Local geometry-aware self-enhancing point cloud segmentation method and storage medium thereof

Abstract

The application discloses a local geometry-aware self-enhanced point cloud segmentation method, comprising the following steps: step 1, obtaining and preprocessing point cloud data, performing initial mapping on input point features, and constructing a local neighborhood through farthest point sampling and ball query; step 2, extracting local surface geometry prior, performing multi-frequency trigonometric position coding on relative position coordinates, and constructing geometry information embedding features; step 3, through an adaptive feature fusion module, dynamically weighting and fusing geometry features and semantic features to form local fusion features; step 4, using a dual attention hybrid pooling module, jointly extracting local significant response and neighborhood context information to enhance local representation capability; step 5, through a multi-branch self-enhanced feature aggregation module, differentiating modeling, weighting aggregation and residual enhancement are performed on local features; step 6, in the decoding stage, through interpolation upsampling and jump connection, point-level resolution is recovered layer by layer, and the semantic class label of each point is output. The method realizes the collaborative modeling of local geometry structure and semantic relationship, thereby improving the segmentation precision of complex scenes and fine-grained objects.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensing, and more particularly to 3D scene understanding and point cloud segmentation technology. Background Technology

[0002] Point cloud semantic segmentation, a key technology for 3D scene understanding, is of great significance in fields such as autonomous driving, robot navigation, 3D reconstruction, and intelligent perception. However, point cloud data itself is characterized by sparseness, disorder, non-uniform distribution, and susceptibility to occlusion and noise, posing significant challenges to the segmentation task. Early methods relied heavily on handcrafted geometric features and traditional classifiers, which have limited feature representation capabilities and are ill-suited to the point-by-point semantic recognition requirements of large-scale, complex scenes.

[0003] With the development of deep learning, point-based methods, represented by PointNet, PointNet++, and PointNeXt, can directly process raw point clouds and capture semantic information through hierarchical sampling and local aggregation. For example, the published patent document CN115527032A, filed on December 27, 2022, entitled "A 3D Point Cloud Classification and Segmentation Method Integrating Dense Connection Mechanism and PointNet++," discloses a data processing method integrating PointNet and cloud classification and segmentation. However, similar existing methods still have significant shortcomings:

[0004] On the one hand, most methods mainly use relative coordinates or direction vectors to describe local geometric relationships, and do not fully explore explicit geometric priors such as local surface normals and curvatures, resulting in unstable representations of boundary regions and complex curved surface regions.

[0005] On the other hand, semantic features and geometric features are usually fused by direct splicing or simple superposition, which can easily cause channel coupling, noise amplification and feature interference. In addition, local aggregation often relies on a single pooling or single-branch mapping strategy, which has limited ability to express fine-grained structure and local difference information, thus affecting the segmentation accuracy and robustness in complex scenes. Summary of the Invention

[0006] This invention addresses the problems of insufficient utilization of local surface geometry priors, coarse fusion of geometric and semantic features, and limited expressive power of local feature aggregation in existing point cloud semantic segmentation methods. It proposes a local geometry-aware self-enhancing point cloud semantic segmentation method. This method achieves collaborative modeling of local geometric structures and semantic relationships through steps such as geometric position embedding, adaptive feature fusion, dual-attention hybrid pooling, and multi-branch self-enhancing feature aggregation, thereby improving the segmentation accuracy of complex scenes and fine-grained objects.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a local geometric awareness self-enhancing point cloud segmentation method, comprising the following steps:

[0008] Step 1: Acquire and preprocess point cloud data, perform initial mapping on input point features, and construct local neighborhoods through farthest point sampling and ball query;

[0009] Step 2: Extract local surface geometric priors, perform multi-frequency trigonometric function position encoding on relative position coordinates, and construct geometric information embedding features;

[0010] Step 3: Through the adaptive feature fusion module, geometric features and semantic features are dynamically weighted and fused to form local fused features;

[0011] Step 4: Utilize the dual attention hybrid pooling module to jointly extract local salient responses and neighborhood context information, thereby enhancing local representation capabilities;

[0012] Step 5: Through the multi-branch self-enhancing feature aggregation module, perform differential modeling, weighted aggregation, and residual enhancement on local features;

[0013] Step 6: In the decoding stage, point-level resolution is restored layer by layer through interpolation upsampling and skip connections, and the semantic category label of each point is output.

