A point cloud data fast labeling method and system

Abstract

The present application relates to the technical field of three-dimensional point cloud data processing, and particularly relates to a point cloud data fast labeling method and system. The method comprises the following steps: obtaining point cloud data and initial point labeling data; constructing a superpoint structure according to the point cloud data to obtain point cloud superpoint structure data; mapping the initial point labeling data to the point cloud superpoint structure data to obtain label mapping data; performing local topological fidelity label propagation on the label mapping data to obtain first label propagation data; performing projection-guided aggregation on the label mapping data to obtain second label propagation data; obtaining label propagation feature data according to the first label propagation data and the second label propagation data; fitting the confidence of the first label propagation data and the second label propagation data according to the label propagation feature data to obtain label propagation data; and performing label reflection on the label propagation data to obtain point cloud labeling data. The present application realizes efficient and accurate expansion of point cloud label information.

Description

Technical Field

[0001] This invention relates to the field of 3D point cloud data processing technology, and in particular to a method and system for rapid annotation of point cloud data. Background Technology

[0002] With the widespread application of 3D laser scanning and multi-view reconstruction technologies, point cloud data has become an important carrier for expressing spatial geometric structures and surface semantic features, and is used in the power industry for tasks such as facility modeling, digital twins, and asset inspection. However, due to the disordered, high-density, complex structure, and lack of topological constraints of point clouds, manual annotation is extremely costly and highly subjective, severely restricting the training and deployment of subsequent algorithm models. In real-world scenarios, power facility components such as power towers, transformers, and insulators often exhibit structural repetition, occlusion, and ambiguous boundaries, making accurate and efficient annotation difficult to achieve solely through manual or single-path propagation. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a method and system for rapid annotation of point cloud data, thereby resolving at least one of the aforementioned technical issues.

[0004] This application provides a method for rapid annotation of point cloud data, the method comprising:

[0005] S1. Acquire point cloud data and initial point annotation data; construct the superpoint structure based on the point cloud data to obtain the point cloud superpoint structure data;

[0006] S2. Map the initial point label data to the point cloud superpoint structure data to obtain label mapping data;

[0007] S3. Perform local topology-fidelity label propagation on the label mapping data to obtain the first label propagation data; perform projection-guided aggregation on the label mapping data to obtain the second label propagation data; extract label propagation features based on the first and second label propagation data to obtain label propagation feature data; perform confidence fitting on the first and second label propagation data based on the label propagation feature data to obtain the label propagation data.

[0008] S4. Perform label reflection on the label propagation data to obtain point cloud labeled data.

[0009] This invention employs a superpoint structure to perform geometrically perceptual segmentation of the original point cloud, effectively improving the structural stability of label propagation. Local topologically faithful label propagation maintains spatial connectivity and semantic boundary stability during label diffusion, preventing mis-propagation at structural abrupt changes. Projection-guided aggregation extracts gravity centers from the global semantic embedding space, enhancing semantic coverage of sparsely labeled regions. By fusing the two types of propagation results and combining propagation path features for confidence fitting, hierarchical discrimination and dynamic reconciliation of conflicting labels can be achieved, improving overall label credibility and robustness. Finally, a label reflection mechanism accurately maps structural-level semantics back to the point level, enabling high-quality, interpretable point cloud semantic annotation output.

[0010] Optionally, the superpoint structure construction includes:

[0011] Structural perception feature extraction is performed on point cloud data to obtain structural perception feature data;

[0012] Heterogeneous fusion of structural sensing feature data is performed to obtain heterogeneous fused data;

[0013] Multi-scale semantic region segmentation is performed on heterogeneous fused data to obtain region segmentation data;

[0014] Superpoint topology graphs are constructed from the region segmentation data to obtain superpoint topology graph data;

[0015] Consistent encoding is performed on the superpoint topology data to obtain the point cloud superpoint structure data.

[0016] This invention effectively captures geometric details such as local normals, curvature, and spatial density in point clouds through structure-aware feature extraction, resulting in stronger shape recognition capabilities for region segmentation. A heterogeneous fusion strategy jointly models geometric attributes and positional features, enhancing the ability to identify semantically similar regions in structural scenes. Multi-scale semantic region segmentation strengthens the adaptive segmentation capability for components of different scales (such as large-volume equipment and finely connected parts). The constructed superpoint topology explicitly expresses the contact relationships and spatial connectivity between regions, which helps maintain the local continuity of the propagation structure. The consistent coding process, by modeling the topological structure and semantic boundaries, provides a stable and highly expressive structural foundation for label propagation, thereby significantly improving overall annotation accuracy and boundary recognition accuracy.

[0017] Optionally, S2 includes:

[0018] A point-to-superpoint mapping index is constructed based on the initial point annotation data and the point cloud superpoint structure data to obtain the mapping index data;

[0019] Tag weight fusion is performed based on the mapping index data to obtain fused tag data;

[0020] Perform super-point confidence initialization on the tag fusion data to obtain super-point confidence data;

[0021] Label validation is performed on the over-confidence data to obtain label validation data;

[0022] The tag mapping data is obtained by performing mapping processing based on the tag verification data and the mapping index data.

[0023] This invention constructs a point-to-superpoint mapping index, achieving efficient association between point-level labels and structural units, which helps to quickly locate the influence range of initial labels in different geometric structures. Label weight fusion, when processing multi-source labels within the same superpoint, combines label distribution intensity and point density information to improve the robustness and expressiveness of label fusion. Superpoint confidence initialization, combined with the fusion results, calculates category confidence, introducing a quantitative reliability assessment indicator for propagation. The label verification step uses features such as neighborhood consistency and geometric coherence to determine the reliability of labels, preventing initial mislabeling from spreading to adjacent structural regions. Through mapping processing, the verified labels are injected into the superpoint structure, forming stable, fully expressive label mapping data with local geometric fidelity, providing a highly reliable and low-conflict initial semantic benchmark for the propagation process.

[0024] Optionally, the local topology fidelity tag propagation includes:

[0025] A local geometric topology graph is constructed from the label mapping data to obtain the local geometric topology graph data;

[0026] Initialize the tag potential field on the local geometric topology map data to obtain the tag potential field data;

[0027] Local topological tag propagation is performed based on the tag potential energy field data to obtain topological propagation data;

[0028] The propagation path structure is extracted based on the topological propagation data to obtain the first label propagation data.

[0029] This invention constructs a local geometric topology map to explicitly model the spatial contact relationships and geometric consistency between adjacent superpoints in the point cloud, providing structured path constraints for label propagation and avoiding the semantic boundary issues caused by traditional unstructured diffusion. The initialization of the label potential field uses existing label mapping results to assign propagation source weights to specific regions, giving the propagation process directionality and source-driven characteristics. Local topology label expansion propagation dynamically adjusts the propagation step size and path accessibility while maintaining local structural coherence, thereby precisely controlling the propagation range of labels in complex structures. The propagation path structure extraction not only records the specific path of label diffusion but also reflects the topological consistency and diffusion intensity of structural propagation, which is helpful for confidence fusion and propagation backtracking processing.

[0030] Optionally, the projection-guided aggregation includes:

[0031] The projection space is constructed from the label mapping data to obtain the projection space data;

[0032] Tag gravity center data is obtained by extracting the tag gravity center from the projected spatial data.

[0033] The label gravity center data is projected onto the non-label gravity center data from the projection space data to obtain non-local projection data;

[0034] The projection-guided aggregation map data is obtained by processing the non-local projection data into a projection-guided aggregation map.

[0035] Multi-label fusion is performed on the projection-guided aggregation map data to obtain fuzzy label data;

[0036] Boundary suppression optimization is performed on the fuzzy label data to obtain the second label propagation data.

[0037] This invention effectively compresses the distribution of point clouds under high-dimensional geometric semantic features by constructing a projection space, enabling similar semantic regions to cluster in low-dimensional space, facilitating label structure analysis. Label gravity center extraction, based on the aggregation characteristics of high-density regions within a category, accurately locates semantic center points, serving as anchors for non-local label inference. Projection relationships are constructed between unlabeled regions and gravity centers, achieving cross-topological connectivity for label-guided diffusion, effectively compensating for gaps in coverage that are difficult to cover with local propagation. The resulting projection-guided aggregation map strengthens the radiative relationships of semantic centers through graph structure, supporting smooth spatial expansion of label confidence. Fuzzy label fusion generates soft label probability distributions in multi-gravity center interference regions, enhancing the elasticity and generalization ability of label expression. Boundary suppression optimization introduces gradient and consistency constraints for high-uncertainty regions, suppressing semantic boundary crossings and label ambiguity, thereby obtaining a second label propagation result with strong structure awareness, wide coverage, and stable confidence.

[0038] Optionally, the tag gravity center extraction includes:

[0039] The label distribution density map is calculated from the projected spatial data to obtain the label distribution density map data;

[0040] Gravitational center candidate field processing is performed on the tag distribution density map data to obtain gravitational center candidate field data;

[0041] Multimodal label center clustering was performed on the candidate field data of the gravitational center to obtain the center clustering data;

[0042] Representative centers are selected based on the central clustering data to obtain the label gravity center data.

