A local-global cooperative catenary fine-grained point cloud semantic segmentation method, system, device and medium

Abstract

The present application relates to the technical field of catenary recognition, in particular to a local-global collaborative catenary fine-grained point cloud semantic segmentation method and system, device and medium, the present application jointly models the catenary point cloud through local geometric flow and global context flow, and introduces a global-local bidirectional interaction mechanism in the coding stage. The local geometric flow uses deformable kernel point convolution to adaptively extract the local geometric differences of components such as wrist arms, positioning tubes and insulators, enhancing the feature response of small components and complex connection areas; the global context flow models the long-distance structural relationship between catenary components through a linear attention mechanism, enabling the network to make judgments based on the spatial position and assembly relationship of the components in the wrist arm system, and using the global context to gate modulate the local geometric features, thereby preserving fine-grained local features while suppressing background noise and similar interference structures, improving overall segmentation stability in complex scenarios.

Description

Technical Field

[0001] This invention relates to the field of overhead contact line identification technology, and more specifically, to a local-global collaborative method, system, device, and medium for fine-grained point cloud semantic segmentation of overhead contact lines. Background Technology

[0002] The railway overhead contact system is a critical infrastructure in the traction power supply system of electrified railways, and its structural condition directly affects the stability of train current collection, power supply reliability, and operational safety. Components such as the overhead contact system cantilever arms, positioning tubes, insulators, droppers, catenary wires, positioning rings, and connectors typically exhibit complex spatial distribution, significant dimensional differences, and close connections. Accurate identification and fine-grained segmentation of these components are crucial prerequisites for conducting overhead contact system geometric parameter detection, defect location, condition assessment, and operation and maintenance decisions.

[0003] Currently, catenary inspection still largely relies on manual visual inspection, vehicle-mounted inspection image analysis, or rule-based geometric processing methods. Manual inspection is labor-intensive, inefficient, and easily affected by personnel experience, site environment, and observation angle. Two-dimensional image detection methods can identify some appearance anomalies, but lack direct three-dimensional spatial structural information, making it difficult to accurately express the spatial relationships between catenary components. Traditional rule-based geometric methods typically rely on manually set thresholds and prior structures, resulting in limited adaptability to different lines, different cantilever types, different acquisition densities, and complex occlusion scenarios. With the development of three-dimensional laser scanning technology, ground-based three-dimensional laser scanning equipment can quickly acquire high-precision catenary point cloud data. However, the raw point cloud usually only contains discrete spatial coordinates and reflection intensity information, lacking high-level semantic information that can be directly used for detection and analysis. Relying on professionals to perform interactive segmentation using point cloud software will face problems such as low processing efficiency, poor result consistency, and difficulty in large-scale application.

[0004] Therefore, researching a fine-grained semantic segmentation method for 3D point clouds of railway catenary with high automation and stable segmentation accuracy, in order to reduce reliance on manual interactive processing, has significant engineering application value. Summary of the Invention

[0005] The purpose of this invention is to provide a local-global collaborative method, system, device, and medium for semantic segmentation of fine-grained point clouds of overhead contact lines, in order to solve the above-mentioned problems in the prior art.

[0006] This invention is achieved through the following technical solution: In a first aspect, the present invention provides a local-global collaborative method for semantic segmentation of fine-grained point clouds in contact networks, comprising: Acquire point cloud data of the overhead contact system, and preprocess the point cloud data to obtain candidate point clouds of the overhead contact system. A deformable kernel-driven dual-stream architecture network model is constructed, which includes a multi-level global-local bidirectional correction coding stage, a dual-path geometric constraint bottleneck layer, and a dual-channel decoding and recovery stage. The multi-level global-local bidirectional correction coding stage includes generating local geometric features based on the candidate point cloud of the contact network through deformable kernel points, aggregating the local geometric features, injecting them into the global context modeling model to obtain updated global context features, generating modulation weights for the updated global context features, and filtering and enhancing the local geometric features. The dual-path geometric constraint bottleneck layer includes a deep semantic aggregation that is jointly constrained by geometric descriptions generated based on deformable kernel points and the relative spatial positions between points. The dual-channel decoding and recovery stage includes selectively recovering the segmented contact wire components to point-by-point resolution based on geometric constraint information through channel attention and point-level spatial attention, thereby obtaining the three-dimensional features of the segmented contact wire components.

