A point cloud data segmentation method and system based on deep learning

Abstract

The application relates to the technical field of point cloud data segmentation, in particular to a point cloud data segmentation method and system based on deep learning. The method comprises the following steps: obtaining point cloud data, performing point cloud feature extraction on the point cloud data to obtain point cloud feature data; performing structure perception embedding on the point cloud feature data to obtain structure perception graph data; performing pseudo label enhancement training on the structure perception graph data to obtain pseudo label enhancement data; performing attention re-correction on the pseudo label enhancement data to obtain re-correction data; and performing enhancement training on the re-correction data to obtain a point cloud segmentation model. By introducing a structure perception graph neural network and a pseudo label enhancement mechanism, the adaptability of the point cloud segmentation model to spatial structure relationships and weak supervision labels is significantly improved. In combination with a multi-branch semantic encoding and boundary perception decoding strategy, the boundary recognition precision and semantic segmentation stability of the model in a working condition scene are enhanced.

Description

Technical Field

[0001] This invention relates to the field of point cloud data segmentation technology, and in particular to a point cloud data segmentation method and system based on deep learning. Background Technology

[0002] With the rapid development of 3D sensing technologies (such as LiDAR, RGB-D cameras, and ToF), point cloud data has become an important form of representing 3D spatial structures. In power systems, with the widespread application of 3D laser scanning and UAV inspection technologies, point cloud data plays a crucial role in tasks such as transmission tower modeling, line facility identification, and structural analysis. Point cloud segmentation, as a core step in realizing semantic understanding and structural reconstruction of power facilities, is of great significance for improving the automation level of power inspection and constructing digital power grid models. Traditional point cloud segmentation methods mostly rely on geometric rules or manually designed features, making it difficult to adapt to high-noise, low-density, or severely occluded environments in actual operating scenarios. While deep learning-based methods have made significant progress, they generally face challenges such as the difficulty in obtaining accurate labels for point cloud data in real-world scenarios, and weak or unsupervised training often leading to performance degradation in segmentation. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a point cloud data segmentation method and system based on deep learning, thereby resolving at least one of the aforementioned technical issues.

[0004] This application provides a point cloud data segmentation method based on deep learning, the method comprising:

[0005] S1. Acquire point cloud data and extract point cloud features from the point cloud data to obtain point cloud feature data; perform structure-aware embedding on the point cloud feature data to obtain structure-aware map data.

[0006] S2. Perform pseudo-label augmentation training on the structure-aware map data to obtain pseudo-label augmented data;

[0007] S3. Perform attention recalibration on the pseudo-labeled augmented data to obtain recalibrated data;

[0008] S4. Enhance the training of the recalibrated data to obtain the point cloud segmentation model.

[0009] This invention utilizes structure-aware embedding to fully model the geometric relationships and local topological features between points in a point cloud, enhancing the expressive power of point cloud data in unstructured spaces. Pseudo-label-enhanced training effectively alleviates the problem of insufficient labeled data and guides the model to perform self-supervised optimization on the structure graph, improving semantic consistency. Attention recalibration dynamically adjusts the semantic responses between points, strengthening the model's ability to distinguish boundary and ambiguous regions. Combining recalibrated data with structure-preserving data-enhanced training constructs a point cloud segmentation model with boundary recognition capabilities and global consistency, thereby improving segmentation accuracy and structural adaptability, and exhibiting stronger robustness in practical applications.

[0010] Optionally, the point cloud feature extraction includes:

[0011] Acquire point cloud data and extract geometric features from the point cloud data to obtain geometric feature data;

[0012] Auxiliary modal feature extraction is performed on the geometric feature data to obtain auxiliary modal feature data;

[0013] Cross-modal attention fusion is performed on auxiliary modal feature data to obtain cross-modal fused data;

[0014] A point-to-point relationship graph is constructed based on cross-modal fusion data to obtain point-to-point relationship graph data;

[0015] Graph neural network processing is performed on the point relationship graph data to obtain point cloud feature data.

[0016] This invention significantly improves the completeness and discriminative power of feature representation by processing raw point cloud data in stages and gradually fusing geometric and multimodal information. Geometric feature extraction fully captures spatial structural information between points, such as curvature, normal vectors, and point density, which is beneficial for constructing local topological relationships. Auxiliary modalities (such as reflection intensity, RGB texture, or timestamps) are used to supplement the data, and a cross-modal attention mechanism adaptively fuses key feature channels between modalities, enhancing the model's ability to recognize material boundaries, reflective areas, and heterogeneous structures. Through the construction of point relationship graphs and graph neural network processing, features achieve efficient propagation and contextual aggregation within the spatial structure, significantly enhancing the spatial consistency and semantic discriminative power of point cloud features, providing structurally sensitive and modally unified feature support for segmentation tasks.

[0017] Optionally, the structure-aware embedding includes:

[0018] The spatial relationship between points is processed on the point cloud feature data to obtain the spatial relationship data between points;

[0019] Structural features are extracted from the spatial relationship data between points to obtain structural feature data between points;

[0020] A structure graph is constructed based on the structural feature data between points to obtain the structure graph data;

[0021] Based on the structure graph data, graph neural network structure embedding propagation is performed on the point cloud feature data to obtain embedding propagation graph data;

[0022] The embedded propagation graph data is subjected to graph attention hierarchical aggregation to obtain structure-aware graph data.

[0023] This invention significantly enhances the model's ability to perceive spatial relationships and local geometric changes by performing layer-by-layer structured modeling and graph neural network propagation on point cloud feature data. The spatial relationship processing between points accurately expresses the relative positions and spatial dependencies between points and their neighbors, laying the foundation for refined structural processing. Subsequent structural feature extraction can uncover potential geometric boundaries, density abrupt changes, and connection patterns in the point cloud, effectively identifying local structural features. By constructing a structural graph, the point cloud is mapped into graph structure data with topological expressive capabilities, making information transmission between points controllable and structurally sensitive. The embedding propagation of the graph neural network allows the features of each point to absorb the semantic and geometric information of its adjacent structures, enhancing the expressive power from local to global. Layered aggregation of graph attention can adaptively emphasize key structural regions, forming clearer responses to different or ambiguous shapes and boundary regions, overall improving the discriminativeness and expressive completeness of the structure-aware graph, providing more context-aware feature inputs for the task.

[0024] Optionally, the graph attention hierarchical aggregation includes:

[0025] Dependency path data is obtained by constructing dependency paths from the embedded propagation graph data.

[0026] Based on the dependency path data, the embedded propagation graph data is processed by graph attention multi-head hierarchical processing to obtain multi-head hierarchical data. The graph attention multi-head hierarchical processing includes local geometric boundary head processing, global category dependency head processing, and node feature change rate head processing.

[0027] Perform skip-layer joins on multi-head hierarchical data to obtain cross-layer joined data;

[0028] Subgraph clustering and fusion are performed on the cross-layer connection data to obtain subgraph fused data;

[0029] Implicit edge backflow processing is performed on the subgraph fusion data to obtain structure-aware graph data.

[0030] This invention achieves refined modeling of different structural levels and semantic changes in point clouds by constructing semantic dependency paths and introducing a multi-dimensional attention mechanism. Dependency path construction clearly defines the strength of contextual relationships between points, providing structural priors for attention weight allocation. Multi-head hierarchical graph attention processing uses different attention heads to perceive local geometric boundaries, global category dependencies, and node feature change rates, capturing key structural changes and category mutations in point clouds from multiple perspectives, effectively alleviating boundary ambiguity and multi-class interference problems. Skip-cross-layer connections allow long-distance semantic information to propagate across layers in the graph structure, enhancing the model's adaptability to structures at different scales. Subgraph clustering and fusion steps can aggregate and optimize structurally similar regions, improving consistency within semantic blocks and suppressing feature discretization. Implicit edge backflow processing can repair information flow breaks or propagation direction imbalances in the graph structure, forming more stable, bidirectional, and coherent structure-aware graph data, significantly improving the accuracy and continuity of point cloud structure representation overall.

[0031] Optionally, the implicit edge reflow processing includes:

[0032] Weak connectivity data is obtained by performing edge similarity analysis based on the subgraph fusion data;

[0033] Candidate implicit edges are generated based on weak connection data to obtain candidate implicit edge data;

[0034] Calculate the backflow edge weights on the candidate implicit edge data to obtain the backflow edge weight data;

[0035] Bidirectional edge symmetry correction is performed based on the backflow edge weight data to obtain symmetry correction data;

[0036] The structure-aware map is generated based on the symmetry correction data, resulting in structure-aware map data.

[0037] This invention effectively improves the integrity and bidirectional consistency of information transmission in point cloud graph structures by introducing weak connectivity analysis and backflow path reconstruction based on subgraph fusion. Weak connectivity regions with semantic gaps or propagation bottlenecks in structural continuity are identified through edge similarity analysis, providing precise localization for graph structure repair. Subsequently generated candidate implicit edges can be used to supplement potential semantic pathways missing in the original graph, enhancing the graph's coverage in areas with blurred edges or category transitions. Backflow edge weight calculation constructs controllable reverse edge connections based on feature similarity and spatial adjacency, ensuring the edge-filling operation is discriminative and semantically reasonable. Through bidirectional edge symmetry correction, a balance of bidirectional information flow in the graph is achieved, avoiding label propagation shifts or structural heterogeneity misjudgments caused by unidirectional dependence. The generated structure-aware graph is not only more robust in terms of topological integrity but also more accurately describes the semantic continuity and local structural relationships of boundary regions, thus providing high-quality graph structure support for segmentation or classification tasks.

[0038] Optionally, S2 includes:

[0039] Lightweight pseudo-labels are generated based on the structure-aware map data to obtain initial pseudo-label data.

[0040] The initial pseudo-label data is propagated using structural similarity propagation to obtain pseudo-label propagation data.

[0041] Based on the pseudo-label propagation data, low-confidence pseudo-labels are corrected to obtain pseudo-label correction data;

[0042] Adversarial training is performed on the pseudo-labeled corrected data to obtain pseudo-labeled enhanced data.

[0043] The pseudo-label enhancement data is obtained by performing a two-branch self-distillation optimization based on the pseudo-label enhancement data.