[0043] This invention calculates the distribution density map of label samples in the projection space, which accurately captures the aggregation characteristics of semantic labels in the low-dimensional embedding space and reveals the structural trends of labels in the projection space. The candidate field processing of gravity centers uses density gradient information to construct potential attraction kernel regions, making label aggregation directional and structurally perceptible. Multimodal label center clustering identifies multiple aggregation patterns of the same label in space, which is suitable for point cloud data with complex structures or semantic ambiguity, and significantly improves the diversity and spatial adaptability of label expression. The representative center selection is based on indicators such as cluster density, coverage, and local geometric stability to select label gravity points with strong representation and structural centrality, thereby providing a stable label emission source with physical and semantic basis for nonlocal propagation, and significantly enhancing the propagation ability and semantic continuity of label diffusion in structural fracture regions or low-density areas.

[0044] Optionally, the label propagation feature extraction includes:

[0045] Based on the first label propagation data and the second label propagation data, a label propagation evolution trajectory map is constructed to obtain propagation trajectory map data;

[0046] Label similarity and label stability were calculated on the propagation trajectory map data to obtain label similarity data and label stability data, respectively.

[0047] Based on label similarity data and label stability data, the propagation trajectory map data is subjected to propagation path label aggregation feature extraction to obtain label propagation feature data.

[0048] This invention constructs a tag propagation evolution trajectory map, explicitly modeling the structural transmission paths and semantic changes of the first and second tag propagation data, thus achieving a temporal and structured expression of propagation behavior. Tag similarity calculation quantifies the consistency of tag prediction results across different propagation paths, helping to determine semantic convergence among multi-source propagation. Tag stability calculation, combining the degree of tag variation and confidence fluctuation across different paths, identifies potentially unstable propagation segments or areas affected by structural mutations during propagation. The propagation path tag aggregation feature extraction performed in this invention combines indicators such as propagation depth, path redundancy, and confidence consistency to form a semantic evolution profile of each structural unit in tag propagation, providing a high-dimensional discriminative basis for confidence fusion and propagation optimization. This invention enhances the interpretability and process tracking capability of the tag propagation mechanism, effectively supporting the refined control of propagation strategies and the construction of data-driven correction mechanisms in actual power facility structures.

[0049] Optionally, the confidence fitting includes:

[0050] The label confidence scores of the first label propagation data and the second label propagation data are calculated based on the label propagation feature data to obtain the label confidence scores.

[0051] Confidence conflict resolution is performed based on the label confidence data to obtain confidence adjustment data;

[0052] A credible probability field is constructed from the confidence-adjusted data to obtain credible probability field data;

[0053] The label distribution of the first label propagation data and the second label propagation data is optimized based on the reliable probability field data to obtain the label propagation data.

[0054] This invention incorporates label propagation feature data to calculate the multi-source confidence of the propagation results of the first and second labels. This allows for the evaluation of the credibility of each label from dimensions such as propagation path, label consistency, and stability, ensuring that the fusion process is based on quantitative indicators. By handling confidence conflicts, high-confidence bias regions between propagation results are identified. A dynamic adjustment strategy is introduced for conflicting labels to effectively avoid interference from low-confidence labels in the final decision. The constructed credibility probability field not only preserves the label distribution information of local propagation but also combines spatial continuity and semantic gradient to perform regularized modeling of label confidence, enabling the label assignment process to have structure awareness. Label distribution optimization, combined with this probability field, performs soft label correction, strengthening label determinism in high-confidence consistency regions and maintaining semantic flexibility in ambiguous boundary regions, thereby improving the robustness, continuity, and interpretability of the overall label fusion result.

[0055] Optionally, S4 includes:

[0056] A tag reflection index mapping table is constructed on the tag propagation data to obtain the tag reflection data;

[0057] Spatial confidence point-level weights are calculated based on the label reflection data to obtain point-level weight data;

[0058] The label boundaries are fine-tuned based on the point-level weight data to obtain point cloud labeled data.

[0059] This invention establishes a precise mapping relationship from superpoint-level label propagation results to original point-level data by constructing a label reflection index mapping table, ensuring the traceability and structural consistency of the label transmission process. The spatial confidence point-level weight calculation is based on superpoint confidence, the positional relationship of points in structural units, and local geometric features, realizing differentiated adjustment of the point-level label assignment process, so that the label results can reflect the importance and representativeness of points in the structure. The label boundary fine-tuning, combined with spatial gradient, neighborhood consistency, and confidence distribution, locally adjusts the labels in the boundary area, effectively eliminating label mismatch problems caused by imprecise structural division or propagation errors, thereby outputting high-quality point cloud annotation results with spatial continuity, clear boundaries, and semantic consistency.

[0060] Optionally, this application also provides a point cloud data rapid annotation system for performing the point cloud data rapid annotation method described above, the point cloud data rapid annotation system comprising:

[0061] The point cloud structure modeling and initial label acquisition module is used to acquire point cloud data and initial point label data; and to construct superpoint structures based on the point cloud data to obtain point cloud superpoint structure data.

[0062] The label mapping and seed annotation extraction module is used to map the initial point annotation data to the point cloud superpoint structure data to obtain label mapping data;

[0063] The dual-path semantic propagation and confidence fusion module is used to perform local topological fidelity label propagation on the label mapping data to obtain the first label propagation data; to perform projection-guided aggregation on the label mapping data to obtain the second label propagation data; to extract label propagation features based on the first and second label propagation data to obtain label propagation feature data; and to perform confidence fitting on the first and second label propagation data based on the label propagation feature data to obtain the label propagation data.

[0064] The label reflection and point-level annotation output module is used to reflect the label propagation data to obtain point cloud annotation data.

[0065] The purpose of this invention is to construct a superpoint structure with geometric perception and topological expression capabilities to achieve structural unitization of the original point cloud, providing a stable and high-fidelity foundation for semantic processing. By mapping initial point labels to the superpoint structure and fusing local weights and geometric features to achieve label confidence initialization, the labeled information can accurately fall into the core of the semantic structure. This invention uses a dual-path propagation strategy of local topological fidelity propagation and projection-guided aggregation to maintain the spatial structural consistency and semantic space coverage breadth of label propagation, respectively. It then extracts consistency features, conflict information, and propagation stability from the propagation path, achieving fine-grained modeling and confidence fusion of multi-source label results. The credible probability field constructed based on confidence fitting enables spatial regularity in label optimization. Through the label reflection process, structural-level semantics are reasonably mapped back to point-level data, and label fine-tuning is performed on boundary regions to ensure that the output labels have spatial continuity, clear boundaries, and global semantic consistency, significantly improving the efficiency and quality of point cloud labeling. Attached Figure Description

[0066] Other features, objects, and advantages of this application will become more apparent from the following detailed description of the non-limiting embodiments, taken with reference to the accompanying drawings:

[0067] Figure 1 A flowchart illustrating the steps of a method for rapid annotation of point cloud data according to an embodiment is shown.

[0068] Figure 2 A flowchart illustrating the steps of a superpoint structure construction method according to an embodiment is shown.

[0069] Figure 3 A flowchart illustrating the steps of a label mapping method according to one embodiment is shown.

[0070] Figure 4 A flowchart illustrating the steps of a local topology-fidelity tag propagation method according to an embodiment is shown.

[0071] Figure 5 A flowchart illustrating the steps of a label reflection method according to an embodiment is shown;

[0072] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0073] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0074] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0075] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0076] Please see Figures 1 to 5 This application provides a method for rapid annotation of point cloud data, the method comprising:

[0077] S1. Acquire point cloud data and initial point annotation data; construct the superpoint structure based on the point cloud data to obtain the point cloud superpoint structure data;

[0078] In one embodiment, the original 3D point cloud dataset is obtained. Each point p i Includes spatial coordinate information (x i ,y i ,z i Synchronously acquire the corresponding user-defined initial annotation point set. Where y i This represents the semantic category manually assigned by the user. (For point cloud datasets) Perform local geometric feature extraction. Based on principal component analysis (PCA), at each point p... i The normal vector of a point is estimated within its fixed neighborhood to characterize the local surface orientation. Simultaneously, geometric indices such as local curvature, point density, and shape index are calculated to construct a feature set reflecting its local geometry. Based on the dual constraints of spatial Euclidean distance and normal angle, a point-level adjacency graph G = (V, E) is constructed, where node V is a set of points, and edge E represents pairs of points that are adjacent and have similar normals. These geometric features are encoded into a high-dimensional feature vector f for each point. i The data is then input into the clustering module, where spectral embedding combined with the KMeans clustering algorithm, or density-based DBSCAN clustering, is used to perform structural consistency aggregation on the point set, thereby generating preliminary hyperpoint partitioning results. Each superpoint S K Given a spatially continuous set of points with a consistent shape, construct a superpoint topology graph G. sp=(S k E kl ), where nodes are superpoint elements, and edges E kl This indicates that there is physical contact or spatial adjacency between two superpoints. The edge weight is calculated by weighting the difference between the contact surface size and the average normal vector between adjacent superpoints using preset weights.

