Feature-enhanced point cloud segmentation method, device and equipment

Abstract

The invention discloses a feature-enhanced point cloud segmentation method, device and equipment. The method comprises the following steps: constructing a self-adaptive local field according to initial point cloud data obtained after screening processing; respectively carrying out geometric and density analysis according to the self-adaptive local field to obtain geometric features and density features; performing fusion enhancement processing on the geometric features and the density features based on a lightweight attention network to obtain enhanced feature descriptors; point hierarchy screening analysis based on corner confidence is carried out according to the enhanced feature descriptors, and a refined corner point set is obtained; and constructing a graph structure according to the refined corner point set, and carrying out corner segmentation operation based on connectivity analysis to obtain a corner point cloud segmentation result. The method can solve the technical problems that the prior art is easily influenced by noise by depending on a single feature, cannot adapt to specific scene data characteristics by depending on fixed threshold analysis, and is lack of accuracy, integrity and reliability in actual segmentation due to the fact that the association relationship among multiple features is easily ignored in the segmentation process.

Description

Technical Field

[0001] This application relates to the field of three-dimensional point cloud data processing, and in particular to a point cloud segmentation method, apparatus and device for enhancing features. Background Technology

[0002] In engineering fields such as power line inspection, industrial 3D measurement, intelligent manufacturing, and substation equipment modeling, 3D point clouds have become the core data form for acquiring spatial structural information. Relying on sensing devices such as LiDAR, structured light scanning, depth cameras, and binocular vision reconstruction, high-density point cloud data of scenes can be quickly acquired, providing a data foundation for applications such as target recognition, size detection, defect analysis, and 3D reconstruction. Point cloud edges and corners, as key features of an object's geometric structure, carry core information such as contours, boundaries, and transitions. Their accurate segmentation and extraction are crucial prerequisites for achieving target localization, component identification, and model matching. The completeness, accuracy, and robustness of the segmentation results directly determine the reliability and efficiency of subsequent tasks such as 3D modeling, condition monitoring, and intelligent detection.

[0003] As scene complexity increases and point cloud data scales expand, traditional processing methods based on manual annotation and simple geometric calculations are no longer sufficient to meet automation demands. Current mainstream point cloud corner segmentation technologies largely revolve around geometric feature calculation and region discrimination. One type of method directly relies on basic geometric attributes of the point cloud, such as local curvature, normal vectors, and the angle between normal vectors, using threshold judgments to filter corner points. Another type combines 2D image information with a dual-modal fusion strategy, integrating image semantic information with point cloud geometric information to improve segmentation consistency. Still other methods optimize feature representation capabilities through local semantic enhancement, ordered feature encoding, and mask self-supervised learning to improve segmentation performance in complex scenes. While these technologies can achieve basic segmentation functions in regular scenes with low noise and uniform density, they generally suffer from insufficient adaptability, poor stability, and limited accuracy in practical industrial and power scenarios.

[0004] Existing technologies suffer from significant technical limitations. First, their feature design is simplistic and lacks robustness. Most algorithms overly rely on single geometric features such as curvature or normal vectors, failing to fully integrate multi-dimensional information like local density and feature change rate. This leads to false positives and false negatives in noisy, occluded, or sparse point cloud regions, and they are weak in distinguishing between gentle and weak corners, resulting in ambiguous boundary localization. Second, their threshold mechanisms are rigid and lack generalization ability. Existing methods generally use manually set fixed thresholds, unable to dynamically adjust based on point cloud density, noise levels, and local structural differences. This leads to incomplete segmentation in sparse regions and over-segmentation in dense regions. Accuracy fluctuates significantly when migrating across devices and scenes, making them unsuitable for adapting to variable environments such as substations and industrial sites. Third, they lack multi-feature correlation and structural integrity. Algorithms often focus on judging local single-point features, ignoring geometric constraints and feature collaboration between regions. This results in broken and discretized continuous corners, and they cannot effectively distinguish between real corners and high-density noise or pseudo-structures, easily generating a large number of invalid candidate points and reducing segmentation reliability. It can be observed that traditional smoothing and filtering operations tend to blur edge details, and the lack of geometry-preserving optimization strategies makes it difficult to achieve a balance between noise reduction and edge sharpness, ultimately affecting model accuracy and detection reliability. Summary of the Invention

[0005] This application provides a point cloud segmentation method, apparatus, and device with enhanced features, which addresses the technical problems of existing technologies that rely on single features and are easily affected by noise, while relying on fixed threshold analysis cannot adapt to specific scene data characteristics, and the segmentation process easily ignores the correlation between multiple features, resulting in a lack of accuracy, completeness, and reliability in actual segmentation.

[0006] In view of this, the first aspect of this application provides a point cloud segmentation method with enhanced features, comprising:

[0007] An adaptive local neighborhood is constructed based on the initial point cloud data obtained after filtering and processing;

[0008] Geometric and density analyses are performed on the adaptive local neighborhood to obtain geometric and density features;

[0009] The geometric features and density features are fused and enhanced using a lightweight attention network to obtain an enhanced feature descriptor;

[0010] Based on the enhanced feature descriptor, a point-level filtering analysis based on corner confidence is performed to obtain a refined corner point set;

[0011] Based on the refined corner point set, a graph structure is constructed, and a corner segmentation operation based on connectivity analysis is performed to obtain the corner point cloud segmentation result.

[0012] Preferably, the step of constructing an adaptive local neighborhood based on the initial point cloud data obtained after filtering includes:

[0013] The K-nearest neighbor algorithm is used to calculate the average distance of the point cloud based on the original point cloud data;

[0014] Outliers are removed from the original point cloud data based on the average point cloud distance to obtain initial point cloud data.

[0015] An adaptive local neighborhood is constructed based on the initial point cloud data, and the adaptive local neighborhood includes a neighborhood radius.

[0016] Preferably, the step of performing geometric and density analyses based on the adaptive local neighborhood to obtain geometric features and density features includes:

[0017] Principal component analysis is performed in the adaptive local neighborhood to calculate the surface normal vector and curvature, thereby obtaining the geometric properties;

[0018] Geometric features are calculated based on the geometric properties and the adaptive local neighborhood, the geometric features including curvature change intensity and normal vector divergence.

[0019] The density features are obtained by performing normalized local density calculation based on the adaptive local neighborhood.

[0020] Preferably, the step of fusing and enhancing the geometric features and the density features based on a lightweight attention network to obtain an enhanced feature descriptor includes:

[0021] The geometric features and the density features are standardized and concatenated to obtain the initial fused features;

[0022] A lightweight attention network is used to generate a fusion weight vector, and an enhanced fusion calculation is performed based on the fusion weight vector and the initial fusion features to obtain an enhanced feature descriptor.

[0023] Preferably, the step of performing a point-level filtering analysis based on corner confidence using the enhanced feature descriptor to obtain a refined corner point set includes:

[0024] Calculate the corner confidence based on the enhanced feature descriptor;

[0025] A coarse point cloud screening is performed based on a preset voxel segmentation threshold and the corner confidence level to obtain a coarse corner point set;

[0026] After removing isolated points from the coarse corner point set, a constraint-based region growing analysis is performed based on the normal vector and the corner confidence to obtain a refined corner point set.

[0027] Preferably, the step of performing edge-corner segmentation based on connectivity analysis to obtain edge-corner point cloud segmentation results by constructing a graph structure using the refined edge-corner point set includes:

[0028] The refined corner point set is subjected to bilateral filtering using a Gaussian kernel function to obtain the filtered corner point set;

[0029] The filtered corner point set is constructed into a graph structure, and path connection completion processing based on fracture endpoint analysis is performed according to the distance threshold to obtain the connected path;

[0030] Calculate the comprehensive saliency score of each of the connected paths, and filter out pseudo-structures based on the score threshold to achieve corner segmentation and obtain the corner point cloud segmentation result.

[0031] A second aspect of this application provides a point cloud segmentation apparatus for enhancing features, comprising:

[0032] The domain construction unit is used to construct an adaptive local domain based on the initial point cloud data obtained after filtering and processing.

[0033] The feature analysis unit is used to perform geometric and density analyses based on the adaptive local neighborhood to obtain geometric features and density features;

[0034] The fusion enhancement unit is used to perform fusion enhancement processing on the geometric features and the density features based on a lightweight attention network to obtain an enhanced feature descriptor.

[0035] The point cloud filtering unit is used to perform point-level filtering analysis based on corner confidence based on the enhanced feature descriptor to obtain a refined corner point set.