[0014] Step 1 includes the following steps:

[0015] Use the original point cloud with consistent data format and number of features as input;

[0016] The input point features are initially mapped using a multilayer perceptron, projecting them onto a unified feature space.

[0017] The set of center points is obtained by sampling the farthest point, and a ball query is used to construct a local region consisting of K neighboring points for each center point. The semantic information and geometric structure of the points are jointly modeled at the local scale.

[0018] Based on the constructed local region, the relative position coordinates of each point in the neighborhood of the central point are modeled, and the spatial distribution characteristics of the local point set are characterized by the covariance matrix.

[0019] For each local region, the geometric information embedding module is used to extract the prior geometric information of the local surface and the trigonometric function encoding information of the relative position, in order to obtain an enhanced geometric representation that combines local structural information and spatial position information.

[0020] Step 2 includes the following steps:

[0021] Calculate the covariance matrix for the relative position coordinates of K points within the local neighborhood of the center point;

[0022] Extract surface geometric features and perform singular value decomposition on the covariance matrix to obtain the principal direction and normal vector;

[0023] Singular value decomposition is performed on the covariance matrix, and the minimum singular value is the pseudo-normal vector;

[0024] The pseudo-curvature is approximated by normalizing the singular value ratio.

[0025] The pseudo-normal vector and pseudo-curvature are projected into a higher-dimensional space using a mapping function, and the relative coordinates are encoded using multi-frequency trigonometric functions.

[0026] Sine and cosine mappings are performed at multiple frequency scales to obtain a hierarchical spatial representation from low to high frequencies;

[0027] The initial position tensor is obtained by concatenating all the sine and cosine of the three spatial axes. Then, the channel dimension is mapped to C dimension through a convolutional layer to obtain the trigonometric function position features.

[0028] The original semantic features, surface geometric features, and trigonometric function position codes of the i-th point in the k-neighborhood of the center point are concatenated, and then refined through a convolutional layer to obtain the final semantic features.

[0029] In step 3, the adaptive feature fusion module is used to perform weighted fusion of local semantic features and enhanced geometric representation to generate local fusion features that simultaneously contain semantic information and geometric perception capabilities.

[0030] By learning adaptive weights for space and channels through a lightweight convolutional network, the contribution ratio of the two features at each position and channel is dynamically adjusted.

[0031] The original geometric features are nonlinearly transformed and mapped to the same feature space as the semantic features through a lightweight convolutional network.

[0032] We use weights to modulate semantic and positional features, giving them their own feature sensitivities in different regions.

[0033] Step 4 includes the following steps:

[0034] An attention mechanism is used to capture the importance of semantic features and positional features respectively, and functions for learning semantic features and positional features are designed for local semantic features and local positional features;

[0035] The learned local semantic features and local location features are weighted and fused through a learnable parameter to obtain the features of the local region. Then, the features are refined through a multilayer perceptron to obtain the integrated features of the local region.

[0036] In step 5, the original features of the center point are obtained by sampling the farthest point, and the local region of each center point is generated by grouping using ball query. This region consists of K points. Then, the geometric information and semantic features contained in the K neighboring points of the center point are mixed to obtain the mixed features.

[0037] Multiple independent feature mapping branches are preset to mix features in local regions for transformation, resulting in differentiated features;

[0038] The attention weights are generated by summing the features extracted from multiple branches element by element.

[0039] The features extracted from the other branch are weighted element-wise using attention weights, and the weighted aggregation is completed by summation to obtain the weighted features of the local region.

[0040] Steps 1-5 are repeated in multiple layers of the encoder. The encoder consists of four cascaded abstract layers. The number of input points in each layer gradually decreases while the feature dimension gradually increases, enabling hierarchical feature extraction from local details to high-level semantics. After multiple layers of encoding, the deepest feature is sent to the decoder.

[0041] In step 6, the point resolution is restored layer by layer by nearest neighbor interpolation or weighted interpolation based on the inverse distance of the three nearest neighbors. The intermediate features retained in the encoding stage are passed to the corresponding layer in the decoding stage by skip connections. After each interpolation and feature concatenation is completed, the features are refined by a multilayer perceptron. Finally, the features are mapped to the category space by a linear classification layer to output the corresponding semantic label for each point in the original point cloud.

