; It is the sigmoid function, that is Then, the attention features are input into the kernel convolution formula to obtain the feature F1, which is as follows:

[0017]

[0018] in, Indicated Convolution, KPConv(f c The ) represents kernel convolution operation, and its specific formula is as follows:

[0019]

[0020] in, It is a point At a fixed radius ( The neighborhood set of ), i.e. , It belongs to a point any subset, It is a subset of Dianyun Corresponding features Represented as a kernel function, its formula is as follows:

[0021]

[0022] 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; correlation function ,in , ,function Used to quantitatively describe the midpoint of a point cloud neighborhood. Relative to target point offset vector , with core point The degree of linear correlation in 3D spatial location, `<h>` is a hyperparameter used to control the effect of kernel point distance; skip connections are added to the AKPConv block to preserve input features; for skip connections, It is an optional max pooling operation, when D in When it is 2D, it will be used, where D represents the size of the input feature dimension. in This indicates the size of the input feature dimension of the AKPCOnv module, and this operation can be represented as:

[0023]

[0024] Finally, the output features of AKPConv It can be represented as:

[0025]

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

[0027] Furthermore, to accumulate and aggregate local features and generate diverse features, a recurrent residual kernel point attention module (RRKA) is composed of a recurrent point convolutional block (RPConv) consisting of AKPConv and a single hidden layer of multilayer perceptron (MLP). The recurrent point convolutional (RPConv) block is mainly used to learn the accumulated neighborhood features of the point cloud, and its formula is expressed as follows:

[0028]

[0029] in Let RPConv be the input for layer l, recursively executed T times, where t = 1, 2, 3, ..., T; AKPConv() represents the attention kernel convolution operation. For RPConv's input at time t, This is the output of the (t-1)th recursion. To ensure the computational efficiency of RPConv, an MLP is first used to compress the feature dimension before using RPConv, and then the MLP is used to restore the feature dimension. Therefore, a compressed loop block (CRB) based on RPConv and MLP is constructed to improve computational efficiency, and its formula is as follows:

[0030]

[0031] Where W1 and W2 are the learnable parameters of the MLP, PRConv() is the recursive point convolution operation, and BN() is the batch normalization operation. For the input point cloud features, a Residual Recursive Kernel Attention (RRKA) module with residual connections was developed based on CRB. This module can effectively perform repeated operations on local features, enhancing the diversity representation capability of point cloud features. The RRKA output can be expressed as:

[0032]

[0033] Where W0 represents the weight learning parameters of the MLP. The input point cloud features are CRB(·), which is the CRB operation.

[0034] Furthermore, after five layers of the aforementioned encoding modules, a decoding operation is performed. After concatenating the low-level and high-level features, the feature tensor is processed using a Global-Local Channel Attention (GLCA) module. First, a fully connected approach is used to fuse the channel information of the low-level and high-level features in the global space. The formula is as follows: as follows:

[0035]

[0036] The input to this module is... N and C are the number of upsampled point clouds and the dimension of features, respectively, and W... g To fuse fully connected weights from low-level and high-level features; attention weights are obtained using average pooling and one-dimensional convolution in local feature computation:

[0037]

[0038] in, For the average pooling operation of the channel, W k conv These are learnable local one-dimensional kernel convolution weights with a size of k=5, expressed by the following formula:

[0039]

[0040] Based on attention weight The following local attention features can be obtained. ( (This represents element-wise multiplication), the formula is as follows:

[0041]

[0042] Output of the Global-Local Cross-Layer Information Interaction Module (GLCA) Represented as:

[0043]

[0044] Furthermore, the feature tensor is passed through two fully connected layers, and the semantic segmentation result obtained by the Sigmoid activation function is used to address the data imbalance problem. A focus loss function is introduced, and the loss function is set as follows:

[0045]

[0046] 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; the model parameters of the semantic segmentation framework are optimized using stochastic gradient descent based on the focus loss function, and after training, the trained semantic segmentation framework is obtained; the trained semantic segmentation framework is used to discriminate the input test samples and output the semantic segmentation results.