[0044] This invention achieves high-quality pseudo-supervised utilization of structure-aware graph data in the absence of manual annotation by constructing a lightweight generation mechanism and a multi-stage label optimization strategy. A lightweight pseudo-label generation module quickly obtains the initial semantic distribution, providing a foundation for propagation and optimization. Through structural similarity propagation, the structural similarity between nodes in the graph is used to consistently diffuse labels, effectively enhancing semantic coherence and improving label coverage in boundary regions. For low-confidence labels generated during propagation, the system employs a confidence-aware correction strategy to fix label drift or semantic confusion issues. Through adversarial training, the pseudo-label generation model establishes an adversarial relationship between pseudo-labels and the true distribution, thereby improving the discriminative and generalization abilities of generated pseudo-labels. Through bi-branch self-distillation optimization, the stable prediction results of the teacher model guide the student model to iteratively update, gradually compressing pseudo-label errors and enhancing the model's responsiveness to multi-scale semantics.

[0045] Optionally, S3 includes:

[0046] Pseudo-label confidence feature maps are constructed from the pseudo-label augmented data to obtain pseudo-label confidence feature map data;

[0047] Semantic attention graphs are constructed based on pseudo-label confidence feature map data to obtain multi-head attention graph data;

[0048] Attention-guided semantic aggregation is performed on multi-head attention graph data to obtain semantic aggregated data.

[0049] By performing boundary region weight enhancement fusion on semantic aggregation data, recalibrated data is obtained.

[0050] This invention effectively improves the discrimination ability and stability of pseudo-labels in boundary and ambiguous regions by introducing pseudo-label confidence perception and multi-head semantic attention mechanisms. The pseudo-label confidence feature map construction module can clearly identify low-reliability regions and potential erroneous label points based on the confidence level of label distribution and adjacent semantic differences, forming a structured expression of label credibility. Through multi-head attention map construction, the model is guided to generate fine-grained semantic perception weights from multiple semantic attention angles (such as local consistency, category co-occurrence relationships, and boundary abrupt response), improving the model's perceptual flexibility in ambiguous structures. Based on this, attention-guided semantic aggregation operations can achieve semantic concentration and feature purification in ambiguous regions, strengthening the local semantic convergence process. Boundary region weight enhancement fusion amplifies the high gradient responses of edge neighborhoods, enabling label reconstruction to more accurately cover the transition region near the boundary line.

[0051] Optionally, S4 includes:

[0052] Data augmentation is performed on the recalibrated data to generate augmented training set data;

[0053] Multi-branch encoding is performed on the enhanced training set data to obtain multi-branch feature data, which includes basic geometric encoding, modality fusion encoding, structural graph embedding encoding, and hierarchical aggregation encoding.

[0054] Boundary-aware decoding is performed on multi-branch feature data to obtain boundary-aware data;

[0055] Supervised training data is obtained by performing multi-scale supervised training based on boundary awareness data.

[0056] The point cloud segmentation model is obtained by adjusting the regional weights of the supervised training data.

[0057] This invention introduces rotational perturbation, density sampling, and label preservation strategies during the data augmentation generation stage, which not only expands the sample space but also enhances the model's generalization ability to different structural perturbations. The multi-branch encoding module integrates basic geometric encoding, modal fusion encoding, structural graph embedding encoding, and hierarchical aggregation encoding to extract semantic features from dimensions such as spatial structure, heterogeneous modalities, graph topological relationships, and multi-scale context, achieving comprehensive modeling of different types of targets in frontline operational scenarios. The boundary-aware decoding module focuses on structurally abrupt regions through difference enhancement and boundary attention mechanisms, enabling the model to achieve higher segmentation accuracy at point cloud boundaries and semantic breakpoints. A multi-scale supervised training mechanism provides multi-point supervision at the backbone layer, boundary layer, and local refinement layer, effectively improving the consistency and robustness of semantic prediction at different scales. A region weight adjustment strategy dynamically optimizes the learning intensity of easily confused and highly uncertain regions, improving the model's adaptability and segmentation performance in real-world scenarios.

[0058] Optionally, the boundary-aware decoding includes:

[0059] Enhance the feature difference between points in multi-branch feature data to obtain feature-enhanced data;

[0060] Structural difference extraction is performed on the feature-enhanced data to obtain structural difference data;

[0061] Local contextual edge processing is performed on the structural difference data to obtain boundary attention data;

[0062] Boundary labels are reconstructed from the boundary attention data to obtain boundary-aware data.

[0063] This invention enhances the feature difference between points by measuring the degree of feature variation of adjacent points, effectively highlighting areas of dramatic change in space and providing preliminary response basis for boundary localization. The structural difference extraction module further mines geometric discontinuities and structural abrupt changes between point pairs from the local topology, strengthening the perception of edge morphological changes. Based on this, local context edge processing combines spatial neighborhood and structurally sensitive regions, dynamically adjusting the response weights of boundary points and non-boundary points through a boundary attention mechanism to form a boundary attention map with discriminative capabilities. In the boundary label reconstruction stage, fine-grained annotation of boundary semantics is completed based on high-confidence regions in the attention map, enabling the model to accurately delineate object boundaries, surface intersections, and category switching areas.

[0064] Optionally, this application also provides a deep learning-based point cloud data segmentation system for performing the deep learning-based point cloud data segmentation method described above, wherein the deep learning-based point cloud data segmentation system includes:

[0065] The structure-aware feature construction module is used to acquire point cloud data, extract point cloud features from the point cloud data to obtain point cloud feature data, and embed the point cloud feature data into structure-aware data to obtain structure-aware map data.

[0066] The pseudo-label augmentation learning module is used to perform pseudo-label augmentation training on the structure-aware map data to obtain pseudo-label augmented data.

[0067] The semantic attention correction module is used to perform attention recorrection on pseudo-label augmented data to obtain recorrected data;

[0068] The boundary awareness training module is used to enhance the training of recalibrated data to obtain a point cloud segmentation model.

[0069] The purpose of this invention is to construct a structure-aware graph in step S1 by effectively fusing geometric information, modal features, and inter-point topological relationships through point cloud feature extraction and structure-aware embedding, providing high-quality input for semantic transmission and graph neural network computation. Step S2 uses the structure-aware graph data to generate initial pseudo-labels and introduces structural similarity propagation, confidence correction, and adversarial training mechanisms to gradually optimize the accuracy and region consistency of the pseudo-labels, thus providing stable training signals even in weakly supervised or semi-supervised environments. Step S3 introduces a semantic attention mechanism, focusing on the ambiguous regions of label boundaries through pseudo-label confidence maps and multi-head semantic attention maps, achieving boundary-sensitive semantic aggregation and re-correction, and enhancing the model's discriminative ability in structures. Step S4 combines structure-preserving data augmentation and multi-branch feature encoding, supplemented by boundary-aware decoders and region weight adjustment strategies, effectively improving the model's performance in structural continuity, boundary accuracy, and semantic discriminability, resulting in a point cloud segmentation model with high robustness and fine semantic segmentation capabilities. Attached Figure Description

[0070] 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:

[0071] Figure 1 A flowchart illustrating the steps of a point cloud data segmentation method based on deep learning in one embodiment is shown.

[0072] Figure 2 A flowchart illustrating the steps of a point cloud feature extraction method according to an embodiment is shown.

[0073] Figure 3 A flowchart illustrating the steps of a pseudo-label reinforcement learning method according to one embodiment is shown.

[0074] Figure 4 A flowchart illustrating the steps of a semantic attention correction method according to an embodiment is shown.

[0075] Figure 5 A flowchart illustrating the steps of a boundary-aware training method according to an embodiment is shown.

[0076] 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

[0077] 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.

[0078] 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.

[0079] 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.

[0080] Please see Figures 1 to 5 This application provides a point cloud data segmentation method based on deep learning, the method comprising:

[0081] S1. Acquire point cloud data and extract point cloud features from the point cloud data to obtain point cloud feature data; perform structure-aware embedding on the point cloud feature data to obtain structure-aware map data.

[0082] In one embodiment, raw point cloud data is acquired. This point cloud data can be collected using sensing devices such as 3D LiDAR and depth cameras. The sampling result is represented as P = {(x...} i ,y i ,z i ,a i )}, where (x i ,y i ,z i Let be the three-dimensional spatial coordinates of the i-th point, and let a be the coordinates of the i-th point. i The additional attribute data corresponding to this point may specifically include modal information such as reflection intensity, RGB texture value, and timestamp. Joint extraction of geometric and auxiliary modal features is performed on the point cloud data. During geometric feature extraction, the geometric attributes of each point are calculated based on local neighborhood construction (e.g., using a fixed radius r neighborhood or k-nearest neighbor search), including point density, normal vector direction, local curvature, and Euclidean distance distribution with neighboring points. During auxiliary modal extraction, histogram statistics, gradient change information, and modal dissimilarity features are extracted from reflection intensity, color texture, and differences between modalities. The extracted geometric and modal features are then fused. Through channel-level attention processing, such as using a Squeeze-and-Excitation (SE) module or a multi-head feature fusion attention module, weights are assigned to each modal channel to generate a fused feature vector f reflecting the degree of semantic response.i Based on the fused point cloud feature data, a graph structure is constructed to model the relationships between points. This is achieved by calculating f between any two points. i with f j The feature similarity between them is calculated by constructing the adjacency matrix A using a Gaussian kernel function. ij Defined as A ij =exp(―||f i ―f j || 2 / σ 2 ), where A ij Let f be the connection strength of point pair (i,j) in the weighted adjacency matrix, exp be the natural exponential function, and f be the connection strength of point pair (i,j). i f is the fused feature vector of the i-th point. j Let A be the fused feature vector of the j-th point, σ be the smoothing control parameter in the feature space, and A be the fused feature vector of the j-th point. ij This represents the connection strength between point i and point j. Based on this adjacency matrix, a structural graph G = (V, E, A) is constructed, where V is the set of points, E is the set of edges, and A is the weighted adjacency matrix. A graph neural network is then used to propagate features from this structural graph. Models such as Graph Convolutional Networks (GCN), EdgeConvolutional Networks (EdgeConv), or Graph Attention Networks (GAT) can be used to fuse the feature vector f on the graph structure. i Embedding propagation is performed to obtain the structure-aware feature representation F. struct .

