Disturbance-resistant three-dimensional object semantic segmentation geometric enhancement method
By extracting robust orientation consistency features based on statistical geometric feature inference and Rodriguez rotation operation, the problem of segmentation accuracy under perturbation in 3D object semantic segmentation methods is solved, and stable segmentation results are achieved in complex scenes.
Patent Information
- Application Number
- CN202511465040.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-14
- Publication Date
- 2025-11-21
AI Technical Summary
Existing 3D object semantic segmentation methods show a significant decrease in segmentation accuracy under point cloud data perturbation and lack sufficient resistance to perturbation, resulting in insufficient robustness in practical applications and difficulty in coping with complex perturbation scenarios.
By using statistical geometric feature reasoning calculations, we obtain the normal vector and curvature information of point clouds. Combined with sparsity smoothing and Rodriguez rotation operations, we extract robust orientation consistency features, enhance the geometric understanding of point clouds, and improve their resistance to disturbances.
It maintains stable segmentation results in complex application scenarios, significantly improves the anti-perturbation ability of 3D object semantic segmentation, and demonstrates segmentation accuracy and robustness superior to mainstream methods.
Smart Images
Figure CN120997513A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of 3D computer vision technology, and in particular relates to a perturbation-resistant geometric enhancement method for semantic segmentation of 3D objects. Background Technology
[0002] 3D object semantic segmentation is a crucial technology for point cloud scene understanding, enabling the processing, analysis, and semantic interpretation of 3D point cloud data. Compared to images, point clouds can fully depict the geometric structure and spatial location of objects, thus finding wide application in fields such as autonomous driving, robotics, surveying, and industrial manufacturing. In 3D object segmentation tasks, point clouds are used for semantic classification of complex objects. For example, in robot grasping, semantic segmentation is needed to identify the operable parts of an object; in industrial inspection, parts need to be segmented to support quality inspection and assembly; in smart cities and surveying, it can model and semantically segment urban point clouds, assisting in digital twins and urban planning. However, point cloud data often suffers from perturbations such as noise, rotation, translation, scaling, and jitter. Point cloud perturbations refer to structural or unstructured changes in the position, quantity, and arrangement of points introduced by factors such as changes in the external environment and sensor errors. While existing methods (such as PointNeXt, PointAugment, and Stratified Transformer) achieve good segmentation results under standard conditions, their accuracy drops significantly under the aforementioned perturbations, revealing the poor perturbation resistance of current 3D object semantic segmentation methods. Poor perturbation resistance leads to insufficient robustness of semantic segmentation methods in practical applications, making them prone to false positives and even false negatives, resulting in significant negative impacts on application deployment. Since real-world point cloud data often exhibits variations such as noise and jitter, enhancing perturbation resistance is crucial for ensuring application security, reliability, and generalization ability. Developing 3D object semantic segmentation methods robust to point cloud perturbations has significant application and practical value.
[0003] Current perturbation-resistant 3D object semantic segmentation methods mainly fall into two technical routes: data augmentation and robust representation extraction. Data augmentation enhances the network's robustness to perturbations by rotating, translating, or jittering the input point cloud during training. For example, KPConv randomly discards point cloud colors during training to enhance robustness to color changes. PointAugment employs an adversarial learning strategy to jointly optimize the amplifier and classifier networks, improving classification robustness. PointCutMax finds the optimal allocation between two point clouds and replaces points in one sample to generate new training data. While these methods improve the method's robustness to perturbations to some extent, their performance is still limited by the finite types and magnitudes of perturbations. Robust representation extraction, on the other hand, designs representation methods to extract perturbation-resistant point cloud representations to further enhance the method's robustness. PAConv constructs convolutional kernels by dynamically learning weight matrices, enriching the local spatial relationships of point clouds and improving the method's robustness. However, focusing on optimizing local spatial relationships, its robustness is not ideal for large-scale translations and scaling caused by changes in the overall structure and long-distance dependencies, making it difficult to handle complex perturbation scenarios. StratifiedTransformer proposes a hierarchical Transformer structure to capture long-distance contextual information of point clouds, improving the method's robustness. However, its robustness decreases significantly when multiple perturbations are superimposed, and it lacks an effective mechanism to deal with extreme perturbations such as severe jitter and heavy noise, so its robustness performance still needs improvement. Spherical CNNS projects point clouds onto a spherical space to capture rotation-resistant features, which is advantageous in scenarios with rotational perturbations. However, it can only improve robustness for rotational perturbations and has no ability to resist other types of perturbations such as noise, translation, scaling, and jitter, making its applicable scenarios extremely limited. RISurConv ensures the robustness of feature extraction under rotation by designing rotation-invariant surface properties such as included angles. However, similar to Spherical CNNS, it is only optimized for rotational perturbations and cannot cope with other perturbations. In practical applications, it has many limitations and is difficult to meet the needs of complex scenarios.