[0007] Preferably, the generation of local geometric states based on the candidate point cloud of the overhead contact system through deformable kernel points includes local geometric flow and global context flow: The local geometric flow includes extracting local geometric features at the contact wire component level and in tubular regions using deformable convolution; the global context flow includes efficient modeling of structural dependencies between distant components in a carpal tunnel system using linear attention.

[0008] Preferably, the local feature aggregation of local geometric features includes:

[0009] In the formula, The features after weighted aggregation For rigid core points, For the first The learnable weight matrix corresponding to each kernel point This is a distance influence function used to determine the contribution weight of a neighboring point to the current kernel point based on the distance from that neighboring point to the deformable kernel point. Where is the convolution radius, For the domain point, With the center point, It is a deformable core point. The input features are for the neighborhood points.

[0010] Preferably, the updated global context features include: Generate query, key, and value features Q, K, and V based on input features, and compute global context features through linear attention; The deformable kernel point set is vectorized and input into a multilayer perceptron to obtain geometric embedding features. The geometric embedding features are stacked and injected into the global context features using a gated residual method to obtain the global context features updated by the global context stream.

[0011] Preferably, the filtering and enhancement of local geometric features includes:

[0012] In the formula, For local update features, To be the feature resulting from stacking the local geometric features of all center points, For activation function, It is a multilayer perceptron. This is the updated global context feature.

[0013] Preferably, the deep semantic aggregation constrained by the geometric description generated based on deformable kernel points and the relative spatial positions between points includes: Extract the local update features of the bottleneck stage and obtain the set of deformable kernel points corresponding to this stage; After vectorizing the deformable kernel point set, the geometric description matrix is ​​obtained through MLP. This matrix is ​​then concatenated with the bottleneck local features, and the bottleneck input features are obtained through linear mapping. The geometric bias is obtained through a multilayer perceptron and is then incorporated into the attention weight calculation.

[0014] Preferably, the selective recovery of geometric constraint information to point-by-point resolution through channel attention and point-level spatial attention includes: Decoding features are obtained by upsampling nearest neighbors, normalized and linearly projected to obtain features for the encoding stage, and then the features obtained by the encoder and the features after decoding upsampling are spliced ​​together by skipping, and the features are fused through the projection layer. The intermediate fused features are obtained by weighting the fused features using channel attention through average pooling and max pooling to generate channel weights. Then, point-level spatial attention is used to generate point-level weights based on the average pooling and max pooling of each point in the channel dimension. The intermediate fused features are then fused based on the point-level weights to obtain the final fused features.

[0015] Secondly, the present invention also provides a local-global collaborative catenary fine-grained point cloud semantic segmentation system for performing the aforementioned local-global collaborative catenary fine-grained point cloud semantic segmentation method, comprising: The data acquisition module is configured to acquire point cloud data of the overhead contact system and preprocess the point cloud data to obtain candidate point clouds of the overhead contact system. The model building module is configured to construct a deformable kernel-driven dual-stream architecture network model. This model includes a multi-level global-local bidirectional correction coding stage, a dual-path geometric constraint bottleneck layer, and a dual-channel decoding and recovery stage. The multi-level global-local bidirectional correction coding stage involves generating local geometric features based on the candidate point cloud of the catenary using deformable kernels. After local feature aggregation, these features are injected into a global context modeling model to obtain updated global context features. Modulation weights are then generated for the updated global context features, and local geometric features are filtered and enhanced. The dual-path geometric constraint bottleneck layer includes deep semantic aggregation constrained by the geometric description generated based on deformable kernels and the relative spatial positions between points. The dual-channel decoding and recovery stage involves selectively recovering the segmented catenary components to point-by-point resolution based on geometric constraint information through channel attention and point-level spatial attention, resulting in the segmented 3D features of the catenary components.

[0016] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described local-global collaborative contact network fine-grained point cloud semantic segmentation method.

[0017] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described local-global collaborative contact network fine-grained point cloud semantic segmentation method.

[0018] The technical solution of the present invention has at least the following advantages and beneficial effects: This invention jointly models the contact network point cloud using local geometric flow and global context flow, and introduces a global-local bidirectional interaction mechanism during the encoding stage. Local geometric flow utilizes deformable kernel convolution to adaptively extract local geometric differences in components such as cantilever arms, positioning tubes, and insulators, enhancing the feature response of small components and complex connection areas. Global context flow models long-distance structural relationships between contact network components through a linear attention mechanism, enabling the network to make judgments based on the spatial position and assembly relationship of components within the cantilever system. The staged interaction mechanism further injects deformable kernel descriptors into global features and uses global context to gate and modulate local geometric features, thereby suppressing background noise and similar interfering structures while preserving fine-grained local features, improving the overall segmentation stability in complex scenes.