[0036] The point cloud segmentation unit is used to perform edge and corner segmentation operations based on connectivity analysis according to the graph structure constructed by the refined edge and corner point set, so as to obtain the edge and corner point cloud segmentation results.

[0037] Preferably, the point cloud filtering unit is specifically used for:

[0038] Calculate the corner confidence based on the enhanced feature descriptor;

[0039] A coarse point cloud screening is performed based on a preset voxel segmentation threshold and the corner confidence level to obtain a coarse corner point set;

[0040] After removing isolated points from the coarse corner point set, a constraint-based region growing analysis is performed based on the normal vector and the corner confidence to obtain a refined corner point set.

[0041] Preferably, the point cloud segmentation unit is specifically used for:

[0042] The refined corner point set is subjected to bilateral filtering using a Gaussian kernel function to obtain the filtered corner point set;

[0043] The filtered corner point set is constructed into a graph structure, and path connection completion processing based on fracture endpoint analysis is performed according to the distance threshold to obtain the connected path;

[0044] Calculate the comprehensive saliency score of each of the connected paths, and filter out pseudo-structures based on the score threshold to achieve corner segmentation and obtain the corner point cloud segmentation result.

[0045] A third aspect of this application provides a point cloud segmentation device with enhanced features, the device including a processor and a memory;

[0046] The memory is used to store program code and transmit the program code to the processor;

[0047] The processor is configured to execute the point cloud segmentation method with enhanced features as described in the first aspect, according to instructions in the program code.

[0048] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0049] This application provides a point cloud segmentation method with enhanced features, comprising: constructing an adaptive local neighborhood based on the initial point cloud data obtained after screening; performing geometric and density analyses on the adaptive local neighborhood to obtain geometric features and density features; fusing and enhancing the geometric features and density features based on a lightweight attention network to obtain an enhanced feature descriptor; performing a point number-level screening analysis based on corner confidence based on the enhanced feature descriptor to obtain a refined corner point set; and constructing a graph structure based on the refined corner point set to perform corner segmentation operations based on connectivity analysis to obtain the corner point cloud segmentation result.

[0050] The enhanced feature point cloud segmentation method provided in this application integrates the geometric and density features of point cloud data to construct a multi-dimensional enhanced feature descriptor, strengthening the feature recognition of corner regions and overcoming the robustness problem caused by existing technologies relying on single features. Furthermore, by using a point-count hierarchical screening analysis based on corner confidence, reliable corner point sets can be quickly and accurately determined, replacing fixed threshold analysis and overcoming the problems of poor adaptability and accuracy in segmentation methods. In addition, graph structure analysis can be used to perform connectivity path analysis on discrete corner points, fully considering the correlation between multiple features and solving integrity problems such as continuous corner fragmentation and false corner misjudgment. Therefore, this application can solve the technical problems of existing technologies relying on single features being easily affected by noise, and relying on fixed threshold analysis failing to adapt to specific scene data characteristics, and the segmentation process easily ignoring the correlation between multiple features, resulting in a lack of accuracy, completeness, and reliability in actual segmentation.

[0051] Specifically, an adaptive local neighborhood is constructed based on the initial point cloud data. This allows for dynamic adjustment of the support range according to the local distribution characteristics of the point cloud, fundamentally improving the poor adaptability of traditional fixed neighborhoods and fixed thresholds. This avoids over-segmentation, incomplete segmentation, and cross-scene accuracy fluctuations caused by differences in point cloud density and noise. Geometric and density feature analysis are simultaneously performed on the adaptive neighborhood. By introducing multi-dimensional complementary features, this directly addresses the problems of insufficient robustness, noise sensitivity, and difficulty in accurately distinguishing smooth corners caused by the over-reliance on single geometric features in existing technologies, providing a more comprehensive and reliable basis for corner discrimination. Next, a lightweight attention network is used to fuse and enhance the two types of features. Adaptive weight allocation strengthens effective features and suppresses interfering information, compensating for the shortcomings of existing technologies that neglect multiple features. To address the shortcomings of insufficient feature association and discriminative power, this method further enhances the feature recognition of corner regions. Subsequently, based on enhanced feature descriptors, a hierarchical screening of corner confidence points is conducted. This hierarchical screening and localized judgment replace the traditional fixed threshold segmentation, effectively reducing false corner misjudgments, isolated point interference, and boundary ambiguity, thus improving the accuracy of corner recognition. Finally, a graph structure is constructed based on a refined corner point set, and connectivity analysis and segmentation operations are performed. Through connected component verification, break completion, and invalid region filtering, the problems of discrete and fragmented corner segmentation and insufficient integrity in existing technologies are solved, ensuring that the final output corner results maintain structural continuity and morphological integrity. Through the sequential connection and synergistic effect of the above steps, the technical pain points of poor robustness, low threshold adaptability, and insufficient segmentation integrity in existing methods are comprehensively solved. Attached Figure Description

[0052] Figure 1 A flowchart illustrating a point cloud segmentation method with enhanced features provided in an embodiment of this application;

[0053] Figure 2 This is a schematic diagram of a point cloud segmentation device for enhancing features, provided in an embodiment of this application. Detailed Implementation

[0054] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0055] For easier understanding, please refer to Figure 1 An embodiment of a point cloud segmentation method with enhanced features provided in this application includes:

[0056] Step 101: Construct an adaptive local neighborhood based on the initial point cloud data obtained after filtering.

[0057] It should be noted that the initial point cloud data in this embodiment has already undergone statistical filtering and outlier removal operations, and has removed obvious noise and redundant points, retaining effective structural information. To overcome the shortcomings of unstable feature estimation in sparse regions and loss of details in dense regions due to fixed neighborhoods, this embodiment adopts a density-adaptive neighborhood construction strategy.

[0058] Specifically, the neighborhood radius is dynamically adjusted based on the local point cloud density and the global average point distance: it automatically expands the neighborhood range in sparse point cloud regions to ensure sufficient local support for the calculation of geometric quantities such as normal vectors and curvature; and it automatically shrinks the neighborhood in dense point cloud regions to prevent structural information from being smoothed out. Neighborhood construction can choose from K-nearest neighbors, dynamic spherical neighborhoods, or hybrid neighborhood methods; this is just an example and not a limitation. The neighborhood size is adaptively determined by the point cloud distribution, without the need for manually preset fixed parameters, providing a stable and accurate local structural foundation for subsequent feature analysis.

[0059] Step 102: Perform geometric and density analyses based on the adaptive local neighborhood to obtain geometric features and density features.

[0060] This embodiment extracts complementary features from two dimensions: structural morphology and spatial distribution, thereby improving the reliability of corner recognition. Geometric features are calculated based on the spatial surface characteristics of the local neighborhood, mainly including normal vector, curvature, rate of change of curvature, and normal vector divergence. These features can accurately characterize the local transitions, abrupt changes, and smooth transitions of point clouds, and have good distinguishing ability for structures such as sharp corners, weak corners, and edges.

[0061] Density features are obtained by statistically analyzing the ratio of the number of effective points in the neighborhood to the volume of the neighborhood space and then performing logarithmic normalization. This describes the density distribution of local point sets and can effectively distinguish between true structural edges and false edges formed by noise aggregation or point cloud accumulation. Geometric features focus on structural abrupt changes, while density features focus on spatial distribution; combining the two can significantly improve the richness and robustness of edge and corner discrimination criteria.

[0062] Step 103: Based on a lightweight attention network, geometric features and density features are fused and enhanced to obtain enhanced feature descriptors.

[0063] First, the geometric and density features are standardized to eliminate dimensional differences and then concatenated into a multidimensional original feature vector. This vector is then input into a lightweight attention network, where adaptive weights are assigned to features of different dimensions through learnable linear transformations and activation functions.

[0064] This network structure is lightweight and computationally inefficient, making it suitable for real-time processing of large-scale point clouds. It can automatically enhance feature dimensions that contribute significantly to corner recognition while suppressing noise-sensitive and low-discrimination feature dimensions, achieving adaptive feature enhancement. After attention-weighted fusion, an enhanced feature descriptor with high discriminative power and low noise interference is obtained. This descriptor can highlight the feature responses of real corner regions and weaken the interference from non-target regions, providing high-quality input for subsequent confidence calculation.

[0065] Step 104: Perform point-level filtering analysis based on corner confidence based on the enhanced feature descriptor to obtain a refined corner point set.