[0047] The beneficial effects of this invention are as follows:

[0048] The attention kernel convolutional representation module proposed in this invention can effectively learn local neighborhood features of point clouds. It weights channel information using a batch normalization scaling factor and uses standard deviation to characterize the importance of channel weights. This invention proposes a recursive residual kernel attention module, focusing on the attention kernel convolutional module, to mine multi-level local information of point clouds and generate discriminative low-level semantic features. The global-local channel attention module proposed in this invention fuses upsampled high-level features with skip-connected low-level features, achieving interaction between global and local information and improving the fine-grained segmentation effect of point clouds. This invention proposes RRDAN, an airborne LiDAR urban point cloud semantic segmentation method based on a recursive residual dual-attention kernel convolutional network. This network focuses on multi-level feature representation learning and has strong representation capabilities for class-imbalanced airborne LiDAR urban point clouds. Experimental results on two airborne LiDAR urban point cloud datasets show that the proposed RRDAN outperforms state-of-the-art airborne LiDAR urban point cloud semantic segmentation methods.

[0049] 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

[0050] 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:

[0051] Figure 1 This is a flowchart of the method of the present invention;

[0052] Figure 2 A schematic diagram of point cloud data collected by an airborne platform;

[0053] Figure 3 A diagram of a Recursive Residual Dual Attention Kernel Point Convolutional Network (RRDAN) for semantic segmentation of urban point clouds using airborne LiDAR;

[0054] Figure 4 This is a structural diagram of the attention kernel convolution module AKPConv of the present invention;

[0055] Figure 5 This is a structural diagram of the Recursive Residual Kernel Point Attention Module RRKA of the present invention;

[0056] Figure 6 This is a structural diagram of the Global-Local Channel Attention Interaction Module (GLCA) of the present invention;

[0057] Figure 7 The figure shows the experimental results of the RRDAN network of this invention;

[0058] Figure 8 This is an error graph showing the experimental results of the RRDAN network of this invention. Detailed Implementation

[0059] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0060] Figure 1 This is a flowchart of the method of the present invention. The present invention provides a method and system for semantic segmentation of urban point clouds using an airborne LiDAR based on a recursive residual dual-attention kernel point convolutional network. As shown in the figure, in the point cloud acquisition stage, a LiDAR mounted on a UAV or manned aircraft can be used to quickly acquire spatial remote sensing information, including geographical location, reflectivity, 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 point clouds. The network consists of an attention kernel convolutional module (AKPConv), a recurrent residual kernel attention module (RRKA), and a global-local channel attention interaction module (GLCA). First, the point cloud data after gamma correction of reflection intensity obtains local features through a three-layer attention kernel convolutional module. Then, the RRKA module extracts and accumulates semantic features with multi-level detailed representation capabilities from these local features. After concatenating low-level and high-level features, the global-local channel attention module (GLCA) is applied to the feature tensor to learn the fused semantic information. The attention kernel convolutional module (AKPConv) designed in this invention is used to extract local features of the point cloud neighborhood, embedding an attention mechanism to weight channel features and adding skip connections to fuse contextual information. The proposed recurrent residual kernel attention module (RRKA) can aggregate diverse features of neighborhood points through iterative cumulative learning. Through the network's dual attention mechanism, subtle feature representations can be enhanced to improve segmentation performance. This invention proposes a multi-layered dual-attention kernel convolutional network that focuses on recursive cumulative learning of key information to mine the intrinsic information of airborne point clouds with unbalanced category distribution and generate high-level semantic features with discriminative power.

[0061] Specifically, the technical solution of the present invention includes the following:

[0062] 1. Data Acquisition: Urban point cloud data acquisition is achieved using a drone or manned aircraft platform equipped with LiDAR. First, based on the requirements of the urban point cloud semantic segmentation task, the target area to be segmented is selected, and the flight parameters of the airborne platform are set, including flight altitude and speed. A flight route is planned, typically a zigzag pattern. Then, the LiDAR scanning parameters are set, considering the resolution of the point cloud density to ensure scanning accuracy. Here, a non-repeating scanning-based LiDAR is selected, with an FOV of 70.4°×77.2°, a ranging accuracy of 3cm to 1000cm, and a maximum multi-echo rate of 480,000 points / second. During the actual acquisition, the airborne platform equipped with LiDAR flies along the set scanning route to acquire point cloud data of the target area, as illustrated in the diagram below. Figure 2 .