[0083] S2. Perform pseudo-label augmentation training on the structure-aware map data to obtain pseudo-label augmented data;

[0084] In one embodiment, the structural feature representation F based on the structure-aware graph is used. struct We use a pre-trained shallow neural network or a clustering algorithm (such as KMeans) to classify the structural features, thus obtaining an initial set of pseudo-labels, denoted as Y. (0) Perform structural similarity propagation. In the constructed graph structure, the initial pseudo-labels are iteratively diffused using the label propagation algorithm. During propagation, the label distribution of each node i is updated by a weighted average of the label distributions of its neighboring nodes, using the following formula: in This represents the label distribution of node i in the (t+1)th iteration. Let A represent the set of adjacent nodes of node i. ij Let be the edge weight between node i and node j (reflecting feature similarity). Let represent the label distribution of node j in the t-th iteration. The label results obtained from propagation are processed to correct for low confidence. By setting a discriminant metric (such as label distribution entropy or maximum prediction probability), label nodes with confidence levels below a threshold τ are selected. For these low-confidence pseudo-label nodes, a boundary label replacement strategy (i.e., replacing them with adjacent high-confidence category labels) or a replacement based on the consistency principle of graph center points can be used to improve the overall accuracy and continuity of the labels. Adversarial training is used to enhance the discrimination quality of pseudo-labels. A generator-discriminator framework (such as using the PatchGAN architecture) is constructed, where the generator generates a pseudo-label graph distribution, and the discriminator distinguishes between the real label distribution and the pseudo-label distribution. The two are trained adversarially to optimize the label generation network, making its output more structurally reasonable and boundary-separable. Self-distillation via a two-branch approach is used to further optimize the pseudo-label results. The system constructs a teacher-student network structure, where the teacher model parameters are synchronously updated from the student network using an exponential moving average (EMA). During training, the student network learns on perturbation data (such as noise, rotation, etc.) and outputs stable pseudo-label augmented data by minimizing the KL divergence between teacher and student outputs.

[0085] S3. Perform attention recalibration on the pseudo-labeled augmented data to obtain recalibrated data;

[0086] In one embodiment, a pseudo-label confidence feature map is constructed. For each graph node, the system calculates its label confidence score based on the pseudo-label distribution, for example, using the maximum output probability of the softmax function as the basic confidence score. Simultaneously, local feature volatility is introduced as an uncertainty indicator to characterize the intensity of semantic feature changes between points, forming a composite confidence feature representation for each point, denoted as C. i =(maxY) i ,Δf i ), where maxY i Δf represents the pseudo-label confidence of node i. i This represents the rate of change of local features at the node. The system constructs three attention heads with different semantic focus directions: ① a geometric boundary head, used to identify regions of abrupt feature gradient changes; ② a category context head, used to enhance the semantic expression of semantically consistent regions; and ③ a feature mutation head, used to highlight regions with high rates of change. Each attention head generates local attention weights based on the feature correlation and structural adjacency relationships between nodes, using weighted dot products and softmax operations. Differentiated responses are made to perceptual features across different semantic dimensions. Attention-guided semantic feature aggregation is performed. The system utilizes the aforementioned multi-head attention graph to perform attention on the set of neighboring nodes of node i. The feature vectors in the dataset are weighted and summed, and the results of all attention heads are then fused to obtain the fused semantic aggregation vector: in The attention weighted aggregation feature of node i (after fusing all attention heads), where H is the total number of attention heads, h is the current attention head number, j is the index of the adjacent node, and f j Let represent the feature vector of adjacent node j. Semantic response enhancement processing is performed in the boundary region. The system identifies the semantic boundary region and introduces an enhancement factor α. bnd ∈[0,1], the semantic aggregation results are reweighted and fused to amplify the influence of boundary regions on features, resulting in the corrected feature representation: Where f i For the original feature vector, α bnd This is a semantic boundary enhancement factor that controls the degree of influence of the boundary region on the features. Features after aggregation To correct the features. Through the above processing, the recognition sensitivity of semantic boundary regions can be enhanced, the ability of pseudo-labels to distinguish regions of local structural abrupt changes can be improved, and the output recorrected data will serve as high-quality input features for the next stage of training.

[0087] S4. Enhance the training of the recalibrated data to obtain the point cloud segmentation model.

[0088] In one embodiment, synchronous enhancement operations are performed on the input point cloud data. These enhancement operations include two categories: spatial transformation and modal perturbation. Spatial transformations include random rotation, scaling, and occlusion simulation; modal perturbation includes applying perturbations to auxiliary modal channels such as RGB color, reflection intensity, or timestamps to simulate sensor noise or information loss. These enhancement operations are performed synchronously while maintaining the pseudo-labels unchanged. The system performs multi-branch feature encoding. The system establishes four parallel encoding branches to extract information features of different dimensions, including a geometric structure branch, which uses PointNet++ or its improved structure to extract local geometric features and enhance spatial topology awareness between points; a modal information branch, which performs convolution processing on auxiliary channels (such as RGB and intensity) to extract modality-related contextual representations; a graph structure embedding branch, which uses a graph convolutional network (GCN) or graph attention network (GAT) to propagate features from the constructed point cloud structure graph and learn inter-point dependencies; and a hierarchical aggregation branch, which uses dilated convolution or multi-scale residual modules (such as ASPP) to fuse deep and shallow multi-level features to achieve global and local multi-scale modeling. The system performs boundary-aware decoding, including calculating local feature differences between adjacent points and extracting structural difference tensors to represent edge-sensitive regions. The system constructs a boundary attention map to enhance boundary recognition capabilities. Based on the boundary map, the prediction confidence of the model output is adjusted to improve the label response performance in edge regions. Multi-scale supervised training is then conducted. Supervision signals are introduced into the main output channel and the outputs of each intermediate layer (e.g., 1x, 0.5x, 0.25x resolution layers), and various loss function combinations are used for training, including basic cross-entropy loss, Lovász-Softmax loss to optimize ranking performance, and a boundary-aware loss function designed for edge recognition tasks. Each loss function is weighted to ensure a balance between overall accuracy and boundary segmentation performance. The system implements a region weight adjustment mechanism. Based on the confidence statistics and error feedback information of each category or spatial region during training, the system dynamically calculates the region weighting factor ω. i This is used to adjust the weights of the loss function at each point, thereby improving the overall segmentation performance and model robustness.

[0089] Optionally, the point cloud feature extraction includes:

[0090] S11. Acquire point cloud data and extract geometric features from the point cloud data to obtain geometric feature data;

[0091] In one embodiment, raw point cloud data is acquired from a 3D LiDAR or depth camera. Each point represents a sampling location in space; a fixed-radius neighborhood is constructed based on each point. or k-nearest neighbor Calculate the following geometric features, including the normal vector n. i The neighborhood covariance matrix was obtained by fitting it using PCA; curvature κ was obtained. i =λ min / (λ1+λ2+λ3), where κ i For point p i The curvature, λ1 is the point p i The largest eigenvalue of the neighborhood covariance matrix represents the degree of variation along the principal direction, where λ² is the value at point p. i The second largest eigenvalue of the neighborhood covariance matrix represents the degree of variation in a moderate direction, where λ3 is the value at point p. i The smallest eigenvalue of the neighborhood covariance matrix represents the degree of change in the normal direction (used to estimate curvature); point density. Where ρ i for, For, p i Let V be the volume of the sphere's neighborhood; mean distance between points, local centroid offset, etc.

[0092] S12. Perform auxiliary modal feature extraction on the geometric feature data to obtain auxiliary modal feature data;

[0093] In one embodiment, auxiliary modal channel values ​​for each point are extracted, including but not limited to reflection intensity, which can be read directly; color information (RGB), which can be obtained by projection mapping if the point cloud comes from a camera-radar joint calibration system; and time parameters, which represent the sampling time of the point and are suitable for dynamic scenes; each modal data is normalized (e.g., Min-Max) and missing value interpolation is performed; and the data are concatenated to form an auxiliary modal feature vector.

[0094] S13. Perform cross-modal attention fusion on the auxiliary modal feature data to obtain cross-modal fused data;

[0095] In one embodiment, a modal channel embedding network is constructed, for example, using a multi-head channel attention module; for the input and Channel encoding is performed separately to obtain key K. i Value V i With query Q i Vectors; attention weights are calculated as follows Where α ij The attention weights represent the degree of attention that point i pays to point j in the modal channel. Softmax is a normalization function used to transform the similarity score into a probability distribution of the attention weights. Q i Let K be the query vector for point i, obtained from geometric features or fused input encoding. jLet be the key vector (Key) for point j, obtained by encoding auxiliary modal features. d is the dimension parameter used for scaling in the attention mechanism, and is the dimension of the Query / Key vector. The resulting features... The weighted fusion result is characterized by the combined focus on geometric and modal information.

[0096] S14. Construct an inter-point relationship graph based on the cross-modal fusion data to obtain the inter-point relationship graph data;

[0097] In one embodiment, for each point p in the point cloud i The system calculates its relationship with other points p. j The three-dimensional Euclidean distance d between them ij To measure proximity in geometric space, the cross-modal fused feature difference between points is calculated, i.e., the Euclidean distance between fused feature vectors. The system constructs the edge weight matrix of the graph using weighted adjacency rules. The edge weight coefficients between each pair of points are... Among them W ij For point p i With point p j The edge weights between the two points reflect the cross-modal feature similarity between them, and exp is the natural exponential function used to define the exponential decay form of the Gaussian kernel. For point p i Cross-modal fusion feature vectors, For point p j The cross-modal fusion feature vector, where σ is the scaling parameter of the Gaussian kernel function, controls the decay rate of the feature distance. As an indicator function, when point p i It is point p j The value is 1 if it is one of the k-nearest neighbors, and 0 otherwise, where j is the index of the neighbor. For point p i The k-nearest neighbor set is selected based on the fused feature distance. A Gaussian kernel function based on feature distance is used to exponentially decay the feature similarity, and a k-nearest neighbor constraint function is applied. The number of connections for each point is controlled to not exceed k. The constructed graph of relationships between points is denoted as G = (V, E, W), where V represents the set of points, containing all input points; E represents the set of edges, including only the pairs of points that satisfy the k-nearest neighbor constraint; and W represents the edge weight matrix, where the weight of each edge reflects its similarity in the feature space.