[0066] This embodiment is based on enhanced feature descriptors and calculates the corner confidence of each point through a nonlinear mapping function. The confidence value directly reflects the probability that the point belongs to a corner. To balance segmentation efficiency and accuracy, this embodiment adopts a coarse-to-fine filtering hierarchical screening mechanism: first, segmentation thresholds are dynamically generated based on the mean and standard deviation of the confidence scores, using local voxels as units, to quickly screen out high-confidence candidate corner points and eliminate a large number of obviously non-corner points; then, through neighborhood point number constraints, spatial distance constraints, and normal vector consistency constraints, isolated points, scattered points, and pseudo-corner points are eliminated, retaining continuous and structurally stable corner segments.

[0067] This dynamic threshold strategy avoids the problem of poor adaptability of fixed thresholds in different scenarios, and can maintain a stable screening effect under complex noise and uneven density conditions, ultimately obtaining a refined set of edge points with continuous distribution and accurate location.

[0068] Step 105: Based on the refined edge and corner point set, construct a graph structure and perform edge and corner segmentation operation based on connectivity analysis to obtain the edge and corner point cloud segmentation result.

[0069] This embodiment abstracts the refined set of corner points into an undirected graph structure, using spatial points as nodes and spatial proximity, normal vector similarity, and confidence correlation as edge weights to construct local topological relationships. Based on the graph structure, connected component analysis is performed to identify broken corners, discrete segments, and isolated fragments. Broken corners are then completed and connected according to spatial distance thresholds and geometric continuity rules, and adjacent segments are merged.

[0070] Simultaneously, comprehensive significance indicators such as the length, mean confidence score, and spatial compactness of each connected component are calculated to filter out pseudo-corner regions that are too short, have low significance, or have no practical structural meaning, while retaining complete, continuous corner targets that conform to physical structure. The final output corner point cloud segmentation result has advantages such as high integrity, few pseudo edges, and good edge sharpness, and can be directly used for subsequent tasks such as 3D modeling, size detection, and defect identification.

[0071] This embodiment achieves high-precision and high-completeness extraction of point cloud corners in complex scenes through adaptive neighborhood construction, multi-dimensional feature extraction, attention feature fusion, hierarchical confidence filtering, and graph structure connectivity segmentation. Furthermore, by employing adaptive neighborhood, multi-feature fusion, dynamic thresholding, and graph structure connectivity optimization, a complete and robust point cloud corner segmentation process is formed. This process can achieve high-precision and high-completeness corner extraction even in real-world scenarios with noise interference, uneven point cloud density, and complex and variable structures, effectively solving problems such as poor robustness, segmentation fragmentation, and false edge interference inherent in traditional methods. In addition, this method is applicable to 3D point clouds acquired by various devices such as LiDAR, depth cameras, and structured light scanning, and can operate stably in scenarios such as substation equipment inspection, industrial parts measurement, and 3D model reconstruction.

[0072] The point cloud segmentation method with enhanced features provided in this application integrates the geometric and density features of point cloud data to construct a multi-dimensional enhanced feature descriptor, strengthening the feature recognition of corner regions and overcoming the robustness problem caused by the reliance on single features in existing technologies. Furthermore, by using a point-count-level filtering analysis based on corner confidence, a reliable set of corner points can be quickly and accurately determined, thus overcoming the poor adaptability and accuracy of segmentation methods by replacing fixed threshold analysis. In addition, the graph structure allows for connectivity path analysis of discrete corner points, fully considering the correlation between multiple features and solving integrity problems such as continuous corner fragmentation and false corner misjudgment. Therefore, this application can solve the technical problems of existing technologies that rely on single features and are easily affected by noise, while relying on fixed threshold analysis cannot adapt to specific scene data characteristics, and the segmentation process easily ignores the correlation between multiple features, resulting in a lack of accuracy, completeness, and reliability in actual segmentation.

[0073] Specifically, an adaptive local neighborhood is constructed based on the initial point cloud data. This allows for dynamic adjustment of the support range according to the local distribution characteristics of the point cloud, fundamentally improving the poor adaptability of traditional fixed neighborhoods and fixed thresholds. This avoids over-segmentation, incomplete segmentation, and cross-scene accuracy fluctuations caused by differences in point cloud density and noise. Geometric and density feature analysis are simultaneously performed on the adaptive neighborhood. By introducing multi-dimensional complementary features, this directly addresses the problems of insufficient robustness, noise sensitivity, and difficulty in accurately distinguishing smooth corners caused by the over-reliance on single geometric features in existing technologies, providing a more comprehensive and reliable basis for corner discrimination. Next, a lightweight attention network is used to fuse and enhance the two types of features. Adaptive weight allocation strengthens effective features and suppresses interfering information, compensating for the shortcomings of existing technologies that neglect multiple features. To address the shortcomings of insufficient feature association and discriminative power, this method further enhances the feature recognition of corner regions. Subsequently, based on enhanced feature descriptors, a hierarchical screening of corner confidence points is conducted. This hierarchical screening and localized judgment replace the traditional fixed threshold segmentation, effectively reducing false corner misjudgments, isolated point interference, and boundary ambiguity, thus improving the accuracy of corner recognition. Finally, a graph structure is constructed based on a refined corner point set, and connectivity analysis and segmentation operations are performed. Through connected component verification, break completion, and invalid region filtering, the problems of discrete and fragmented corner segmentation and insufficient integrity in existing technologies are solved, ensuring that the final output corner results maintain structural continuity and morphological integrity. Through the sequential connection and synergistic effect of the above steps, the technical pain points of poor robustness, low threshold adaptability, and insufficient segmentation integrity in existing methods are comprehensively solved.

[0074] As a further preferred embodiment of the previous one, step 101 includes:

[0075] The K-nearest neighbor algorithm is used to calculate the average distance of the point cloud based on the original point cloud data;

[0076] Outliers in the original point cloud data are removed based on the average distance of the point cloud, resulting in the initial point cloud data.

[0077] An adaptive local neighborhood is constructed based on the initial point cloud data. The adaptive local neighborhood includes the neighborhood radius.

[0078] It should be noted that the original point cloud data in this embodiment can come from different 3D sensing devices, such as LiDAR scanning, structured light measurement, binocular vision reconstruction, depth camera acquisition, or industrial CT scanning, etc., and the specific source is not limited here. The original point cloud data is an unordered set of points, which can be denoted as... Each point contains three-dimensional coordinate information of x, y, and z; basic preprocessing of the raw point cloud data can facilitate subsequent data analysis and calculation.

[0079] In this embodiment, the K-nearest neighbor algorithm is used to calculate the average distance between the original point cloud data, which can reflect the sparsity density of the point cloud data. This allows us to identify point cloud data that deviates from the target data and remove these outliers to obtain point cloud data with higher density.

[0080] The core principle of the K-Nearest Neighbors algorithm is as follows: For each data point in the point cloud, search for the K nearest neighbors in the spatial range, where K is a preset number of neighboring points, typically 8, 16, or 32, which can be flexibly chosen according to the density of the point cloud. For each point, calculate the Euclidean distance between it and its K nearest neighbors, and take the average to obtain the local average distance of that point; then, perform global statistics on the local average distances of all points to obtain the overall average distance of the point cloud. This average distance of the point cloud can objectively reflect the overall spatial distribution density of the point cloud and is a key benchmark parameter for subsequent outlier removal and neighborhood radius determination, effectively avoiding the instability caused by parameters set by human experience.

[0081] Specifically, the point cloud average distance is calculated by taking the average distance between each point and its K nearest neighbors:

[0082] ;

[0083] in, The average distance of the point cloud. For the first The first point K is the nearest neighbor point, where K is the nearest neighbor point. The total number of nearest neighbors of a point.

[0084] Based on the calculated average distance of the point cloud, distance-based filtering can be performed to remove outliers whose average distance exceeds X times the standard deviation, thus obtaining the point set composed of the initial point cloud data. This point set can be used in subsequent calculation and analysis tasks.

[0085] This embodiment implements outlier removal based on statistical filtering principles. Using the average distance of the point cloud as a basis, it calculates the mean and standard deviation of the local distances of all points to construct dynamic discrimination rules. Points whose local distances significantly deviate from the global distribution are identified as outliers and removed. These outliers are typically caused by factors such as equipment noise, environmental interference, and abnormal reflections, which can severely affect the accuracy of geometric calculations such as normal vectors and curvature.

[0086] Outlier removal can employ two optional strategies: the first is a global fixed-multiple threshold method, suitable for point clouds with uniform density; the second is a local region adaptive threshold method, suitable for scenarios with large density differences. Both methods use the average distance of the point cloud as a benchmark, which can retain effective structural points while removing noise points to the greatest extent, ultimately obtaining clean, stable initial point cloud data without obvious outliers, providing high-quality input for subsequent neighborhood construction.