[0019] The bottleneck geometry enhancement module fuses local deformation descriptors with high-level semantic features and introduces geometric biases formed by the relative spatial positions of point pairs in the self-attention computation. This process enables the network to explicitly consider three-dimensional spatial relationships during high-level semantic aggregation, which helps to distinguish spatially adjacent and small-scale components such as positioning loops, connections, and caps.

[0020] The adaptive feature fusion mechanism at the decoding end recalibrates upsampled features and skip connection features through channel attention and point-level spatial attention. This mechanism can enhance the response of boundary points, small component points, and connection region points, and reduce the interference of large-sized structures on small auxiliary components, thereby improving the problems of missing segmentation of small components, blurred boundaries, and adhesion of adjacent components.

[0021] Therefore, compared with the point cloud segmentation methods for the carpal tunnel based on concavity / convexity, supervoxel clustering, and line fitting, this invention reduces the reliance on manual thresholds and fixed geometric rules; compared with two-dimensional image detection methods, this invention directly outputs point-by-point semantic labels for three-dimensional point clouds, which can preserve the true spatial location, geometric boundaries, and morphological information of components, providing a more stable three-dimensional semantic data foundation for contact network component extraction, parameter verification, and digital operation and maintenance. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a diagram of the main network architecture of the present invention; Figure 2 This is a structural diagram of the dual-stream encoder of the present invention; Figure 3 This is a diagram of the geometric bottleneck layer structure of the present invention; Figure 4 This is a diagram of the adaptive feature fusion structure of the decoder in this invention; Figure 5 This is a semantic segmentation result diagram of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0025] The independently described modules or sub-modules may or may not be physically separated; they may be implemented in software or hardware, and some modules or sub-modules may be implemented in software, with the processor calling the software to implement the function of these modules or sub-modules, while other modules or sub-modules may be implemented in hardware, such as through hardware circuits. Furthermore, some or all of the modules can be selected to achieve the purpose of this application's solution according to actual needs.

[0026] In this embodiment, the contact network point cloud contains numerous slender components, oblique components, tubular components, small connectors, and complex intersection regions. These components often exhibit high similarity in local geometry, but their classification depends on their global structural location and the connection relationships with adjacent components. Existing point cloud segmentation networks typically treat local neighborhood feature extraction, global context modeling, and decoding recovery as relatively independent processes, which easily leads to the following problems: when relying solely on local convolution, the model struggles to distinguish components with similar local morphologies but different structural classifications; when relying solely on global attention, the local geometric differences of small connectors, boundary points, and complex intersection regions are easily weakened during deep semantic aggregation.

[0027] To address the aforementioned issues, this invention proposes a feature learning network based on deformable kernel point geometric priors. This network does not simply parallelize local and global flows, nor does it directly superimpose attention mechanisms into the encoder-decoder structure. Instead, it directly applies deformable kernel points to the encoding and bottleneck stages to generate locally deformable geometric features, which are further transformed into deep semantic aggregation constraints at the bottleneck layer. In the decoding stage, the deformable kernel points are not recalculated; instead, the geometric-semantic features constrained by the deformable kernel points are received from the encoding stage via skip connections and selectively restored. Thus, local geometric deformation information can indirectly participate in point-by-point prediction through feature representation.

[0028] This invention addresses the problems of existing technologies, such as reliance on manual geometric rules, lack of three-dimensional spatial information in two-dimensional image detection, easy omission of small contact wire components, and difficulty in distinguishing the boundaries of adjacent components. Compared to existing methods based on concavity / convexity, supervoxel clustering, RANSAC fitting, or two-dimensional target detection, the difference in this invention lies not in the individual use of deformable kernel points, linear attention, or decoding attention, but in the clear information dependency among these three elements: deformable kernel points generate local deformation states, local deformation states constrain the global context, the global context inversely modulates local geometric features, the bottleneck layer further utilizes local deformation descriptions and relative spatial positions to constrain deep semantic aggregation, and the decoding stage selectively recovers the aforementioned encoded features. This information chain enables the modules to work synergistically.