[0004] To address the above shortcomings, this invention proposes a disturbance-resistant geometric enhancement method for semantic segmentation of three-dimensional objects. Summary of the Invention
[0005] The purpose of this invention is to provide a disturbance-resistant geometric enhancement method for semantic segmentation of three-dimensional objects, aiming to solve the problems mentioned in the background art.
[0006] The objective of this invention is achieved through the following technical solution: A perturbation-resistant geometric enhancement method for semantic segmentation of 3D objects includes the following steps: Acquire the input 3D point cloud data; Statistical geometric feature inference calculations are performed on 3D point cloud data to obtain point cloud geometric features. The specific process is as follows: First, a decentralized sub-point cloud vector set is constructed. The distribution and clustering degree of local sub-point clouds in different directions are quantified by sum of squares statistics. The solution of the normal vector of the local sub-point cloud is transformed into an optimization problem to obtain the normal vector of the neighborhood. Then, the initial curvature is calculated by combining the eigenvalues. The initial curvature is corrected by introducing a sparsity smoothing factor and a sparsity adjustment temperature to obtain the curvature after sparsity smoothing. The geometric features of the point cloud are subjected to geometrically enhanced directional consistency feature extraction. The specific process is as follows: First, the obtained normal vector and curvature information are embedded into scalar features to form geometrically enhanced scalar features; then, the scalar features are expanded into independent spatial vectors through adaptive feature vectorization operation; next, the spatial vectors are processed by Rodriguez rotation operation, and a relative direction refinement mechanism is introduced to adjust the relative direction of the features; finally, the number of scalar channels is restored by grouped convolution, and feature aggregation and point cloud disorder representation are achieved through summation operation to obtain a robust point cloud representation for semantic segmentation of 3D objects.
[0007] Furthermore, the formula for constructing the decentralized sub-point cloud vector set is as follows: ; in, For a decentralized sub-point cloud vector set; For the first A decentralized point; For the neighboring region One point; Representing the neighborhood The average position of all points in; The index of the point; Indicates the number of points in the neighborhood; This refers to the local neighborhood of the point cloud.
[0008] Furthermore, the formula for transforming the solution of the normal vector of the local sub-point cloud into an optimization problem is as follows: ; in, To optimize the objective; Let be the unit normal vector of the neighborhood to be determined; It is a three-dimensional Cartesian space; The number of points in the neighborhood; for transpose; It is the product of the decentralized sub-point cloud vector set and its transpose; right The formula for decomposition is: ; in, and All are 3x3 square matrices; It is a diagonal matrix, and the three values of its diagonal are... Three eigenvalues; The The list is The Middle eigenvalues Corresponding feature vector .
[0009] Furthermore, the formula for calculating the initial curvature is: ; in, The initial curvature; It is the smallest eigenvalue; It is the largest eigenvalue; These are intermediate eigenvalues; It is a very small positive integer.
[0010] Furthermore, the formula for obtaining the curvature after sparsity smoothing is as follows: ; in, The curvature after sparsity smoothing; This indicates the sparsity adjustment temperature used to adjust the strength of smoothness. The number of points in the neighborhood; The index of the point; Center point; for , neighborhood points; For point The initial curvature at that point.
[0011] Furthermore, the formula for the adaptive feature vectorization operation is as follows: ; in, For scalar characteristics Vectorization operations; Features that are vectorized; It is a scalar feature; , and It is a set of learnable parameters; This is a matrix transpose operation.
[0012] Furthermore, the formula for the relative direction refinement mechanism is as follows: ; in, To Relative direction refinement operation; These are the feature vectors refined in relative directions; For the cropping operation, it will Limited to Within the range; From the input The neural network was trained to obtain it; It is a set of triples containing the neighborhood The average of the absolute values of the features of different feature channels of the points in the middle on three orthogonal bases; for Features obtained after Rodriguez rotation.