[0098] S15. Perform graph neural network processing on the point relationship graph data to obtain point cloud feature data.

[0099] In one embodiment, a graph neural network model is selected, including but not limited to EdgeConv, graph attention networks, or dynamic graph convolutional neural networks. In each graph convolutional layer, the system updates the features of each node through the following operations. in Let be the output feature vector of node i in the (l+1)th layer, which integrates its own and neighborhood features. φ is the node aggregation function, and is a multilayer perceptron (MLP) with residual structure. Let be the input feature vector of node i in the l-th layer. This is a max pooling operation for adjacent edge features, where ψ is the edge feature mapping function, representing the feature relationship between node i and its adjacent node j (e.g., mapping after difference through MLP). Let v be the input feature vector of node j in the l-th layer; for the current node v i eigenvectors The system traverses its adjacent nodes. Calculate the edge feature mapping function This function can be represented as the difference between two feature vectors input into a multilayer perceptron (MLP) for mapping, reflecting the semantic differences between nodes. Subsequently, max pooling is performed on the edge feature responses of all neighboring nodes to retain the most significant neighbor response information, and this is fused with the current node's own features. The fused result is then input into the node aggregation function φ(·), and an MLP with a residual connection structure is used to obtain the updated node feature vectors. By stacking multiple graph neural processing modules, this network progressively extracts features from local neighborhoods to higher-order topologies, achieving layer-by-layer propagation and global integration of node features. The output point cloud feature set is denoted as... Each of them This represents the depth feature representation of the i-th point after fusing geometric, modal, and structural information.

[0100] Optionally, the structure-aware embedding includes:

[0101] The spatial relationship between points is processed on the point cloud feature data to obtain the spatial relationship data between points;

[0102] In one embodiment, input point cloud feature data Each f i This includes location coordinates, geometric attributes, and fused modal features; a local adjacency set is constructed for each point using spherical neighborhoods (e.g., radius r) or k-nearest neighbors (e.g., k=16). For each pair of adjacent points Calculate the following spatial relationships, including the relative position vector Δp. ij =p j ―p i p j Let p be the position coordinates of the j-th point. i Let be the coordinates of the i-th point; d be the Euclidean distance between the points. ij =||Δp ij ||;Point to included angle θij =arccos(n i ·n j (If a normal vector is provided), arccos is the inverse cosine function used to find the angle between two unit vectors, n i Let n be the normal vector of the i-th point. j Let j be the normal vector of the j-th point; output the set of structural relation vectors.

[0103] Structural features are extracted from the spatial relationship data between points to obtain structural feature data between points;

[0104] In one embodiment, the spatial relationship (Δp) ij ,d ij ,θ ij The data is fed into the structure encoding function ψ(·) to construct the structure difference vector: s ij =ψ(Δp) ij ,d ij ,θ ij )=MLP([Δp ij ||d ij ||θ ij ]), where s ij Let ψ be the structural difference vector from point i to point j, which is the encoded result and reflects the geometric / spatial structural changes between point pairs. ψ is the structural encoding function, used to fuse multiple spatial geometric information to generate a structural difference representation. Δp ij Let d be the spatial position difference vector between point i and point j. ij Let θ be the Euclidean distance between point i and point j. ij For the angle information between point pairs (e.g., the angle between surface normals or the direction difference), the MLP concatenates various features and inputs them into a multilayer perceptron (multilayer feedforward neural network) to obtain structural features; through local comparison, such as s ij and Compare, The mean of the structural features of point i and all its neighbors k is used to obtain local structural perturbation information. The output structural feature tensor S = {s ij} represents the structural response characteristics between each pair of points.

[0105] A structure graph is constructed based on the structural feature data between points to obtain the structure graph data;

[0106] In one embodiment, a weighted graph structure G = (V, E, W) is constructed. ij ), where the vertex set V = {f i}, edge set Edge weight W ij =exp(―||s ij || 2 / σ2 ), where W ij Let be the weight of edge (i,j), representing the structural similarity between node pairs, exp be the natural exponential function, and s be the weight of edge (i,j). ij σ represents the structural feature difference (e.g., position difference, normal difference, or feature space distance) between nodes i and j, where σ is the scaling parameter of the Gaussian kernel function, controlling the sensitivity of edge weights to structural differences and representing structural similarity. If the edge weight is less than the threshold δ, pruning is performed to reduce redundant connections, resulting in a sparse structure graph representation that includes topological strength and spatial dissimilarity. Output the structure graph data.

[0107] Based on the structure graph data, graph neural network structure embedding propagation is performed on the point cloud feature data to obtain embedding propagation graph data;

[0108] In one embodiment, a graph neural network (such as GAT, EdgeConv, or GraphSAGE) is used to propagate the embeddings to the structure graph: in The output features of node i after the update at layer l+1, ψ is the feature combination function (e.g., concatenated and processed by MLP), α ij The structure-aware attention score depends on W. ij With feature difference, Let i be the input feature of node i in the l-th layer. For the input features of node j in layer l, s ij Provide structural auxiliary information (such as spatial distance, normal angle difference, edge connection weight, etc.) between nodes i and j. Let φ be the set of neighboring nodes of node i, and let φ be the node update function, such as nonlinear activation with residuals. A multi-layer stacking approach (L = 2-3 layers) is used to represent the structure from local to large-scale structures. The output is embedded propagation graph data, i.e., high-order node features containing structure-aware semantics.

[0109] The embedded propagation graph data is subjected to graph attention hierarchical aggregation to obtain structure-aware graph data.

[0110] In one embodiment, a multi-head graph attention module is constructed, with each head focusing on different semantic dimensions. Head 1 focuses on geometric boundaries (regions with high edge weight gradients), constructing boundary enhancement masks based on edge weight gradients and assigning higher attention weights to regions with large gradient changes in structural edges, such as obtaining attention weights through normalization based on edge weight gradients. Head 2 focuses on global class consistency (regions with high class relevance), calculating the structural path dependency strength between nodes of the same class and using it as global attention guidance. Head 3 focuses on feature volatility (regions with large rates of change), i.e., calculating volatility to adjust attention weights, such as performing weighted calculations based on volatility on top of existing attention weights. Each attention head uses a different parameter set (Q...). h ,K h Vh ), calculate attention output in Let be the attention-aggregated feature vector of node i under the h-th attention head, and j be the index of node i's neighbor node. Let Q be the set of neighboring nodes of node i, and let softmax be the normalization function for the attention scores of neighboring nodes. h Let f be the query weight matrix for the h-th attention head. i K represents the original input features of node i. h Let h be the key weight matrix of the h-th attention head. Let V be the feature transpose vector (key vector) of neighbor node j, d be the scaling factor for the embedding dimension, and V be the key vector. h Let f be the weight matrix of the h-th attention head. j The original input features of neighbor node j are used; the outputs of multiple attention heads are fused (concatenated or weighted averaged): or in Let h be the aggregated feature (structure-aware feature) of node i, and let Concat be the output concatenation operation of multiple attention heads. Let h be the index of the attention head, and H be the total number of attention heads. Let w be the feature vector of node i after attention aggregation under the h-th attention head. h The fusion weights (obtained by weighted average) for the h-th attention head are used to output the structure-aware map data.

[0111] Optionally, the graph attention hierarchical aggregation includes:

[0112] Dependency path data is obtained by constructing dependency paths from the embedded propagation graph data.

[0113] In one embodiment, the input is embedded propagation graph data. And the adjacency matrix of the structure graph A = [W ij On graph G = (V, E), use a shortest path algorithm (such as Dijkstra's algorithm) or a maximum weight path algorithm to calculate the weights of the dependent paths between points; construct the dependent path matrix. Its weight can be expressed as in Let W be the total weight (i.e., path strength) of the dependency path from node i to node j, u be the starting node in the path, v be the ending node in the path, Path(i,j) be the set of edges traversed by the path from node i to node j in the graph, and W be the total weight (i.e., path strength) of the dependency path from node i to node j. uv Let the structural edge weight of edge (u,v) in the path represent the structural dependency strength, and set a maximum hop limit T. max This avoids calculating long paths across the entire graph. Output dependency path data.

[0114] Based on the dependency path data, the embedded propagation graph data is processed by graph attention multi-head hierarchical processing to obtain multi-head hierarchical data. The graph attention multi-head hierarchical processing includes local geometric boundary head processing, global category dependency head processing, and node feature change rate head processing.

[0115] In one embodiment, three semantic attention heads are constructed and processed separately, including a Local Geometric Boundary Header (Head-GEO): which uses the gradient of the edge weights in the structure graph as the basis for attention weights, emphasizes abrupt changes in boundary position features, and calculates the edge weight gradients. in W represents the edge weight gradient, indicating the intensity of the change in connection weight from node i to j. ij W represents the edge weight from node i to node j in the structure graph. ji The edge weights from node j to node i in the structure graph are used as indicators of boundary mutation; gradient-adjusted attention scoring incorporates boundary weight factors. like in For local geometric attention weights, sim(f) i ,f j ) is the feature similarity function, f i Let f be the embedded feature vector of node i. j The embedded feature vector of node j; the output feature after weighted aggregation of boundary region features; Global Class Dependency Header (Head-SEM): based on dependency path Construct a global attention mechanism that focuses on semantically consistent regions and uses path relevance weights. Construct a global attention graph (such as a fully connected subgraph and path weights); perform Query-Key interactions on nodes along the path to adjust attention: in The global semantic attention weights are weighted based on path relevance and feature similarity. To determine the semantic path association between nodes i and j, sim(f) is used. i ,f j f represents the feature similarity between nodes i and j. i Let f be the embedded feature vector of node i. j The embedded feature vector of node j; Head-VAR: Calculated by measuring the feature differences over time or spatial scales (e.g., ...). Where Δf i Let i be the characteristic rate of change of node i. Let i be the feature vector of node i in the l-th layer of the graph neural network. (The feature vector of node i in the (l-1)th layer of the graph neural network) is used as the attention enhancement factor. As a Score adjustment item in the Query-Key, or to directly guide Dropout / weight adjustments, an attention enhancement graph is generated; a Transformer-style multi-head attention formula is used. in Let be the attention weights from node i to neighbor node j in the h-th attention head, and softmax be the weight normalization operation over all neighbor nodes. Let be the query vector (Query) of node i in the h-th head, obtained by multiplying the input features by the query weight matrix. Let be the transpose of the key vector (Key) of node j in the h-th head, where d is the dimension of the key vector (used as a scaling factor). Let j be the aggregated feature result output by the h-th attention head, and j be the graph neighbor nodes of node i. Let $\frac{h}{j}$ be the value vector of node $j$ in the $h$-th head. This vector is used to generate a new representation. Each head corresponds to a structural semantic dependency type, and output features are generated accordingly. The output is multi-head hierarchical data.