[0079] S2. Map the initial point label data to the point cloud superpoint structure data to obtain label mapping data;

[0080] In one embodiment, a mapping index table between points and superpoints is established. All labeled points (p i ,y i Mapped to its superpoint S k Build a tag set The voting results and confidence scores for superpoint labels are calculated as follows: Main label in For the super point S k The main tag, To select the most frequent label from all candidate labels, `count` is a function to count the number of occurrences, and `y` is the label data. Given a set of labels; calculate the initial confidence level. For conf k Points below a threshold (e.g., 0.6) are marked as candidate labels; the preliminary label mapping results are output.

[0081] S3. Perform local topology-fidelity label propagation on the label mapping data to obtain the first label propagation data; perform projection-guided aggregation on the label mapping data to obtain the second label propagation data; extract label propagation features based on the first and second label propagation data to obtain label propagation feature data; perform confidence fitting on the first and second label propagation data based on the label propagation feature data to obtain the label propagation data.

[0082] In one embodiment, the system is based on initial label mapping data. A dual-path propagation approach combining local topology-fidelity label propagation and projection-guided aggregation is employed to generate first-label propagation data and second-label propagation data, respectively. Local topology-fidelity label propagation (i.e., the first propagation path) forms a local propagation graph G based on the previously constructed superpoint topology. local The edge weights of this graph are calculated by weighting the difference in the normal angles between superpoints and the boundary contact degree, with the weights being preset weights. In this graph, all seed superpoints S are... k Potential energy value φ k Initialize to 1; unmarked superpoint potential energy is initialized to 0. Perform multiple rounds of potential energy diffusion based on the following propagation equation: in Z represents the tag potential energy value of the (t+1)th round of potential energy diffusion. i Let be the normalization factor for node i, and j be the superpoint order. Let w be the set of adjacent superpoints of node i. ij To propagate edge weights in the local topology, Let be the tag potential energy value of the t-th round of potential energy diffusion. The propagation process terminates after reaching a stability threshold or a preset number of rounds T, and tags are assigned to superpoints based on the location of the maximum response of the potential energy field. If the confidence level of its propagation is higher than a set threshold (such as 0.7), the label is accepted; otherwise, it is marked as "pending confirmation".

[0083] Projection-guided aggregation (i.e., the second propagation path) emphasizes the dense aggregation of labels in the semantic space. The system represents the feature of all labeled superpoints f k The input is fed into a dimensionality reduction algorithm (such as t-SNE), projected into a two-dimensional semantic space, and density estimation ρ is performed on similar label points. y (x), extract the set of gravitational center points C for each type of label. y For all unlabeled superpoints S j After projecting, it is compared with each class C. y Calculate the similarity weight w jy Construct a nonlocal projection aggregation graph and perform soft label aggregation according to the following formula: Where P(y) j () is the superpoint S j The label probability distribution (second propagation path), w j For the super point S j The total weight of the aggregated tags, w jy For the super point S j To Category Center C y The similarity weight, where y is the candidate label category. The predicted label results for the second propagation path. To select the operation with the highest probability among all candidate labels, y j For the super point S j The predicted label category (an intermediate variable under unconfirmed or multi-label probabilities) is determined. If a point exhibits high entropy in label aggregation (e.g., multiple label probabilities are close, entropy value is greater than threshold δ), a boundary suppression strategy is implemented, and no label is assigned temporarily. The system integrates the output results of two propagation paths, constructs a superpoint propagation evolution graph, records the propagation trajectory, path depth, and label changes of each superpoint in the two paths, and extracts the following propagation consistency feature: label consistency difference Δy = |y (1) -y (2) | where Δy is the label consistency difference, y (1)For the predicted label results of the superpoints in the first propagation path, y (2) The predicted label results for the superpoints in the second propagation path, with path propagation depth d. k Confidence difference Δconf=|conf (1) -conf (2) | etc., where Δconf is the confidence difference, conf (1) For the label confidence of the first propagation path, conf (2) Calculate the label confidence for the second propagation path. Use an ensemble learning model (such as Gradient Boosting Tree (GBDT) or Soft Voting model) to fuse the results of the first and second label propagation paths, outputting the final label y for each superpoint. k and its confidence level conf k The final label propagation result is obtained.

[0084] S4. Perform label reflection on the label propagation data to obtain point cloud labeled data.

[0085] In one embodiment, the system constructs a superpoint-to-point reflection index mapping table. Each superpoint S k Corresponding to the set of points contained within it Where p i Let p be the i-th point in the point cloud. For each point p... i The system determines the superpoint S to which it belongs. k Directly inherit the tag y of the superpoint k and the corresponding confidence level conf k The system generates preliminary point-level label prediction results. To avoid label discontinuities or mismatches within the super-point boundary region, the system at point p... i In cases involving multiple adjacent superverts, a soft label fusion mechanism is executed. (System statistics p) i Set of adjacent superpoints The point-level label probability distribution is calculated based on the following weighted fusion formula: Where P(y) i Let p be a point. i The probability distribution of the fused labels, S k For the kth superpoint, For point p i The set of adjacent superpoints, ω ik For point p i For the superpoint S k The weights, based on p i Euclidean distance to the center of the superpoint and confidence level of the superpoint label (conf) k In the joint calculation, a higher weight indicates that the superpoint contributes more to the label of the point. This represents the indicator vector or probabilistic form of the superpoint label. The system avoids label breaks or semantic conflicts during soft label fusion by smoothing label boundaries. The system then determines the fusion result P(y) based on the fusion result. i Generate each point p i Most likely tags ( For point p i The predicted label, arg max is the maximum probability selection operation, P(y i Let p be a point. i (The probability distribution of the fused labels) and its confidence level conf i =maxP(y i (conf) i For point p i The label confidence score, where max is the maximization operator, P(y i Let p be a point. i The fusion label probability distribution is used to output the point cloud annotation result dataset. As a semantic annotation output.

[0086] Optionally, the superpoint structure construction includes:

[0087] S11. Extract structure-aware features from point cloud data to obtain structure-aware feature data;

[0088] In one embodiment, the geometric attributes of each point are extracted. For each point p in the point cloud... i Build a neighborhood based on k nearest neighbors or sphere radius r Calculate the covariance matrix C within the neighborhood. i Solve for the eigenvalues ​​λ1≥λ2≥λ3; calculate the normal vector. Corresponding to the direction of the minimum eigenvalue; calculate curvature Calculate density V i For the neighborhood volume, For point p i The neighborhood point set (based on k-nearest neighbor or radius r search).

[0089] S12. Perform heterogeneous fusion on the structure-aware feature data to obtain heterogeneous fused data;

[0090] In one embodiment, the system selects feature information from multiple sources as fusion input, including spatial location features, i.e., point p. i Coordinate information in three-dimensional space is represented as The absolute position information of the points is used to preserve the geometric structural features, which are the output of the structure-aware module in the previous step and are represented as follows: It includes local geometric properties such as normal vectors, curvature, and density; contextual auxiliary features, including auxiliary perceptual information carried in the point cloud, such as color values ​​(RGB), laser reflection intensity, or echo count, collectively denoted as... After concatenating location features, geometric features, and auxiliary features into a unified vector according to their dimensions, dimensionality compression and feature fusion are performed using a linear mapping method or a multilayer perceptron (MLP). Specifically, an affine transformation is applied to the concatenated vector using a weight matrix, and a nonlinear activation function (such as ReLU) is introduced to generate the final fused feature vector. For point p... i Its fused feature vector is Where σ represents the activation function, such as the Modified Linear Unit (ReLU); W is the fusion weight matrix; b is the bias term; || represents the vector concatenation operation; and [·] represents the original total feature vector after concatenation. The system output is a fusion feature matrix containing all points.

[0091] S13. Perform multi-scale semantic region segmentation on the heterogeneous fused data to obtain region segmentation data;

[0092] In one embodiment, the system acquires a set of superpoint structures in the point cloud. Each superpoint S k It contains a subset of the original point set. Simultaneously acquire initial point label data. Where p i For points with manual annotations, y i The system iterates through the corresponding semantic categories. Each point p in i Search for the structural unit S to which it belongs in the superpoint set. k That is, to find conditions that satisfy the criteria. Super point S k For mapping relationships that meet the conditions, the system constructs an index mapping table from point to superpoint. in For point p i The mapping index key. In some cases, point p i Located in a boundary region with two or more superpoints, the system sets a priority strategy for conflict resolution, for example, based on p i Choose the closest Euclidean distance among the centroids of each candidate superpoint; or, allow p i Simultaneously, points are mapped to multiple superpoints for fusion or voting weighting. After mapping all initial points, the system calculates the value of each superpoint S. k The received initial list of labels forms a label attribution table. in for, This represents the set of labels corresponding to all labeled points within the superpoint.