[0063] 2. Data Preprocessing: The acquired point cloud of the target region is preprocessed, mainly including point cloud registration, noise removal, and radiometric correction. Then, training and test samples are extracted from the point cloud data in blocks, and semantic labels are assigned to the training samples. Specifically, to reduce the influence of the long-tailed distribution of the point cloud echo intensity, a gamma transform is used to correct the point cloud echo intensity data to a normal distribution to improve the network's robustness. The formula is as follows:

[0064]

[0065] Where I is the collected echo intensity; I γ The echo intensity after gamma correction ranges from 0 to 255; γ is a parameter with a value range of 0 ≤ γ ≤ 1; this formula can be used to map the original echo intensity to the image space.

[0066] 3. Input the gamma-corrected point cloud training data into the attention kernel point convolution block (AKPConv), such as... Figure 4 As shown, point cloud features are learned to obtain attention kernel point convolutional features; during downsampling, the AKPConv module is used to downsample the number of points in the point cloud. Specifically, the attention weights are first obtained using the scaling factor in the sample normalization layer, resulting in the channel attention features f. c ,as follows:

[0067]

[0068] in, Input point cloud sample features; for The mean, for The variance; To avoid a denominator of 0, set it to a small constant, 1 × 10. -5 ; Features of normalized point cloud samples; It is a product of elements; It is the sigmoid function, that is .

[0069] Then, the attention features are input into the kernel convolution formula to obtain the feature F1, as follows:

[0070]

[0071] in, Indicated Convolution, KPConv() represents kernel-point convolution operation, and its specific formula is as follows:

[0072]

[0073] in, It is a point At a fixed radius ( The neighborhood set of ), i.e. , It belongs to a point any subset, It is a subset of Dianyun Corresponding features, kernel function The formula is as follows:

[0074]

[0075] 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; correlation function , This is a hyperparameter used to control the influence of the kernel point distance.

[0076] To preserve input features, skip connections are added to the AKPConv block; for skip connections, It is an optional max pooling operation. When D in When it is 2D, it will be used, and the operation can be represented as:

[0077]

[0078] Finally, the output features of AKPConv It can be represented as:

[0079]

[0080] Here, ReLU(x) = max(0, x) represents the activation function.

[0081] 4. To accumulate and aggregate local features and generate diverse features, a recurrent residual kernel point attention module (RRKA) is composed of a recurrent point convolutional block (RPConv) consisting of AKPConv and a single hidden layer of multilayer perceptron (MLP), as shown below. Figure 5 As shown. The recursive point convolution (RPConv) block is mainly used to learn the cumulative neighborhood features of the point cloud, and its formula is expressed as follows:

[0082]

[0083] in Let RPConv be the input for layer l, recursively executed T times, where t = 1, 2, 3, ..., T; AKPConv() represents the attention kernel convolution operation. For RPConv's input at time t, This is the output of the (t-1)th recursion.

[0084] To improve computational efficiency, an MLP is first used to compress the feature dimensions before RPConv, and then the MLP is used to restore the feature dimensions. Therefore, a compressed circular block (CRB) based on RPConv and MLP is constructed to improve computational efficiency, and its formula is as follows:

[0085]

[0086] Where W1 and W2 are the learnable parameters of the MLP, PRConv() is the recursive point convolution operation, and BN() is the batch normalization operation.

[0087] Based on CRB, a Residual Recursive Kernel Attention (RRKA) module with residual connections was developed. This module can effectively perform repeated operations on local features, enhancing the diversity representation capability of point cloud features. The RRKA output can be represented as:

[0088]

[0089] Where W0 represents the weight learning parameters of the MLP. The input point cloud features are used, and CRB() is the CRB operation.

[0090] 5. After five layers of the above encoding modules, the decoding operation is performed. After concatenating the low-level and high-level features, the feature tensor is processed using a Global-Local Channel Attention (GLCA) module, such as... Figure 6 As shown.

[0091] A fully connected approach is used to fuse channel information of low-level and high-level features in the global space. The formula is as follows: as follows:

[0092]

[0093] The input to this module is... N and C are the number of upsampled point clouds and the dimension of features, respectively, and W... g Fully connected weights that integrate features from both low-level and high-level layers.

[0094] In local feature computation, attention weights are obtained using average pooling and one-dimensional convolution:

[0095]

[0096] in, For the average pooling operation of the channel, W k These are learnable local one-dimensional kernel convolution weights with a size of k=5, and their formula is as follows:

[0097]

[0098] Based on attention weight The following local attention features can be obtained. ( (This represents element-wise multiplication), the formula is as follows:

[0099]

[0100] Output of the Global-Local Cross-Layer Information Interaction Module (GLCA) Represented as:

[0101]

[0102] 6. The semantic segmentation result obtained by passing the feature tensor through two fully connected layers and the sigmoid activation function is then processed. To address the imbalance problem, a focus loss function is introduced, and the loss function is set as follows:

[0103]

[0104] Wherein, λ is generally 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. The model parameters of the semantic segmentation framework are optimized using stochastic gradient descent based on the focal loss function. After training, a trained semantic segmentation framework is obtained. The trained semantic segmentation framework is then used to discriminate input test samples, and the semantic segmentation result is output.

