Complex terrain point cloud segmentation method based on multi-scale space Transform network

Abstract

The invention relates to the technical field of computer vision, and discloses a complex terrain point cloud segmentation method based on a multi-scale space Transform network, and the method comprises the following steps: obtaining original ALS point cloud data; performing data preprocessing operation on the original ALS point cloud data, and constructing a BCE geometric enhancement feature based on a point neighborhood; the method comprises the following steps: constructing a Point-SCT network based on a multi-scale space Convolution-Swin Transform, and constructing the Point-SCT network based on the multi-scale space Convolution-Swin Transform; a training task is carried out on Point-SCT by using the ALS point cloud data, and ground points are segmented from the ALS point cloud data. The model designed by the invention can effectively combine the relation modeling of local feature aggregation and global context, and meanwhile, the heavy calculation caused by applying global attention to the whole point cloud is avoided. Experiments prove that in a ground point extraction task oriented to a large-scale remote sensing scene with complex terrains and vegetation environments, the Point-SCT can robustly segment the ground point cloud, and the precision and generalization both reach the most advanced level.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a complex terrain point cloud segmentation method based on a multi-scale spatial Transformer network. Background Art

[0002] In large-scale remote sensing scenarios with complex terrain, geological structures and dense vegetation, the complexity of the natural environment leads to frequent disasters. In order to improve prevention and control, it is particularly important to continuously monitor disaster-prone areas and conduct accurate pre-disaster risk assessment and post-disaster damage assessment. However, such environments often pose high challenges to conventional ground measurement and data acquisition methods.

[0003] As an active remote sensing technology, airborne laser scanning (ALS) can continuously carry out large-scale dynamic three-dimensional observation of the ground. The collection of its dense spatial point cloud has the characteristics of dynamic flexibility, high resolution and accuracy, and strong penetration of vegetation cover, and provides rich spatial attribute information, which can be effectively applied to disaster monitoring, mining subsidence and landslide detection and other activities. The above applications are mainly realized by completely extracting discrete ground points in the ALS point cloud through the point cloud filtering method, and constructing related fine digital models, and then studying and managing the spatial terrain change laws and natural resources of the target area in the same space and different time dimensions, so as to effectively prevent geological disasters and risk assessment. Therefore, ground point cloud filtering has become a prerequisite for ALS point cloud data preprocessing and related applications.

[0004] However, in such complex large-scale scenes, there is a complex geometric topological structure between ground points (GR) and non-ground points (NG), with large out-of-class differences and similarities between classes. This makes data compression-based methods challenging in the task of fine separation of GR and NG, and the requirements for the automation, efficiency and accuracy of filtering methods are gradually increasing.

[0005] In the past research on ground point filtering methods, traditional filtering methods can be divided into three types in terms of the geometric properties of point clouds and filtering rules: slope-based, surface-based, and morphological methods. When such methods are applied to relatively smooth or flat terrain, they will obtain relatively ideal filtering results, but there are problems such as failure to extract objects and features in complex terrains with steep slopes, gullies, and dense vegetation. In recent years, deep learning (DL) has developed exponentially in the past few years, especially in the field of computer vision. The core advantage of DL models is that they can directly learn the laws and key features related to specific tasks from raw data, simplify the complex feature design work in the early stage, emphasize the focus on multi-scale and multi-dimensional feature information, and show significant advantages in accuracy and performance. However, most visual models based on the convolutional neural network (CNN) architecture are designed to process regularized images, grids, or voxelized data. Since point cloud data is disordered, density-unbalanced, and unstructured, point cloud data needs to be appropriately compressed and regularized. This also leads to the point compression-based approach losing both the elevation semantic information and spatial geometric structure characteristics of the point cloud.

[0006] To address the above challenges, many studies have proposed a variety of point-based deep learning processing methods for point clouds, which can be divided into CNN-based and Transformer-based methods according to the attention capabilities of local and global learning.

[0007] Point CNN-based methods use multiple or shared point convolution operators to directly perform one-dimensional operations on points, and then classify or segment the output points. This type of method shows broader prospects in the expansion of small target segmentation or small-scale application scenarios due to its end-to-end concise, efficient and scalable network architecture. However, most CNN-based methods focus on aggregating local features and fail to explicitly model global dependencies in multi-level features. This limitation leads to poor performance in capturing global contextual relationships on larger-scale data, and therefore cannot achieve ideal results when processing large-scale outdoor point clouds.