[0087] Taking one method as an example, for each point in the initial point cloud data Constructing an adaptive local neighborhood Set the nearest neighbor search number K=200, calculate the average distance between each point and its 200 nearest neighbors, and remove points that deviate from the search. Outliers whose mean exceeds 2 standard deviations are identified; then, all means are used as the global average distance, and a reference radius of 3 times the average distance is set. The number of neighborhood points of each point within this radius is counted, and the global average density is calculated. Then, using the average point distance as a temporary search radius, the number of local neighborhood points for each point is counted to obtain the local density estimate. Finally, an adjustable scaling factor is selected, and the obtained parameters are substituted into the formula to finally obtain the radius of the adaptive local neighborhood for each point:

[0088] ;

[0089] in, For global average density, This is a local density estimate. It is an adjustable scaling factor, with a value range of 2.0 to 3.0.

[0090] It is understood that the adaptive local neighborhood in this embodiment refers to the neighborhood range dynamically adjusted by each data point based on its local density, with the core being the adaptive neighborhood radius. The construction principle is as follows: taking each point in the initial point cloud as the center, the neighborhood radius is automatically calculated based on the local point cloud density and the average distance to the global point cloud. Specifically, two methods can be used: The first is a density-based adaptive radius algorithm, where the lower the local density, the larger the neighborhood radius, ensuring sufficient support for geometric calculations in sparse regions; conversely, the higher the local density, the smaller the radius, avoiding excessive smoothing of edge details. The second is a hybrid neighborhood algorithm combining K-nearest neighbors and spherical neighborhoods. The initial neighborhood is first determined by K-nearest neighbors, and then the range is constrained by the spherical radius, balancing stability and consistency. Both algorithms do not require manual radius adjustment and can automatically adapt to complex scenarios such as sparse, dense, and unevenly noisy environments, ensuring that each point has a neighborhood range best suited to its own structure. The adaptive local neighborhood constructed through this step can significantly improve the accuracy and robustness of subsequent calculations of features such as normal vectors, curvature, and density, laying the foundation for accurate corner segmentation. In actual scenarios, you can choose according to your needs; this is just an example and not a limitation.

[0091] As a further preferred embodiment of the previous one, step 102 includes:

[0092] Principal component analysis is performed in the adaptive local neighborhood to calculate the surface normal vector and curvature, thereby obtaining the geometric properties;

[0093] Geometric features are calculated based on geometric properties and adaptive local neighborhood, including curvature variation intensity and normal vector divergence.

[0094] The density features are obtained by performing normalized local density calculation based on adaptive local neighborhood.

[0095] An adaptive local neighborhood constructed for the initial point cloud data points can expand stable estimates in sparse regions and shrink to preserve details in dense regions. The normal vector of the neighborhood surface can be calculated using principal component analysis. and curvature This is the key geometric property for calculating geometric features. The geometric features in this embodiment include, but are not limited to, the intensity of curvature change and the divergence of the normal vector. The specific calculation process is expressed as follows:

[0096] ;

[0097] ;

[0098] in, The intensity of curvature change, This represents the divergence of the normal vector.

[0099] The process of calculating normalized local density based on adaptive local neighborhood is as follows:

[0100] ;

[0101] in, This refers to density characteristics.

[0102] It should be noted that the adaptive local neighborhood in this embodiment has completed the dynamic radius construction based on the point cloud density, which can provide reliable support for local surface estimation. Principal component analysis is a classic algorithm for local geometric estimation of point clouds. Its principle is to construct a three-dimensional covariance matrix for the point set in the neighborhood, and obtain three orthogonal eigenvectors and their corresponding eigenvalues ​​through eigenvalue decomposition. Among them, the eigenvector corresponding to the smallest eigenvalue is the normal vector of that point, which is used to characterize the orientation of the local surface; the ratio of the three eigenvalues ​​can be further calculated to obtain the curvature. The larger the curvature, the more curved the local surface is and the closer it is to the edge and corner structure. When calculating the normal vector, normal vector uniformity processing can be selected, and it can be uniformly corrected according to the direction of the point cloud external normal to avoid the direction jump of the neighborhood normal vector; the curvature calculation can use various forms such as Gaussian curvature, average curvature, or principal curvature. This embodiment prefers principal curvature, which is more sensitive to abrupt changes in edge and corner and has more stable geometric properties.

[0103] Furthermore, the geometric features in this embodiment are used to quantify the degree of change in local structure, serving as a key indicator for distinguishing between planes, gently curved surfaces, and corners. The calculation principle for the curvature change intensity is as follows: taking the current point as the center, traversing all points in the adaptive neighborhood, calculating the absolute difference between the curvature of the current point and the curvature of all points in the neighborhood, and taking the average; the larger the value, the more drastic the local structural change, and the more likely it belongs to a corner region. The normal vector divergence is obtained by calculating the cosine of the angle between the normal vector of the current point and the average normal vector of all points in the neighborhood, used to characterize the consistency of the neighborhood normal vectors: normal vectors in planar regions are highly consistent, with minimal divergence; normal vectors in corner regions rapidly change direction, with significantly increased divergence. These two features can be used complementaryly, describing the structural abrupt change characteristics from the perspectives of numerical change and spatial orientation, respectively, effectively improving the ability to identify weak corners and gently curved corners.

[0104] Local density describes the spatial distribution of point clouds and can effectively distinguish between real structural corners and pseudo-corners formed by noise aggregation. The calculation principle is as follows: count the number of effective points in the adaptive neighborhood and combine it with the sphere volume corresponding to the neighborhood radius to obtain the initial local density; to eliminate the magnitude difference caused by the overall density of the point cloud, the initial density is logarithmically normalized to compress the value to a stable range, resulting in the normalized local density.

[0105] This density feature has a clear physical meaning: the edges of real structures are usually distributed in a regular linear pattern with uniform and stable density; noise and anomalous reflection points tend to cluster together, with abnormally high density. This step provides two optional calculation methods: one is density calculation based on the volume of the sphere's neighborhood, which has higher accuracy and is suitable for high-precision scenarios; the other is fast density calculation based on the proportion of neighborhood points, which is faster and suitable for real-time processing scenarios. The normalized density feature and geometric feature are independent of each other and complement each other, which can significantly improve the anti-interference ability in subsequent edge and corner discrimination and reduce the risk of misjudgment caused by false edges and corners and noise points.

[0106] This embodiment, supported by adaptive neighborhood, utilizes PCA geometric estimation, structural change calculation, and normalized density statistics to output highly robust geometric and density features. This enables the system to maintain stable and reliable feature representation capabilities even in scenarios with complex noise, uneven density, and variable structures, providing high-quality input for subsequent feature fusion enhancement and edge / corner filtering.

[0107] As a further preferred embodiment of the previous one, step 103 includes:

[0108] The geometric features and density features are standardized and concatenated to obtain the initial fused features;

[0109] A lightweight attention network is used to generate a fusion weight vector, and the fusion is enhanced by calculating the fusion weight vector and the initial fusion features to obtain the enhanced feature descriptor.

[0110] It should be noted that since geometric features include the intensity of curvature change and the divergence of the normal vector, while density features are normalized local densities, the numerical range, physical meaning, and dimensions of the two types of features differ significantly. Direct concatenation would lead to high-amplitude features dominating the calculation and low-amplitude features being weakened. Therefore, standardization must be performed first. Standardization uses zero-mean unit variance standardization. The principle is to subtract the global mean from each feature dimension and then divide by the global standard deviation, so that all features are uniformly mapped to a similar numerical range, eliminating the calculation bias caused by differences in dimensions and amplitudes. After standardization, the geometric feature vector and the density feature vector are concatenated end-to-end in dimensional order to form a one-dimensional feature vector with a fixed dimension, which is the initial fused feature. The concatenation method can be either serial direct concatenation or cross-dimensional concatenation. The former has a clear structure and is easy to implement, while the latter has tighter feature coupling and stronger discrimination potential. This embodiment prioritizes serial direct concatenation to balance stability and computational efficiency. The initial fused feature achieves preliminary integration of multi-feature information, but it does not distinguish the contribution of each dimension to corner discrimination and still needs further enhancement and optimization.