[0093] S14. Construct a superpoint topology graph from the region segmentation data to obtain superpoint topology graph data;

[0094] In one embodiment, a contact structure is established between semantic regions to explicitly express the internal structural relationships of the point cloud. Each region R... j Consider it as a superpoint S j If R i With R j If the distance to the point cloud boundary is less than δ, where δ is a preset threshold, then an edge (S) is established. i ,S j ); Edge weight w ij Defined as Where exp is the exponential function used to represent the decay of edge weights. For the super point S i The average normal vector For the super point S j The average normal vector, σ n The Gaussian kernel bandwidth of the normal vector difference controls the degree of weight attenuation of the normal component, c. i For the super point S i The geometric centroid position, c j For the super point S j The geometric centroid position, σ p The Gaussian kernel bandwidth is the distance from the geometric centroid, which controls the influence of positional differences on edge weights; the output is the superpoint topology graph structure.

[0095] S15. Perform consistency coding based on the superpoint topology map data to obtain the point cloud superpoint structure data.

[0096] In one embodiment, the system uses the tag fusion data obtained in the previous processing stage. For each superpoint S k Perform the following confidence calculation: conf k =α1·s k +α2·(1-H(P k )), where conf k For the super point S k The label confidence value, α1 is the label support weight data, with a value of 0.6, s k main tag At super point S k The label support in H(P) is given by α2, which is the information entropy weight data with a value of 0.4. k () is the superpoint S k Label probability distribution P k The information entropy of (y) is used to calculate the confidence value conf. kIt can be used to determine the reliability of the label results. If it is less than a preset threshold θ (e.g., 0.6), then the superpoint S is considered reliable. k Labeled as a low-confidence candidate label. The system outputs the initial super-point confidence data structure.

[0097] Optionally, S2 includes:

[0098] S21. Construct a point-to-superpoint mapping index based on the initial point annotation data and the point cloud superpoint structure data to obtain the mapping index data;

[0099] In one embodiment, a fast mapping relationship is established from initial point labels to the superpoint structure. Given the point cloud superpoint structure... Each superpoint S k Containing point set Traverse the initial point label data For each p i Find its superpoint S k Build a mapping index table If p i Located at the intersection of multiple superpoints, it can set a priority strategy (such as the nearest neighbor method) or bidirectional mapping; it outputs mapping index data, representing the initial label list received by each superpoint.

[0100] S22. Perform label weight fusion based on the mapping index data to obtain label fusion data;

[0101] In one embodiment, multiple tags received at each superpoint are fused to determine primary tag candidates and their support. For each superpoint S k Statistical analysis of its tag frequency in For label y in super point S k The number of times it appears in For indicator functions: when y i The value is 1 when y = 0, otherwise it is 0, and i is the superpoint S. k The index of the point in the label set, where y is the label category. To map to superpoint S k The set of tags corresponding to the points; define the fusion tags. Highest frequency tag in For the super point S k The main tag, i.e., the tag that appears most frequently. The most frequently occurring tag category, The frequency counter is used, k is the superpoint index, and y is the candidate label category; support is also calculated. Where s k For the super point S kMain tag support, Count is the main tag The number of times it appears, For the super point S k Main tags The number of times it appears, For the super point S k The set of labels for all marked points within the range; a label probability distribution vector P can be constructed from this set. k (y) to support fuzzy annotation: Output the label fusion data, which contains the frequency statistics of label y.

[0102] S23. Perform super-point confidence initialization on the tag fusion data to obtain super-point confidence data;

[0103] In one embodiment, the system, based on the label fusion data obtained in the previous processing stage, performs processing on each superpoint S. k Perform the following confidence calculation: conf k =α1·s k +α2·(1-H(P k )), where conf k For the super point S k The label confidence value, α1 is the label support weight data, with a value of 0.6, s k main tag At super point S k The label support in H(P) is given by α2, which is the information entropy weight data with a value of 0.4. k () is the superpoint S k Label probability distribution P k The information entropy of (y). The calculated confidence value conf k It can be used to determine the reliability of the label results. If it is less than a preset threshold θ (e.g., 0.6), then the superpoint S is considered reliable. k It was marked as a low-confidence candidate label.

[0104] S24. Perform label verification on the super-point confidence data to obtain label verification data;

[0105] In one embodiment, the system targets each superpoint S k Perform the following verification procedure: Calculate the geometric consistency index σ inside the superpoint. n,k That is, the variance of the set of normal vectors of its contained points, σ n,k A smaller value indicates less susceptibility to change and a more regular structure; calculate the neighborhood label consistency index η. k Its definition is related to S k The proportion of adjacent superpoints whose labels match their main labels, i.e. Where η kFor the super point S k The neighborhood label consistency ratio, where j is the index variable of the adjacent supernode. For the super point S k The set of adjacent superpoints, This is an indicator function; it takes a value of 1 if the labels match, and 0 otherwise. For adjacent super point S j The main tag, For the current superpoint S k The main tag, For the super point S k The set of adjacent superpoints. Set triple validation thresholds θ1 (confidence threshold), θ2 (normal variance threshold), and θ3 (neighborhood consistency threshold), for example, 0.6, 0.1, and 0.7 respectively; superpoints S satisfying the following logical conditions... k The verification was considered passed by conf k >θ1∧σ n,k <θ2∧η k >θ3,conf k For the super point S k The label confidence value is set to "verified" for verified points; otherwise, it is set to "pending verification", indicating that the label is not yet stable.

[0106] S25. Perform mapping processing based on the tag verification data and mapping index data to obtain tag mapping data.

[0107] In one embodiment, the verified tags and states are rebound to structural superpoints to generate propagable tag mapping data. For all S k Only retain the "verified" status label as the starting point for propagation; Mapping index data to superpoint label set P can be included at the same time k (y) serves as a soft label to support distribution; a label state graph is constructed synchronously for propagation path control (such as whether it can continue to spread, whether it participates in propagation weight).

[0108] Optionally, the local topology fidelity tag propagation includes:

[0109] S31. Construct a local geometric topology graph from the label mapping data to obtain local geometric topology graph data;

[0110] In one embodiment, the system assigns each superpoint unit S k Abstracted as a node in a graph. The set of points represented by each node is denoted as . That is, the set of original points contained in the superpoint. To establish topological connections between nodes, the system uses the following spatial adjacency criterion: for any two superpoints S... k With Sl , if there are point pairs within them where p i is the original point belonging to the superpoint S k , is the set of original points contained in the superpoint S k , p j is the original point belonging to the superpoint S l , is the set of original points contained in the superpoint S l such that their distance satisfies |p i - p j | < r (where r is the set threshold of the spatial neighborhood radius), then it is considered that there is a local contact relationship between these two superpoints, and an edge connection can be established in the graph. For the superpoint pair (S k , S l ) that satisfies the adjacency condition, the system calculates its edge weight w kl . The calculation of the edge weight adopts the form of a combined Gaussian kernel function, and is calculated by combining the normal vector difference and the centroid position difference where w kl is the edge weight between the superpoint S k and S l , exp is the natural exponential function, used for kernel weight calculation, is the average normal vector of the superpoint S k , is the average normal vector of the superpoint S l , σ n is the Gaussian kernel bandwidth of the normal difference, with a value of 0.3, c k is the geometric centroid position coordinate of the superpoint S k , c l is the geometric centroid position coordinate of the superpoint S l , σ p is the Gaussian kernel bandwidth of the centroid difference, with a value of 0.5; when the system constructs the topological graph, it performs edge screening. If the calculated edge weight w kl is greater than the set threshold ∈ (such as ∈ = 0.2), then the edge (S k , S l ) is retained; if there is an obvious normal vector inversion phenomenon (such as the included angle approaching 180°) or the semantic boundary belongs to different categories between the two regions connected by the edge, the system prohibits the establishment of this edge to avoid the error propagation path. The system outputs the completed local geometric topological graph.

[0111] S32. Initialize the label potential field for the local geometric topological graph data to obtain the label potential field data;

[0112] In one embodiment, initialize the seed label as the propagation potential source to form a drivable label diffusion potential field. For each graph node Sk Set the potential energy value φ k : For the seed label set, i.e., the initially labeled set of nodes, set the initialization with confidence weights. conf k For node S k The label confidence score is usually derived from preliminary predictions or human-annotated confidence scores, while simultaneously constructing a label matrix. φ k The propagation model is used as the initial state input, and the propagation path empty set of each label potential source is recorded.