[0008] The core of the visual Transformer method is to use self-attention that focuses on global dependencies to perform a set of operations on the sequence and cardinality invariance of the input elements, and to provide input for subsequent processing modules by calculating the similarity between any embedded vectors and weighted summing them, thereby establishing connections with all lower-level input vectors. This process enables the model to effectively establish long-distance position and feature information associations, which is crucial for point cloud classification and segmentation tasks in large-scale scenes, but it also ignores the spatial detail features of dense point areas, and a large number of linear projection modules increase the model's computational complexity and memory resource requirements.

[0009] In addition, the superposition of point clouds will inevitably introduce a lot of noise when obtaining RGB, and enhance the feature difference between the point and the surrounding neighborhood.

[0010] We designed a feature combination of BCE as input features based on the geometric relationship of point neighborhoods, thereby enhancing the difference of surface geometric features and highlighting the outer boundaries of the class. Summary of the invention

[0011] The purpose of the present invention is to solve the problems existing in the prior art and to propose a complex terrain point cloud segmentation method based on a multi-scale spatial Transformer network.

[0012] In order to achieve the above object, the present invention adopts the following technical solutions:

[0013] A complex terrain point cloud segmentation method based on a multi-scale spatial Transformer network includes the following steps:

[0014] Get the original ALS point cloud data;

[0015] Perform data preprocessing on the original ALS point cloud data and construct BCE geometric enhancement features based on point neighborhood;

[0016] Construct a Point-SCT network based on multi-scale spatial Convolution-SwinTransformer;

[0017] Use ALS point cloud data to train Point-SCT and segment ground points from ALS point cloud data.

[0018] Preferably, the specific process of performing data preprocessing operation on the original ALS point cloud data is:

[0019] Remove noise, spots or abnormal points in the original point cloud;

[0020] Get the data of the number of ground points and the number of non-ground points in the profile buffer point cloud data;

[0021] Divide the entire terrain into regular blocks;

[0022] Add BCE features to the points in the regular block instead of RGB as the input of the model.

[0023] Preferably, the specific process of constructing the Point-SCT network based on the multi-scale spatial Convolution-Swin Transformer is: constructing the Point-SCT network into a multi-level hierarchical structure, and combining it with the spatial Convolution-Swin Transformer, the Point-SCT network adopts an encoder-decoder framework design, and the encoder integrates an MLFA module, random sampling and SSwin Transformer module; the decoder contains Layer Normalization (LN), Linear and Interpolation operation modules, the decoder and the encoder adopt a cross-layer link architecture, and are cascaded with the feature output points corresponding to the encoder.

[0024] Preferably, the MLFA module comprises:

[0025] Local geometry structure encoding module (LGSE module), used to fuse multi-dimensional spatial feature vectors with different structural information;

[0026] The dilute pooling module (DP module), used as an attention neuron module based on weakly supervised learning, automatically learns potential local geometric features for vector weight allocation, and realizes effective dilution of high-density features by constraining feature vectors with weights.

[0027] Preferably, the SSwin Transforme module is based on the self-attention method of the dynamic spatial sliding window. On the basis of realizing local feature aggregation in the MLFA module, the SSwin Transformer constructs long-range spatial information connections between local rich geometric features to realize the learning of global point geometric features.

[0028] Preferably, the step of performing data preprocessing on the original ALS point cloud data further includes:

[0029] The data with added BCE features are divided into training set, validation set and test set. The training set is used to train the model, while the validation set and test set are used to verify and test the model performance.

[0030] Preferably, the data with added BCE features is divided into a training set, a validation set and a test set in a ratio of 7:2:1.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. This paper proposes a point-based multi-scale spatial Convolution–Swin Transformer (Point-SCT) method, which can effectively combine local feature aggregation and global context relationship modeling, while avoiding the heavy calculation caused by applying global attention to the entire point cloud. Experiments have shown that Point-SCT can robustly segment ground point clouds in large-scale remote sensing scenes with complex terrain and vegetation environments, and its accuracy and generalization have reached the most advanced level.

