Method, system and device for segmenting three-dimensional point cloud of power line and storage medium

Abstract

The invention relates to the technical field of power line point cloud segmentation, and discloses a power line three-dimensional point cloud segmentation method, system and device and a storage medium, and the method comprises the steps: obtaining original point cloud data of a power line, inputting the original point cloud data into a multi-scale spatial feature extraction model, and obtaining spatial features; according to the spatial features, performing foreground probability prediction on the original point cloud data to obtain a foreground point set, performing multi-radius sphere neighborhood aggregation on the foreground point set to obtain local features, and performing super-point aggregation on the foreground point set to obtain global features; fusing the local features and the global features according to a feature fusion model to obtain a fused query vector; and performing classification probability prediction on the fusion query vector to obtain a point cloud segmentation result of the power line. According to the method, on the basis of keeping low calculation cost and low model complexity, the semantic recognition capability of the sparse structure target of the power line can be effectively improved, and a high-precision and high-efficiency semantic segmentation effect is achieved.

Description

Technical Field

[0001] This invention relates to the field of power line point cloud segmentation technology, and in particular to a method, system, device and storage medium for segmenting three-dimensional point clouds of power lines. Background Technology

[0002] Currently, in the inspection and equipment condition monitoring of power systems, drones equipped with lidar and image sensors are widely used to acquire 3D point cloud and image data of transmission lines. This multi-source sensing data provides important support for applications such as 3D modeling of transmission channels, automated defect detection, and environmental analysis.

[0003] Conventional deep learning-based 3D point cloud segmentation methods are mostly optimized for dense point cloud environments such as urban areas or indoor spaces. When applied to power line inspection tasks, they have certain limitations. On the one hand, power cables and some insulation structures are often represented by only a few sparse points in the point cloud, exhibiting a long-distance linear distribution and significant "low density, high continuity" characteristics. Existing methods are mostly based on local neighborhood perception mechanisms, which are difficult to capture the cross-scale structural features of such targets, easily leading to problems such as missed detections, fragmented segmentation, and target fragmentation. On the other hand, there are stable spatial topologies and connection patterns between poles and cables, such as "suspended harnesses" and "multiple poles sharing cables." These structural priors are of great value for recognizing coherent target semantics. However, existing methods generally lack mechanisms for explicitly modeling structural dependencies, failing to effectively capture long-distance dependencies between targets, resulting in fragmented, inconsistent, and structurally misjudged prediction results. In addition, power inspection systems usually run on UAV platforms with limited computing power, which imposes strict constraints on the computational efficiency and model size of segmentation methods. Existing high-precision models often rely on complex Transformer structures or high-dimensional feature stacking, which incur high computational and memory overhead and lack real-time processing capabilities, thus limiting their deployment and implementation in actual inspection tasks. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method, system, device, and storage medium for segmenting three-dimensional point clouds of power lines. This addresses the shortcomings of existing methods in modeling sparse linear targets and continuous structures in power scenarios, thereby improving the segmentation efficiency and recognition accuracy of three-dimensional point clouds of power lines.

[0005] In a first aspect, the present invention provides a method for segmenting a three-dimensional point cloud of a power line, the method comprising: The raw point cloud data of the power line is acquired, and the raw point cloud data is input into a preset multi-scale spatial feature extraction model to obtain the spatial features of the raw point cloud data. The multi-scale spatial feature extraction model is constructed based on the sparse three-dimensional U-Net model. Based on the spatial features, foreground probability prediction is performed on the original point cloud data to obtain a set of foreground points. Multi-radius sphere neighborhood aggregation is performed on the set of foreground points to obtain local features. Superpoint aggregation is performed on the set of foreground points to obtain global features. The local features and the global features are fused according to a preset feature fusion model to obtain a fused query vector. The feature fusion model is constructed based on a large language model. The fusion query vector is used to perform classification probability prediction to obtain the category probability vector of the original point cloud data. Based on the category probability vector, the point cloud segmentation result of the power line is obtained.

[0006] Further, the step of performing foreground probability prediction on the original point cloud data based on the spatial features to obtain a set of foreground points includes: The spatial features are input into the foreground probability prediction head to obtain the foreground probability of each point. The foreground probability prediction head consists of a linear layer and an activation function. The foreground probability is compared with a foreground probability threshold, and the points corresponding to the foreground probabilities that are greater than the foreground probability threshold are taken as foreground points, thus obtaining a set of foreground points.

[0007] Furthermore, the step of performing multi-radius spherical neighborhood aggregation on the set of foreground points to obtain local features includes: Based on multiple preset scale radii, a spherical neighborhood search is performed on each foreground point in the set of foreground points within each scale radius to obtain multiple first neighborhoods; Average pooling is used to aggregate the spatial features of all neighboring points in each first neighborhood to obtain the scale local features corresponding to each scale radius. The local features corresponding to each scale radius are spliced ​​and linearly fused to obtain the local features corresponding to each foreground point.