[0113] S33. Perform local topological tag extension propagation based on the tag potential energy field data to obtain topological propagation data;

[0114] In one embodiment, potential-driven label propagation is performed on the topological graph, and semantic coherence and boundary suppression are controlled during the propagation process. Potential diffusion propagation iterations are performed (based on graph Laplacian smoothing). in For node S k Label confidence after the (t+1)th round of propagation, Z k For w kl The sum, where l is the order of adjacent superpoints. For node S k The set of adjacent nodes, w kl For the super point S k Its adjacent superpoint S l The edge weights between them represent their geometric or semantic similarity. For adjacent node S l Label confidence in round t; node confidence φ updated in each round of propagation. k After each round of propagation, determine whether a label can be assigned. Where y k For the current node S k The propagation label (determined by weighted majority), To select the label with the highest y, l represents the adjacent superpoint order. For node S k The set of adjacent nodes, As an indicator function, when the adjacent node S l The value is 1 when the label is y, and 0 otherwise. l For adjacent node S l The tag categories, y is the enumerated tag categories, w kl For node S k With adjacent node S l The edge weights (propagation weights) between them, φ l For adjacent node Sl The label confidence score; if the propagation label score > δ, then it is marked as a propagation label; if and The difference is greater than θ n Propagation is prohibited if there is a high curvature jump at the boundary point of the super-point region; propagation iteration termination conditions include reaching the maximum propagation round number T; or no new label assignment is generated in this round.

[0115] S34. Extract the propagation path structure based on the topology propagation data to obtain the first label propagation data.

[0116] In one embodiment, for each supernode S that obtains a label through propagation k The system traces the complete path from the initial label point (i.e., the seed point) to form the label source chain. This path is represented as... ,in This represents the starting node of the propagation, i.e., the original label seed point; S k The target superpoint for receiving labels; intermediate nodes These represent the transition nodes in the propagation process along the graph structure. For each propagation path, the system calculates the cumulative edge weight of the path, representing the sum of the propagation intensity along that path. The calculation formula is as follows: PS k To propagate from the tag source to Superpoint S k The propagation path strength (cumulative edge weight), where (i,j) is any pair of adjacent nodes in the label propagation path, Path k For the super point S k The label propagation path (node ​​sequence), w ij The edge connection weights between adjacent node pairs in the path are calculated; the average propagation confidence is calculated by statistically analyzing the average potential value of nodes on the path to represent propagation stability; the path length, i.e., the number of nodes (or edges) in the path, represents the layer-level distance traversed by the propagation. If a target superpoint S... k There are multiple propagation labels originating from different paths (i.e., they may propagate to the same node from multiple seed points). Compare all labels pointing to S. k The propagation path score (e.g., path strength, average confidence) is calculated; the propagation path with the highest score is retained as the recorded path, while other redundant paths are ignored to ensure the uniqueness and optimality of the propagation record. The system organizes the above information into standard structural units, each structural record containing the following fields, including the superpoint number S. k ; propagate label value y k That is, the label obtained by the node's propagation; propagation confidence φ k That is, the final potential energy value of the node; propagation path. k , represents the propagation path chain of the tag; path strength PSk , represents the cumulative propagation intensity along the path; path length L, represents the number of edges the propagation crosses.

[0117] Optionally, the projection-guided aggregation includes:

[0118] The projection space is constructed from the label mapping data to obtain the projection space data;

[0119] In one embodiment, for each superpoint S k The system extracts its structural feature vector F. k This includes, but is not limited to, the following components, such as the average normal vector. This indicates the overall directionality of the normal distribution of the point cloud within the superpoint; the centroid position c k That is, the geometric center of all points constituting the superpoint; point cloud density index and curvature statistics; preliminary label confidence φ. k This characterizes the credibility of the existing labels on the superpoint. After concatenating and encoding the above structural feature vectors, a superpoint structural feature matrix F = [F1; F2; ...; F...] is constructed. K [ ], where K represents the total number of hyperpoints. The system performs embedding processing on the feature matrix F to obtain the low-dimensional coordinates z of each hyperpoint in the projection space. k ,Right now or in The embedded two-dimensional projected coordinates, The embedded 3D projection coordinates are denoted by ε(·), which represents the dimensionality reduction embedding function. ε is the preset graph convolutional network parameter, which is encoded through graph convolution. Feature aggregation is achieved by combining the adjacency matrix, and the embedding matrix is ​​output. The adjacency probability distribution is constructed using the inner product between the embedding vectors. The predicted adjacency relationships are compared and learned with the true adjacency matrix AAA using the binary cross-entropy loss function to optimize the graph embedding results.

[0120] Tag gravity center data is obtained by extracting the tag gravity center from the projected spatial data.

[0121] In one embodiment, the system uses the already labeled superpoints y k For the embedded projected coordinates z k Perform label classification and construct a set of projected coordinates for each label category, denoted as . in Let z be the set of all superpoints with label y in the projected space. k The projected coordinates (position in the low-dimensional projected space) of the labeled superpoint are given by y. k For super point z k The tag category, where y is a specific tag category. Given a known set of labels, the system provides the point set for each label category y. Density estimation is performed to construct the density function ρ in the label space. y (z). The density estimation method used is kernel density estimation or Gaussian mixture model. The density function ρ is obtained... y After (z), the density-weighted mean point is calculated to determine whether convergence has occurred. If convergence is found to the density peak, a peak search operation is performed in the density field to extract the set of candidate gravitational centers. Each center This corresponds to a peak position of the label density. If the number of candidate centers is greater than 1, the system will categorize the centers according to their density values. Sort the data and select the one with the highest density as the principal gravitational center of label y. And calculate its mass fraction. The system can calculate the support radius r of the gravitational center. y , which represents the effective range of label aggregation that the center can attract in the projection space, defined by the isodense profile or Gaussian variance.

[0122] The label gravity center data is projected onto the non-label gravity center data from the projection space data to obtain non-local projection data;

[0123] In one embodiment, for each unlabeled superpoint S j Extract its embedded coordinates in the projection space. Subsequently, the system iterates through all label categories y and calculates the gravitational force between the point and the center of gravity of each label. The spatial similarity weights between them are defined using a Gaussian kernel function as follows: Where w jy For the super point S j To the center of gravity of the label The initial similarity weights (calculated using the Gaussian kernel function), z j For the unlabeled superpoint S j The embedded coordinates in the d-dimensional projective space, where exp is the natural exponential function. Let y be the embedding coordinates of the gravitational center of label y in the projected space, and σ be the Gaussian kernel bandwidth parameter, controlling the degree of distance attenuation. The system applies mass weighting to the initial weights to obtain the corrected gravitational weights. in For the super point S j The mass-weighted gravitational weight of label y, q y ∈[0,1] represents the gravitational center mass score of label y, w jy For the super point S j To the center of gravity of the label Initial similarity weights (calculated using a Gaussian kernel function). Adjusted weights for all labels. After normalization, the unlabeled superpoint S is obtained. j Normalized gravitational distribution probability vector P for various labels j (y): Where P j (y) represents the unlabeled superpoint S. j The normalized gravitational probability value for label y. For the super point S j The gravity weights are weighted by the quality of label y, and y′ is the label index variable used to traverse label categories during normalization. For the super point S j The mass-weighted gravitational weights for label y′. If there exists a label y that satisfies the maximum gravitational probability... (Where δ is a preset confidence threshold, such as 0.7), then the superpoint is considered a potential label point. The output non-local projection data is... This dataset records the probability distribution of each unlabeled superpoint's affiliation in the labeled gravitational space.

[0124] The projection-guided aggregation map data is obtained by processing the non-local projection data into a projection-guided aggregation map.

[0125] In one embodiment, all unlabeled superpoints S j As ordinary nodes in the graph, and associated with nodes satisfying the gravitational probability condition P j Virtual connections are made between the labeled gravitational centers (y) > ∈ (where ∈ is a preset lower limit of gravity), forming a graph structure G. proj The tag gravity center participates in propagation as a virtual tag node in the graph. Each connecting edge is based on the tag gravity probability P calculated in the previous stage. j (y) Define the initial edge weight w jy The system normalizes the weight matrix and generates unlabeled node S. j The propagation transfer weight vector T j This is used to control the propagation weight distribution of each label at that point. On this guide graph structure, the system can optionally execute one or more propagation processes, for example, denoted by propagation time step t, where the label fusion score vector in the (t+1)th round... Based on the current propagation weight and the previous round of label score The weighted summation is represented as Where y represents the label category corresponding to the label gravity center node, and T jy For the super point S j Normalized propagation transition weights for labeled node y. If an unlabeled point S... jIf multiple label scores are close and all exceed a preset threshold, the node is marked as having an ambiguous label state. The system outputs the constructed projection-guided aggregation map data. This data structure includes each unlabeled node S j The candidate label gravity probability distribution P j (y), and its list of connection label nodes in the graph.