[0111] The lightweight attention network in this embodiment is a miniature neural network specifically designed for point cloud edge computing. Its core principle is to automatically learn the importance of different feature dimensions through a self-attention mechanism, assigning high weights to features that contribute significantly to edge and corner discrimination, and low weights to features that are sensitive to noise or have weak discriminative power, thus achieving adaptive feature enhancement. The network structure contains only one linear transformation layer, one activation function layer, and one softmax normalization layer, without deep convolutions or complex modules. It has extremely small parameters, fast computation speed, and can run in real-time on embedded and edge devices.

[0112] In the implementation process, after the initial fused features are input into a lightweight attention network, they are first transformed linearly to complete feature mapping, then an activation function is used to introduce non-linear expressive power, and finally a fused weight vector with values ​​between 0 and 1, summing to 1, is output by a softmax function. The weight vector has the same dimensions as the initial fused features, achieving one-to-one weighting across dimensions. The enhanced fusion calculation adopts an element-wise multiplication method, multiplying each dimension of the initial fused features by its corresponding weight to obtain the weighted feature vector, which is the enhanced feature descriptor.

[0113] This embodiment provides two optional attention implementation forms: the first is channel-based attention, which globally assigns weights to the overall feature dimension, offering strong stability and high versatility; the second is adaptive feature attention, which dynamically adjusts weights for different local structures, making it more adaptable to complex scenarios. After enhanced fusion, the feature descriptor significantly highlights the feature responses of corner regions, weakens noise and interference from non-target regions, and greatly improves discriminative power and noise resistance. It can effectively solve the problems of poor robustness of single features and lack of synergistic utilization of multiple features, providing high-quality and robust feature inputs for subsequent corner confidence calculation and hierarchical selection.

[0114] Specifically, geometric features , With density characteristics The process of concatenating the standardized features into the initial fused feature vector is as follows:

[0115] ;

[0116] in, , , These are the geometric and density features after standardization.

[0117] The lightweight attention network in this embodiment is designed based on a lightweight network architecture and introduces an attention mechanism for feature enhancement and fusion. Specifically, it generates a fusion weight vector based on the lightweight attention network, expressed as follows:

[0118] ;

[0119] in, , These are all learnable network parameters. The process of feature enhancement fusion based on this fusion weight vector is expressed as follows:

[0120] ;

[0121] in, Calculate dot products in the network; This is known as enhanced feature descriptor, a feature vector that characterizes the edge and corner attributes of each point. This mechanism can dynamically adjust the contribution of each feature for different local structures, thereby providing more reliable and accurate feature data for subsequent calculations and analyses.

[0122] As a further preferred embodiment of the previous one, step 104 includes:

[0123] Calculate corner confidence based on enhanced feature descriptors;

[0124] Based on the preset voxel segmentation threshold and corner confidence, a coarse screening of the point cloud is performed to obtain a coarse screening corner point set;

[0125] After removing isolated points from the coarse-screened corner point set, a constraint-based region growing analysis is performed based on the normal vector and corner confidence to obtain a refined corner point set.

[0126] In this embodiment, the corner confidence score is a core indicator used to quantify the probability that each point belongs to a true corner. Its value is typically normalized to between 0 and 1, with higher values ​​indicating a greater likelihood that the point is a corner. The calculation principle is as follows: A highly discriminative enhanced feature descriptor is input into a preset linear mapping function, which undergoes a non-linear transformation via a Sigmoid activation function, mapping the multi-dimensional features into a single confidence scalar. Essentially, this process utilizes the correlation between features and corner structures to complete probability scoring, making the confidence scores for planar and noisy regions approach 0, and the confidence scores for abrupt corner changes approach 1, thus achieving a preliminary distinction between corner and non-corner points. This step provides two optional calculation methods: one is fixed-parameter linear mapping, which is computationally efficient and suitable for embedded deployments; the other is lightweight non-linear mapping, which offers higher accuracy and is suitable for complex scenarios, allowing for flexible selection based on system computing power.

[0127] The core principle of voxel coarse screening is to divide the 3D point cloud space into a uniform, regular voxel grid, and perform local dynamic threshold judgment on a voxel-by-voxel basis, replacing the traditional global fixed threshold and improving adaptability to complex scenes. In practice, the voxel size is first set according to the average distance of the point cloud. Within each voxel, the mean and standard deviation of the confidence score are calculated to generate a dynamic voxel segmentation threshold specific to that voxel. Points with a confidence score greater than the corresponding voxel threshold are identified as candidate corner points, retained, and aggregated into a coarse-screened corner point set. This operation can quickly filter out a large number of points with low confidence scores that are clearly not corners, significantly reducing subsequent computational load. It also avoids the problems of missed segmentation in sparse areas and over-segmentation in dense areas caused by global thresholds, balancing screening efficiency and reliability.

[0128] Isolated points refer to noise points or pseudo-corner points that have very few neighboring points and are discretely distributed in space, significantly affecting the integrity of subsequent region growth and segmentation. The removal principle is as follows: for each point in the coarse-screened corner point set, the number of effective neighboring points within its adaptive local neighborhood is counted. If the number of points in the neighborhood is less than a preset minimum point count threshold, it is determined to be an isolated point and directly removed. The minimum point count threshold can be set to 2, 3, or 5 depending on the point cloud density; a smaller value is used for sparse scenes, and a larger value is used for dense scenes. By removing isolated points, pseudo-corner interference can be eliminated, and noise residual points can be filtered out, making the coarse-screening results more regular and providing stable seed points for subsequent region growth.

[0129] Region growing is a key algorithm for achieving continuous corner extraction. This step employs a dual-constraint growth mechanism to ensure the growth result closely matches the real physical structure. Specifically, the remaining points after removing isolated points are used as initial seed points. The algorithm expands outwards from the seed point to its neighborhood, simultaneously satisfying two constraints: first, the angle between the normal vector of the neighboring point and the current point is less than the maximum angle threshold, ensuring geometric consistency and avoiding erroneous cross-region growth; second, the corner confidence of the neighboring point is higher than the minimum confidence threshold, ensuring only high-confidence corner points are retained. Points satisfying these constraints are added to the current growth region, and the expansion is iterative until no new points can be added. After growth, adjacent connected regions are merged, and fragmented regions with too few points are filtered out, ultimately resulting in a refined set of corner points with continuous structure, clear boundaries, and minimal artifacts. This process effectively solves the problems of easy breakage, discreteness, and high misjudgment rate in traditional single-point judgment, significantly improving the completeness and accuracy of corner extraction.

[0130] It should be noted that the core logic of coarse and fine selection of points based on enhanced feature descriptors is as follows: First, by combining a linear scoring function with Sigmoid activation, the high-dimensional feature vector of the descriptor is mapped to a corner confidence score in the range of 0 to 1. The higher the score, the greater the probability that the point is a corner point. In the coarse selection stage, based on the threshold dynamically generated by voxels, points with scores higher than the threshold are retained as candidate corner point sets, quickly filtering out a large number of non-corner points. In the fine selection stage, isolated points with too few candidate points in the neighborhood of the candidate set are first removed, and then the remaining points are retained. Region growing is performed by combining the confidence threshold and the angle constraint between the normal vectors to merge adjacent regions and filter out tiny regions with too few points, finally obtaining a refined corner point set.

[0131] Specifically, this embodiment employs a hierarchical screening mechanism to perform dual screening of the enhanced feature descriptors, resulting in a more refined set of corner points with higher density. The initial screening stage requires calculating the corner confidence of the enhanced feature descriptors based on a linear scoring function. :

[0132] ;

[0133] in, For the Sigmoid function, , These are trainable parameters.

[0134] Then, the point cloud data points are divided into a voxel grid, and in each voxel grid... The voxel segmentation threshold is dynamically generated internally, i.e., the preset voxel segmentation threshold is:

[0135] ;

[0136] in, , These are the mean and standard deviation of the confidence scores for the interior angles of the voxel lattice, respectively. This is the global sensitivity coefficient, which typically ranges from 0.8 to 1.2.

[0137] Then, based on the preset voxel segmentation threshold, and corner confidence Perform coarse point cloud screening to filter out The points are used as the coarse selection of the corner point set. ,in, The voxel segmentation threshold is the value corresponding to the i-th point.

[0138] Based on the obtained coarsely screened corner point set, further screening of corner points is carried out, mainly to remove isolated points from the point set. First, the coarsely screened corner point set... The number of candidate points in the middle field is less than Isolated points are removed, among which, The seed is a small integer, which can be set according to actual needs, such as 2, 3, or 5. Then, the remaining coarsely screened corner points are used as seeds. Region growing is performed based on constraints configured according to the normal vector and corner confidence. This allows merging of adjacent regions with fewer points, generating a refined corner point set. The constraint is that the angle between the normal vectors is less than... Angle confidence level greater than , Typically, a small number is set, such as between 0.2 and 0.4, and the resulting refined set of corner points can be expressed as: .