[0116] Perform skip-layer joins on multi-head hierarchical data to obtain cross-layer joined data;

[0117] In one embodiment, the features output by each head are treated as different structural layers. Through cross-layer connections, each node feature supports self-connection and skip connections (e.g., (GEO→VAR)); cross-layer connections employ... in For the skip-cross-layer fusion feature of node i, Let γ be the output feature of node i in the mutation response feature layer, and γ1 be the weighted fusion coefficient of the semantic layer features. Let γi be the output feature of node i in the semantic response feature layer, and γ2 be the weighted fusion coefficient of the geometric layer features. The output features of node i at the geometric structure feature layer, or the outputs of the three layers fused through a gating mechanism. GRU stands for Gated Recurrent Unit, and skip connections enhance the fusion capability of long-range dependencies and boundary anomaly responses.

[0118] Subgraph clustering and fusion are performed on the cross-layer connection data to obtain subgraph fused data;

[0119] In one embodiment, the system constructs a feature similarity matrix Sim(i,j), whose elements are defined as follows: Sim(i,j) is the semantic similarity score between node i and node j in the cross-layer feature space. This represents the cross-layer skip connection feature representation of node i (such as a multi-scale semantic feature vector obtained by concatenating or aggregating different layers). Let be the cross-layer skip connection feature representation of node j, and let cos(·,·) represent the cosine similarity function, used to measure the semantic similarity between point pairs in the embedding space. Based on the above similarity matrix, spectral clustering or structural consistency clustering algorithms (such as DBSCAN combined with feature embedding) are used to divide the entire graph into sub-regions, dividing all points into several sub-graphs that are highly similar in structure and semantics. The system performs local feature aggregation operations within each subgraph to generate representative fused features. The aggregation method includes extracting the maximum response value of features at each point within the subgraph; and for each node within the subgraph... Assign attention weights β ij Based on this, a weighted summation is performed to obtain the representative features after fusion. in The fused representative features are represented by j, which is the index of a node currently traversed within the subgraph. Let β be the set of all node indices contained in the k-th subgraph to which node i belongs. ij The attention weights from node j within the subgraph to the center node i are generated by standardization based on the similarity between nodes or the semantic response strength. This represents the cross-layer feature vector of node j within the subgraph. Output the fused feature set of the subgraph.

[0120] Implicit edge backflow processing is performed on the subgraph fusion data to obtain structure-aware graph data.

[0121] In one embodiment, for nodes at the boundary of a subgraph, the similarity Sim(i,j) between their neighboring nodes is calculated to be consistent with the label; if Sim(i,j) > τ and there are no direct edges connecting them, an implicit backflow edge is added; the weight of the backflow edge is estimated through feature similarity and path dependency compensation. in Let be the edge weight of the implicit backflow edge from node j to node i (used to supplement unconnected but similar node pairs), and λ be a weighting coefficient used to balance the influence of path dependency and feature similarity. The path dependency compensation metric from node i to node j can be represented as a penalty term or access probability score based on the length of the subgraph connection path. This is the node feature similarity term, representing the degree of proximity between two nodes in the feature space. Let be the fused feature vector of node i. Given the fused feature vector of node j, construct the enhanced symmetric graph adjacency matrix W′=W+W revW′ is the adjacency matrix of the constructed structure-aware graph, which is equal to the sum of the matrices of the original graph and the backflow edges. W is the adjacency matrix of the original graph structure (which may be sparse or contain only explicit connections). rev The structure-aware graph is obtained by completing the adjacency matrix composed of implicit edges (backflow edges).

[0122] Optionally, the implicit edge reflow processing includes:

[0123] Weak connectivity data is obtained by performing edge similarity analysis based on the subgraph fusion data;

[0124] In one embodiment, the input is subgraph fusion feature data. Let N be the subgraph fusion feature vector of the i-th node, where N is the total number of nodes, i is the node index, and subgraph partitioning information is also provided. in For the first subgraph, For the k-th subgraph; for each subgraph boundary node For subgraph boundary operators, it indicates how to select the set of boundary nodes. For the currently processed subgraph, find its neighborhood from other subgraphs. node v j Let b be the index of the adjacent subgraph (satisfying b≠a), and a be the index of the current subgraph, forming candidate pairs across subgraph boundaries; calculate feature similarity metrics, such as using cosine similarity. Sim(i,j) is the feature similarity index between nodes i and j, calculated using cosine similarity. Let i be the feature vector of the subgraph fusion for node i. The subgraph fusion feature vector for node j. Let L2 norm be the feature vector of node i. Let Sim(i,j) be the L2 norm of the feature vector of node j, if Sim(i,j)>τ weak (τ weak The weak connection similarity threshold is set to filter similar but unconnected node pairs. Pairs of nodes without explicit graph edges are marked as weak connection candidates. The output is the weak connection dataset. (i,j) represents candidate weakly connected node pairs, and E is the set of edges in the original graph structure, used to determine whether there is already a connection.

[0125] Candidate implicit edges are generated based on weak connection data to obtain candidate implicit edge data;

[0126] In one embodiment, for weakly connected node pairs Construct a new candidate edge The direction indicates the information flow or attention propagation path from node j to node i. The system assigns a confidence status label to each candidate edge based on the following factors: node confidence, which refers to the stability or deterministic score of the current node in clustering, classification, or other reasoning tasks, calculated as the confidence score of the current prediction results of nodes i and j (e.g., the maximum classification probability or the inverse ratio of the Softmax output entropy); intra-class aggregation, which refers to the degree of aggregation between the node and other nodes in its subgraph in the feature space, calculated by the average feature distance between node i and its neighbors of the same class; the smaller the distance, the higher the aggregation; and structural redundancy, which refers to whether the edge is redundant or duplicates information in the existing structural graph, checking whether the candidate edge exists in the reachable path of the original graph (e.g., using BFS to determine if it is a duplicate information path). If so, the confidence level is reduced. Candidate states can be divided into multiple levels such as "high confidence," "medium confidence," and "pending confirmation." For each boundary node i, only its top k most similar weakly connected candidate nodes are retained (i.e., Top-k similarity filtering is performed) as the priority objects for edge generation. The candidate implicit edge dataset is output.

[0127] Calculate the backflow edge weights on the candidate implicit edge data to obtain the backflow edge weight data;

[0128] In one embodiment, for each candidate edge Calculate the backflow edge weight The weights can be defined using the following weighting formula. in For the return edge The edge weight is the validity score of the edge, α is the semantic similarity weighting coefficient, Sim(i,j) is the semantic similarity measurement result between node i and node j, i is the node order term, j is the other node order term, and β is the structural path potential weighting coefficient. The structural propagation potential score represents the estimated ability of node j to propagate to node i via the structural path, γ is the confidence difference weighting coefficient, and Δ conf (i,j) represents the confidence difference between pseudo-labels. If If the edge threshold is weak, it will not be added to the graph structure.

[0129] Bidirectional edge symmetry correction is performed based on the backflow edge weight data to obtain symmetry correction data;

[0130] In one embodiment, each pair is checked Does the original graph already have a positively directed edge e? ij If it exists, then perform symmetric fusion: in Let e ​​be the new edge weight after bidirectional fusion, and let e be the edge weight after symmetric correction. ij The weight value, For edge e in the original graphij The original edge weights, For the return edge The edge weights, i.e., the edge scores obtained during backpropagation, are used to maintain bidirectional information flow; if they do not exist, a new symmetric edge is added. and assign values To create a new symmetrical edge from node v when it does not exist in the original graph. i Point to v j , Create new edges for symmetry The assigned edge weight is taken from its corresponding backflow edge. For the return edge The edge weights, i.e. the edge scores obtained during backpropagation, are used to set an asymmetric retention mechanism for high-discrepancy regions (such as boundary regions), which only backflows weak connections rather than being completely symmetric.

[0131] The structure-aware map is generated based on the symmetry correction data, resulting in structure-aware map data.

[0132] In one embodiment, the original graph structure is represented as a triple G = (V, E) orig W orig ), where V represents the point set; E orig W represents the original edge set; orig This represents the corresponding edge weight matrix. The system uses a modified symmetric edge set. Merge it with the original edge set to construct edge set E. * =E orig ∪ε sym And reconstruct the weighted adjacency matrix A accordingly. final .

[0133] Optionally, S2 includes:

[0134] S21. Generate lightweight pseudo-labels based on the structure-aware map data to obtain initial pseudo-label data;

[0135] In one embodiment, the system performs preliminary pseudo-label generation based on the structure-aware graph data to support subsequent weakly supervised learning or self-supervised optimization processes. The input data includes a structure-aware graph G = (V, E, F), where V represents the set of nodes and E represents the set of edges. This represents the structure-aware features of each node. The system utilizes a lightweight point classifier to classify each node v. iFor category prediction, the classifier can employ a one- or two-layer Multilayer Perceptron (MLP) structure or a simple Graph Convolutional Network (GCN). During classification, the features of each node are forward-propagated to obtain the predicted probability distribution of the corresponding category; then, the category with the highest probability is selected as the initial pseudo-label for the node. The system records the maximum value of this predicted probability as the classification confidence score (i.e., the reliability of the prediction result) for the node. The system sets a confidence threshold θ. init Only nodes with a confidence level higher than the threshold are included in the initial pseudo-label set.