[0013] Furthermore, the final point cloud representation output by the geometrically enhanced orientation consistency feature extraction step is calculated using the following formula: ; in, This is a characteristic of directional consistency; This represents a grouped convolution that restores the feature vector to a scalar number of channels; This represents the summation operation used to achieve feature aggregation and the representation of the disordered nature of point clouds; These are the feature vectors refined in relative directions; It is a scalar feature.
[0014] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.
[0015] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the steps of the method described above.
[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention effectively enhances the robustness of 3D object semantic segmentation methods, enabling them to maintain stable and accurate segmentation results even in complex application scenarios. Technically, this invention first transforms the calculation of normal vectors and curvature of discrete point clouds into a statistical optimization problem through a statistical geometric feature inference calculation step. The normal vectors of the neighborhood are obtained by solving this problem, and the initial curvature is calculated using eigenvalues. Sparsity smoothing is introduced to correct curvature errors in sparse neighborhoods, accurately capturing the geometric spatial information of the point cloud and significantly enhancing the method's geometric understanding of perturbed point clouds. Subsequently, through a geometrically enhanced direction consistency feature extraction step, the obtained normal vectors and curvature information are embedded into scalar features. Adaptive feature vectorization operations expand the scalar features into independent spatial vectors, and Rodriguez rotation is used to achieve vector rotation. Combined with a relative direction refinement mechanism, point cloud perturbations are accurately modeled, and finally, robust direction consistency representations are extracted, directly improving the method's robustness against perturbations from a technical perspective. In terms of performance verification, test results based on the S3DIS dataset show that the present invention maintains stable segmentation performance under various perturbation scenarios such as rotation, translation, scaling, and jitter. In the rotation invariance test, it also shows better adaptability than mainstream comparison methods. At the same time, the visualization effect of the semantic segmentation results is better in terms of overall integrity and detail accuracy. These results further confirm the actual effect of the present invention in stable and accurate segmentation in complex application scenarios, and fully support the technical goal of improving anti-perturbation capability. Attached Figure Description
[0017] Figure 1 This is a schematic diagram of the overall method of the present invention; wherein: in the statistical geometric feature reasoning calculation step, the normal vector and curvature are calculated by statistically analyzing the distribution of local point clouds in different directions, the red dot represents the projection of the point in the direction of the normal vector, the red arrow represents the normal vector, and the different colors around it (yellow, red halos) represent the curvature values near the point; in the geometrically enhanced direction consistency feature extraction step, robust direction consistency features are extracted through feature rotation and relative direction refinement mechanisms, and the green text represents key operations.
[0018] Figure 2 This is a partial semantic segmentation result image of a 3D object. Detailed Implementation
[0019] In order to provide a clearer understanding of the technical features, objectives and beneficial effects of the present invention, the technical solution of the present invention will now be described in detail below, but it should not be construed as limiting the scope of implementation of the present invention.
[0020] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.
[0021] This invention provides a perturbation-resistant geometric enhancement method for semantic segmentation of 3D objects, the flowchart of which is shown below. Figure 1 As shown, the process begins with statistical geometric feature inference calculations to process the input 3D point cloud data and obtain point cloud geometric features. This step transforms the calculation of normal vectors and curvature of discrete point clouds into a statistical optimization problem by statistically analyzing the distribution of local point sets in different directions. Considering the low local geometric information content of sparse point sets and the potential for misleading geometric representations, sparsity is introduced to smooth the curvature of sparse regions and enhance the reliability of curvature. Next, geometrically enhanced directional consistency feature extraction is performed, and the features undergo in-depth processing. This step first embeds the normal vectors and curvature information obtained from the statistical geometric feature inference calculation into scalar features, forming geometrically enhanced scalar features. Geometric prior information helps the method understand the relationship between shallow geometric information and deep semantic information, enabling the method to more efficiently perceive the spatial distribution and shape changes of the point cloud, thus making it more robust to perturbations. To extract robust directional consistency representations, this step adaptively expands the scalar features into independent spatial vectors, giving the features directional expressions while maintaining the relative position information of the original features, thereby improving the information content and flexibility of 3D object semantic segmentation feature extraction. Subsequently, a feature rotation mechanism is used to capture consistent patterns in the anisotropic feature orientations, thereby better perceiving the directional information of the features. Further refinement of the relative orientation of each feature vector is performed to precisely adjust the relative directional relationships of features in the feature space, thus more accurately modeling the perturbations of the point cloud in physical space. Finally, scalar channel numbers are recovered through grouped convolution, and feature aggregation and point cloud disorder representation are achieved through summation operations, resulting in a robust point cloud representation for 3D object semantic segmentation.