[0126] Multi-label fusion is performed on the projection-guided aggregation map data to obtain fuzzy label data;

[0127] In one embodiment, the label probability distribution is preserved when multiple labels are possible, forming fuzzy labels. For each unlabeled superpoint S to be processed... j The system directly uses the tag gravity probability distribution P calculated in the previous stage. j (y) serves as the basis for its multi-label fusion. Where P j (y) represents the gravitational probability of label y for this superpoint, reflecting the relative attraction during label propagation. The system determines whether this superpoint is in a definite label state. The criterion is that if the maximum label probability of this superpoint (i.e., ...) is... If the value is greater than the preset confidence threshold δ, the label status is marked as clear; if the value is lower than δ, the label is considered ambiguous and its status is marked as vague. The system calculates the entropy index of the label distribution, and for each label distribution P... j (y), calculate its information entropy H j =-∑ y P j (y)log2P j (y), where H j For the super point S j The information entropy of the label distribution, where y is the label category, and P is the label distribution information entropy. j (y) represents the label y for the superpoint S. j The gravitational probability (i.e., the degree of attraction for label propagation) is calculated. A higher entropy value indicates a greater likelihood that the point may belong to multiple label categories. The system organizes the above multi-label fusion results into structured data records, each containing a superpoint identifier, label probability distribution, label distribution entropy value, and a state label (clear or fuzzy).

[0128] Boundary suppression optimization is performed on the fuzzy label data to obtain the second label propagation data.

[0129] In one embodiment, the system identifies boundary points in a blurred region. The specific identification criteria include two categories: one is that the label distribution entropy is higher than a set threshold θ. HThe first type of region represents areas with significant label uncertainty; the second type contains adjacent superpoint pairs with abrupt changes in label probability, indicating significant differences in label distributions and constituting potential boundary regions. For these identified fuzzy boundary regions, the system enhances spatial continuity through label regularization, including but not limited to conditional random field models or graph-based regularization models. One optimization method uses an optimization energy function defined as: calculating the total label energy... Where E(P) is the total energy function value of all superpoints, j is the superpoint order term, and H(P) is the total energy function value of all superpoints. j () is the superpoint S j The label distribution entropy, λ is the weight of the balance term, and (i,j) are adjacent superpoint pairs. For the adjacency relationship between superpoints, P i For the super point S i The label prediction probability, P j For the super point S j The label prediction probability is calculated. Another optimization approach is based on a graph neural network module, which performs neighborhood confidence reconstruction and smooth propagation on fuzzy labels within a graph structure. Through a multi-layer graph message passing mechanism, context enhancement is applied to fuzzy labels, resulting in more robust label confidence predictions. After label optimization, the system, based on the updated label distribution probability vector, calculates the label prediction probability for each superpoint S. j The label category corresponding to the highest probability is selected as its propagation label, i.e. Where y j For the super point S j The final predicted label category, To make the label probability P j (y) The largest category y, P j (y) is the superpoint S j The predicted label probability for category y is calculated, and the updated confidence score (conf) is used based on the maximum label probability value. j The system updates the state labels (such as labeled or stable labels) based on their entropy or propagation state. The system outputs the second-stage label propagation dataset, including the final label, corresponding confidence score, and state identifier for each participating superpoint.

[0130] Optionally, the tag gravity center extraction includes:

[0131] The label distribution density map is calculated from the projected spatial data to obtain the label distribution density map data;

[0132] In one embodiment, the system first acquires a set of points in the two-dimensional projection space. Each point z i The system obtains the projected coordinates of its original superpoint or pixel. iThis represents the semantic category labeling result. The system classifies and aggregates all points according to the label category y, forming the set of projected coordinates Z corresponding to each label category. y ={z i ∣y i =y}, that is, the set of all projection points of label y is Z. y The system has a point set Z for each type of label. y Perform two-dimensional density function estimation. Density function ρ y (z) represents any point in the projection space. The spatial distribution intensity of label y at a given location is defined as follows: Where ρ y (z) represents the distribution density value of label y at the projected coordinate z, |Z y | represents the number of sample points under label y, z i Let Z be the two-dimensional projected coordinates of the i-th point. y Let be the set of all projected coordinates corresponding to label y, exp be the natural exponential function, z be the two-dimensional coordinate position in any projected space, and σ be the kernel bandwidth parameter used to control the smoothness of the density distribution. This parameter can be set empirically (e.g., σ = 0.5) or adaptively determined using the Silverman rule. The system will assign the density function ρ to each label category y. y (z) is mapped to a 2D density map. The density map output resolution is set to 128×128. A separate density map function is output for each label class.

[0133] Gravitational center candidate field processing is performed on the tag distribution density map data to obtain gravitational center candidate field data;

[0134] In one embodiment, candidate points that may become semantic centers are extracted from the density map to form a gravitational center field. For the density map ρ y (z) Perform local extremum detection: For each grid point z i,j Check if it is a local maximum within a 3×3 neighborhood; exclude values ​​with a density less than the threshold ρ. min (e.g., 0.05) weak centers; each detected extreme point Each candidate point (center) is considered a candidate center, and its location and local density value are recorded. For each candidate point (center), the local attraction radius is calculated, which is the region where the density around it drops to half its maximum value. Alternatively, calculate the decay distance along the direction of the fastest descent of the density gradient; to obtain the central candidate field. For the k-th candidate center point Density values ​​at a given location (i.e., local peaks). Output candidate field data for the gravitational center.

[0135] Multimodal label center clustering was performed on the candidate field data of the gravitational center to obtain the center clustering data;

[0136] In one embodiment, local candidate centers are aggregated to avoid redundant centers or noise peaks under the same label, and the main mode with a clear structure is extracted. All candidate center points are then... As sample points, DBSCAN (Density Peak Clustering) or the system calculates local kernel density around each point to obtain new density-weighted centers, and iterates and updates them until the positional change is less than a threshold. Distance is then used to determine the centers of all converged points. If the Euclidean distance between two converged points is less than the merging threshold, they are considered the centers of the same cluster, and the principal modes are merged. The recommended distance threshold is δ = 0.1 (based on the scaling ratio of the projected spatial coordinates). Each cluster... For each potential mode: if the number of clusters is less than 2, it indicates that the label of that class is unimodal; otherwise, it is a multimodal label center; calculate the centroid (such as the mean or density-weighted centroid) for each cluster. in Let i be the k-th candidate gravity center point belonging to label m, and let i be the cluster. All sample point indices, Let c be the m-th cluster (modality), which consists of a set of candidate centroids. i Let ρ be the coordinate vector of the i-th candidate centroid in the cluster. i The density value of the i-th candidate center point (which can be the local density of the point set or the response intensity) is output as the center clustering data, which contains multiple candidate main centers.

[0137] Representative centers are selected based on the central clustering data to obtain the label gravity center data.

[0138] In one embodiment, one or more of the most representative cluster centers are selected from multiple cluster centers to guide label aggregation. For each modality center... Calculate the following metrics, including peak density. ρ y Let be the kernel density function for category y. Let the coordinates of the m-th modality center of category y be given; calculate the coverage. The proportion of points within its radius of influence to the total number of points of this type; calculate the representative influence function: in The data consists of center influence data, with α representing peak density weighting (α = 0.6) and β representing coverage weighting (β = 0.4). Based on the center influence data, the top 1-2 centers with the highest scores are selected as gravitational centers. If multimodal propagation is required, all modal centers can be retained and propagated separately. The output is labeled gravitational center data.

[0139] Optionally, the label propagation feature extraction includes:

[0140] Based on the first label propagation data and the second label propagation data, a label propagation evolution trajectory map is constructed to obtain propagation trajectory map data;

[0141] In one embodiment, the system uses point or super-point units S in the point cloud. i As graph nodes, all units participating in the label propagation process are selected to form the propagation node set V. For each node S... i The system establishes a dual-path label record table, which includes the label results obtained from the first label propagation path. The label results obtained from the second label propagation path Tag propagation source path identifier in and This indicates the label source location of the node in both types of propagation (such as its original seed node number or path position code). When constructing the graph structure, the system builds the edge set E based on the adjacency relationships between supernodes or the continuity of propagation paths. Each edge e... i,j ∈E indicates that there is a label propagation association between two propagation nodes, and the edge attributes include the following indicators, including label variability. Let y be an indicator function, if y i ≠y j The value is 1 if it is y, otherwise it is 0. i For node S i The tag, y j For node S j Tags; Source Difference d src =dist(src) i ,src j ), dist is the distance between source path locations, src i For node S i The location of the propagation source path, src j For node S j The location of the propagation source path is determined, and this distance is calculated based on the Euclidean distance of the tag source point in the propagation graph, the hierarchical difference, or the number of path hops. A tag propagation evolution trajectory graph is constructed. Where A represents the set of propagation attribute information attached to each edge in the graph, including label evolution indicators, path source information, and propagation variation characteristics.

[0142] Label similarity and label stability were calculated on the propagation trajectory map data to obtain label similarity data and label stability data, respectively.