[0136] S22. Perform structural similarity propagation on the initial pseudo-label data to obtain pseudo-label propagation data;

[0137] In one embodiment, the system assigns each currently unlabeled target node v j Perform the following operation to find node v in the graph. j The set of adjacent nodes Only nodes that already contain pseudo-labels are retained to form the propagation candidate set. For each node v in the propagation candidate set... k Calculate its relationship with node v j Structural feature similarity, Where S jk For node v j With v k Structural feature similarity score, F j For node v j The structural feature vector represents its embedding or semantic representation in the graph, F. k For node v k Structural feature vectors, ||F j ||For node v j The L2 norm of the eigenvectors, ||F k ||For node v k The L2 norm of an eigenvector measures the directional consistency between the eigenvectors of two nodes. If, in the propagation candidate set, there exists a eigenvector with the same directional consistency as v... j The node with the highest similarity v k Furthermore, their similarity value is higher than the preset threshold δ. sim Then the pseudo-label category v of this node k Assign target node v j As its dissemination label By combining similarity scores and the confidence of neighbor pseudo-labels, a credibility score for propagated labels is calculated, serving as a quantitative indicator of node label quality.

[0138] S23. Based on the pseudo-label propagation data, perform low-confidence pseudo-label correction to obtain pseudo-label correction data;

[0139] In one embodiment, the system sets a confidence correction threshold θ. refine For nodes with a confidence level below the threshold, corrections are made as follows: for low-confidence nodes v... i traverse its set of adjacent nodes Analyze the distribution of pseudo-labels among adjacent nodes, and use the category with the highest frequency as the corrected label value. This strategy enables label updates based on local consistency and is suitable for label correction in regions with ambiguous class boundaries. Alternatively, to prevent the corrected results from excessively participating in model training, the system reduces the confidence score of the corrected nodes. This is achieved by multiplying the original confidence score by a shrinkage factor α∈(0,1) to obtain the new corrected confidence score value. in The corrected pseudo-label confidence score represents the new weighted confidence score after shrinkage processing, where α is the confidence shrinkage factor, with a value such as 0.95. For node v i The original pseudo-label confidence values ​​are then calculated. This operation can dynamically weight the values ​​in the loss function, reducing the negative impact of potentially erroneous pseudo-labels on the model. After the above corrections are completed, the system outputs the pseudo-label corrected dataset.

[0140] S24. Perform adversarial training on the pseudo-labeled corrected data to obtain pseudo-labeled enhanced data;

[0141] In one embodiment, the system is enhanced through adversarial training by constructing a generator and a discriminator. The generator G is constructed as a shallow or deep fully connected neural network (such as an MLP), and the input is the feature vector F of each node. i The output is a probability distribution vector matching the number of categories (which can be activated via Softmax). The generator is constructed using real pseudo-labels. As a reference, the output of G is guided to be as close as possible to these reference labels; the loss function is the adversarial loss (as shown in the above equation), and can be combined with cross-entropy loss for co-optimization; the optimization objective is to make G(F) i The generator G receives node features F to "fool" the discriminator D with a higher probability, even if D determines it to be real. i As input, generate the predicted label distribution G(F) i The goal is to approximate the distribution of real pseudo-labels. The discriminator D can also use a fully connected structure, with input F. i Corresponding label (real label y) i Or generate label G(F) iThe concatenated vector of the input label (D) outputs a probability value between 0 and 1, representing whether the label is trustworthy. When inputting real label pairs, D outputs 1; when inputting generated label pairs, D outputs 0. Training is performed using binary cross-entropy loss or adversarial loss. The training method is "alternating optimization," meaning D is updated once, followed by an update to G. The discriminator D determines whether the input label comes from the set of real and fake labels (i.e.,...). The results are generated by either the generator or the algorithm and used for adversarial signal modeling. The training objective function is... in Let G represent the min-max optimization objective in adversarial training, where the generator G attempts to minimize the loss, and the discriminator D attempts to maximize the loss. To represent the distribution of real pseudo-labeled samples Below, D is the logarithmic expectation of the discriminator's prediction of the joint distribution of the true label y and feature F, where D(y|F) is the probability that the discriminator classifies the input feature F and label y as "true pseudo-labels". Let D(G(F)) represent the log-expected probability that the discriminator identifies the generator's output pseudo-label G(F) as "forged" under the true feature distribution. D(G(F)) is the probability that the discriminator classifies the generator's generated label G(F) as a "real pseudo-label". The discriminator D aims to maximize its ability to distinguish real labels; the generator G aims to minimize the probability that its generated result is identified as "forged". Together, they constitute an adversarial objective function, whereby the generator attempts to "deceive" the discriminator, while the discriminator continuously improves its ability to distinguish between real and pseudo labels. This training iteratively optimizes the G and D modules by alternately updating them, achieving alignment between the pseudo-label distribution and the real distribution in the feature space. The system selects adversarially stable nodes as pseudo-label reinforcement points.

[0142] S25. Perform dual-branch self-distillation optimization based on the pseudo-label enhancement data to obtain pseudo-label enhancement data.

[0143] In one embodiment, the system constructs a dual-branch network structure, including a shared feature extraction backbone network and two decoder branches (Branch1 and Branch2) to handle pseudo-label training and teacher output, respectively. Branch1 is used to receive pseudo-label reinforcement data and perform regular supervised training; Branch2 serves as the teacher model, updating parameters through an exponential moving average (EMA) mechanism to avoid direct impact from gradient fluctuations. Its parameter update rule is to smoothly update the parameters of the current student branch (Branch1) to the teacher branch at a fixed ratio. in Here, λ represents the updated parameters for the teacher branch (exponential moving average result), λ is the exponential moving average weighting factor, controlling the proportion of teacher and student parameters fused together, and θ... teacherFor the teacher branch's current parameters (before updating), θ student These are the current parameters for the student branch (i.e., Branch2). During training, the system introduces a knowledge distillation loss between the two branches, which minimizes the difference between the output probability distributions of the two branches for the same node; this difference is measured by calculating the KL divergence. in Here, represents the knowledge distillation loss function value, used to measure the difference between the two-branch output distributions; 'i' is the index, indicating the current node number being processed; and 'KL' is the Kullback-Leibler divergence function, used to measure the distance between two probability distributions. Let be the predicted probability distribution of branch 1 (student branch) for the i-th node. Let be the predicted probability distribution for the i-th node in branch 2 (the student branch). For identified category edge nodes (such as those based on confidence gradient or category switching rate detection), the system reweights them, assigning higher weights to them in the loss function; this strategy effectively enhances the learning ability for ambiguous or semantically transitional regions. The system outputs a pseudo-label augmented dataset.

[0144] Optionally, S3 includes:

[0145] S31. Construct a pseudo-label confidence feature map from the pseudo-label augmented data to obtain pseudo-label confidence feature map data;

[0146] In one embodiment, the system uses all nodes V = {v i Let} be the set of vertices of the graph; construct the edge set E of the graph based on spatial adjacency relations (such as adjacency within the radius or fixed k nearest neighbors), connecting locally neighboring pairs of nodes; for each pair of adjacent nodes (v i ,v j ), calculate its edge weight Φ ij As a joint measure of label consistency and confidence similarity, it is defined as follows: Where Φ ij For node v i With v j The edge weights between the two values ​​represent the combined strength of pseudo-label consistency and confidence similarity; exp is an exponential function used to map confidence differences to a decay coefficient; α is a sensitivity adjustment factor for confidence differences; Conf i For node v i The false label confidence score, Conf j For node v j The false label confidence score, This is an indicator function that takes the value 1 only if the labels match, and 0 otherwise. For node v i pseudo-label categories, For node v j The system identifies the pseudo-label category and outputs a pseudo-label confidence feature map.

[0147] S32. Construct a semantic attention graph based on the pseudo-label confidence feature map data to obtain multi-head attention graph data;

[0148] In one embodiment, in Based on this, multi-head attention processing is performed. Each attention head h... k Learning different attention weight functions like in Let be the attention weights from point j to point i at the k-th attention head, exp be the exponential function, and LeakyReLU be the leakage-corrected linear unit activation function. W is the attention score vector for the k-th attention head. k Let F be the linear transformation matrix of the k-th attention head. i Let F be the input feature vector of point i. j Let j be the input feature vector of point j, and j′ be the index of any neighboring point in the neighborhood of point i. Let F be the set of adjacent nodes of point i. j′ Given the input feature vector of any point j′ in the neighborhood, the attention head types include: confidence difference guide head, which emphasizes regions with drastic changes in confidence; structural boundary probe head, which learns regions with abrupt changes in local structure; and intra-class consistency head, which emphasizes the stable propagation of continuous regions of the class. Each head outputs a different region of interest map, which is then concatenated to obtain the attention map tensor.

[0149] S33. Perform attention-guided semantic aggregation on the multi-head attention graph data to obtain semantic aggregated data;

[0150] In one embodiment, an attention weight graph is used. Guided feature propagation and fusion in For the multi-head attention aggregation feature of the i-th point, To concatenate the output features of H attention heads, where j is the index number within the neighborhood, Let i be the neighborhood set of point i, containing its local neighbors. W represents the attention weights from point j to point i on the k-th attention head. k For, F j To prevent aggregation bias caused by uneven class sampling, class weights are adjusted. in The aggregated feature of the i-th point after category weighting adjustment. The category label for point i The corresponding category feature scaling factor, The multi-head attention aggregation feature is used for the i-th point; the original feature and the aggregation feature are fused. in Let be the semantic feature vector after fusion for the i-th point, BN be the batch normalization operation used to standardize the feature distribution, and γ be the fusion weight coefficient between the original features and the aggregated features. Let i be the original input feature vector of the i-th point. Let be the multi-head attention aggregation feature of the i-th point.

[0151] S34. Perform boundary region weight enhancement fusion on the semantic aggregation data to obtain recalibrated data.