[0139] This embodiment uses confidence level as the core, voxel coarse screening to accelerate the process, and constrained growth to ensure precision, to achieve a complete closed loop from coarse localization to fine extraction. The output refined corner point set can be directly used for subsequent graph structure construction and connectivity segmentation, providing solid support for the final high-quality corner segmentation results.

[0140] As a further preferred embodiment of the previous one, step 105 includes:

[0141] A Gaussian kernel function is used to perform bilateral filtering on the refined corner point set to obtain the filtered corner point set;

[0142] The filtered corner point set is constructed into a graph structure, and path connection completion processing based on fracture endpoint analysis is performed according to the distance threshold to obtain the connected path;

[0143] The comprehensive saliency score of each connected path is calculated, and pseudo-structures are filtered out based on the score threshold to achieve corner segmentation and obtain the corner point cloud segmentation result.

[0144] This embodiment aims to improve the completeness and structural rationality of point cloud segmentation. Therefore, it first uses a Gaussian kernel function to perform bilateral filtering on the refined corner point set. The point position update method during the filtering process is as follows:

[0145] ;

[0146] in, , Let these represent the spatial Gaussian kernel function and the eigendomain Gaussian kernel function, respectively. Used to measure the similarity between the spatial locations of two points. It is used to measure the matching degree of two-point enhanced feature descriptors; the Gaussian kernel function can smooth noise while preserving edge sharpness; by updating the position of each point, the updated set of filtered corner points can be obtained.

[0147] Then we can use the refined corner point set Construct a graph structure and detect pairs of points whose distance is less than a distance threshold. The fracture endpoint, where the distance to the threshold Distance from the average point cloud Related, usually set to The distance is between 3 and 10 times that of the original point cloud data. Based on the detected fracture endpoints, geometric paths are searched in the initial point cloud data, and path completion processing is performed to obtain multiple connected paths.

[0148] The specific path completion process is as follows: taking the detected broken endpoints as the starting and ending points, search for intermediate points in the original point cloud that meet the constraints. The constraints are that the angle between the normal vectors of the two points is no greater than 15° and the similarity of the enhanced feature descriptors is no less than 0.7. If an intermediate point that meets the conditions is found, it is directly connected to form a complete path. If no intermediate point is found, 3 to 10 transition points are generated by linear interpolation based on the spatial coordinates and feature attributes of the two endpoints. The transition points must meet the constraints of the two endpoints. Finally, the endpoints and transition points are connected to form a continuous connected path.

[0149] Next, the overall significance score of each connected path is calculated:

[0150] ;

[0151] in, , , This represents the weighting coefficient, with a value between 0 and 1, and satisfies the following conditions: For example, it can take values , , ; This represents the average corner confidence score of all points on a connected path. This represents the length of the connected path, in mm. This indicates the deviation of a connected path, which is the average distance between the path and the ideal straight line, in mm.

[0152] Then compare the comprehensive significance score with the score threshold. ,if If so, the corresponding pseudo-structure is removed; scoring threshold This value is usually determined through experimental analysis of training data or typical scenarios; it is an empirical value. For example, it can be set by observing the actual distribution of corners. The corner point cloud segmentation result obtained from corner segmentation is expressed as... ;Segmenting the corner point set It is the set of all points identified as corners in the original point cloud dataset. In terms of data structure, this set is represented as: a set of three-dimensional coordinates of M points. ,and Each point represents the coordinates of a corresponding point in the original point cloud dataset or its new coordinates obtained through geometric optimization.

[0153] The point cloud segmentation method with enhanced features provided in this application integrates the geometric and density features of point cloud data to construct a multi-dimensional enhanced feature descriptor, strengthening the feature recognition of corner regions and overcoming the robustness problem caused by the reliance on single features in existing technologies. Furthermore, by using a point-count-level filtering analysis based on corner confidence, a reliable set of corner points can be quickly and accurately determined, thus overcoming the poor adaptability and accuracy of segmentation methods by replacing fixed threshold analysis. In addition, the graph structure allows for connectivity path analysis of discrete corner points, fully considering the correlation between multiple features and solving integrity problems such as continuous corner fragmentation and false corner misjudgment. Therefore, this application can solve the technical problems of existing technologies that rely on single features and are easily affected by noise, while relying on fixed threshold analysis cannot adapt to specific scene data characteristics, and the segmentation process easily ignores the correlation between multiple features, resulting in a lack of accuracy, completeness, and reliability in actual segmentation.

[0154] Specifically, an adaptive local neighborhood is constructed based on the initial point cloud data. This allows for dynamic adjustment of the support range according to the local distribution characteristics of the point cloud, fundamentally improving the poor adaptability of traditional fixed neighborhoods and fixed thresholds. This avoids over-segmentation, incomplete segmentation, and cross-scene accuracy fluctuations caused by differences in point cloud density and noise. Geometric and density feature analysis are simultaneously performed on the adaptive neighborhood. By introducing multi-dimensional complementary features, this directly addresses the problems of insufficient robustness, noise sensitivity, and difficulty in accurately distinguishing smooth corners caused by the over-reliance on single geometric features in existing technologies, providing a more comprehensive and reliable basis for corner discrimination. Next, a lightweight attention network is used to fuse and enhance the two types of features. Adaptive weight allocation strengthens effective features and suppresses interfering information, compensating for the shortcomings of existing technologies that neglect multiple features. To address the shortcomings of insufficient feature association and discriminative power, this method further enhances the feature recognition of corner regions. Subsequently, based on enhanced feature descriptors, a hierarchical screening of corner confidence points is conducted. This hierarchical screening and localized judgment replace the traditional fixed threshold segmentation, effectively reducing false corner misjudgments, isolated point interference, and boundary ambiguity, thus improving the accuracy of corner recognition. Finally, a graph structure is constructed based on a refined corner point set, and connectivity analysis and segmentation operations are performed. Through connected component verification, break completion, and invalid region filtering, the problems of discrete and fragmented corner segmentation and insufficient integrity in existing technologies are solved, ensuring that the final output corner results maintain structural continuity and morphological integrity. Through the sequential connection and synergistic effect of the above steps, the technical pain points of poor robustness, low threshold adaptability, and insufficient segmentation integrity in existing methods are comprehensively solved.

[0155] For easier understanding, please refer to Figure 2 This application provides an embodiment of a point cloud segmentation apparatus for enhancing features, comprising:

[0156] Domain construction unit 201 is used to construct an adaptive local domain based on the initial point cloud data obtained after filtering and processing.

[0157] The feature analysis unit 202 is used to perform geometric and density analyses based on the adaptive local neighborhood to obtain geometric features and density features;

[0158] The fusion enhancement unit 203 is used to perform fusion enhancement processing on geometric features and density features based on a lightweight attention network to obtain enhanced feature descriptors;

[0159] The point cloud filtering unit 204 is used to perform point number-level filtering analysis based on corner confidence based on the enhanced feature descriptor to obtain a refined corner point set;

[0160] Point cloud segmentation unit 205 is used to perform edge and corner segmentation operation based on connectivity analysis based on the graph structure constructed by the refined edge and corner point set to obtain edge and corner point cloud segmentation results.

[0161] Furthermore, the point cloud filtering unit 204 is specifically used for:

[0162] Calculate corner confidence based on enhanced feature descriptors;

[0163] Based on the preset voxel segmentation threshold and corner confidence, a coarse screening of the point cloud is performed to obtain a coarse screening corner point set;

[0164] After removing isolated points from the coarse-screened corner point set, a constraint-based region growing analysis is performed based on the normal vector and corner confidence to obtain a refined corner point set.

[0165] Furthermore, the point cloud segmentation unit 205 is specifically used for:

[0166] A Gaussian kernel function is used to perform bilateral filtering on the refined corner point set to obtain the filtered corner point set;

[0167] The filtered corner point set is constructed into a graph structure, and path connection completion processing based on fracture endpoint analysis is performed according to the distance threshold to obtain the connected path;

[0168] The comprehensive saliency score of each connected path is calculated, and pseudo-structures are filtered out based on the score threshold to achieve corner segmentation and obtain the corner point cloud segmentation result.