[0143] In one embodiment, label similarity is calculated for each propagation node S. iThe system calculates the label results obtained in the two types of propagation paths. and similarity index sim i If the labels are in hard label form (i.e., discrete category labels), define a label consistency discriminant function. Where sim i For node S i Tag similarity metrics, This is an indicator function; the value is 1 if the labels are the same, and 0 otherwise. If soft labels or probability distribution labels are used... and Similarity can then be measured using distribution distance, such as Where ||·||1 represents the L1 norm (i.e., the sum of absolute values). At the spatial scale, a local label similarity map can be constructed. Simultaneously, the overall propagation label consistency index is calculated. Sim global Here, |V| represents the global label consistency index (average label similarity), |V| represents the total number of propagation nodes, and i represents the node order. Label stability calculation assesses the fluctuation of label evolution at each node during propagation. If intermediate label states exist in the propagation path (e.g., intermediate label sequences in multiple rounds of propagation), the system can calculate the rate of change or number of jumps in the label sequence; if a node label comes from the aggregation of multiple neighboring nodes, the variance or entropy of the label aggregation result can be calculated. The stability index stab is defined as follows: i =1-Var(y i (src)) or stab i =1-Entropy(p i ), where stab i For node S i The label stability index, Var(y) i (src)) represents the variance of the tag results aggregated from the source path, and Entropy(p) represents the variance of the tag results aggregated from the source path. i The value represents the entropy of the current label distribution. Higher label stability indicates stronger consistency and aggregation among nodes during label propagation.

[0144] Based on label similarity data and label stability data, the propagation trajectory map data is subjected to propagation path label aggregation feature extraction to obtain label propagation feature data.

[0145] In one embodiment, a propagation trajectory diagram is used. Based on this, where V represents the set of propagation nodes, E represents the propagation edges between nodes, and A is the set of attributes of the propagation edges. The system assigns each node S... i Construct the propagation aggregation vector f ∈V iThis vector integrates multiple propagation-related features, including the label similarity index sim. i ; Tag stability index stab i Node degree deg(i), representing the number of connections a node has in the propagation graph; label confidence. These correspond to the label confidence levels of the node in the first and second propagation paths, respectively. The node propagation aggregation vector is defined as follows: The system constructs an edge attribute vector φ for each pair of adjacent nodes (i,j)∈E. ij It includes the following features, such as label variants. Determine if there is a jump in the labels of two nodes; similarity difference |sim i -sim j |, Measures the change in consistency of label propagation; Differences in propagation probability distribution ||p i -p j ||2, The L2 norm is used to measure the distance between label probability distributions. The edge propagation jump feature is defined as φ. ij =[δ ij ,|sim i -sim j |,||p i -p j ||2]. The system uses a graph neural network (such as GCN) to perform path convolution aggregation on the node and edge features to extract a high-dimensional propagation fusion feature map, and then integrates the propagation information using aggregation operations (such as average pooling or max pooling). The system performs clustering on highly consistent regions in the propagation map (i.e., subgraphs with highly consistent node labels and low jump rates) to extract structurally reliable clusters. The aforementioned integration yields the label propagation feature data.

[0146] Optionally, the confidence fitting includes:

[0147] The label confidence scores of the first label propagation data and the second label propagation data are calculated based on the label propagation feature data to obtain the label confidence scores.

[0148] In one embodiment, for each structural unit (e.g., superpoint) S i The system obtains the label prediction probability distributions obtained through two propagation paths (e.g., geometric propagation and projection potential propagation): Where C is the number of label categories. This represents the label prediction probability distribution along the path k∈{1,2}. The system extracts structural unit S. i The propagation characteristic data, including the label similarity index sim i This indicates the consistency of the prediction between the unit and neighboring or similar labeled nodes; the propagation stability index stab iThis indicates the degree of change in the predicted label value over multiple propagations; higher stability indicates a more reliable prediction. The path prediction confidence term... This represents the maximum component in the label distribution along path k, reflecting the strength of label preference. For each path propagation result, the system calculates the confidence level of the structural unit along that path. in for, To determine the propagation stability weight, a value of 0.3 is assigned. This is the tag similarity weight, with a value of 0.4. The weight for predicting preference intensity is set to 0.3. Let be the largest component (representing preference strength) in the label probability distribution of path k. To take the maximum value, Let be the predicted probability value under category c, and represent the largest component in the label probability distribution. Optionally, after obtaining the individual confidence scores of the two paths, the path with the higher confidence score can be selected, or the confidence scores of the two paths can be normalized and fused to obtain the label confidence dataset.

[0149] Confidence conflict resolution is performed based on the label confidence data to obtain confidence adjustment data;

[0150] In one embodiment, the system identifies two propagation paths. Structural units where there are discrepancies in the label output results. Define the label confidence conflict region. A set of structural elements that satisfy the following two conditions: in A confidence conflict region is a set of structural units in two paths that have different label categories but similar confidence levels. i Let be a structural unit (such as a point or superpoint), and arg max be the label category with the highest predicted probability on the path. Structural unit S i Label prediction probability distribution on path 1 Structural unit S i Label prediction probability distribution on path 2 The confidence score of the label corresponding to path 1. δ represents the label confidence level on path 2. c The set confidence difference tolerance threshold (e.g., 0.2). For those falling into... For structural units within the region, the system performs confidence fusion and reweighting calculations, including but not limited to methods such as using the propagation stability index stab. i Using [0,1] as the fusion weight, the label fusion distribution is calculated as follows: in To predict the probability distribution of the fused labels, `normalize` is used to standardize the fused probability distribution to ensure it is a valid probability vector. `stab` i Structural unit S i The indicators of propagation stability Structural unit S i Label prediction probability distribution on path 1 Structural unit S i The system predicts the probability distribution of labels on path 2; it increases the weight of labels corresponding to paths with higher confidence, for example, by using soft weighting to increase the probability value of labels on high-confidence paths, thus reducing the interference from the other path. For paths not belonging to... The system directly retains the label output of paths with higher prediction confidence, i.e., The system updates the confidence score of the fused label prediction results. For the fused structural unit S i Label confidence score, The probability distribution of the fused labels is predicted by using the largest component in the fused label distribution as the label preference strength, which reflects the prediction confidence.

[0151] A credible probability field is constructed from the confidence-adjusted data to obtain credible probability field data;

[0152] In one embodiment, the system S combines all structural units (such as points or superpoints) i Treat them as nodes in a probability field and construct a spatial adjacency graph G = (V, E), where V = {S} i} represents the set of all nodes; adjacency relationships E can be determined in the following ways: based on spatial K nearest neighbor relationships (e.g., setting k=10), find the k nearest neighbors of each node in space; or define the connection relationships based on an existing structural graph (e.g., a superpoint graph, a boundary topology graph). For each graph node S i Define the initial state of the probability field as follows, such as the label prediction distribution. Fusion prediction derived from the confidence adjustment phase; label confidence score This serves as a confidence measure for the prediction result; both together constitute the probability field initialization conditions. Graph propagation modeling includes the following two methods, for each node S... i Its updated label prediction distribution It can be represented as in For node S i The updated label prediction probability distribution after graph propagation, where j is the index order item. For node S i The set of adjacent nodes, wij For adjacent node S j For node S i Confidence contribution For node S i The initial label prediction probability distribution. Where w ij For adjacent node S j For node S i The confidence contribution weight, For adjacent node S j The initial confidence score, where k is the K-nearest neighbor parameter used to construct spatial adjacency relationships (e.g., k = 10). For node S i The set of adjacent nodes, For adjacent node S k The initial confidence score. Alternatively, construct a Laplace propagation model and optimize the continuous label field by solving the linear system. in Let P be the Laplacian matrix of the graph, encoding adjacency relationships and weights; let P be the smoothed probability distribution matrix to be solved; and let Q be the source term matrix containing trusted node guidance terms, defining strong constraints or weak supervision guidance for some high-confidence nodes. The system outputs trusted probability field data, including nodes in the probability field, the label prediction distribution after graph domain propagation / smoothing, and the corresponding updated confidence scores, calculated based on the post-propagation probability peak or adjacency consistency.

[0153] The label distribution of the first label propagation data and the second label propagation data is optimized based on the reliable probability field data to obtain the label propagation data.

[0154] In one embodiment, the propagation results of the first and second tags are globally optimized based on a confidence probability field, making the tag distribution more continuous, stable, and highly confident. For each S... i Fusion tag output or in The output of the fused labels represents the category with the highest probability. For the midpoint or superpoint S of the credible probability field i The label prediction distribution vector, for The category index with the highest probability value, if S i If it is located in the edge region (determined by the variance of the neighboring labels), then its label is smoothed; if S i If the confidence level is too low, you can revert to using the initial path labels. Output label propagation data.

[0155] Optionally, S4 includes:

[0156] S41. Construct a tag reflection index mapping table for the tag propagation data to obtain tag reflection data;

[0157] In one embodiment, an efficient mapping relationship is established between superpoints and the original points in the point cloud for back-projection of the labeling results. During the superpoint construction phase (see Superpoint Topology Graph Construction), each point p has already been mapped... i Assigned to its superpoint unit S j At this point, an index mapping table can be constructed. Some points may fail to match superpoints (such as isolated points or boundary points). Nearest neighbor or multi-scale KNN matching is used to find the nearest superpoint. in To find the superpoint index that minimizes the distance, dist is the distance function, and p i For the i-th point in the original point cloud, centroid(S) k () is the superpoint S k The geometric center (centroid) will propagate the label data. Each superpoint label y in j Through mapping Project back to each p i Up. Output label reflection data.