[0033] 2. The present invention proposes a new MLFA mechanism, which embeds the LGSE and DP modules to achieve high-density aggregation and adaptive optimization retention of point neighborhood detail features.

[0034] 3. The present invention proposes a combination of enhanced features BCE suitable for the model structure design to replace the RGB input, which solves the problem of inevitably introducing a large amount of noise when obtaining RGB by overlaying point cloud images, and enhances the difference in characteristics between ground points and surrounding low vegetation. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is the framework diagram of the Point-SCT model proposed in the present invention;

[0036] Figure 2 This is a framework diagram of the MLFA module proposed in the present invention;

[0037] Figure 3 This is a point-based spatial Swin Transformer framework diagram introduced by the present invention;

[0038] Figure 4 The results of ground point segmentation on the experimental test set data are compared between the present invention and the current advanced point-based and Transformer-based methods. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0040] A complex terrain point cloud segmentation method based on a multi-scale spatial Transformer network includes the following steps:

[0041] S1. Perform data preprocessing on the original ALS large-scale field point cloud data and construct BCE geometric enhancement features based on point neighborhood;

[0042] S2, constructing a Point-SCT network based on multi-scale spatial Convolution-Swin Transformer;

[0043] S3, use the data in S1 to train Point-SCT and segment ground points from large-scale ALS point clouds;

[0044] Furthermore, the specific steps in S1 are as follows:

[0045] S11. First, use point cloud processing software to manually remove noise, spots or abnormal points in the original point cloud;

[0046] S12. Through the section buffer, the point cloud is manually classified and labeled, and the number of ground points is n. GR and the number of non-ground points n NG data.

[0047] S13. In order to reduce the memory and computing resources required for a single forward pass of the network and adapt the learning and stability of the local features of the network receptive field, we divide the entire terrain into n regular blocks. Each regular block has P = {p1, p2, ..., p i} i＝N , where N is the number of points in the rule block;

[0048] S14. To avoid the inevitable introduction of a large amount of noise and calibration errors when ALS point cloud overlay and multi-source image registration are used to obtain RGB, and to enhance the spatial geometric feature differences between points and their neighborhoods, the points in these regular blocks are added with BCE features instead of RGB as the input of the model. The specific steps for calculating each feature are as follows:

[0049] S141, Boundary Detector: Point p i =(x i ,y i , z i ) as the query point, first use the k-dimensional (kd) tree to obtain the k nearest neighbor points P k ={p1, p2, ..., p j} j＝k The position and distance from the center point are then calculated by using the Gaussian kernel function G to obtain the relationship between the average inverse density value and k to obtain the centroid C i Finally, the distance d between its centroid and mass center is calculated to identify the edge area R i ;

[0050]

[0051] R i =min||Ci -p i ||

[0052] S142, Surface Curvature: First, through p i To query k nearest neighbor points P k , and both belong to Euclidean space finite subset of , and calculate the covariance V of these neighboring points i . Secondly, the covariance needs to be decomposed into eigenvalues ​​to obtain eigenvalues ​​λ1, λ2, λ3, and λ1<λ2<λ3. Finally, the eigenvalues ​​are used to calculate the curvature δ i .

[0053]

[0054] S143, Average Elevation: By comparing p i Neighbor point set P k The point elevations in the grid are averaged to obtain the neighborhood average elevation E. i .

[0055]

[0056] S15. Calculate the enhanced features of the points in each regular block after classification under S14, and construct the input representation of the model, that is, (x, y, z, C i , δ i , E i ), and the data is divided into training set, validation set and test set in the ratio of 7:2:1.

[0057] S16, the test set data is a challenging area with complex geometry and high vegetation coverage.

[0058] Furthermore, the specific steps of S2 are as follows:

[0059] S21 and Point-SCT are multi-level hierarchical structures, combined with spatial Convolution-Swin Transformer, and designed in an encoder-decoder framework. The encoder integrates MLFA modules, random sampling, and SSwin Transformer modules; the decoder contains Layer Normalization (LN), Linear, and lnterpolation operation modules. The decoder and encoder adopt an architecture similar to FPN cross-layer linking, and are cascaded with the feature output points corresponding to the encoder to achieve multi-scale global feature fusion.

