A method for optimizing boundary semantic segmentation of a three-dimensional point cloud

Abstract

The application discloses a three-dimensional point cloud semantic segmentation method with optimized boundaries, belongs to the field of automatic driving, computer vision and deep learning, and particularly relates to a three-dimensional point cloud semantic segmentation method with optimized boundaries. The segmentation method comprises the following steps: step one, three-dimensional point cloud preprocessing; step two, an encoding layer is composed of a local feature fusion module and a random sampling module; step three, a decoding layer: taking the feature encoding obtained in step two as input, sampling the number of three-dimensional point clouds, and performing feature dimension and contrast boundary learning feature dimension, encoding layer reverse order jump layer into shared MLP fusion features; after completing feature fusion, performing full connection dimension lifting, then performing random inactivation operation, finally reducing the dimension to the same dimension of the semantic label number through shared MLP, and finally outputting a three-dimensional point cloud semantic segmentation result; and step four, contrast boundary optimization. The application aims to improve the precision of large-scale three-dimensional point cloud semantic segmentation.

Description

Technical Field

[0001] This invention belongs to the fields of autonomous driving, computer vision, and deep learning, and specifically relates to a method for semantic segmentation of 3D point clouds with optimized boundaries. Background Technology

[0002] 3D point clouds are three-dimensional data points attached to the surface of an object in three-dimensional space, generated through various methods such as radar, image processing, and data synthesis. Based on the scale of the generated 3D point clouds, they can be divided into object-level 3D point clouds and scene-level 3D point clouds. Object-level 3D point clouds represent a single object in three-dimensional space as three-dimensional data points. Scene-level 3D point clouds represent multiple different objects existing within a certain spatial range, also represented as three-dimensional data points. Scene-level 3D point clouds are further divided into indoor-level 3D point clouds and outdoor large-scale 3D point clouds based on their size. Indoor-level 3D point clouds are obtained through image generation, small-scale LiDAR, and data synthesis of indoor scenes. Outdoor large-scale point clouds are obtained through vehicle-mounted or airborne LiDAR, representing 3D point clouds covering a larger spatial area.

[0003] 3D point cloud segmentation is a process that involves processing acquired 3D point cloud data. It separates 3D point clouds based on their different features and uses different annotations to represent different regions. 3D point cloud segmentation can be further divided into partial segmentation (fine-level), instance segmentation (object-level), and semantic segmentation (scene-level).

[0004] 3D point cloud semantic segmentation involves taking data points within a scene and their inherent feature information as input, performing a series of processing steps, and finally assigning a semantic label to each 3D data point. 3D point cloud semantic segmentation is mainly used in autonomous driving, robotics, drones, railway track inspection, 3D modeling, and the entertainment industry. The performance requirements for 3D point cloud semantic segmentation vary depending on the application scenario. In indoor scenes, the primary performance requirement is segmentation accuracy; for large-scale outdoor scenes, while maintaining accuracy, it is also necessary to minimize the semantic segmentation processing time.

[0005] 3D point cloud semantic segmentation can be broadly categorized into two types based on the different feature extraction methods. The first type uses manually generated features and classifiers for semantic segmentation of 3D point clouds. In the first stage, these methods employ clustering or classification, introducing manually generated features to detect and identify objects within a region. This allows for the classification of scattered points, linear points, and planar points in the data, directly assigning semantic labels. In the second stage, to consider the relationships within the local context, Markov Random Fields (MRF) are introduced to model the context of the data points. The second type uses deep learning-based methods for 3D point cloud semantic segmentation. These methods are mainly divided into three types: multi-view methods, 3D point cloud voxelization methods, and methods based on the original 3D point cloud. Multi-view methods involve capturing multiple images in 3D space or projecting a 3D model onto a 2D image, then extracting features through convolutional and pooling layers, and finally aggregating the features before inputting them into the network to return the semantically segmented 3D point cloud result. This method is not suitable for large-scale spatial scenes, and the segmentation accuracy is related to the viewpoint selection. The 3D point cloud voxelization method converts the 3D point cloud into a 3D mesh and then performs 3D convolution to extract features. However, the problem is that there may be multiple semantic labels within a voxel, and the 3D point cloud semantic segmentation performance is severely affected by the boundaries between each object. This causes the network model to be unable to separate the boundaries of different objects within the same voxel, which in turn affects the accuracy of 3D point cloud semantic segmentation.