[0029] For details, please refer to Figures 1-5 This invention provides a local-global collaborative method for semantic segmentation of fine-grained point clouds in contact networks, comprising: S1: Acquire point cloud data of the overhead contact system, and preprocess the point cloud data to obtain candidate point clouds of the overhead contact system; The invention uses a ground laser scanner to acquire point cloud data of a three-dimensional contact network system. The scanner is deployed within a range of 5-10m from the contact network support.

[0030] Coordinate Alignment: This invention utilizes the direction of the railway track below the scanner for coordinate correction. Since the scanner scans between two railway tracks, the region x∈(-2,2), y∈(-2,2) is first extracted, and the z-values ​​of the points are analyzed. The point z_max with the highest z-value distribution represents the ballast, and the railway track is located 0.1m-0.3m above the ballast. Therefore, z∈[z_max+0.1, z_max+0.3] is extracted and clustered, retaining the largest cluster, i.e., finding one railway track. Subsequently, Principal Component Analysis (PCA) is applied based on this largest cluster, taking the largest eigenvector as the railway extension direction, and calculating the angle between this direction and the x-axis to obtain a rotation matrix. This matrix is ​​applied to the original point cloud, making the direction of the railway track parallel to the x-axis.

[0031] 2. Region of Interest Extraction: Rotated point cloud data typically contains numerous environmental factors (such as surrounding vegetation, buildings, etc.). To reduce the interference of the surrounding environment on subsequent network training and inference, spatial clipping is first performed based on the height of the catenary arm system, the scanner position, and prior knowledge of the railway structure. Using the scanner position as the starting point of the 3D coordinates (0, 0, 0), the original point cloud is spatially clipped, selecting points x∈(-15, 15), y∈(-15, 15), and z∈(-3, 12). Points satisfying these conditions are retained to obtain the candidate point cloud for the catenary. The obtained candidate point cloud includes the catenary arm, contact wire, supports, droppers, railway tracks, and other railway facilities.

[0032] S2: Construct a deformable kernel-driven dual-stream architecture network model, which includes a multi-level global-local bidirectional correction coding stage, a dual-path geometric constraint bottleneck layer, and a dual-channel decoding and recovery stage. The multi-level global-local bidirectional correction coding stage includes generating local geometric features based on the candidate point cloud of the catenary through deformable core points, aggregating local geometric features, injecting them into the global context modeling model to obtain updated global context features, generating modulation weights for the updated global context features, and filtering and enhancing local geometric features; the dual-path geometric constraint bottleneck layer includes deep semantic aggregation constrained by the geometric description generated based on deformable core points and the relative spatial positions between points; the dual-channel decoding and recovery stage includes selectively recovering to point-by-point resolution based on geometric constraint information through channel attention and point-level spatial attention to obtain the segmented three-dimensional features of the catenary components.

[0033] In this embodiment, the network adopts an encoder-decoder architecture, including a multi-level global-local bidirectional correction coding stage, a dual-path geometric constraint bottleneck layer, and a dual-channel decoding and recovery stage.

[0034] The core idea is that deformable kernel points are not only used for local neighborhood feature aggregation, but also for generating local geometric states that characterize the current neighborhood morphological changes. These local geometric states first participate in local feature aggregation, and are then injected into the global context modeling process to constrain semantic associations between distant components. The updated global context features further generate modulation weights to filter and enhance local geometric features. At the bottleneck layer, the geometric descriptions formed by deformable kernel points and the relative spatial positions between points jointly constrain deep semantic aggregation. In the decoding stage, the aforementioned geometric constraint information is selectively restored to point-by-point resolution through channel attention and point-level spatial attention.

[0035] The network forms a learning process of "local deformation perception - global structure correction - bottleneck geometric constraints - decoding boundary recovery", thereby improving the model's ability to identify complex components, small connectors, boundary points and complex intersection areas of the overhead contact system.

[0036] In one exemplary embodiment of the present invention, the generation of local geometric states based on candidate point clouds of overhead contact lines through deformable kernel points includes local geometric flow and global context flow: A two-stream encoder architecture is employed in the encoding phase, comprising a local geometry stream and a global context stream. The local geometry stream primarily utilizes deformable convolution to extract local geometric features at the contact wire component level and in tubular regions; the global context stream uses linear attention to efficiently model structural dependencies between distant components in the carousel system. To minimize the loss of long-range semantic information while capturing local geometric features, the local geometry stream interacts with the global context stream in each two-stream encoding phase.