[0042] The storage medium is a computer-readable storage medium for storing software program code, which is used to execute the local geometry-aware self-enhancing point cloud segmentation method.

[0043] The local geometry-aware self-enhancing point cloud semantic segmentation method proposed in this invention has the following advantages compared with existing semantic segmentation methods:

[0044] 1. Point cloud semantic segmentation methods extract pseudo-normal vectors and pseudo-curvatures in local neighborhoods and combine them with multi-frequency sine and cosine position encoding to construct geometric features, which can more fully characterize the local surface shape, boundary changes and multi-scale spatial relationships of point clouds, thereby improving the local geometric modeling capability.

[0045] 2. The point cloud semantic segmentation method dynamically weights semantic and geometric features through adaptive feature fusion and dual attention hybrid pooling mechanism, and jointly preserves the overall distribution information of the neighborhood and local salient features, thereby reducing feature interference and noise propagation caused by simple splicing, and improving the stability and discriminativeness of local aggregation;

[0046] 3. Point cloud semantic segmentation methods extract local features differentially through multiple parameter-independent mapping branches, and combine adaptive weights and residual enhancement to complete multi-branch self-enhancing feature aggregation. This can better capture subtle differences between edges, textures and similar geometric structures, thereby improving the segmentation effect of complex scenes and fine-grained objects. Attached Figure Description

[0047] The following is a brief explanation of the content represented by each figure in this specification:

[0048] Figure 1 This is a flowchart illustrating the point cloud semantic segmentation method proposed in this invention;

[0049] Figure 2 This is a network structure diagram of the point cloud semantic segmentation method proposed in this invention;

[0050] Figure 3 This is a schematic diagram of the geometric information embedding module structure proposed in this invention;

[0051] Figure 4 This is a schematic diagram of the dual attention hybrid pooling module structure proposed in this invention;

[0052] Figure 5 This is a schematic diagram of the multi-branch self-enhancing feature aggregation module structure proposed in this invention;

[0053] Figure 6 This is a visualization of the method of the present invention on the S3DIS dataset;

[0054] Figure 7 This is a visualization of the method of the present invention on the ScanNetV2 dataset. Detailed Implementation

[0055] The following description, with reference to the accompanying drawings, details the specific implementation of the present invention, including the shape and structure of each component, the relative positions and connections between the parts, the function and working principle of each part, the manufacturing process, and the operation and use methods, to help those skilled in the art to have a more complete, accurate, and in-depth understanding of the inventive concept and technical solution of the present invention.

[0056] This invention relates to a scheme for 3D scene understanding and point cloud segmentation, specifically a local geometry-aware self-enhancing point cloud segmentation method. The following is a detailed description of this method based on the accompanying drawings:

[0057] Figure 1This is a schematic diagram of the overall process of the proposed point cloud semantic segmentation method. The process mainly includes the following steps: point cloud data input and initial feature construction, local region division and neighborhood feature extraction, geometric information embedding, geometric semantic adaptive fusion, dual attention hybrid pooling, multi-branch self-enhancing feature aggregation, and decoding prediction output. First, the input original 3D point cloud is subjected to initial feature mapping, and local region representations corresponding to each sampling point are constructed through hierarchical sampling and neighborhood query. Then, for each local region, the geometric information embedding module is used to extract local surface geometric prior information and trigonometric function encoding information of relative position, thereby obtaining an enhanced geometric representation that combines local structural information and spatial position information. Then, the adaptive feature fusion module is used to perform weighted fusion of local semantic features and enhanced geometric representation to generate local fusion features that simultaneously contain semantic information and geometric perception capabilities. On this basis, the dual attention hybrid pooling module is further used to jointly extract key response information and statistical features in the local neighborhood to enhance the expression of salient features; subsequently, the multi-branch self-enhancing feature aggregation module is used to perform multi-path interaction modeling and aggregation enhancement of local features, thereby improving the richness and discriminative ability of feature representation. The above process is repeated in the multi-layer encoding stage of the network to gradually extract higher-level and more representative point cloud features. Finally, through feature propagation and layer-by-layer upsampling operations in the decoding and prediction module, combined with the shallow detail information passed in the encoding stage, the point-level resolution is gradually restored, and the semantic category label corresponding to each point in the original point cloud is finally output, realizing end-to-end point cloud semantic segmentation.