[0006] Methods based on raw 3D point clouds directly extract feature information from the input 3D point cloud for segmentation, thus minimizing the loss of original feature information. They typically perform well in indoor-scale 3D point cloud semantic segmentation. However, performing semantic segmentation of 3D point clouds at different scales depends on the structure of the network model. Some existing network models can only perform 3D point cloud semantic segmentation for indoor scenes. This is because 3D point clouds are unordered, making it difficult for methods based on raw 3D point clouds to directly extract the relationships between data points; preprocessing is required before model training.

[0007] Many network models based on raw 3D point clouds, capable of handling scenes of varying scales, exhibit poor segmentation performance in local areas, particularly at boundaries where they struggle to separate semantic labels for different objects. This is primarily due to issues such as occlusion, sparsity, and discontinuity during 3D point cloud acquisition, which prevent the network model from clearly learning and separating boundary points during training. Therefore, different methods yield varying results for 3D point clouds of different scales. Summary of the Invention

[0008] This invention provides a method for optimizing the boundary of 3D point cloud semantic segmentation. Its purpose is to improve the accuracy of large-scale 3D point cloud semantic segmentation, solve the problems that some deep learning-based 3D point cloud semantic segmentation methods cannot handle large-scale 3D point clouds, and the low time efficiency of other semantic segmentation methods, while accelerating the semantic segmentation speed.

[0009] Technical solution:

[0010] A semantic segmentation method for 3D point clouds with optimized boundaries, characterized in that: the segmentation method includes:

[0011] Step 1: 3D point cloud preprocessing: 3D point cloud preprocessing is performed only once before the encoding layer. The feature dimension of the input 3D point cloud is set to 8 dimensions by using a shared multilayer perceptron (MLP).

[0012] Step 2: The encoding layer consists of a local feature fusion module and a random sampling module; the local feature fusion module consists of two consecutive sets of local spatial encoding and an attention pooling part;

[0013] 1. Local Spatial Coding: Local feature coding consists of three parts: finding points in local regions, encoding the relative positions of points, and enhancing point features;

[0014] 2. Attention pooling;

[0015] Attention scores for each local feature point are learned using a function consisting of Conv2d and softmax. Then, the features of each local region point are multiplied by each attention score, and the sum is fed into a shared MLP composed of Batch Normalization (BN) and LeakyReLU. This yields a feature vector f. i ;

[0016] Step 3: Decoding Layer: The feature encoding obtained in Step 2 is used as the input of the decoding layer. The number of 3D point clouds is sampled by the nearest neighbor difference. The feature dimension is learned by learning the feature dimension for the contrast boundary. The encoding layer is connected in reverse skip layer and fed into the shared MLP to fuse features. After feature fusion, the dimension is increased by full connection, then random deactivation is performed. Finally, the dimension is reduced to the same dimension as the number of semantic labels through the shared MLP, and the final output is the 3D point cloud semantic segmentation result.

[0017] Step 4: Contrast Boundary Optimization: Contrast boundary optimization employs a contrast boundary learning framework, which mainly consists of two modules:

[0018] 1. Boundary sub-scene mining: Extract ground truth labels from 3D point cloud sampling at different scales. The number of sampled point clouds in each layer is the same as the random sampling result of the corresponding coding layer. As the scale of the 3D point cloud decreases, the specific labels of the sampled points at the boundary may become uncertain. Boundary sub-scene mining in the CBL framework is used to determine the semantic labels of each layer of point cloud.

[0019] 2. Contrastive Boundary Learning: During model training, the boundary points of each layer are compared with the semantic labels of the corresponding layers obtained from the boundary sub-scenes. Contrastive learning uses InfoNCE loss and its generalization to optimize the objective function defining the boundary points, so that the final results of the boundary points are closer to the neighboring points from the same category. Finally, the 3D point cloud features of the corresponding decoding layer are optimized.

[0020] In step two, (1) local spatial encoding, points in local regions are found: local regions controlled by scene scale are based on Euclidean distance d, and nearest neighbor points are collected using the KNN algorithm, with the center point of each local region as i. th Sort the data; the nearest points within each local region are denoted by K, which is... Where the subscript i represents the index of the center point, and the superscript k represents the index of the points within the local area of ​​the center point i. A KD-tree index is constructed using KNN.