[0037] For the center point The network in its neighborhood The offset of the rigid kernel points is learned internally, enabling the original fixed kernel points to adaptively deform based on the directional distribution, scale changes, and connectivity of the point cloud in the current neighborhood. Let the number of rigid kernel points be... The corresponding rigid core point is ,for The learned offset is Then the deformable kernel point is represented as , can be obtained from A set of core points consisting of deformable core points Unlike conventional local convolution which only uses kernel points as sampling or aggregation locations, the deformable kernel point set in this invention... This is further used as a description of the local deformation state of the current neighborhood. This set is used both for local feature aggregation and as a geometric prior for subsequent global context correction, bottleneck geometric constraints, and decoded feature recovery. Features are obtained by weighted aggregation of neighborhood points:

[0038] In the formula, The features after weighted aggregation, For rigid core points, For the first The learnable weight matrix corresponding to each kernel point This is a distance influence function used to determine the contribution weight of a neighboring point to the current kernel point based on the distance from that neighboring point to the deformable kernel point. Where is the convolution radius. For the domain point, With the center point, It is a deformable core point. The input features are for the neighborhood points.

[0039] After stacking the local geometric features of all center points, we get This process enhances the model's ability to geometrically represent diagonal rods, tubular structures, small connectors, and complex connection regions.

[0040] In this way, the model obtains not ordinary neighborhood features, but geometric feature representations carrying local deformation states, which provides a basis for subsequent differentiation of catenary components with similar local morphology but different global structural affiliations.

[0041] In one exemplary embodiment of the present invention, to avoid the weakening of local geometric information during deep semantic modeling, a global-local bidirectional correction module is set up at each encoding stage. This module does not simply concatenate local and global features, but rather establishes a bidirectional correction relationship between local geometric features and global contextual features, centered around the local deformation state represented by the deformable kernel point.

[0042] Generate query, key, and value features Q, K, and V based on input features, and compute global context features through linear attention;

[0043] in, Represents the positive kernel eigenmap. Represents a vector consisting entirely of 1s. Let T be a small constant used to ensure numerical stability, and T be the transpose. The global context flow consists of a linear attention layer and a feedforward network, ultimately yielding the global feature representation. ; Deformable kernel set After vectorization, the data is input into a multilayer perceptron to obtain geometric embedding features. These geometric embedding features are then stacked. By injecting global context features using a gated residual method, the updated global context features can be obtained.

[0044]

[0045] In the formula, For the updated global context features, This is a learnable weight matrix.

[0046] Then, the updated global context features are used to generate modulation weights to adaptively enhance local geometric features:

[0047] In the formula, For local update features, To be the feature resulting from stacking the local geometric features of all center points, For activation function, It is a multilayer perceptron. This is the updated global context feature.

[0048] The modulation weights range from (0, 2). These weights enhance useful local responses based on global structural relationships and suppress background noise or similar interfering structures. Through this bidirectional interaction, the model can fuse local geometric information with global structural information at each encoding stage, thereby improving its ability to identify complex components and small connectors in the overhead contact system.

[0049] In one exemplary embodiment of the present invention, the deep semantic aggregation constrained by the geometric description generated based on deformable kernel points and the relative spatial positions between points includes: After multi-level downsampling, the deepest features of the encoder have strong semantic expressive power, but the point cloud resolution is low. Contact wire components with similar local morphology but different structural affiliations are easily misassociated during deep semantic aggregation, leading to a decrease in the spatial topological discrimination ability of bottleneck features. To address this, this invention sets a dual-path geometric constraint bottleneck layer at the encoder bottleneck position, ensuring that the deep semantic aggregation process continues to be constrained by both local deformation states and spatial relationships between points.

[0050] First, extract the local update features of the bottleneck stage and obtain the set of deformable kernel points corresponding to this stage; after vectorizing the set of deformable kernel points, obtain the geometric description matrix through MLP, concatenate it with the local features of the bottleneck, and then obtain the bottleneck input features through linear mapping.

[0051] In the formula, To obtain the bottleneck input features through linear mapping, For linear models, This is the geometric description matrix. This operation combines deep semantic features with local deformation descriptions, enabling the bottleneck layer to preserve the structural affiliation differences between components when performing high-level semantic aggregation, and reducing the probability of incorrectly aggregating components with similar local morphology but different spatial connections.