[0022] The method of the present invention will be described in detail below: (a) Statistical-based geometric feature reasoning and calculation; To obtain accurate geometric information of perturbed point clouds, the geometric meaning of normal vectors and curvature is extended to a three-dimensional local discrete space. This transforms the problem of calculating the normal vectors and curvature of discrete point clouds in three-dimensional object semantic segmentation tasks into a statistical optimization problem, namely, solving for the normal vectors. n This makes the projection of points in the neighborhood most concentrated in this direction.
[0023] First, construct a decentralized sub-point cloud vector set. The dot product of this vector set can directly reflect the projection values of a point in different directions.
[0024] ; in, For the first A decentralized point; For the neighboring region One point; Representing the neighborhood The average position of all points in; The index of the point; Indicates the number of points in the neighborhood; This refers to the local neighborhood of the point cloud.
[0025] The distribution and clustering of local sub-point clouds in different directions are quantified using the sum of squares statistical method. The problem of finding the normal vectors of a local sub-point cloud is transformed into the optimization problem shown below. : ; in, To optimize the objective; Let be the normal vector of the neighborhood to be determined; It is a three-dimensional Cartesian space; The number of points in the neighborhood; for transpose; It is the product of the decentralized sub-point cloud vector set and its transpose.
[0026] When it is Minimum eigenvalue When the corresponding feature vector is, Obtain the optimal solution. It can be decomposed as follows: ; in, and All are 3x3 square matrices; It is a diagonal matrix, and the three values of its diagonal are... The three eigenvalues reflect the neighborhood. The degree of dispersion of the distribution of points in the three eigenvector directions; The The list is The Middle eigenvalues Corresponding feature vector Based on this, the minimum eigenvalue is... The initial curvature is estimated by the ratio of the sum of the three eigenvalues. At the same time, a very small positive integer is introduced. To avoid calculation errors caused by the denominator approaching zero, this can be formulated as follows: ; in, It is the largest eigenvalue; These are intermediate eigenvalues; It is a very small positive integer.
[0027] To avoid introducing erroneous spatial geometric information into the semantic segmentation of 3D objects, this invention proposes a sparse curvature smoothing method, which introduces... p The sparsity smoothing factor at the point, combined with the sparsity adjustment temperature used to adjust the strength of the smoothness. The initial curvature is corrected to reduce fluctuations and errors in the curvature of the sparse neighborhood, ultimately yielding the curvature after sparsity smoothing, which can be formulated as follows: ; in, The curvature after sparsity smoothing; This indicates the sparsity adjustment temperature used to adjust the strength of smoothness. The number of points in the neighborhood; The index of the point; Center point; for , neighborhood points; For point The initial curvature at that point.
[0028] (ii) Geometrically enhanced directional consistency feature extraction; First, the normal vector and curvature information obtained from the statistical geometric feature inference calculation steps are embedded into scalar features. Then, the scalar features are expanded into independent spatial vectors through adaptive feature vectorization operations to enrich the semantic segmentation representation of 3D objects.
[0029] Embedding operations can be formulated as follows: ; in, This refers to the splicing operation along the feature channel dimension; This refers to scalar features embedded with normal vectors and curvature information. Initial scalar features; It is the normal vector; The curvature is the curvature after sparsity smoothing.
[0030] Adaptive feature vectorization operation It can be formalized as follows: ; in, scalar features Vectorization operations; Features that are vectorized; It is a scalar feature; , and It is a set of learnable parameters; This is a matrix transpose operation.
[0031] Then, a Rodriguez rotation (rotation around three orthogonal axes in three-dimensional Cartesian space) operation is performed. Vectorized features Processing is performed to obtain the rotated features. .