[0060] S22. First, the Stem Projection module normalizes the input point cloud and maps it into a unified high-dimensional feature space, using nonlinear transformation to enhance feature expression and obtain a high-dimensional feature vector (N×D1), where N is the number of input points.

[0061] S23. Secondly, the high-dimensional projection feature vector is sent to the encoder stage (Encoder Stage1, EncoderStage2, Encoder Stage3) and the decoder stage (Decoder Stage1, Decoder Stage2, DecoderStage3). The specific operations of each stage are:

[0062] S231, in the MLFA module of Encoder Stage1, given a high-dimensional space input point set P = {p1, p2, ..., p i} i＝N ,in Each point carries its enhanced feature BCE, forming a feature set F = {f1, f2, ..., f i} i＝N , In detail:

[0063] (1) First, we define K-Nearest Neighbors (KNN) based on Euclidean distance, with the effective radius r as the neighborhood radius, and search for neighboring points point by point for P to obtain the center point p i The spatial index corresponding to the k neighborhood points.

[0064] (2) In the LGSE module, the neighborhood aggregation function g is used to capture the center point neighborhood coordinates and features and Calculate the center point Its neighboring points The relative spatial coordinates and Euclidean distance of the center point are used to construct a position encoding vector, which contains the spatial geometric structure reference information of the center point and the neighborhood. Then, it is embedded in the center point as a feature vector to obtain

[0065]

[0066] (3) On this basis, we transform the feature vector with strong geometric shape information into Embed the center point and get This prevents the gradual weakening or loss of features of sparse point clouds during multi-stage encoding.

[0067] (4) Secondly, As the given input of the dilution module, its shape is set to (i, j). Through the weakly supervised neuron attention function Attn(·), the center point feature vector is mapped to the attention weight, and the vector set Calculate and generate the feature vector attention distribution matrix W ij .

[0068]

[0069]

[0070] (5) Furthermore, in order to pass the constraints of the current weight on the features to the neurons in the next encoding stage, the feature dilution matrix is ​​processed by the Conv function, and high scores are given to feature units with strong geometric expressions, while low-score feature vectors are ignored. At the same time, the activation unit ReLU is used in the Conv layer to correct and amplify the diluted feature attention map, and the diluted feature set is calculated. in And passed as input to the next processing unit.

[0071]

[0072] S232, the random sampling layer, uses a multi-level sampling rate of 1 / 4 to quickly reduce the density of high-dimensional feature vectors, so that the encoder achieves an effective balance between sampling efficiency and retention of local feature details.

[0073] S233, in the SSwin-Transformer module of Encoder Stage1, dynamic spatial patches are used to divide the input point spatial feature vector, and the unique query point corresponding to each window is used to focus on all points in the window. All windows independently perform shared multi-head self-attention, and interact with the output features between different spatial windows. In detail:

[0074] (1) Taking the sampling point feature vector set As input, since the number of points in each window is different, we will Represents the number of feature points in the i-th window, and given the number of heads in the self-attention n , each head dimension is H d The query Q, key K and value V matrices are obtained through a learnable linear projection module, where d a and d b are the vector dimensions of query, key and value respectively.

[0075]

[0076] (2) Q, K, and V are fed into the sparse vector self-attention Offset-Attention (OA) operation with an offset operator, and attention weights are calculated only between Q and K that may be correlated. According to the window and aggregate feature contribution, dot product is performed with the value matrix V to obtain the aggregate feature

[0077]

[0078] (3) Global self-attention sharpening feature Feed it into the LBR (LBR module embeds Linear, Batch Norm and ReLU) network and apply the global input features point by point Further projection to obtain the output feature map

[0079]

[0080] S234. In Decoder Stage 1, the multi-scale feature vectors extracted by the above modules are aggregated layer by layer through LN, Linear and Interpolation operation modules to comprehensively capture a wider range of fine contextual spatial semantic information.

[0081] S235, the subsequent Encoder Stage, and the 2nd and 3rd stages of the Decoder Stage are similar to the operations of the first stage. The cardinality of the point feature vector set generated by each stage in the encoder is [N, 1 / 16N, 1 / 64N, 1 / 256N], and the decoder is the opposite.

[0082] S24. Finally, the SegHead layer performs classification prediction on the upsampled spatial vector.