[0052] Subsequently, self-attention computation with geometric bias is performed at the bottleneck layer. For any two points... and And calculate its relative spatial position. Geometric bias is obtained through a multilayer perceptron. This geometric bias is added to the attention weight calculation:

[0053] in, This is the result after calculating the attention weights. , , These represent the query, key, and value vectors, respectively. This represents the dimension of the query and the key. (This is achieved by introducing...) The model can explicitly consider the three-dimensional spatial relationships between point pairs during high-level semantic aggregation. Finally, the module outputs enhanced bottleneck features through residual connections, normalization layers, and feedforward networks. This module is mainly used to improve the structural discriminativeness of bottleneck stage features and enhance the model's ability to distinguish complex connected regions and geometrically similar components.

[0054] In one exemplary embodiment of the present invention, the selective recovery to point-by-point resolution based on geometric constraint information through channel attention and point-level spatial attention includes: The encoding and bottleneck stages have already formed geometric-semantic features constrained by deformable kernel points. However, during the stepwise upsampling process, fine-grained boundary information and small-scale component responses may still be weakened, leading to unclear boundary localization and insufficient small component responses in the pointwise prediction stage. To address this, this invention sets up an adaptive recovery module guided by geometric constraint features in the decoding stage, which selectively enhances the local geometric features transmitted from the encoding end while restoring the point cloud resolution.

[0055] In the In each decoding stage, the decoding features are first obtained through nearest neighbor upsampling. The skip connection features corresponding to the encoding stage are obtained by concatenating the updated local geometric features and global context features, and then normalizing and linearly projecting them to obtain the features of the encoding stage. The features obtained from the encoder and the upsampled features from the decoder are then spliced ​​together in a skip pattern, and the features are fused through a projection layer. ,in, For learnable projection layers, This is a feature of fusion.

[0056] To further improve the response of fused features to small components and boundary points, this invention introduces channel attention and point-level spatial attention sequentially. Channel attention generates channel weights through average pooling and max pooling. And weight the features: Point-level spatial attention further generates point-level weights based on the average pooling and max pooling of each point along the channel dimension. The final fusion features are obtained: .in, Used to select feature channels that are more important for class discrimination. Used to highlight boundary points, small connector points, and complex intersection areas.

[0057] Through the above-mentioned adaptive recovery at the decoding end, the local geometric constraints formed by deformable core points in the encoding stage can be carried over to the point-by-point prediction stage, thereby reducing the coverage of semantic features of large-sized rods on small connectors and boundary points, and improving the integrity and boundary clarity of fine-grained component segmentation.

[0058] In one exemplary embodiment of the present invention, to alleviate the problem of uneven class distribution in the contact wire point cloud, the present invention employs a cross-entropy loss function with class weights. Let the frequency of the c-th class point in the training set be:

[0059] in, For the first Points For the frequency of the c-th class point in the training set, This represents the number of points in the k-th class in the training set.

[0060] Define smooth inverse frequency weights :

[0061] And normalize it:

[0062] in, To prevent small constants from being divided by zero, For the number of semantic categories, for The result after normalization.

[0063] For a training batch Each point represents the final segmentation loss. for:

[0064] in, For the first The true category label of each point, The network predicts the probability that the point belongs to the true class. Finally, the network updates the model parameters through backpropagation using normalized smooth inverse frequency-weighted cross-entropy as the segmentation loss.

[0065] In one exemplary embodiment of the present invention, the proposed network is trained with a batch size of 2 and 300 training epochs. To optimize the network parameters, the AdamW optimizer is used, with an initial learning rate set to... The number of kernel points in the network was set to H=15, and the deformable convolution radius was set to 5.0. To alleviate the problem of uneven distribution of samples of different semantic categories in the contact network point cloud, a weighted cross-entropy loss function with normalized smooth inverse frequency category weights was used during training, so that low-frequency small component categories receive more sufficient supervision during parameter updates. After a set number of rounds of forward propagation, loss calculation, backpropagation, and parameter updates, the network gradually converged, and finally the trained GA-GLNet model weight file was output and saved. This trained model can be used for automated end-to-end semantic segmentation and inference of railway contact network 3D point clouds.