[0032] Considering that translation or rotation changes the point cloud as a whole, while scaling or jitter alters the relative positions or distances of points, this can lead to unstable feature representations in the feature space for 3D object semantic segmentation methods, easily causing false detections in practical applications. Therefore, this invention introduces a relative direction thinning mechanism. It is formalized as follows: ; in, To Relative direction refinement operation; These are the feature vectors refined in relative directions; For the cropping operation, it will Limited to Within the range; From the input The neural network was trained to obtain it; It is a set of triples that contains the neighborhood. The average of the absolute values of the features of different feature channels of the points in the middle on three orthogonal bases; for Features obtained after Rodriguez rotation.
[0033] Finally, the refined feature vectors are restored to their original scalar channel count through grouped convolution, and then summed to output the final orientation consistency features. It is formalized as follows: ; in, This represents a grouped convolution that restores the feature vector to a scalar number of channels; This represents the summation operation used to achieve feature aggregation and the representation of the disordered nature of point clouds; These are the feature vectors refined in relative directions; It is a scalar feature.
[0034] Geometrically enhanced directional consistency representation enables the method to maintain robust and accurate representation of point clouds under perturbation conditions, thereby improving the perturbation resistance of 3D object semantic segmentation methods.
[0035] Example 1: Performance Verification; To verify the stability of the proposed method for 3D object semantic segmentation under various point cloud perturbations, the S3DIS dataset was used as the test data. Eight mainstream 3D semantic segmentation methods, including PointNet++, MinkowskiNet, Point Transformer, PAConv, StratifiedTransformer, PointNeXt, PointVector, and EP-Net, were selected as comparison objects. Nine test scenarios were designed, including one unperturbed scenario (basic performance test), three rotational perturbation scenarios (rotation of 90°, 180°, and 270° respectively), two translational perturbation scenarios (translation of 0.2m and -0.2m respectively), two scaling perturbation scenarios (scaling by 0.8x and 1.2x respectively), and one jitter perturbation scenario (adding random jitter noise). The segmentation accuracy was used as the evaluation index to compare the performance of each method under different scenarios. The results are shown in Table 1.
[0036] Table 1. Comparison of the robustness test performance of semantic segmentation of 3D objects (bolded optimal results)
[0037] As shown in Table 1, in the perturbation resistance test of 3D object semantic segmentation on the S3DIS dataset, the method of this invention maintains high accuracy and stability under various rotation, translation, scaling and jitter perturbations, which is significantly better than the comparative methods such as PointNet++, MinkowskiNet, and Point Transformer.
[0038] To further verify the rotation invariance of the method of this invention, the S3DIS dataset was used as the test data. Six mainstream methods, including PointNet, RIConv, RandLA-Net, Point Transformer, RRI, and PointNeXt, were selected as comparison objects. Three types of rotation invariance test scenarios were designed: z / z scenario (both the input and target point clouds rotate along the z-axis), z / SO3 scenario (the input point cloud rotates along the z-axis, and the target point cloud rotates along any axis in three-dimensional space), and SO3 / SO3 scenario (both the input and target point clouds rotate along any axis in three-dimensional space). The segmentation accuracy was used as the evaluation index to compare the adaptability of each method under different rotation scenarios. The results are shown in Table 2.
[0039] Table 2. Performance Comparison of Rotation Invariance Test for Semantic Segmentation of 3D Objects (Optimal Result in Bold)
[0040] As shown in Table 2, the method of the present invention performs best in different rotation scenarios, and its segmentation accuracy is higher than that of the comparative methods such as PointNet, RIConv, and RRI, reflecting its excellent semantic segmentation performance against disturbances in 3D objects.
[0041] In addition, visualizations of some 3D object semantic segmentation results (see...) Figure 2 This further confirms the performance advantages of the method of the present invention. From the overall segmentation integrity to the detail accuracy, its segmentation effect is better than other methods, and it has excellent semantic segmentation performance.
[0042] Based on the above verification results, it can be seen that the method of the present invention not only has significant performance advantages in the semantic segmentation of three-dimensional objects, but also has excellent point cloud anti-disturbance performance, and can effectively deal with various point cloud disturbance problems in practical applications.
[0043] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.