[0083] Furthermore, the specific steps of S3 are as follows:

[0084] Experiments based on multi-scale spatial features Point-SCT are carried out using the training set and validation set data with enhanced features BCE in S1.

[0085] The specific implementation process is as follows:

[0086] The multi-scale spatial Point-SCT model architecture of this embodiment can be seen Figure 1 The convolution-based multi-scale local feature aggregation module MLFA framework can be seen as follows Figure 2 The specific operation process is as follows:

[0087] 1. Ground point cloud extraction of the experimental area

[0088] The point cloud data of the area was obtained by using a laser scanning system mounted on a drone at a flight altitude of 120m. In the early stage of data preprocessing, in order to ensure the quality of the training set data, we first used LiDAR360 point cloud processing software to preprocess the point cloud, and then manually finely classified and annotated it through the profile moving buffer, obtaining data with 1,813,506 ground points and 1,933,801 non-ground points. Secondly, in order to reduce the memory and computing resources required for a single forward pass of the network, and at the same time adapt to the learning and stability of the network receptive field in local features, we divided the entire terrain into 100 50m×50m square blocks.

[0089] 2. Perform data preprocessing on the original ALS large-scale field point cloud data and construct BCE geometric enhancement features based on point neighborhood

[0090] According to the BCE calculation method classification, the point cloud is enhanced with additional features, and the calculation methods are:

[0091] Boundary Detector: Point p i =(x i ,y i , z i ) as the query point, first obtain k neighboring points P k ={p1, p2, ..., p j} j＝k The position and distance from the center point are then calculated using the Gaussian kernel function G to obtain the centroid C i Finally, the edge area R is identified by the distance d between the centroid and the centroid i ;

[0092] Surface Curvature: First, through p i To query k nearest neighbor points P k , calculate the covariance V of these adjacent points i Secondly, the covariance is decomposed by eigenvalues ​​and the curvature δ is calculated using the eigenvalues. i .

[0093] Average Elevation: By i Neighbor point set P k The point elevations in the grid are averaged to obtain the neighborhood average elevation E. i .

[0094] Finally, we construct the input dimensions required for the model, namely (x, y, z, C i , δ i , E i ), and the data is divided into training set, validation set and test set in the ratio of 7:2:1.

[0095] 3. Construct a Point-SCT network based on multi-scale spatial Convolution-Swin Transformer;

[0096] like Figure 1 As shown in the figure. Assuming a single aircraft target as an example, the mechanism of action of each module in the model and the change of feature scale are demonstrated. The framework directly processes the original point cloud containing coordinates (xyz) and derived geometric features (Boundary Detector, Curvature and Average Elevation). The model framework adopts an encoder-decoder structure, in which the encoder integrates a multi-scale local feature aggregation mechanism, including MLFA module, random sampling layer and spatial SwinTransformer (SSwin Transformer) module.

[0097] First, the Stem Projection module normalizes the input point cloud and maps it into a unified high-dimensional feature space, using nonlinear transformation to enhance the feature expression. Subsequently, the projected features are fed into the encoder, and the local and global spatial features of the point cloud are asynchronously modeled in the same stage through MLFA and SSwinTransformer. In the MLFA mechanism, we encode and embed the high-dimensional neighborhood geometric features of the shallow input points to aggregate the shallow high-dimensional local feature semantic information, which helps the model converge faster. Specifically, Figure 2 As shown:

[0098] LGSE: The local geometric structure encoding module fuses multi-dimensional spatial feature vectors with different structural information, significantly enhancing the nonlinear mapping capability of the feature space, helping neurons to accurately capture more subtle local geometric changes and their complex spatial relationships.

[0099] DP: The dilution pooling module becomes an attention neuron module based on weakly supervised learning, which automatically learns the potential local geometric features for vector weight allocation and realizes the effective dilution of high-density features by constraining the feature vectors with weights.

[0100] SSwin Transformer: SSwin Transformer is a self-attention method based on dynamic spatial sliding windows. On the basis of local feature aggregation achieved in the previous module, it further builds long-range spatial information connections between local rich geometric features to achieve learning of global point geometric features. This module consists of a multi-head self-attention module with an offset operator and an LBR module, such as Figure 3 shown.