[0152] In one embodiment, the set of boundary points is identified based on the characteristic mutation rate. Where Δ i Let be the local feature mutation rate at point i, representing the maximum feature difference between this point and its neighboring points. Let i be the neighborhood set of point i The operation of taking the maximum value from the middle. The final feature after semantic aggregation of the i-th point, Let |j| be the semantic aggregated features of neighboring point j, and |j||2 be the Euclidean distance (norm 2). The set of boundary points contains all local feature mutation rates exceeding the boundary threshold δ. edge The point, v i Let δ be the i-th node in the graph, corresponding to a point in the point cloud. edge This is the boundary abrupt change threshold, used to identify boundary points. For all nodes identified as boundary points, boundary enhancement weights are calculated. in λ is the feature enhancement weight for the i-th boundary point. edge σ represents the boundary augmentation amplitude control coefficient, σ is the activation function (such as Sigmoid), and Δ i Let δ be the local feature mutation rate at point i. edge For all nodes, a weighted feature adjustment is performed based on whether they are boundary points; for boundary points, the recorrected feature is the original feature multiplied by the enhancement weight, i.e. in The recorrected features of the i-th point Enhance the weights of the features at the i-th boundary point. For. System construct boundary mask. Used to mark whether each node belongs to a boundary point. A value of 1 indicates a boundary point, and a value of 0 indicates a non-boundary point.

[0153] Optionally, S4 includes:

[0154] S41. Perform data augmentation on the recalibrated data to generate augmented training set data;

[0155] In one embodiment, the input data is recalibrated point cloud feature data; data augmentation strategies include, but are not limited to, random rotation augmentation, randomly rotating around the principal axis θ∈[―π / 8,π / 8]; scale perturbation augmentation, multiplying each point coordinate by a perturbation factor s∈[0.9,1.1]; and boundary point inversion sampling, applying boundary masks. Local perturbation sampling is performed on the points to construct boundary discrimination training hard examples; modal occlusion enhancement is performed by randomly occluding RGB or reflection intensity modes to train the network robustness; for each group of enhanced samples, the original labels, perturbation parameters and perturbation mask are maintained to form training tuples.

[0156] S42. Perform multi-branch encoding on the enhanced training set data to obtain multi-branch feature data, wherein the multi-branch encoding includes basic geometric encoding, modality fusion encoding, structural graph embedding encoding, and hierarchical aggregation encoding.

[0157] In one embodiment, the basic geometric coding branch is the branch based on three-dimensional coordinate information (x... i ,y i ,z i The input consists of a multilayer perceptron (MLP) network combined with a local normalization mechanism to encode the geometric features of the point cloud, capturing local topological structure and shape change features, thereby constructing a low-level perception capability of the original geometric morphology of the point cloud. The modal fusion encoding branch processes multiple sensor modal data (e.g., RGB color information, laser reflection intensity, timestamps, etc.) attached to the point cloud. A cross-modal attention fusion module is used to dynamically weight and fuse information between different modalities, extracting complementary modal features to enhance the representation capability of the point cloud under different observation conditions. The structural graph embedding encoding branch takes the previously constructed structural perception graph data as input, including the connection relationships between points and graph edge weights. A graph neural network (e.g., graph attention network GAT or graph convolutional network GCN) is used to aggregate information from the node neighborhood and extract structural semantic features, realizing topological modeling of the point cloud in spatial relationships. The hierarchical aggregation encoding branch constructs an encoder stack structure with three or more layers (e.g., a Transformer encoding module) to model the contextual relationships of the point cloud from a global perspective. Each layer connects the current feature with the embedded output of the previous stage's structure graph to achieve semantic hierarchical progression and the aggregation capability of global feature awareness. All branch outputs are concatenated into a unified feature vector.

[0158] S43. Perform boundary-aware decoding on the multi-branch feature data to obtain boundary-aware data;

[0159] In one embodiment, the system calculates the degree of feature difference between each point and its neighboring points as a boundary response index. The local feature difference is defined as the maximum feature distance between the current point and all its neighbors, used to characterize whether the point is in a semantically abrupt region. Combining the adjacency relationships in the aforementioned structure-aware graph, the system models the latent boundary flow, i.e., by mapping the partial derivatives of the feature change trends of adjacent nodes in the graph, constructing a local structural difference map to refine the feature representation of boundary points and identify structural abrupt cues. A context encoder combining dilated convolution and depthwise separable convolution is used to perform context aggregation on the edge regions of the point cloud with a larger receptive field, extracting semantic patterns related to the boundary. A pre-trained dedicated boundary classification subnetwork is used as an auxiliary branch to perform binary classification prediction on whether each point belongs to a boundary region, generating a boundary mask B. i ∈{0,1}. During training, a loss function is used to balance the training difficulty caused by the sparse distribution of boundary points, and to strengthen the discriminative learning of boundary categories, i.e. in The boundary enhancement feature for the i-th point is the boundary-aware feature used in downstream tasks. Let B be the multi-branch fusion feature of the i-th point, which is the original fusion representation without boundary enhancement processing. λ is the boundary enhancement coefficient, which controls the weight contribution of the boundary enhancement feature in the overall feature. i This is the boundary label mask for the i-th point, with a value of 0 or 1, used to indicate whether the point is determined to be a boundary point. Let be the local feature difference response of the i-th point, representing the intensity of feature change between it and its neighboring points. The system performs boundary enhancement processing on the original multi-branch features. The boundary mask is used as weights to perform weighted fusion of the boundary response features, resulting in the boundary-enhanced feature representation. Formally, the boundary-enhanced semantic features are a proportional combination of the original features and the difference features; the proportionality factor is adjustable to control the range of boundary influence.

[0160] S44. Perform multi-scale supervised training based on boundary awareness data to obtain supervised training data;

[0161] In one embodiment, the system introduces multiple scale branches (e.g., three levels) at the end of the network, with each scale corresponding to a prediction subheading module used to output the pseudo-label prediction results at that scale. in Head represents the class prediction result (pseudo-label output) for the i-th point at the s-th scale. s For the classification prediction submodule used at the s-th scale (which is either a fully connected layer or a convolutional layer), The boundary enhancement feature at point i is used as the input for predictions at all scales, where s is the scale index, representing the s-th layer in the multi-scale structure. The prediction output at each scale is based on the boundary-enhanced feature input, while retaining its corresponding intermediate response features. The prediction result at each scale is aligned with the corresponding annotation to construct the loss function. Let w be the total supervised loss function, s be the scale index, representing the s-th layer in the multi-scale structure, and w be the total supervised loss function. s These are the loss weighting coefficients corresponding to scale s, used to balance the contributions of losses at different scales. The semantic segmentation loss at scale s is a weighted combination of cross-entropy loss and lovasz loss. A boundary-specific supervised loss, such as Dice Loss or Focal Loss, is used to enhance boundary recognition capabilities. The semantic segmentation loss employs a combination of cross-entropy loss and lovasz loss, balancing pixel-level accuracy with region boundary continuity. Lovasz loss optimizes the IOU (Intersection over Union) metric and is more sensitive to edge structures. Simultaneously, a dedicated loss function for boundary regions is used to improve the model's ability to recognize boundary points. This boundary loss can be Dice Loss (to measure region overlap) or Focal Loss (to address the imbalance problem of boundary classes), and its supervision target is the aforementioned boundary mask. The scale losses and boundary losses are combined in a weighted manner to form the overall training loss function. Different weight factors can be configured for each scale loss to reflect the contribution of different layers in learning. The boundary loss is added as an independent term to strengthen the constraint. During training, the predicted outputs at each scale and their corresponding intermediate hidden layer feature responses are retained.

[0162] S45. Adjust the region weights on the supervised training data to obtain the point cloud segmentation model.

[0163] In one embodiment, the regional imbalance weight matrix W is calculated based on the predicted heatmap and label distribution. region : Among them W c The region imbalance weight for category c is used to adjust the importance of different categories in the loss calculation, ∈ is the smoothing constant, c is the category index, and freq is the region imbalance weight for category c. c The model is designed to measure the frequency of occurrence of class c in the training set; a region weight adjustment factor is introduced for each class for logits recalibration; model parameters include encoder parameters (multi-branch); decoder parameters (boundary-aware); classifier head parameters (multi-scale); and a region weight table; the model output can be deployed as an inference API or integrated into a backend analytics engine.

[0164] Optionally, the boundary-aware decoding includes:

[0165] Enhance the feature difference between points in multi-branch feature data to obtain feature-enhanced data;

[0166] In one embodiment, for each point's fused feature vector F i and its neighborhood set Calculate the feature difference between each point and its neighboring points. Where δ i Let be the feature dissimilarity of point i, representing the maximum difference between this point and other points in its neighborhood in the feature space. To perform a maximum value operation on the feature differences of all points j in the neighborhood set of point i, F i Let F be the fused feature vector of point i, representing the original feature representation of that point after multi-branch encoding. j Let ||||2 be the fused feature vector of neighboring node j, ||||2 is the L2 norm (Euclidean distance), and the dissimilarity δ is used. i Constructing inter-point difference enhancement factors in Let F be the enhanced feature vector of point i, representing the feature representation after difference enhancement. i Let be the fused feature vector of point i, represent the fused feature vector of the original feature representation of that point after multi-branch encoding, represent the original feature representation of that point after multi-branch encoding, α is the scaling coefficient of the difference enhancement, is a trainable weight parameter, and δ i Let F be the feature dissimilarity of point i, representing the maximum difference between this point and other points in its neighborhood in the feature space, mean(F j Let be the average of the characteristics of all neighboring points in the neighborhood of point i, and j be the index of the neighboring point. Let i be the neighborhood set of point i, containing all feature points adjacent to point i (e.g., constructed through K-nearest neighbors or spatial adjacency).