[0058] Figure 2 This is the network structure diagram for this method. The original point cloud with consistent data format and feature number is taken as input. The point cloud data can be represented as follows: ,in Represents the three-dimensional coordinates of the i-th point. Additional attribute features can be represented, including color information, reflectance intensity, or other sensory attributes. After the original point cloud is input into the encoder network, the input point features are first initially mapped using a multilayer perceptron, projecting them onto a unified feature space. Subsequently, a set of center points is obtained through farthest point sampling, and a sphere query is used to construct a local region consisting of K neighboring points for each center point, thereby jointly modeling the semantic information and geometric structure of the points at a local scale.

[0059] After the local region is constructed, the relative position coordinates of each point in the neighborhood of the center point are first modeled, and the spatial distribution characteristics of the local point set are characterized by the covariance matrix. Then, for each local region, the geometric information embedding module is used to extract the prior geometric information of the local surface and the trigonometric function encoding information of the relative position, thereby obtaining an enhanced geometric representation that combines local structural information and spatial position information.

[0060] Figure 3 This is a structural diagram of the geometric information embedding module. First, the relative position coordinates of K points within the local neighborhood of the center point are determined. First calculate the covariance matrix. The formula is expressed as:

[0061]

[0062]

[0063] in, The covariance matrix represents the local centroid. The eigenvectors of the covariance matrix represent the principal directions of the local point distribution, and the eigenvalues ​​measure the intensity of point distribution along different directions. The eigenvector corresponding to the smallest eigenvalue is the normal vector direction, representing the minimum change of the point cloud in the direction perpendicular to the surface.

[0064] To further extract surface geometric features, singular value decomposition (SVD) is performed on the covariance matrix to obtain the principal directions and normal vectors. The SVD of the covariance matrix yields:

[0065]

[0066] in, , It is an orthogonal matrix whose column vectors are left / right singular vectors. = and satisfy When the local point cloud is approximately located on a plane, the variance of the point distribution along the normal direction is minimized. Therefore, the pseudo-normal vector corresponds to the minimum singular value. eigenvectors ,in It can be represented as

[0067]

[0068] The core of curvature is to quantify the degree of bending of a local surface through the distribution of singular values ​​in the covariance matrix. Therefore, pseudo-curvature can be approximated by normalizing the proportion of singular values, as shown in the formula:

[0069]

[0070] in, This represents the pseudo-curvature of the k-th sampling point. The normalized scale reflects local geometric characteristics. Since pseudo-normal vectors and pseudo-curvatures are low-dimensional, in order to embed them into high-dimensional geometric features, we use a mapping function to project them into a higher-dimensional space, i.e.:

[0071]

[0072] However, relying solely on surface geometric priors is insufficient to characterize multi-scale spatial relationships. Therefore, multi-frequency trigonometric function position encoding is further applied to the relative coordinates. This involves considering the relative position information of each neighborhood point. Sine and cosine mappings are performed at multiple frequency scales to obtain a hierarchical spatial representation from low to high frequencies, with the following encoding form:

[0073]

[0074] Where α controls the scaling intensity of the input coordinates, β controls the wavelength growth rate, and the frequency index... , for ,index It is a multi-scale coding system from low frequency to high frequency.

[0075] Subsequently, the three axes By splicing together all the sine and cosines, we get The initial position tensor is then used to map the channel dimension to C dimensions through a convolutional layer, resulting in trigonometric function position features, as shown in the following formula:

[0076] Finally, the original semantic features of the i-th point in the k-neighborhood of the center point are... Surface geometric features and trigonometric function position encoding The features are concatenated and refined through convolutional layers to obtain the final semantic features:

[0077]

[0078] Where A is a convolutional layer. This is for splicing operations.

[0079] In this way, the geometric position embedding module can simultaneously enhance the ability to perceive the local surface shape of point clouds and the ability to express multi-scale spatial relationships.

[0080] Then, the local semantic features and the enhanced geometric representation are weighted and fused using an adaptive feature fusion module to generate local fused features that simultaneously contain semantic information and geometric perception capabilities.