[0105] like Figure 7 These are the experimental results of the RRDAN semantic segmentation network described in this invention on an open-source airborne city point cloud dataset ISPRS. Figure 8 The error map shows that the nine categories in the test area are well segmented. The segmentation performance of this invention can be further illustrated through comparative experiments. The method of this invention was compared with other existing methods such as LUH, RIT_1, alsNet, KPConv, DPE, GANet, DANCE-NET, D-FCN, RandLANet, GraNet, and LGENet on the ISPRS dataset, as shown in Table 1. Overall Accuracy (OA) and Average F1 Index (Avg. F1) were calculated respectively. A higher OA indicates a higher proportion of correct predictions among all correctly predicted results; a higher Avg. F1 index indicates a better overall evaluation. Table 1 presents the values ​​of various indicators for the detection results of different methods.

[0106] Table 1. Comparison of RRADN with various methods on the ISPRS dataset.

[0107] As can be seen, the method of this invention achieves the best OA and Avg. F1 on this dataset. Furthermore, the performance of the method described in this invention outperforms other airborne LiDAR urban point cloud semantic segmentation methods. The method proposed in this invention can achieve better results and has advantages over other methods in categories such as cars, roofs, facades, low shrubs, and trees.

[0108] 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 LiDAR urban point clouds using a recursive residual dual-attention kernel point convolutional network, characterized in that: The method includes the following steps: S1: Obtain the point cloud of the target area, wherein the point cloud of the target area is spatial remote sensing information, including geographical location or reflectance intensity; S2: Preprocess the acquired point cloud of the target area to obtain training sample data and test sample data; S3: Input a portion of the labeled samples into a recursive residual dual-attention kernel convolutional network for training; S4: After training is complete, perform semantic segmentation on the test samples and obtain the results; In step S1, a UAV or manned aircraft platform equipped with LiDAR is used to collect urban point cloud data. Specifically, this includes: S11, selecting the target area to be segmented based on the task requirements of urban point cloud semantic segmentation, setting the flight parameters of the airborne platform, including but not limited to flight altitude and speed, and planning the flight route, using a "Z" shaped flight route; S12, setting the scanning parameters of the LiDAR, considering the overlap rate of the point cloud to ensure the accuracy of the point cloud scanning, selecting a non-repeating scanning-based LiDAR with an FOV of 70.4°×77.2°, a ranging accuracy of 3cm~1000cm, and a maximum multi-echo rate of 480,000 points / second; S13, during the actual data collection, the airborne platform equipped with LiDAR flies along the set scanning route to acquire point cloud data of the target area. In step S2, the acquired target region point cloud is preprocessed, mainly including point cloud registration, noise removal, and radiometric correction. Then, training and test samples are extracted from the point cloud data in blocks, and semantic labels are assigned to the training samples. Specifically, to reduce the influence of the long-tailed distribution of the point cloud echo intensity, a gamma transform is used to correct the point cloud echo intensity data to a normal distribution to improve the network's robustness. The formula is as follows: Where I is the collected echo intensity; I γ The echo intensity after gamma correction ranges from 0 to 255; γ is a parameter with a value range of 0 ≤ γ ≤ 1; this formula can be used to map the original echo intensity to the image space. In step S3, the gamma-corrected point cloud training data is input into the attention kernel point convolutional block (AKPConv) to learn the point cloud features and obtain the attention kernel point convolutional features. During downsampling, the AKPConv module can be used to downsample the number of points in the point cloud. Specifically, the attention weights are first obtained using the scaling factor in the sample normalization layer, resulting in the channel attention features f. c ,as follows: in, Input point cloud sample features; for The mean, for The variance; To avoid a denominator of 0, set it to a small constant, 1 × 10. -5 ; Features of normalized point cloud samples; It is a product of elements; It is the sigmoid function, that is Then, the attention features are input into the kernel convolution formula to obtain the feature F1, which is as follows: in, Indicated Convolution, KPConv() represents kernel-point convolution operation, and its specific formula is as follows: in, It is a point At a fixed radius ( The neighborhood set of ), i.e. , It belongs to a point any subset, It is a subset of Dianyun Corresponding features Represented as a kernel function, its 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, It is the weight matrix corresponding to the kernel points; correlation function ,in , ,function Used to quantitatively describe the midpoint of a point cloud neighborhood. Relative to target point offset vector , with core point The degree of linear correlation in 3D spatial location, `<h>` is a hyperparameter used to control the effect of kernel point distance; skip connections are added to the AKPConv block to preserve input features; for skip connections, It is an optional max pooling operation, when D in When it is 2D, it will be used, where D represents the size of the input feature dimension. in This indicates the size of the input feature dimension of the AKPCOnv module, and this operation can be represented as: Finally, the output features of AKPConv It can be represented as: Here, ReLU(x) = max(0, x) represents the activation function.