[0167] Structural difference extraction is performed on the feature-enhanced data to obtain structural difference data;

[0168] In one embodiment, a structure guide diagram is constructed. Where the point set V corresponds to the feature points, and the edge set ε comes from spatial adjacency (such as KNN); for each edge e ij ∈ε, calculate the structural difference in Let be the structural difference between point i and point j, measuring the Euclidean difference in the enhanced features between the two points. Let be the enhanced feature vector of point i, representing its feature representation after feature difference enhancement. Let be the enhanced feature vector of point j, ||||2 is the L2 norm operator (Euclidean distance), and the difference is greater than the threshold θ. struct Edges are marked as potential boundary edges; weighted aggregation of all edges forms a structural difference response map. φ(·) is the activation function, such as ReLU, where Let be the structural difference response value of point i, representing the intensity of its structural abrupt change in its local neighborhood, and j be the index of the neighboring point of point i. Let K be the set of adjacent points of point i, determined by K nearest neighbors or spatial radius. The structural difference between point i and point j measures the Euclidean difference in the enhancement features between the two points.

[0169] Local contextual edge processing is performed on the structural difference data to obtain boundary attention data;

[0170] In one embodiment, the input is a structural difference response map, where the difference intensity of each point represents the degree of structural change in its neighborhood. To obtain the contextual relationships of boundary points in the local structure, context-aware convolutional operations are used for feature aggregation. The convolution method can be dilated point convolution or gated graph convolution to enhance the perception of cross-scale features in edge regions. In a specific implementation, for any point, all its neighboring points within a neighborhood of radius r in space are selected to form a context window, and the structural differences within the neighborhood are input to the convolution module for processing. in The boundary attention score for the i-th point represents the boundary saliency of that point in the context of structural abrupt change. Conv dilated This is an expanded convolution operation used to increase the receptive field and capture sparse structural features within the neighborhood. Let be the structural difference response value of the j-th point, representing the degree of structural difference between it and its neighboring points, where j is the index of any point within the neighborhood. Let be the set of spatial neighborhoods centered at point i with radius r. The output of this convolution operation serves as the initial boundary attention score for that point; a higher value indicates that the point is located in a more significant structural abrupt change region. To normalize the boundary attention score across the entire neighborhood, a Softmax function is introduced to normalize the boundary scores of all adjacent points, generating boundary attention weights. The output boundary attention data represents the boundary response intensity carried by each point and its relative weight distribution within the neighborhood.

[0171] Boundary labels are reconstructed from the boundary attention data to obtain boundary-aware data.

[0172] In one embodiment, a binary classification module for boundary point discrimination is constructed based on the boundary attention features obtained in the previous steps. This module can employ a shallow multilayer perceptron structure (such as a fully connected neural network containing one or two layers) to predict whether each point belongs to the boundary region. in For, σ is, W is, Let b be the boundary attention feature vector for each point, and let b be the probability value of that point being a boundary point. Specifically, the input is the boundary attention feature vector for each point, and the output is the probability value of that point being a boundary point, ranging from 0 to 1. This probability value is normalized using the Sigmoid activation function to ensure it has probabilistic meaning. A fixed boundary discrimination threshold is set (e.g., 0.5). If the boundary probability value of a point is not less than this threshold, it is determined to be a boundary point and assigned a value of 1; otherwise, it is determined to be a non-boundary point and assigned a value of 0. A boundary label is generated for each point. The original enhanced features of each point are concatenated with their corresponding boundary probability values ​​to form new boundary-aware features. This concatenation operation can be achieved through feature-level connections, that is, incorporating boundary information as an additional dimension into the original feature vector. The output boundary-aware data consists of two parts: the boundary label for each point and the concatenated boundary-aware features.

[0173] Optionally, this application also provides a deep learning-based point cloud data segmentation system for performing the deep learning-based point cloud data segmentation method described above, wherein the deep learning-based point cloud data segmentation system includes:

[0174] The structure-aware feature construction module is used to acquire point cloud data, extract point cloud features from the point cloud data to obtain point cloud feature data, and embed the point cloud feature data into structure-aware data to obtain structure-aware map data.

[0175] The pseudo-label augmentation learning module is used to perform pseudo-label augmentation training on the structure-aware map data to obtain pseudo-label augmented data.

[0176] The semantic attention correction module is used to perform attention recorrection on pseudo-label augmented data to obtain recorrected data;

[0177] The boundary awareness training module is used to enhance the training of recalibrated data to obtain a point cloud segmentation model.

[0178] 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.

[0179] 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 point cloud data segmentation method based on deep learning, characterized in that, The method includes: S1. Acquire point cloud data and extract point cloud features from the point cloud data to obtain point cloud feature data; perform structure-aware embedding on the point cloud feature data to obtain structure-aware map data. S2. Perform pseudo-label augmentation training on the structure-aware map data to obtain pseudo-label augmented data; S3. Construct pseudo-label confidence feature maps from the pseudo-label augmented data to obtain pseudo-label confidence feature map data; construct semantic attention maps from the pseudo-label confidence feature map data to obtain multi-head attention map data; perform attention-guided semantic aggregation on the multi-head attention map data to obtain semantic aggregation data; perform boundary region weight enhancement fusion on the semantic aggregation data to obtain recalibrated data. S4. Perform data augmentation on the recalibrated data to generate augmented training set data; perform multi-branch encoding on the augmented training set data to obtain multi-branch feature data, where multi-branch encoding includes basic geometric encoding, modality fusion encoding, structure graph embedding encoding, and hierarchical aggregation encoding; perform boundary-aware decoding on the multi-branch feature data to obtain boundary-aware data; perform multi-scale supervised training based on the boundary-aware data to obtain supervised training data; adjust the region weights on the supervised training data to obtain the point cloud segmentation model.

2. The method according to claim 1, characterized in that, The point cloud feature extraction includes: Acquire point cloud data and extract geometric features from the point cloud data to obtain geometric feature data; Auxiliary modal feature extraction is performed on the geometric feature data to obtain auxiliary modal feature data; Cross-modal attention fusion is performed on auxiliary modal feature data to obtain cross-modal fused data; A point-to-point relationship graph is constructed based on cross-modal fusion data to obtain point-to-point relationship graph data; Graph neural network processing is performed on the point relationship graph data to obtain point cloud feature data.

3. The method according to claim 1, characterized in that, The structure-aware embedding includes: The spatial relationship between points is processed on the point cloud feature data to obtain the spatial relationship data between points; Structural features are extracted from the spatial relationship data between points to obtain structural feature data between points; A structure graph is constructed based on the structural feature data between points to obtain the structure graph data; Based on the structure graph data, graph neural network structure embedding propagation is performed on the point cloud feature data to obtain embedding propagation graph data; The embedded propagation graph data is subjected to graph attention hierarchical aggregation to obtain structure-aware graph data.

4. The method according to claim 3, characterized in that, The graph attention hierarchical aggregation includes: Dependency path data is obtained by constructing dependency paths from the embedded propagation graph data. Based on the dependency path data, the embedded propagation graph data is processed by graph attention multi-head hierarchical processing to obtain multi-head hierarchical data. The graph attention multi-head hierarchical processing includes local geometric boundary head processing, global category dependency head processing, and node feature change rate head processing. Perform skip-layer joins on multi-head hierarchical data to obtain cross-layer joined data; Subgraph clustering and fusion are performed on the cross-layer connection data to obtain subgraph fused data; Implicit edge backflow processing is performed on the subgraph fusion data to obtain structure-aware graph data: for nodes at the subgraph boundaries, the similarity between their neighboring nodes is calculated. Consistency with the label; if If there are no direct edges connecting them, an implicit backflow edge is added; the weight of the backflow edge is estimated using feature similarity and path dependency compensation. ,in For the node To the node The edge weight of the implicit backflow edge. These are weighting coefficients. For the node To the node The path dependency compensation metric can be represented as a penalty term or access probability score based on the length of the subgraph connection path. This is the node feature similarity term, representing the degree of proximity between two nodes in the feature space. For nodes The fused feature vector, For nodes The fused feature vectors are used to construct an enhanced symmetric graph adjacency matrix. , The adjacency matrix of the constructed structure-aware graph is equal to the sum of the matrices of the original graph and the backflow edges. This is the adjacency matrix of the original graph structure. The structure-aware graph is obtained by constructing a complete adjacency matrix consisting of implicit edges.

5. The method according to claim 4, characterized in that, The implicit edge backflow processing includes: Weak connectivity data is obtained by performing edge similarity analysis based on the subgraph fusion data; Candidate implicit edges are generated based on weak connection data to obtain candidate implicit edge data; Calculate the backflow edge weights on the candidate implicit edge data to obtain the backflow edge weight data; Bidirectional edge symmetry correction is performed based on the backflow edge weight data to obtain symmetry correction data; The structure-aware map is generated based on the symmetry correction data, resulting in structure-aware map data.

6. The method according to claim 1, characterized in that, S2 include: Lightweight pseudo-labels are generated based on the structure-aware map data to obtain initial pseudo-label data. The initial pseudo-label data is propagated using structural similarity propagation to obtain pseudo-label propagation data. Based on the pseudo-label propagation data, low-confidence pseudo-labels are corrected to obtain pseudo-label correction data; Adversarial training is performed on the pseudo-labeled corrected data to obtain pseudo-labeled enhanced data. The pseudo-label enhancement data is obtained by performing a two-branch self-distillation optimization based on the pseudo-label enhancement data.

7. The method according to claim 1, characterized in that, The boundary-aware decoding includes: Enhance the feature difference between points in multi-branch feature data to obtain feature-enhanced data; Structural difference extraction is performed on the feature-enhanced data to obtain structural difference data; Local contextual edge processing is performed on the structural difference data to obtain boundary attention data; Boundary labels are reconstructed from the boundary attention data to obtain boundary-aware data.

8. A point cloud data segmentation system based on deep learning, characterized in that, For performing the deep learning-based point cloud data segmentation method as described in claim 1, the deep learning-based point cloud data segmentation system comprises: The structure-aware feature construction module is used to acquire point cloud data, extract point cloud features from the point cloud data to obtain point cloud feature data, and embed the point cloud feature data into structure-aware data to obtain structure-aware map data. The pseudo-label augmentation learning module is used to perform pseudo-label augmentation training on the structure-aware map data to obtain pseudo-label augmented data. The semantic attention correction module is used to perform attention recorrection on pseudo-label augmented data to obtain recorrected data; The boundary awareness training module is used to enhance the training of recalibrated data to obtain a point cloud segmentation model.

Images