[0021] (2) Relative Position Encoding: To enable the entire network structure to better understand local features, the encoding method involves concatenating four parts: the coordinates of the center point of each local region, the coordinates of neighboring points within the local region, the difference between the coordinates of neighboring points and the center point of the local region, and the Euclidean distance of neighboring points relative to the center point of the local region. This is then processed using an MLP (Multi-Level Processing) algorithm. This method encodes the relative position of each local region... of A point is associated with the center point of this local region to improve the local performance of the overall network structure;

[0022] (3) Point feature enhancement: The features of the K nearest neighbor points in each local region are F = {f i 1 …f i k The relative positional codes of nearby points are concatenated with the original features of that point. A new feature is constructed from the positional codes and the original features, and the concatenated feature is the result. The subscript i represents the index of the center point, and the superscript k represents the index of the points in the neighborhood of the center point i.

[0023] This invention has the following advantages and positive effects:

[0024] This invention uses a deep learning-based method for original 3D point cloud sampling. Compared to other deep learning methods, this method has a time complexity of O(1) for 3D point cloud sampling, enabling faster processing of large-scale 3D point cloud tasks. A local feature expansion module is used during local feature extraction to increase the receptive field of each point, thereby expanding the neighborhood. The proposed 6-layer downsampling structure and comparative boundary optimization optimize the boundaries at different scales while sampling 3D point clouds on a large scale, resulting in clearer boundary segmentation of the 3D point cloud semantic segmentation network model and providing effective assistance for tasks such as object detection. Attached Figure Description

[0025] Figure 1 This is an overall structural diagram of a three-dimensional point cloud semantic segmentation method with optimized boundaries according to the present invention;

[0026] Figure 2 This is an explanatory diagram of the 1D feature extraction module CBL1d in the 3D point cloud semantic segmentation method for optimizing boundaries of the present invention;

[0027] Figure 3 This is an illustration of the local feature fusion module in a 3D point cloud semantic segmentation method for optimizing boundaries according to the present invention;

[0028] Figure 4 This invention relates to the 2D feature extraction module CBL2d in a 3D point cloud semantic segmentation method for optimizing boundaries.

[0029] Figure 5 This is an illustration of the local spatial coding module in a three-dimensional point cloud semantic segmentation method for optimizing boundaries according to the present invention;

[0030] Figure 6 This is the attention pooling module in the 3D point cloud semantic segmentation method for optimizing boundaries according to the present invention. Detailed Implementation

[0031] To improve the accuracy of semantic segmentation for large-scale 3D point clouds and address the limitations of some deep learning-based 3D point cloud semantic segmentation methods in handling large-scale 3D point clouds, as well as the low time efficiency of other semantic segmentation methods, this paper proposes a boundary-optimized 3D point cloud semantic segmentation method. This method improves the accuracy and processing efficiency of 3D point cloud semantic segmentation under different scales of 3D point clouds. This method uses random sampling to reduce resource consumption during sampling; it incorporates attention pooling in local aggregation to enhance the receptive field of local regions; and it introduces contrastive boundary learning during network model training, building neighborhoods only in edge regions to reduce resource consumption. The results of the boundary point contrastive learning are then used to optimize the model, improving the network model's boundary handling capability. These methods can solve semantic segmentation problems in different scene levels, improve the boundary accuracy of 3D point cloud semantic segmentation, and enhance the overall processing efficiency of the network model.

[0032] The invention will be further described below with reference to the accompanying drawings:

[0033] A semantic segmentation method for 3D point clouds with optimized boundaries includes 3D point cloud preprocessing, an encoding layer, a decoding layer, and contrastive boundary optimization. The encoding layer includes a local feature fusion module and a random sampling module, and the contrastive boundary optimization includes contrastive boundary learning and boundary sub-scene mining. The method's processing flow is as follows: Figure 1 As shown, the specific steps include the following:

[0034] Step 1: 3D Point Cloud Preprocessing: Because different datasets provide different original feature dimensions (such as R, G, B, illumination, coordinates, number of channels, normal vectors, etc.), preprocessing of the dataset is necessary. A 3D point cloud preprocessing step is performed before the encoding layer. First, the acquired raw 3D point cloud is processed by a shared MLP to extract features from each point. The total number of extracted 3D point clouds is N. A CBL1d module, composed of one-dimensional convolution (Conv1d), batch normalization (Batch Normalization), and leaky linear units (LeakyReLU), is used as the shared MLP for feature extraction. Figure 2 As shown, this makes the output 3D point cloud feature dimension 8.