Claims
1. A perturbation-resistant geometric enhancement method for semantic segmentation of three-dimensional objects, characterized in that, Includes the following steps: Acquire the input 3D point cloud data; Statistical geometric feature inference calculations are performed on 3D point cloud data to obtain point cloud geometric features; The specific process is as follows: First, construct a decentralized sub-point cloud vector set. Then, use the sum of squares to statistically quantify the distribution and clustering degree of the local sub-point cloud in different directions. Transform the solution of the normal vector of the local sub-point cloud into an optimization problem and solve it to obtain the normal vector of the neighborhood. Then, the initial curvature is calculated by combining the eigenvalues, and the initial curvature is corrected by introducing the sparsity smoothing factor and the sparsity adjustment temperature to obtain the curvature after sparsity smoothing. Geometrically enhanced orientation consistency feature extraction is performed on the geometric features of the point cloud; The specific process is as follows: First, the obtained normal vector and curvature information are embedded into scalar features to form geometrically enhanced scalar features; then, the scalar features are expanded into independent spatial vectors through adaptive feature vectorization operations; next, the spatial vectors are processed by Rodriguez rotation operations, and a relative direction refinement mechanism is introduced to adjust the relative direction of the features; finally, the number of scalar channels is recovered through grouped convolution, and feature aggregation and point cloud disorder representation are achieved through summation operations to obtain a robust point cloud representation for semantic segmentation of 3D objects.
2. The disturbance-resistant geometric enhancement method for semantic segmentation of three-dimensional objects according to claim 1, characterized in that, The formula for constructing the decentralized sub-point cloud vector set is as follows: ； in, For a decentralized sub-point cloud vector set; For the first A decentralized point; For the neighboring region One point; Representing the neighborhood The average position of all points in; The index of the point; Indicates the number of points in the neighborhood; This refers to the local neighborhood of the point cloud.
3. The disturbance-resistant geometric enhancement method for semantic segmentation of three-dimensional objects according to claim 1, characterized in that, The formula for transforming the solution of the normal vector of the local sub-point cloud into an optimization problem is as follows: ； in, To optimize the objective; Let be the unit normal vector of the neighborhood to be determined; It is a three-dimensional Cartesian space; The number of points in the neighborhood; for transpose; It is the product of the decentralized sub-point cloud vector set and its transpose; right The formula for decomposition is: ； in, and All are 3x3 square matrices; It is a diagonal matrix, and the three values of its diagonal are... Three eigenvalues; The The list is The Middle eigenvalues Corresponding feature vector .
4. The disturbance-resistant geometric enhancement method for semantic segmentation of three-dimensional objects according to claim 1, characterized in that, The formula for calculating the initial curvature is: ； in, The initial curvature; It is the smallest eigenvalue; It is the largest eigenvalue; These are intermediate eigenvalues; It is a very small positive integer.
5. The disturbance-resistant geometric enhancement method for semantic segmentation of three-dimensional objects according to claim 1, characterized in that, The formula for obtaining the curvature after sparsity smoothing is: ； in, The curvature after sparsity smoothing; This indicates the sparsity adjustment temperature used to adjust the strength of smoothness. The number of points in the neighborhood; The index of the point; Center point; for , neighborhood points; For point The initial curvature at that point.
6. The disturbance-resistant geometric enhancement method for semantic segmentation of three-dimensional objects according to claim 1, characterized in that, The formula for the adaptive feature vectorization operation is: ； in, For scalar characteristics Vectorization operations; Features that are vectorized; It is a scalar feature; , and It is a set of learnable parameters; This is a matrix transpose operation.
7. The disturbance-resistant geometric enhancement method for semantic segmentation of three-dimensional objects according to claim 1, characterized in that, The formula for the relative direction refinement mechanism is: ； in, To Relative direction refinement operation; These are the feature vectors refined in relative directions; For the cropping operation, it will Limited to Within the range; From the input The neural network was trained to obtain it; It is a set of triples containing the neighborhood The average of the absolute values of the features of different feature channels of the points in the middle on three orthogonal bases; for Features obtained after Rodriguez rotation.
8. The disturbance-resistant geometric enhancement method for semantic segmentation of three-dimensional objects according to claim 1, characterized in that, The final point cloud representation output by the geometrically enhanced orientation consistency feature extraction step is calculated using the following formula: ； in, This is a characteristic of directional consistency; This represents a grouped convolution that restores the feature vector to a scalar number of channels; This represents the summation operation used to achieve feature aggregation and the representation of the disordered nature of point clouds; These are the feature vectors refined in relative directions; It is a scalar feature.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-8.
10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-8.