[0066] Secondly, the present invention also provides a local-global collaborative catenary fine-grained point cloud semantic segmentation system for performing the aforementioned local-global collaborative catenary fine-grained point cloud semantic segmentation method, comprising: The data acquisition module is configured to acquire point cloud data of the overhead contact system and preprocess the point cloud data to obtain candidate point clouds of the overhead contact system. The model building module is configured to construct a deformable kernel-driven dual-stream architecture network model. This model includes a multi-level global-local bidirectional correction coding stage, a dual-path geometric constraint bottleneck layer, and a dual-channel decoding and recovery stage. The multi-level global-local bidirectional correction coding stage involves generating local geometric features based on the candidate point cloud of the catenary using deformable kernels. After local feature aggregation, these features are injected into a global context modeling model to obtain updated global context features. Modulation weights are then generated for the updated global context features, and local geometric features are filtered and enhanced. The dual-path geometric constraint bottleneck layer includes deep semantic aggregation constrained by the geometric description generated based on deformable kernels and the relative spatial positions between points. The dual-channel decoding and recovery stage involves selectively recovering the segmented catenary components to point-by-point resolution based on geometric constraint information through channel attention and point-level spatial attention, resulting in the segmented 3D features of the catenary components.

[0067] A global-local dual-stream semantic segmentation framework for 3D point clouds of railway catenary systems is proposed. The core technology of this invention is the construction of a dual-stream semantic segmentation network for railway catenary point clouds, where one branch extracts local geometric details and the other modeles the global structural context. This framework can simultaneously handle the fine-grained local morphology of catenary components and the long-range structural relationships in the cantilever system, solving the component confusion problem caused by traditional geometric rule-based methods that rely solely on local shapes or linear fitting.

[0068] A staged global-local feature interaction mechanism is introduced. This network adds an interaction mechanism between the two streams, injecting deformable kernel point descriptors obtained from the local geometric stream into the global context features, while simultaneously using the updated global context features to perform gated modulation on the local geometric features. This mechanism enables the layer-by-layer fusion of local geometric details and global structural relationships, improving the discrimination ability of small components, boundaries between adjacent components, and complex connection regions.

[0069] A local geometric descriptor for dual-stream interaction and bottleneck enhancement. This method uses the deformation kernel point set as a local geometric descriptor to characterize the geometric deformation state of local regions in the contact network point cloud. This descriptor can be used for local-to-global geometric injection during the encoding stage, as well as for high-level semantic feature enhancement in the bottleneck layer. Through this design, the network can explicitly propagate local geometric changes to the global structure modeling and bottleneck semantic reasoning processes, improving the recognition ability of complex connectors and small components.

[0070] Bottleneck Layer Feature Enhancement Based on Geometric Bias Self-Attention. This invention introduces a geometric bias self-attention module at the encoder bottleneck location, fusing local deformation descriptors with high-level semantic features and incorporating geometric biases generated from the relative 3D positions of point pairs into the attention calculation. This module can explicitly consider the 3D spatial relationships between components during the high-level semantic aggregation stage, thereby enhancing the separability between spatially adjacent components and geometrically similar components.

[0071] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0072] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. This computer software product, stored in a storage medium, includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0073] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A local-global collaborative method for semantic segmentation of fine-grained point clouds in contact networks, characterized in that, include: Acquire point cloud data of the overhead contact system, and preprocess the point cloud data to obtain candidate point clouds of the overhead contact system. A deformable kernel-driven dual-stream architecture network model is constructed, which includes a multi-level global-local bidirectional correction coding stage, a dual-path geometric constraint bottleneck layer, and a dual-channel decoding and recovery stage. The multi-level global-local bidirectional correction coding stage includes generating local geometric features based on the candidate point cloud of the contact network through deformable kernel points, aggregating the local geometric features, injecting them into the global context modeling model to obtain updated global context features, generating modulation weights for the updated global context features, and filtering and enhancing the local geometric features. The dual-path geometric constraint bottleneck layer includes a deep semantic aggregation that is jointly constrained by geometric descriptions generated based on deformable kernel points and the relative spatial positions between points. The dual-channel decoding and recovery stage includes selectively recovering the segmented contact wire components to point-by-point resolution based on geometric constraint information through channel attention and point-level spatial attention, thereby obtaining the three-dimensional features of the segmented contact wire components.