[0081] Directly concatenating positional and semantic features forces information from different sources to align along the channel dimension. This rigid fusion mechanism causes geometric and semantic information to interfere with each other during transmission, making it difficult to distinguish regions with similar semantics but different geometric structures. To address this, we propose an adaptive feature fusion module. This module learns adaptive weights for the space and channels through a lightweight convolutional network, dynamically adjusting the contribution ratio of the two features at each position and channel.

[0082] Specifically, the original geometric features are first subjected to a nonlinear transformation, and then a lightweight convolutional network is used to map the original 3D position coordinates to the same feature space as the semantic features:

[0083] in It consists of two convolutional layers, which include batch normalization and the ReLU activation function. These are the original geometric features. They are then concatenated with semantic features and a fusion weight is generated using the Sigmoid activation function.

[0084]

[0085] in For the Sigmoid function, It is a single linear layer. This is for splicing operations.

[0086] Finally, we use weights to modulate semantic and location features, giving them their own feature sensitivities in different regions. This process can be described as follows:

[0087]

[0088] This module preserves richer geometric details in structurally complex regions (such as edges and corners), while focusing on utilizing high-level semantic information in semantically clear regions (such as flat surfaces), thereby improving the discriminativeness and robustness of the feature representation as a whole. After obtaining the fused local features, this invention further achieves local feature aggregation through a dual-attention hybrid pooling module.

[0089] Figure 4 This is a diagram of the dual-attention hybrid pooling module. Traditional max pooling, which integrates neighboring features, may neglect geometric context, while simple attention pooling may lose important feature details. This module simultaneously utilizes relative positional information and semantic features to generate attention weights, and adaptively fuses the results of attention pooling and max pooling through learnable parameters.

[0090] First, an attention mechanism is used to capture the importance of semantic features and positional features respectively, focusing on local semantic features. and local location features A function was designed respectively sum function To learn semantic and positional features, the formula is as follows:

[0091] Among them, the function sum function It consists of 1×1 convolutional layers and a softmax activation function. For semantic feature scores, The location feature score is then used. The semantic feature score is subsequently used. and location feature scores The elements are added together to obtain the comprehensive attention weight of the point in the neighborhood. Then, the original semantic features are weighted and summed according to this weight. The process is as follows:

[0092]

[0093] in, This indicates element-wise multiplication. To preserve the overall distribution and contextual information of neighborhood features, we employ attention pooling. Simultaneously, we introduce max pooling to extract the most salient local features, as shown in the following formula:

[0094]

[0095] in, This is for max pooling.

[0096] Finally, the two are weighted and fused using a learnable parameter α to obtain the features of the local region. These features are then refined using multilayer perceptrons (MLPs) to obtain the integrated features of the local region. :

[0097]

[0098] Here, α is a learnable parameter activated by the sigmoid function, used to adaptively balance the contributions of the two pooling strategies.

[0099] This pooling method not only takes into account the most salient local features, but also preserves the overall distribution of the neighborhood and the geometric context, making the model more discriminative in detail.

[0100] After completing local feature pooling, in order to further explore the complementary relationship between different feature patterns in the neighborhood, this invention performs multi-branch parallel modeling of the pooled local features through a multi-branch self-enhancing feature aggregation module.

[0101] Figure 5 This is a structural diagram of a multi-branch self-enhanced feature aggregation module. Compared to single-path feature transformation, the multi-branch structure can characterize the detailed changes and geometric patterns within a local region from different perspectives, thus obtaining richer feature representations. Specifically, it uses farthest point sampling... Obtain the original features of the center point Using ball queries, a local region R is generated for each center point, consisting of K points. The hybrid features of the K neighboring points of the center point, containing both geometric and semantic information, are represented as follows: This module is designed with multiple independent feature mapping branches ( , , ) local region mixing features By performing a transformation, we obtain the differentiated features:

[0102]

[0103] in, , , It consists of lightweight convolutional layers with independent parameters to ensure differentiated capture of local details and geometry. Features extracted from multiple branches are utilized. and After element-wise summation, attention weights are generated using softmax.

[0104]

[0105] in, It can adaptively measure the importance of each point within a local region to the current feature representation, and integrates... and The dual-view information it carries avoids feature blurring caused by a single viewpoint.