[0169] It should be noted that the original point cloud data in this embodiment can come from different 3D sensing devices, such as LiDAR scanning, structured light measurement, binocular vision reconstruction, depth camera acquisition, or industrial CT scanning, etc., and the specific source is not limited here. The original point cloud data is an unordered set of points, which can be denoted as... Basic preprocessing of raw point cloud data can facilitate subsequent data analysis and calculation.

[0170] The K-nearest neighbor algorithm is used to calculate the average distance between raw point cloud data, which reflects the sparsity density of the point cloud data. This helps identify point cloud data that deviates significantly from the target data. Based on this, these outliers can be removed, resulting in denser point cloud data. Specifically, the average distance is calculated by taking the average distance between each point and its K nearest neighbors.

[0171] ;

[0172] in, The average distance of the point cloud. For the first The first point K is the nearest neighbor point, where K is the nearest neighbor point. The total number of nearest neighbors of a point.

[0173] Based on the calculated average distance of the point cloud, distance-based filtering can be performed to remove outliers whose average distance exceeds X times the standard deviation, thus obtaining the point set composed of the initial point cloud data. This point set can be used in subsequent calculation and analysis tasks.

[0174] For each point in the initial point cloud data Constructing an adaptive local neighborhood Set the nearest neighbor search number K=200, calculate the average distance between each point and its 200 nearest neighbors, and remove points that deviate from the search. Outliers whose mean exceeds 2 standard deviations are identified; then, all means are used as the global average distance, and a reference radius of 3 times the average distance is set. The number of neighborhood points of each point within this radius is counted, and the global average density is calculated. Then, using the average point distance as a temporary search radius, the number of local neighborhood points for each point is counted to obtain the local density estimate. Finally, an adjustable scaling factor is selected, and the obtained parameters are substituted into the formula to finally obtain the radius of the adaptive local neighborhood for each point:

[0175] ;

[0176] in, For global average density, This is a local density estimate. It is an adjustable scaling factor, with a value range of 2.0 to 3.0.

[0177] An adaptive local neighborhood constructed for the initial point cloud data points can expand stable estimates in sparse regions and shrink to preserve details in dense regions. The normal vector of the neighborhood surface can be calculated using principal component analysis. and curvature This is the key geometric property for calculating geometric features. The geometric features in this embodiment include, but are not limited to, the intensity of curvature change and the divergence of the normal vector. The specific calculation process is expressed as follows:

[0178] ;

[0179] ;

[0180] in, The intensity of curvature change, This represents the divergence of the normal vector.

[0181] The process of calculating normalized local density based on adaptive local neighborhood is as follows:

[0182] ;

[0183] in, This refers to density characteristics.

[0184] Geometric features , With density characteristics The process of concatenating the standardized features into the initial fused feature vector is as follows:

[0185] ;

[0186] in, , , These are the geometric and density features after standardization.

[0187] The lightweight attention network in this embodiment is designed based on a lightweight network architecture and introduces an attention mechanism for feature enhancement and fusion. Specifically, it generates a fusion weight vector based on the lightweight attention network, expressed as follows:

[0188] ;

[0189] in, , These are all learnable network parameters. The process of feature enhancement fusion based on this fusion weight vector is expressed as follows:

[0190] ;

[0191] in, Calculate dot products in the network; This is known as enhanced feature descriptor, a feature vector that characterizes the edge and corner attributes of each point. This mechanism can dynamically adjust the contribution of each feature for different local structures, thereby providing more reliable and accurate feature data for subsequent calculations and analyses.

[0192] It should be noted that the core logic of coarse and fine selection of points based on enhanced feature descriptors is as follows: First, by combining a linear scoring function with Sigmoid activation, the high-dimensional feature vector of the descriptor is mapped to a corner confidence score in the range of 0 to 1. The higher the score, the greater the probability that the point is a corner point. In the coarse selection stage, based on the threshold dynamically generated by voxels, points with scores higher than the threshold are retained as candidate corner point sets, quickly filtering out a large number of non-corner points. In the fine selection stage, isolated points with too few candidate points in the neighborhood of the candidate set are first removed, and then the remaining points are retained. Region growing is performed by combining the confidence threshold and the angle constraint between the normal vectors to merge adjacent regions and filter out tiny regions with too few points, finally obtaining a refined corner point set.

[0193] Specifically, this embodiment employs a hierarchical screening mechanism to perform dual screening of the enhanced feature descriptors, resulting in a more refined set of corner points with higher density. The initial screening stage requires calculating the corner confidence of the enhanced feature descriptors based on a linear scoring function. :

[0194] ;

[0195] in, For the Sigmoid function, , These are trainable parameters.

[0196] Then, the point cloud data points are divided into a voxel grid, and in each voxel grid... The voxel segmentation threshold is dynamically generated internally, i.e., the preset voxel segmentation threshold is:

[0197] ;

[0198] in, , These are the mean and standard deviation of the confidence scores for the interior angles of the voxel lattice, respectively. This is the global sensitivity coefficient, which typically ranges from 0.8 to 1.2.

[0199] Then, based on the preset voxel segmentation threshold, and corner confidence Perform coarse point cloud screening to filter out The points are used as the coarse selection of the corner point set. ,in, The voxel segmentation threshold is the value corresponding to the i-th point.

[0200] Based on the obtained coarsely screened corner point set, further screening of corner points is carried out, mainly to remove isolated points from the point set. First, the coarsely screened corner point set... The number of candidate points in the middle field is less than Isolated points are removed, among which, The seed is a small integer, which can be set according to actual needs, such as 2, 3, or 5. Then, the remaining coarsely screened corner points are used as seeds. Region growing is performed based on constraints configured according to the normal vector and corner confidence. This allows merging of adjacent regions with fewer points, generating a refined corner point set. The constraint is that the angle between the normal vectors is less than... Angle confidence level greater than , Typically, a small number is set, such as between 0.2 and 0.4, and the resulting refined set of corner points can be expressed as: .

[0201] Furthermore, this embodiment aims to improve the completeness and structural rationality of point cloud segmentation. Therefore, a Gaussian kernel function is first used to perform bilateral filtering on the refined corner point set. The point position update method during the filtering process is as follows:

[0202] ;

[0203] in, , Let these represent the spatial Gaussian kernel function and the eigendomain Gaussian kernel function, respectively. Used to measure the similarity between the spatial locations of two points. It is used to measure the matching degree of two-point enhanced feature descriptors; the Gaussian kernel function can smooth noise while preserving edge sharpness; by updating the position of each point, the updated set of filtered corner points can be obtained.

[0204] Then we can use the refined corner point set Construct a graph structure and detect pairs of points whose distance is less than a distance threshold. The fracture endpoint, where the distance to the threshold Distance from the average point cloud Related, usually set to The distance is between 3 and 10 times that of the original point cloud data. Based on the detected fracture endpoints, geometric paths are searched in the initial point cloud data, and path completion processing is performed to obtain multiple connected paths.

[0205] The specific path completion process is as follows: taking the detected broken endpoints as the starting and ending points, search for intermediate points in the original point cloud that meet the constraints. The constraints are that the angle between the normal vectors of the two points is no greater than 15° and the similarity of the enhanced feature descriptors is no less than 0.7. If an intermediate point that meets the conditions is found, it is directly connected to form a complete path. If no intermediate point is found, 3 to 10 transition points are generated by linear interpolation based on the spatial coordinates and feature attributes of the two endpoints. The transition points must meet the constraints of the two endpoints. Finally, the endpoints and transition points are connected to form a continuous connected path.

[0206] Next, the overall significance score of each connected path is calculated:

[0207] ;

[0208] in, , , This represents the weighting coefficient, with a value between 0 and 1, and satisfies the following conditions: For example, it can take values , , ; This represents the average corner confidence score of all points on a connected path. This represents the length of the connected path, in mm. This indicates the deviation of a connected path, which is the average distance between the path and the ideal straight line, in mm.

[0209] Then compare the comprehensive significance score with the score threshold. ,if If so, the corresponding pseudo-structure is removed; scoring threshold This value is usually determined through experimental analysis of training data or typical scenarios; it is an empirical value. For example, it can be set by observing the actual distribution of corners. The corner point cloud segmentation result obtained from corner segmentation is expressed as... ;Segmenting the corner point set It is the set of all points identified as corners in the original point cloud dataset. In terms of data structure, this set is represented as: a set of three-dimensional coordinates of M points. ,and Each point represents the coordinates of a corresponding point in the original point cloud dataset or its new coordinates obtained through geometric optimization.