[0158] S42. Calculate the spatial confidence point-level weights based on the label reflection data to obtain the point-level weight data;

[0159] In one embodiment, each original point p i Its corresponding superpoint S j The mapping relationship between them is established using label reflection data, denoted as Where conf j Indicates the superpoint S j The label confidence level. The system uses conf j As the initial confidence weight for this point, denoted as (where p) i ∈S j ),in For point p i The initial confidence weights are determined. The system uses a boundary structure sensitivity factor to characterize the label stability of the local region where the point is located. Point p is defined. i The local neighborhood is Calculate the variance σ of the label categories of points within its neighborhood. i This indicates the degree of label variation. Where σ i For point p i The label variance index represents the degree of consistency of label categories within its neighborhood. Var is the variance function, and y k For the neighborhood point p kThe tag category, p k Let p be the k-th point in the original point cloud. i The adjacent nodes; To maximize the global label variance and avoid the influence of outliers, the system uses a weighted fusion method to integrate the two factors mentioned above, obtaining the point-level label confidence weights. Where λ∈[0,1] is the fusion weight coefficient (e.g., a value of 0.7).

[0160] S43. Fine-tune the label boundaries based on the point-level weight data to obtain point cloud annotation data.

[0161] In one embodiment, for each point cloud point p i Define its label neighborhood inconsistency index η i , Where η i For point p i The label neighborhood inconsistency index p is the number of neighboring points. k For a certain adjacent point in the neighborhood, For point p i The neighborhood set, Let y be an indicator function. k For adjacent point p k The tag category, y i For point p i Current tag category. When η i When the threshold τ is set to 0.3, point p i These are marked as label boundary points. For the identified boundary points, a label re-evaluation mechanism based on point-level weights is used for fine-tuning, such as performing weighted k-nearest neighbor voting to select the category with the highest weighted frequency as the new label. in For point p i The re-evaluated new label categories To find the category that maximizes the expression, p k For a certain adjacent point in the neighborhood, For point p i The neighborhood point set (e.g., constructed based on sphere radius or k-nearest neighbors), This is an indicator function; it takes a value of 1 if the condition is met, and 0 otherwise. k For the neighborhood point p k The weights. Alternatively, calculate the weighted average of the neighborhood label distributions to form a new label probability vector. in For point p i The re-estimated label probability distribution is obtained by taking the weighted average of the label probabilities within the neighborhood, p. k For a certain adjacent point in the neighborhood, For point p i The neighborhood point set, p k For adjacent point p k The label probability distribution vector, w k For the neighborhood point p k The weights are determined, and the category corresponding to the maximum value is taken as the label. in For point p i The re-evaluated new label categories To select the tag category with the highest probability, To re-estimate the probability value corresponding to label y in the probability vector, for points not belonging to the boundary region, their original labels remain unchanged, and only their label confidence weights are regularized. Morphological operations (such as dilation-erosion) are applied to the initially corrected boundary results to complete label hole filling and boundary closure within a small area.

[0162] Optionally, this application also provides a point cloud data rapid annotation system for performing the point cloud data rapid annotation method described above, the point cloud data rapid annotation system comprising:

[0163] The point cloud structure modeling and initial label acquisition module is used to acquire point cloud data and initial point label data; and to construct superpoint structures based on the point cloud data to obtain point cloud superpoint structure data.

[0164] The label mapping and seed annotation extraction module is used to map the initial point annotation data to the point cloud superpoint structure data to obtain label mapping data;

[0165] The dual-path semantic propagation and confidence fusion module is used to perform local topological fidelity label propagation on the label mapping data to obtain the first label propagation data; to perform projection-guided aggregation on the label mapping data to obtain the second label propagation data; to extract label propagation features based on the first and second label propagation data to obtain label propagation feature data; and to perform confidence fitting on the first and second label propagation data based on the label propagation feature data to obtain the label propagation data.

[0166] The label reflection and point-level annotation output module is used to reflect the label propagation data to obtain point cloud annotation data.

[0167] Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended application documents rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the application documents be incorporated into the invention.

[0168] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for rapid annotation of point cloud data, characterized in that, The method includes: S1. Acquire point cloud data and initial point annotation data; construct the superpoint structure based on the point cloud data to obtain the point cloud superpoint structure data; S2. Map the initial point label data to the point cloud superpoint structure data to obtain label mapping data; S3. Construct a local geometric topology graph from the tag mapping data to obtain local geometric topology graph data; initialize the tag potential field from the local geometric topology graph data to obtain tag potential field data; perform local topology tag extension propagation based on the tag potential field data to obtain topology propagation data; extract the propagation path structure from the topology propagation data to obtain first tag propagation data; construct a projection space from the tag mapping data to obtain projection space data; extract the tag gravity center from the projection space data to obtain tag gravity center data; project the tag gravity center data onto the non-tag gravity center data from the projection space data to obtain non-local projection data; perform projection-guided aggregation graph processing on the non-local projection data to obtain projection-guided aggregation graph data; perform multi-tag fusion on the projection-guided aggregation graph data to obtain fuzzy tag data; perform boundary suppression optimization on the fuzzy tag data to obtain second tag propagation data; extract tag propagation features from the first and second tag propagation data to obtain tag propagation feature data; perform confidence fitting on the first and second tag propagation data based on the tag propagation feature data to obtain tag propagation data. S4. Perform label reflection on the label propagation data to obtain point cloud labeled data; The extraction of the tag's gravitational center includes: The label distribution density map is calculated from the projected spatial data to obtain the label distribution density map data; the label distribution density map data is processed into a candidate field of gravitational centers to obtain candidate field data of gravitational centers; the candidate field data of gravitational centers is clustered into multimodal label centers to obtain center cluster data; and representative centers are selected based on the center cluster data to obtain the label gravitational center data.

2. The method according to claim 1, characterized in that, The superpoint structure construction includes: Structural perception feature extraction is performed on point cloud data to obtain structural perception feature data; Heterogeneous fusion of structural sensing feature data is performed to obtain heterogeneous fused data; Multi-scale semantic region segmentation is performed on heterogeneous fused data to obtain region segmentation data; Superpoint topology graphs are constructed from the region segmentation data to obtain superpoint topology graph data; Consistent encoding is performed on the superpoint topology data to obtain the point cloud superpoint structure data.

3. The method according to claim 1, characterized in that, S2 include: A point-to-superpoint mapping index is constructed based on the initial point annotation data and the point cloud superpoint structure data to obtain the mapping index data; Tag weight fusion is performed based on the mapping index data to obtain fused tag data; Perform super-point confidence initialization on the tag fusion data to obtain super-point confidence data; Label validation is performed on the over-confidence data to obtain label validation data; The tag mapping data is obtained by performing mapping processing based on the tag verification data and the mapping index data.

4. The method according to claim 1, characterized in that, The tag propagation feature extraction includes: Based on the first label propagation data and the second label propagation data, a label propagation evolution trajectory map is constructed to obtain propagation trajectory map data; Label similarity and label stability were calculated on the propagation trajectory map data to obtain label similarity data and label stability data, respectively. Based on label similarity data and label stability data, the propagation trajectory map data is subjected to propagation path label aggregation feature extraction to obtain label propagation feature data.

5. The method according to claim 1, characterized in that, The confidence level fitting includes: The label confidence scores of the first label propagation data and the second label propagation data are calculated based on the label propagation feature data to obtain the label confidence scores. Confidence conflict resolution is performed based on the label confidence data to obtain confidence adjustment data; A credible probability field is constructed from the confidence-adjusted data to obtain credible probability field data; The label distribution of the first label propagation data and the second label propagation data is optimized based on the reliable probability field data to obtain the label propagation data.

6. The method according to claim 1, characterized in that, S4 includes: A tag reflection index mapping table is constructed on the tag propagation data to obtain the tag reflection data; Spatial confidence point-level weights are calculated based on the label reflection data to obtain point-level weight data; The label boundaries are fine-tuned based on the point-level weight data to obtain point cloud labeled data.

7. A rapid point cloud data annotation system, characterized in that, For performing the point cloud data rapid annotation method as described in claim 1, the point cloud data rapid annotation system includes: The point cloud structure modeling and initial label acquisition module is used to acquire point cloud data and initial point label data; and to construct superpoint structures based on the point cloud data to obtain point cloud superpoint structure data. The label mapping and seed annotation extraction module is used to map the initial point annotation data to the point cloud superpoint structure data to obtain label mapping data; The dual-path semantic propagation and confidence fusion module is used to perform local topological fidelity label propagation on the label mapping data to obtain the first label propagation data; to perform projection-guided aggregation on the label mapping data to obtain the second label propagation data; to extract label propagation features based on the first and second label propagation data to obtain label propagation feature data; and to perform confidence fitting on the first and second label propagation data based on the label propagation feature data to obtain the label propagation data. The label reflection and point-level annotation output module is used to reflect the label propagation data to obtain point cloud annotation data.

Images