[0035] Step 2: Encoding layer: such as Figure 1As shown, a point cloud with initial 8-dimensional features is input into six consecutive coding layers, with random sampling between each coding layer. In the first three coding layers, the number of randomly sampled point clouds is 1 / 4 of the number of point clouds in the previous layer, and in the last three coding layers, the number of randomly sampled point clouds is 1 / 2 of the number of point clouds in the previous layer. A local feature fusion module is used between each layer to perform dimensionality upscaling, doubling the feature dimension between each layer, i.e., the dimensionality of the 3D point cloud changes from (8, 16, 32, 64, 128, 256, 512) after batch normalization. After the local feature fusion operation, another random sampling operation is performed at each layer. 25% of the random samples from the first three layers are retained, and 50% are retained from the last three layers, resulting in a total number of 3D point clouds.

[0036] Local feature fusion module: such as Figure 3 As shown, it consists of two local spatial encodings and two attention pooling operations stacked together. First, the 3D point cloud features are fed into the CBL2d module for dimensionality halving. The CBL2d module is composed of 2D convolution (Conv2d), batch normalization, and leaky linear units (LeakyReLU). The CBL2d module is as follows... Figure 4 As shown in the diagram, the local spatial coordinates are then concatenated with the local features, and the result is processed through a local spatial encoding module and attention pooling to obtain a local attention feature. This feature is then concatenated with the local spatial coordinates again, and processed through the local spatial encoding module and attention pooling once more. The resulting local attention feature undergoes dimensionality upscaling via the CB2d module, while the input 3D point cloud features also undergo dimensionality upscaling via the CB2d module. Finally, the two features of the same dimension are summed and then activated using LeakyReLU. The CB2d module consists of two-dimensional convolution (Conv2d) and batch normalization.

[0037] (1) Local spatial coding:

[0038] like Figure 5 As shown. The local feature fusion part first collects local region points within the Euclidean distance range using KNN, based on the 3D point cloud features with coordinate information. The center point of each local region is i. th Sort the data so that the K points within each local region are... And P i ∈χ nThe subscript represents the center point number, and the superscript represents the point number within the local region. The coordinates of the local region's center point, the coordinates of its neighboring points within the local region, the difference between the coordinates of its neighboring points and the center point, the Euclidean distance of its neighboring points relative to the center point, and the estimated normal vector of each point are concatenated and fed into the CBL2d module to obtain the relative position code of the local region, which is r. i k This can be expressed as formula (1):

[0039]

[0040] in This indicates a splicing operation; MLP is a CBL2d multilayer perceptron. This represents the difference between the coordinates of neighboring points within a local region and the coordinates of the center point of that local region. This indicates the extraction of the Euclidean distance from the neighboring points to the center point. This represents the estimate of the normal vector at each point within a local region.

[0041] The characteristics of the K nearest neighbors within each local region are F = {f} i 1 …f i k}, where F represents the feature set of neighboring points within the local region; f i k This represents the features of neighboring points within a local region. The superscript indicates the index of the neighboring point within the local region, and the subscript indicates the index of the local region. The relative position codes of the neighboring points within the local region are concatenated with the features of the neighboring points within that local region to construct a new feature, which is a concatenation of the feature and the point position coordinate codes. in This represents the new set of features of neighboring points within the local region after construction; This represents the new features of neighboring points within a local region after construction. The superscript indicates the index of the neighboring point within the local region, and the subscript indicates the index of the local region. Its formula is (2):

[0042]

[0043] (2). Attention pooling:

[0044] like Figure 6 As shown, first calculate the attention score for each point, and then construct the new features. The attention score for each point is calculated by using the shared weights W as hyperparameters in the attention score calculation formula. The superscript indicates the index of the nearest point within the local area, and the subscript indicates the index of the local area. Its expression formula is (3):

[0045]

[0046] The g() function is a shared function consisting of a two-dimensional convolution and a softmax function.

[0047] The obtained attention score is compared with the newly constructed features. Perform a dot product operation to obtain the attention feature of each point. Sum the attention features of each point in the local area and input them into the CBL2d module as the attention feature of that local region. Its expression formula is (4):

[0048]

[0049] Step 3: Decoding layer: such as Figure 1 As shown, after 6 encoding layers, 10 decoding layers are performed. The first decoding layer concatenates the feature encoding output from the last encoding layer with the features optimized by contrastive boundary learning, and then inputs it into the CBL2d module, which maintains the dimensionality at 512. Decoding layers 2 through 7 employ the following method: first, the KNN algorithm is used to find the nearest neighbor, and then the number of points in the point cloud is upsampled using nearest neighbor interpolation. The number of 3D point clouds is increased by a factor of 2 from layers 2 to 4, and by a factor of 4 from layers 5 to 7, resulting in a total number of 3D point clouds. First, nearest neighbor interpolation is performed, and then shared MLP is performed. After nearest neighbor interpolation is performed in each layer from layer 2 to layer 7, the 3D point cloud features obtained from the previous decoding layer, the 3D point cloud features optimized by the corresponding layer's boundary learning, and the 3D point cloud features of the corresponding number of point clouds in the encoding layer are concatenated and then input into the CBL2d module to reduce the feature dimension to half of the previous layer. Its expression is (5):