[0106] Finally, attention weights are utilized. Features extracted from another branch Element-wise weighting is performed, and weighted aggregation is completed by summation to obtain the weighted features of the local region. The specific process is as follows:

[0107]

[0108] To preserve the original feature information of the center point and enhance the feature representation, we will... The data is mapped to the same dimension as the local region features using a multilayer perceptron, and then weighted together with the local region features. The enhanced features of the module's final output are obtained through residual connection fusion. The entire process can be represented as:

[0109]

[0110] The aforementioned operations, such as geometric position embedding, adaptive feature fusion, dual attention hybrid pooling, and multi-branch self-enhancing feature aggregation, are repeatedly performed across multiple layers of the encoder. In one embodiment, the encoder consists of four cascaded ensemble abstraction layers, with the number of input points gradually decreasing and the feature dimension gradually increasing in each layer, thereby achieving hierarchical feature extraction from local details to high-level semantics. After multi-layer encoding, the deepest layer features are fed into the decoder, where the point resolution is restored layer by layer through nearest neighbor interpolation or weighted interpolation based on the inverse distance of the three nearest neighbors. Skip connections are then used to pass the intermediate features retained in the encoding stage to the corresponding layer in the decoding stage. After each interpolation and feature concatenation, a multilayer perceptron is used to refine the features, and finally, a linear classification layer maps the features to a category space, outputting a corresponding semantic label for each point in the original point cloud.

[0111] The system configuration used for point cloud semantic segmentation in this invention includes an NVIDIA GeForce RTX 4090 graphics processing unit, 24.0 GB of video memory, and a 13th generation Intel Core i9-13900KF central processing unit. Experimental results on the S3DIS and ScanNetV2 datasets show that the proposed method can achieve good segmentation performance in complex indoor scenes, achieving segmentation results of 78.0% and 72.0% on the S3DIS and ScanNetV2 datasets, respectively, indicating that the method has good scene adaptability.

[0112] Figure 6 The figure shows the semantic segmentation performance of this method on the S3DIS dataset. As can be seen from the figure, the present invention exhibits good segmentation ability in both boundary and non-boundary regions of objects such as doors, pillars, blackboards, and sofas, especially in edge regions with significant geometric changes, where it can reduce missegmentation between adjacent categories.

[0113] Figure 7 This paper presents the semantic segmentation performance of the proposed method on the ScanNetV2 dataset. The segmentation results are closer to the baseline values. For objects with similar colors, appearances, or geometric structures, such as tables, doors, and walls, this invention can construct more accurate local spatial relationships through explicit geometric priors and a self-reinforcing aggregation mechanism, thereby improving the separability between different objects. This intuitively demonstrates that the method has good segmentation performance in different scenarios.

[0114] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.

Claims

1. A local geometric awareness self-enhancing point cloud segmentation method, characterized in that, Includes the following steps: Step 1: Acquire and preprocess point cloud data, perform initial mapping on input point features, and construct local neighborhoods through farthest point sampling and ball query; Step 2: Extract local surface geometric priors, perform multi-frequency trigonometric function position encoding on relative position coordinates, and construct geometric information embedding features; Step 3: Through the adaptive feature fusion module, geometric features and semantic features are dynamically weighted and fused to form local fused features; Step 4: Utilize the dual attention hybrid pooling module to jointly extract local salient responses and neighborhood context information, thereby enhancing local representation capabilities; Step 5: Through the multi-branch self-enhancing feature aggregation module, perform differential modeling, weighted aggregation, and residual enhancement on local features; Step 6: In the decoding stage, point-level resolution is restored layer by layer through interpolation upsampling and skip connections, and the semantic category label of each point is output.

2. The local geometric awareness self-enhancing point cloud segmentation method according to claim 1, characterized in that: Step 1 includes the following steps: Use the original point cloud with consistent data format and number of features as input; The input point features are initially mapped using a multilayer perceptron, projecting them onto a unified feature space. The set of center points is obtained by sampling the farthest point, and a ball query is used to construct a local region consisting of K neighboring points for each center point. The semantic information and geometric structure of the points are jointly modeled at the local scale. Based on the constructed local region, the relative position coordinates of each point in the neighborhood of the central point are modeled, and the spatial distribution characteristics of the local point set are characterized by the covariance matrix. For each local region, the geometric information embedding module is used to extract the prior geometric information of the local surface and the trigonometric function encoding information of the relative position, in order to obtain an enhanced geometric representation that combines local structural information and spatial position information.