2. The local-global collaborative contact network fine-grained point cloud semantic segmentation method according to claim 1, characterized in that, The generation of local geometric state based on candidate point clouds of overhead contact lines through deformable kernel points includes local geometric flow and global context flow: The local geometric flow includes extracting local geometric features at the contact wire component level and in tubular regions using deformable convolution; the global context flow includes efficient modeling of structural dependencies between distant components in a carpal tunnel system using linear attention.

3. The local-global collaborative contact network fine-grained point cloud semantic segmentation method according to claim 1, characterized in that, The local feature aggregation of local geometric features includes: In the formula, The features after weighted aggregation For rigid core points, For the first The learnable weight matrix corresponding to each kernel point This is a distance influence function used to determine the contribution weight of a neighboring point to the current kernel point based on the distance from that neighboring point to the deformable kernel point. Where is the convolution radius, For the domain point, With the center point, It is a deformable core point. The input features are for the neighborhood points.

4. The local-global collaborative contact network fine-grained point cloud semantic segmentation method according to claim 1, characterized in that, The updated global context features include: Generate query, key, and value features Q, K, and V based on input features, and compute global context features through linear attention; The deformable kernel point set is vectorized and input into a multilayer perceptron to obtain geometric embedding features. The geometric embedding features are stacked and injected into the global context features using a gated residual method to obtain the global context features updated by the global context stream.

5. The local-global collaborative method for semantic segmentation of fine-grained point clouds in contact networks according to claim 1, characterized in that, The filtering and enhancement of local geometric features includes: In the formula, For local update features, To be the feature resulting from stacking the local geometric features of all center points, For activation function, It is a multilayer perceptron. This is the updated global context feature.

6. The local-global collaborative contact network fine-grained point cloud semantic segmentation method according to claim 1, characterized in that, The deep semantic aggregation constrained by the geometric description generated based on deformable kernel points and the relative spatial positions between points includes: Extract the local update features of the bottleneck stage and obtain the set of deformable kernel points corresponding to this stage; After vectorizing the deformable kernel point set, the geometric description matrix is ​​obtained through MLP. This matrix is ​​then concatenated with the bottleneck local features, and the bottleneck input features are obtained through linear mapping. The geometric bias is obtained through a multilayer perceptron and is then incorporated into the attention weight calculation.

7. The local-global collaborative method for semantic segmentation of fine-grained point clouds of contact networks according to claim 1, characterized in that, The selective recovery of geometrically constrained information to point-by-point resolution via channel attention and point-level spatial attention includes: Decoding features are obtained by upsampling nearest neighbors, normalized and linearly projected to obtain features for the encoding stage, and then the features obtained by the encoder and the features after decoding upsampling are spliced ​​together by skipping, and the features are fused through the projection layer. The intermediate fused features are obtained by weighting the fused features using channel attention through average pooling and max pooling to generate channel weights. Then, point-level spatial attention is used to generate point-level weights based on the average pooling and max pooling of each point in the channel dimension. The intermediate fused features are then fused based on the point-level weights to obtain the final fused features.

8. A local-global collaborative contact network fine-grained point cloud semantic segmentation system, characterized in that, A method for performing a local-global collaborative contact network fine-grained point cloud semantic segmentation as described in any one of claims 1-7 includes: The data acquisition module is configured to acquire point cloud data of the overhead contact system and preprocess the point cloud data to obtain candidate point clouds of the overhead contact system. The model building module is configured to construct a deformable kernel-driven dual-stream architecture network model. This model includes a multi-level global-local bidirectional correction coding stage, a dual-path geometric constraint bottleneck layer, and a dual-channel decoding and recovery stage. The multi-level global-local bidirectional correction coding stage involves generating local geometric features based on the candidate point cloud of the catenary using deformable kernels. After local feature aggregation, these features are injected into a global context modeling model to obtain updated global context features. Modulation weights are then generated for the updated global context features, and local geometric features are filtered and enhanced. The dual-path geometric constraint bottleneck layer includes deep semantic aggregation constrained by the geometric description generated based on deformable kernels and the relative spatial positions between points. The dual-channel decoding and recovery stage involves selectively recovering the segmented catenary components to point-by-point resolution based on geometric constraint information through channel attention and point-level spatial attention, resulting in the segmented 3D features of the catenary components.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the local-global collaborative contact network fine-grained point cloud semantic segmentation method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements a local-global collaborative contact network fine-grained point cloud semantic segmentation method as described in any one of claims 1-7.

Images