[0050]

[0051] Where f i d This represents the 3D point cloud features of the current decoding layer. This represents the features of the upper-level 3D point cloud. f represents the 3D point cloud feature of the encoding layer corresponding to the number of point clouds. i C This represents the 3D point cloud features optimized through boundary learning. Layer 8 of the decoding layer inputs the complete 3D point cloud with 8-dimensional features into the CBL2d module, increasing the dimensionality to 64. Layer 9 of the decoding layer performs random deactivation with a deactivation rate of 50%, then inputs the feature back into the CBL2d module, reducing the 3D point cloud feature dimension to 32. Layer 10 of the decoding layer uses CBL2d to set the dimension to the same number as the semantic labels, finally predicting the semantic segmentation result.

[0052] Step 4: Boundary Optimization Comparison

[0053] Structure as Figure 1 As shown in the figure. This algorithm adopts the CBL 3D point cloud semantic segmentation and contrast boundary framework, which consists of two parts: sub-scene boundary mining and contrast boundary learning.

[0054] 1. Sub-scene boundary mining: The purpose is to find the boundary part of the ground truth in the training process, and the boundary mining structure of each layer can be used as the ground truth of the next layer. The number of sub-scene boundary mining layers is the same as that of the encoding layer, and each sub-scene boundary mining layer shares random sampling with each layer of the encoding layer. The ground truth of the dataset is introduced into the model training process for comparative boundary learning. Boundary points are the points that distinguish different objects. The search method is to find points with different semantic labels from neighboring points in the 3D point cloud. The purpose is to more clearly separate different objects in the end-to-end network model and improve the semantic segmentation accuracy of 3D point cloud. The boundary neighborhood is selected as a sphere with a scale of 0.1% of the overall 3D point cloud, and its expression formula is (6):

[0055]

[0056] Where χ represents the entire 3D point cloud, B represents the set of boundary points, and N... i Indicates x i The set of neighborhood points, x j Indicates that in N i Within the neighborhood of x i Points of semantic label inconsistency, where l represents the semantic label, and B n ∈χ n , where n represents the corresponding coding layer number. The superscript indicates the sampling layer number, and the subscript indicates the order of the points in the neighborhood, which will not be elaborated further.

[0057] Each layer samples the same points as its corresponding randomly sampled points; the first layer directly uses the ground truth labels from the dataset. and The method for determining the truth labels of the boundaries after the second layer is to average the labels of all points in a local region centered on the boundary point of the previous layer. Its expression formula is (7):

[0058]

[0059] Where AVG() represents average pooling. This represents the set of neighborhood points centered at the boundary point in the presampling layer.

[0060] 2. Contrastive Boundary Learning: First, determine the boundary point for contrastive learning according to Formula 6. Then, denote the opposite point in the neighborhood centered on the boundary point as {x}. j ∈N i ∧li =l j}, the negative pair is {x} j ∈N i ∧l i ≠l j}. Contrastive training uses InfoNCE loss (Info Noise Contrastive Estimation loss) for generalization, where the temperature coefficient τ is introduced as a hyperparameter. The aim is to optimize the boundary points to be closer to the opposing points. The boundary features obtained after contrastive learning are then used to optimize the boundary point features in the upsampling stage, thus enhancing the boundary region constraint and enabling the network model to recognize boundary features. The generalization function of InfoNCE loss is expressed as (8):

[0061]

[0062] Where L B f represents the contrast boundary learning loss. i Represents the corresponding boundary point x i Features, f j Indicates with x i Features of points with inconsistent semantic labels, where the subscript k represents all points in a local region, τ is the temperature coefficient as a hyperparameter, and d(.,.) represents the distance measurement result.

[0063] The global loss function calculated based on the entire training process is (9):

[0064]

[0065] L CE λ represents the cross-entropy loss, and λ is the loss weight for learning the boundary conditions.