3. The local geometric awareness self-enhancing point cloud segmentation method according to claim 1, characterized in that: Step 2 includes the following steps: Calculate the covariance matrix for the relative position coordinates of K points within the local neighborhood of the center point; Extract surface geometric features and perform singular value decomposition on the covariance matrix to obtain the principal direction and normal vector; Singular value decomposition is performed on the covariance matrix, and the minimum singular value is the pseudo-normal vector; The pseudo-curvature is approximated by normalizing the singular value ratio. The pseudo-normal vector and pseudo-curvature are projected into a higher-dimensional space using a mapping function, and the relative coordinates are encoded using multi-frequency trigonometric functions. Sine and cosine mappings are performed at multiple frequency scales to obtain a hierarchical spatial representation from low to high frequencies; The initial position tensor is obtained by concatenating all the sine and cosine of the three spatial axes. Then, the channel dimension is mapped to C dimension through a convolutional layer to obtain the trigonometric function position features. The original semantic features, surface geometric features, and trigonometric function position codes of the i-th point in the k-neighborhood of the center point are concatenated, and then refined through a convolutional layer to obtain the final semantic features.

4. The local geometric awareness self-enhancing point cloud segmentation method according to claim 1, characterized in that: In step 3, the adaptive feature fusion module is used to perform weighted fusion of local semantic features and enhanced geometric representation to generate local fusion features that simultaneously contain semantic information and geometric perception capabilities. By learning adaptive weights for space and channels through a lightweight convolutional network, the contribution ratio of the two features at each position and channel is dynamically adjusted. The original geometric features are nonlinearly transformed and mapped to the same feature space as the semantic features through a lightweight convolutional network. We use weights to modulate semantic and positional features, giving them their own feature sensitivities in different regions.

5. The local geometric awareness self-enhancing point cloud segmentation method according to claim 1, characterized in that: Step 4 includes the following steps: An attention mechanism is used to capture the importance of semantic features and positional features respectively, and functions for learning semantic features and positional features are designed for local semantic features and local positional features; The learned local semantic features and local location features are weighted and fused through a learnable parameter to obtain the features of the local region. Then, the features are refined through a multilayer perceptron to obtain the integrated features of the local region.

6. The local geometry-aware self-enhancing point cloud segmentation method according to claim 1, characterized in that: In step 5, the original features of the center point are obtained by sampling the farthest point, and the local region of each center point is generated by grouping using ball query. This region consists of K points. Then, the geometric information and semantic features contained in the K neighboring points of the center point are mixed to obtain the mixed features. Multiple independent feature mapping branches are preset to mix features in local regions for transformation, resulting in differentiated features; The attention weights are generated by summing the features extracted from multiple branches element by element. The features extracted from the other branch are weighted element-wise using attention weights, and the weighted aggregation is completed by summation to obtain the weighted features of the local region.

7. The local geometric awareness self-enhancing point cloud segmentation method according to any one of claims 1-6, characterized in that: Steps 1-5 are repeated in multiple layers of the encoder. The encoder consists of four cascaded abstract layers. The number of input points in each layer gradually decreases while the feature dimension gradually increases, enabling hierarchical feature extraction from local details to high-level semantics. After multiple layers of encoding, the deepest feature is sent to the decoder.

8. The local geometric awareness self-enhancing point cloud segmentation method according to claim 1, characterized in that: In step 6, the point resolution is restored layer by layer by nearest neighbor interpolation or weighted interpolation based on the inverse distance of the three nearest neighbors. The intermediate features retained in the encoding stage are passed to the corresponding layer in the decoding stage by skip connections. After each interpolation and feature concatenation is completed, the features are refined by a multilayer perceptron. Finally, the features are mapped to the category space by a linear classification layer to output the corresponding semantic label for each point in the original point cloud.

9. A storage medium, wherein the storage medium is a computer-readable storage medium for storing software program code, characterized in that: The software program code is used to execute the local geometry-aware self-enhancing point cloud segmentation method as described in any one of claims 1-8.

Images