[0101] 4. The training set and validation set data of the ALS point cloud data obtained in the above steps are used to train Point-SCT. In detail, we implemented the construction of the Point-SCT model based on the Pytorch framework, and trained the data set for 300 epochs on the NVIDIARTX 3090GPU. Among them, the optimizer selects AdamW to adjust the parameter weights, with a weight decay of 0.01 and a momentum of 0.9. The initial learning rate is set to 0.001, and the cosine annealing method is used to periodically adjust the learning rate. Since the outdoor point cloud we provide is large in volume and the point density is uneven, the Batch and Grid Size are modulated to 2 and 0.02. In order to ensure that each spatial sampling window always has a certain number of points in the multi-level encoding stage, and to control the operation within the effective video memory and maximize the performance, we finally modulated the maximum input point limit to 8000, the k value of the nearest neighbor search to 32, the multi-level sampling rate to 0.25, and the spatial window size to 4. In addition, the depth of the encoding sampling stage of the model framework is [2, 4, 2], and given the number of input point clouds N, the cardinality of the point feature vector set generated by each stage of the encoder is [N, 1 / 16N, 1 / 64N, 1 / 256N].

[0102] To further evaluate the performance of Point-SCT, we combined three other mainstream point convolution-based models, two point Transformer-based models, and Point-SCT, selected challenging and representative terrain in the region as the test set, and conducted a qualitative analysis and comparison of the ground point cloud extraction performance of the above models. Figure 4 shown.

[0103] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A complex terrain point cloud segmentation method based on a multi-scale spatial Transformer network, characterized by: The following steps are involved: Get the original ALS point cloud data; Perform data preprocessing on the original ALS point cloud data and construct BCE geometric enhancement features based on point neighborhood; Construct a Point-SCT network based on multi-scale spatial Convolution-Swin Transformer; Use ALS point cloud data to train Point-SCT and segment ground points from ALS point cloud data.

2. According to claim 1, a complex terrain point cloud segmentation method based on a multi-scale spatial Transformer network is characterized in that: The specific process of performing data preprocessing operation on the original ALS point cloud data is as follows: Remove noise, spots or abnormal points in the original point cloud; Get the data of the number of ground points and the number of non-ground points in the profile buffer point cloud data; Divide the entire terrain into regular blocks; Add BCE features to the points in the regular block instead of RGB as the input of the model.

3. The complex terrain point cloud segmentation method based on a multi-scale spatial Transformer network according to claim 2, characterized in that: The specific process of constructing the Point-SCT network based on the multi-scale spatial Convolution-Swin Transformer is as follows: the Point-SCT network is constructed as a multi-level hierarchical structure and combined with the spatial Convolution–Swin Transformer. The Point-SCT network adopts an encoder-decoder framework design, and the encoder integrates an MLFA module, random sampling, and an SSwin Transformer module; the decoder contains Layer Normalization (LN), Linear and Interpolation operation modules. The decoder and the encoder adopt a cross-layer link architecture and are cascaded with the feature output points corresponding to the encoder.

4. The complex terrain point cloud segmentation method based on a multi-scale spatial Transformer network according to claim 3, characterized in that: The MLFA module includes: Local geometry structure encoding module (LGSE module), used to fuse multi-dimensional spatial feature vectors with different structural information; The dilute pooling module (DP module), used as an attention neuron module based on weakly supervised learning, automatically learns potential local geometric features for vector weight allocation, and realizes effective dilution of high-density features by constraining feature vectors with weights.

5. The complex terrain point cloud segmentation method based on a multi-scale spatial Transformer network according to claim 4, characterized in that: The SSwin Transforme module is based on the self-attention method of the dynamic spatial sliding window. On the basis of realizing local feature aggregation in the MLFA module, the SSwin Transformer constructs long-range spatial information connections between local rich geometric features to realize the learning of global point geometric features.

6. The complex terrain point cloud segmentation method based on a multi-scale spatial Transformer network according to claim 5, characterized in that: The step of performing data preprocessing on the original ALS point cloud data also includes: The data with added BCE features are divided into training set, validation set and test set. The training set is used to train the model, while the validation set and test set are used to verify and test the model performance.

7. The complex terrain point cloud segmentation method based on a multi-scale spatial Transformer network according to claim 6, characterized in that: The data with added BCE features are divided into a training set, a validation set, and a test set in a ratio of 7:2:1.

Images