[0210] It should be noted that the hardware of this embodiment is based on a high-performance embedded processing platform and equipped with an AI acceleration chip to meet the real-time computing needs of large-scale point clouds. The point cloud acquisition module is used to acquire raw 3D point cloud data and output an unordered set of points containing spatial coordinate information; the high-speed storage module is used to cache the raw point cloud, intermediate features and segmentation results; the communication interface supports transmission methods such as Gigabit Ethernet and USB 3.0, and can interact with host computers, cloud platforms and detection equipment; the human-machine interaction terminal is used for parameter configuration, process monitoring and result visualization.

[0211] After power-on, the device automatically initializes. The original point cloud undergoes preprocessing, including statistical filtering and outlier removal, to obtain a clean initial point cloud. Subsequently, it sequentially enters each functional unit to perform calculations, ultimately outputting standardized corner segmentation results. The hardware features low power consumption, high stability, and strong anti-interference capabilities, enabling continuous and stable operation in complex environments such as outdoor substations and industrial sites.

[0212] Furthermore, the fusion enhancement unit 203 in this embodiment is deployed on the hardware AI acceleration engine, undertaking the core feature fusion calculation task. It supports offline training and online inference, and parameter optimization can be completed before leaving the factory. The optimal model is directly loaded at runtime, ensuring a balance between accuracy and speed.

[0213] After the device is started, the hardware modules and software units operate collaboratively in a data flow-driven mode: the point cloud acquisition module acquires raw data and completes preprocessing; the neighborhood construction unit 201 generates adaptive local neighborhoods; the feature analysis unit 202 outputs geometric and density features; the fusion enhancement unit 203 generates enhanced feature descriptors; the point cloud screening unit 204 obtains a refined set of corner points through confidence calculation, coarse screening, removal of isolated points, and region growing; the point cloud segmentation unit 205 outputs the final result through filtering, graph construction, break completion, and pseudo-structure filtering; the results are uploaded to the host computer or cloud via the communication interface and visualized in real time on the interactive terminal. The low coupling and high cohesion of each unit, the absence of redundant computation, and the hardware parallel acceleration and software modular design enable the device to possess high real-time performance, high stability, and strong scalability.

[0214] This device embodiment fully implements all the functions of the method embodiment, transforming the algorithm flow into an engineering system with significant advantages: adaptive neighborhood and dynamic threshold adapt to point clouds of different densities and noise levels, ensuring stable accuracy across scenarios; multi-feature fusion and attention mechanisms enhance anti-interference capabilities, effectively distinguishing between real corners and pseudo-structures; graph connectivity analysis and break completion ensure continuous and complete corners, avoiding breaks and omissions; modular design facilitates hardware portability and functional upgrades, allowing deployment on various devices such as embedded terminals, edge computing boxes, and mobile inspection platforms. The overall system requires no manual parameter tuning and boasts a high degree of automation, effectively addressing the technical pain points of traditional point cloud corner segmentation, such as poor robustness, low adaptability, and incomplete segmentation, meeting the practical needs of industrial and power scenarios.

[0215] This application also provides a point cloud segmentation device with enhanced features, the device including a processor and a memory;

[0216] The memory is used to store program code and transfer the program code to the processor;

[0217] The processor is used to execute the point cloud segmentation method with enhanced features in the above method embodiments according to the instructions in the program code.

[0218] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0219] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0220] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0221] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for executing all or part of the steps of the methods described in the various embodiments of this application through a computer device (which may be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0222] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A point cloud segmentation method with enhanced features, characterized in that, include: An adaptive local neighborhood is constructed based on the initial point cloud data obtained after filtering and processing; Geometric and density analyses are performed on the adaptive local neighborhood to obtain geometric and density features; The geometric features and density features are fused and enhanced using a lightweight attention network to obtain an enhanced feature descriptor; Based on the enhanced feature descriptor, a point-level filtering analysis based on corner confidence is performed to obtain a refined corner point set; Based on the refined corner point set, a graph structure is constructed, and a corner segmentation operation based on connectivity analysis is performed to obtain the corner point cloud segmentation result.

2. The point cloud segmentation method with enhanced features according to claim 1, characterized in that, The construction of an adaptive local neighborhood based on the initial point cloud data obtained after filtering includes: The K-nearest neighbor algorithm is used to calculate the average distance of the point cloud based on the original point cloud data; Outliers are removed from the original point cloud data based on the average point cloud distance to obtain initial point cloud data. An adaptive local neighborhood is constructed based on the initial point cloud data, and the adaptive local neighborhood includes a neighborhood radius.

3. The point cloud segmentation method with enhanced features according to claim 1, characterized in that, The step of performing geometric and density analyses based on the adaptive local neighborhood to obtain geometric features and density features includes: Principal component analysis is performed in the adaptive local neighborhood to calculate the surface normal vector and curvature, thereby obtaining the geometric properties; Geometric features are calculated based on the geometric properties and the adaptive local neighborhood, the geometric features including curvature change intensity and normal vector divergence. The density features are obtained by performing normalized local density calculation based on the adaptive local neighborhood.

4. The point cloud segmentation method with enhanced features according to claim 1, characterized in that, The enhanced feature descriptor is obtained by fusing and enhancing the geometric features and density features based on a lightweight attention network, including: The geometric features and the density features are standardized and concatenated to obtain the initial fused features; A lightweight attention network is used to generate a fusion weight vector, and an enhanced fusion calculation is performed based on the fusion weight vector and the initial fusion features to obtain an enhanced feature descriptor.

5. The point cloud segmentation method with enhanced features according to claim 1, characterized in that, The step of performing a point-level filtering analysis based on corner confidence using the enhanced feature descriptor to obtain a refined corner point set includes: Calculate the corner confidence based on the enhanced feature descriptor; A coarse point cloud screening is performed based on a preset voxel segmentation threshold and the corner confidence level to obtain a coarse corner point set; After removing isolated points from the coarse corner point set, a constraint-based region growing analysis is performed based on the normal vector and the corner confidence to obtain a refined corner point set.

6. The point cloud segmentation method with enhanced features according to claim 1, characterized in that, The corner segmentation operation based on connectivity analysis, performed according to the refined corner point set to construct the graph structure, yields the corner point cloud segmentation result, including: The refined corner point set is subjected to bilateral filtering using a Gaussian kernel function to obtain the filtered corner point set; The filtered corner point set is constructed into a graph structure, and path connection completion processing based on fracture endpoint analysis is performed according to the distance threshold to obtain the connected path; Calculate the comprehensive saliency score of each of the connected paths, and filter out pseudo-structures based on the score threshold to achieve corner segmentation and obtain the corner point cloud segmentation result.

7. A point cloud segmentation device with enhanced features, characterized in that, include: The domain construction unit is used to construct an adaptive local domain based on the initial point cloud data obtained after filtering and processing. The feature analysis unit is used to perform geometric and density analyses based on the adaptive local neighborhood to obtain geometric features and density features; The fusion enhancement unit is used to perform fusion enhancement processing on the geometric features and the density features based on a lightweight attention network to obtain an enhanced feature descriptor. The point cloud filtering unit is used to perform point-level filtering analysis based on corner confidence based on the enhanced feature descriptor to obtain a refined corner point set. The point cloud segmentation unit is used to perform edge and corner segmentation operations based on connectivity analysis according to the graph structure constructed by the refined edge and corner point set, so as to obtain the edge and corner point cloud segmentation results.

8. The point cloud segmentation apparatus with enhanced features according to claim 7, characterized in that, The point cloud filtering unit is specifically used for: Calculate the corner confidence based on the enhanced feature descriptor; A coarse point cloud screening is performed based on a preset voxel segmentation threshold and the corner confidence level to obtain a coarse corner point set; After removing isolated points from the coarse corner point set, a constraint-based region growing analysis is performed based on the normal vector and the corner confidence to obtain a refined corner point set.

9. The point cloud segmentation apparatus for enhancing features according to claim 7, characterized in that, The point cloud segmentation unit is specifically used for: The refined corner point set is subjected to bilateral filtering using a Gaussian kernel function to obtain the filtered corner point set; The filtered corner point set is constructed into a graph structure, and path connection completion processing based on fracture endpoint analysis is performed according to the distance threshold to obtain the connected path; Calculate the comprehensive saliency score of each of the connected paths, and filter out pseudo-structures based on the score threshold to achieve corner segmentation and obtain the corner point cloud segmentation result.

10. A point cloud segmentation device with enhanced features, characterized in that, The device includes a processor and a memory; The memory is used to store program code and transmit the program code to the processor; The processor is configured to execute the point cloud segmentation method with enhanced features as described in any one of claims 1-6 according to the instructions in the program code.

Images