Claims

1. A semantic segmentation method for 3D point clouds with optimized boundaries, characterized in that: The segmentation method includes: Step 1: 3D point cloud preprocessing: 3D point cloud preprocessing is performed only once before the encoding layer. The feature dimension of the input 3D point cloud is set to 8 dimensions by using a shared MLP. Step 2: The encoding layer consists of a local feature fusion module and a random sampling module; the local feature fusion module consists of two consecutive sets of local spatial encoding and an attention pooling part; (2.1) Local spatial coding: Local feature coding consists of three parts: finding points in local regions, encoding the position of relative points, and enhancing point features; (2.2) Attention pooling; Attention scores for each local feature point are learned using a function consisting of Conv2d and softmax. Then, the features of each local region point are multiplied by each attention score, and the sum is fed into a shared MLP composed of Batch Normalization (BN) and LeakyReLU. This yields a feature vector. ; Step 3: Decoding Layer: The feature encoding obtained in Step 2 is used as the input of the decoding layer. The number of 3D point clouds is sampled by the nearest neighbor difference. The feature dimension and the contrast boundary are used to learn the feature dimension. The reverse skip connection of the encoding layer is fed into the shared MLP to fuse features. After the feature fusion is completed, the dimension is increased by full connection, and then random deactivation is performed. Finally, the dimension is reduced to the same dimension as the number of semantic labels through the shared MLP, and the final output is the 3D point cloud semantic segmentation result. Step 4: Contrast Boundary Optimization: Contrast boundary optimization employs a contrast boundary learning framework, consisting of two modules: (4.1) Boundary sub-scene mining: Extract ground truth labels from 3D point cloud sampling at different scales. The number of sampled point clouds in each layer is the same as the random sampling result of the corresponding coding layer. As the scale of the 3D point cloud decreases, the specific labels of the points sampled at the boundary will be uncertain. Boundary sub-scene mining in the CBL framework is used to determine the semantic labels of each layer of point cloud. (4.2) Contrastive boundary learning: During the model training process, the boundary points of each layer are compared with the semantic labels of the corresponding layers obtained from the boundary sub-scenes. Contrastive learning uses InfoNCE loss and its generalization to optimize the objective function that defines the boundary points, so that the final result of the boundary points is closer to the neighboring points from the same category. Finally, the 3D point cloud features of the corresponding decoding layer are optimized.

2. The three-dimensional point cloud semantic segmentation method with optimized boundaries according to claim 1, characterized in that: Step 2 (2.1) includes the following parts: (2.1.1) In local spatial encoding, finding points in local regions: Local regions controlled by scene scale are based on Euclidean distance d, and the nearest neighbor points are collected using the KNN algorithm, with the center point of each local region as the nearest neighbor. Sort; neighboring points within each local region are sorted. This means, that is ; where subscript Indicates the index of the center point, superscript Indicates the center point The index of points within a local region is used to construct a KD-tree index using KNN; (2.1.2) Relative Position Encoding: To enable the entire network structure to better understand local features, the encoding method involves concatenating four parts: the coordinates of the center point of each local region, the coordinates of neighboring points within the local region, the difference between the coordinates of neighboring points within the local region and the coordinates of the center point of the local region, and the Euclidean distance of neighboring points within the local region relative to the center point of the local region. This is then processed by MLP. This method encodes the relative position of each local region... of A point is associated with the center point of this local region to improve the local performance of the overall network structure; (2.1.3) Point feature enhancement: within each local region The feature of the nearest neighbors is... The relative positional codes of nearby points are concatenated with the original features of that point. A new feature is constructed from the positional codes and the original features. The concatenated feature is then the result. Subscript Indicates the index of the center point, superscript Indicates the center point The index of the point in the neighborhood.

3. The three-dimensional point cloud semantic segmentation method with optimized boundaries according to claim 1, characterized in that: The number of layers for boundary mining in step four is the same as that of the encoding layer, and each layer of boundary mining and each layer of the encoding layer share random sampling; the ground truth of the dataset is introduced during model training to compare boundary learning; boundary points are the points that distinguish different objects. The search method is to find points in the 3D point cloud that have different semantic labels from their neighbors. The purpose is to more clearly separate different objects in the end-to-end network model and improve the semantic segmentation accuracy of the 3D point cloud; the boundary neighborhood is selected as a sphere with a scale of 0.1% of the overall 3D point cloud, and its expression formula is (6): （6） in Represents the entire 3D point cloud. Represents the set of boundary points. Indicates The set of neighborhood points, Indicates in Neighborhood and Points of semantic label inconsistency Represents semantic tags, and , The superscript indicates the number of the corresponding coding layer; the subscript indicates the sampling layer number and the subscript indicates the order of the positions of the points in the neighborhood.

Images