2. The semantic segmentation method for airborne LiDAR urban point clouds using a recursive residual dual-attention kernel point convolutional network according to claim 1, characterized in that: To accumulate and aggregate local features and generate diverse features, a recurrent residual kernel point attention module (RRKA) is composed of a recurrent point convolutional block (RPConv) consisting of AKPConv and a single hidden layer of multilayer perceptron (MLP). The recurrent point convolutional (RPConv) block is mainly used to learn the accumulated neighborhood features of the point cloud, and its formula is expressed as follows: in Let RPConv be the input for layer l, recursively executed T times, where t = 1, 2, 3, ..., T; AKPConv() represents the attention kernel convolution operation. For RPConv's input at time t, This is the output of the (t-1)th recursion. To ensure the computational efficiency of RPConv, an MLP is first used to compress the feature dimension before using RPConv, and then the MLP is used to restore the feature dimension. Therefore, a compressed loop block (CRB) based on RPConv and MLP is constructed to improve computational efficiency, and its formula is as follows: Where W1 and W2 are the learnable parameters of the MLP, PRConv() is the recursive point convolution operation, and BN() is the batch normalization operation. For the input point cloud features, a Residual Recursive Kernel Attention (RRKA) module with residual connections was developed based on CRB. This module can effectively perform repeated operations on local features, enhancing the diversity representation capability of point cloud features. The RRKA output can be expressed as: Where W0 represents the weight learning parameters of the MLP. The input point cloud features are CRB(·), which is the CRB operation.

3. The semantic segmentation method for airborne LiDAR urban point clouds using a recursive residual dual-attention kernel point convolutional network according to claim 2, characterized in that: After five layers of the above encoding modules, the decoding operation is performed. After concatenating the low-level and high-level features, the feature tensor is processed using a Global-Local Channel Attention (GLCA) module. First, a fully connected approach is used to fuse the channel information of low-level and high-level features in the global space. The formula is as follows: as follows: The input to this module is... N and C are the number of upsampled point clouds and the dimension of features, respectively, and W... g To fuse low-level and high-level features, fully connected weights are used; in local feature computation, average pooling and one-dimensional convolution are used to obtain attention weights. in, For the average pooling operation of the channel, W k conv These are learnable local one-dimensional kernel convolution weights with a size of k=5, expressed by the following formula: Based on attention weight The following local attention features can be obtained. The formula is as follows: in, This represents the element-wise multiplication, and is the output of the Global-Local Cross-Layer Information Interaction (GLCA) module. Represented as: 。 4. The semantic segmentation method for airborne LiDAR urban point clouds using a recursive residual dual-attention kernel point convolutional network according to claim 3, characterized in that: The semantic segmentation result obtained by passing the feature tensor through two fully connected layers and the sigmoid activation function is then used. To address the imbalance problem, a focus loss function is introduced, which is set as follows: Here, λ is typically set to 2. This represents the category weight parameter, where N represents the number of point clouds. Let represent the probability that the j-th sample belongs to category c; optimize the recursive residual dual-attention kernel convolutional network using stochastic gradient descent based on the focus loss function, and obtain the trained recursive residual dual-attention kernel convolutional network after training; use the trained recursive residual dual-attention kernel convolutional network to discriminate the input test samples and output the semantic segmentation result.

5. An airborne LiDAR urban point cloud semantic segmentation system based on a recursive residual dual-attention kernel point convolutional network, characterized in that: The system employs the method described in any one of claims 1 to 4.

Images