[0008] Furthermore, the step of performing superpoint aggregation on the foreground point set to obtain global features includes: Within a preset superpoint radius, a spherical neighborhood search is performed on each foreground point in the set of foreground points to obtain a second neighborhood, wherein the superpoint radius is greater than the scale radius; The second neighborhood is segmented using a region growing method to obtain multiple super-point regions; Global pooling is used to aggregate the spatial features of all neighboring points within each superpoint region to obtain a superpoint feature set. The superpoint feature set contains the superpoint features corresponding to each superpoint region and is used as the global feature set.

[0009] Furthermore, the feature fusion model includes a local feature fusion module, a global feature fusion module, and a query vector fusion module, wherein the local feature fusion module and the global feature fusion module are respectively connected to the query vector fusion module; Both the local feature fusion module and the global feature fusion module are constructed using the RWKV model, and the query vector fusion module is constructed based on a gating fusion mechanism.

[0010] Further, the step of aggregating the local features and the global features according to a preset feature fusion model to obtain a fused query vector includes: The local features are input into the local feature fusion module to obtain the local query vector; The global features are input into the global feature fusion module to obtain the global query vector; The local query vector and the global query vector are input into the query vector fusion module. The local query vector and the global query vector are weighted and fused according to the gating fusion mechanism to obtain the fused query vector.

[0011] Furthermore, the step of weightedly fusing the local query vector and the global query vector according to the gating fusion mechanism to obtain the fused query vector includes: Based on the multilayer perceptron, uncertainty prediction is performed on the local query vector and the global query vector respectively to obtain branch uncertainty values, which include local branch uncertainty values ​​and global branch uncertainty values; Based on the branch uncertainty value, the feature weights are calculated, and the local query vector and the global query vector are weighted and fused according to the feature weights to obtain the fused query vector.

[0012] Secondly, the present invention provides a segmentation system for three-dimensional point clouds of power lines, the system comprising: The spatial feature extraction module is used to acquire the original point cloud data of the power line, input the original point cloud data into a preset multi-scale spatial feature extraction model, and obtain the spatial features of the original point cloud data. The multi-scale spatial feature extraction model is constructed based on the sparse three-dimensional U-Net model. The feature aggregation module is used to perform foreground probability prediction on the original point cloud data based on the spatial features to obtain a set of foreground points, perform multi-radius spherical neighborhood aggregation on the set of foreground points to obtain local features, and perform superpoint aggregation on the set of foreground points to obtain global features. The parallel perceptual fusion module is used to fuse the local features and the global features according to a preset feature fusion model to obtain a fused query vector. The feature fusion model is constructed based on a large language model. The classification prediction module is used to perform classification probability prediction on the fused query vector to obtain the category probability vector of the original point cloud data, and obtain the point cloud segmentation result of the power line based on the category probability vector.

[0013] Thirdly, embodiments of the present invention also provide a computer 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 steps of the above-described method.

[0014] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0015] This invention provides a method, system, device, and storage medium for segmenting 3D point clouds of power lines. The invention efficiently extracts multi-scale local structural information of the point cloud through sparse convolution, reducing redundant computation and ensuring effective perception of low-density regions. Foreground point filtering and multi-scale sphere query effectively improve the separability of slender targets. Superpoint partitioning based on spatial proximity and feature similarity compresses the point cloud size and improves global semantic consistency. The parallel RWKV architecture effectively integrates local fine-grained features with global structural features, significantly improving prediction accuracy at semantic boundaries. This invention effectively enhances semantic recognition capabilities for sparsely structured targets while maintaining low computational cost and low model complexity. It is suitable for efficient deployment on resource-constrained platforms and has significant practicality and scalability. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the method for segmenting the three-dimensional point cloud of power lines in an embodiment of the present invention. Figure 2 This is a schematic diagram comparing the evaluation indicators of the experimental results in the comparative experiment in the embodiments of the present invention; Figure 3 This is a schematic diagram of the structure of the three-dimensional point cloud segmentation system for power lines in an embodiment of the present invention; Figure 4 This is an internal structural diagram of the computer device in an embodiment of the present invention; Figure label: 10. Spatial feature extraction module; 20. Feature aggregation module; 30. Parallel perception fusion module; 40. Classification prediction module. Detailed Implementation

[0017] 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, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 The first embodiment of the present invention proposes a method for segmenting three-dimensional point clouds of power lines, including steps S10 to S40: Step S10: Obtain the original point cloud data of the power line, input the original point cloud data into a preset multi-scale spatial feature extraction model to obtain the spatial features of the original point cloud data. The multi-scale spatial feature extraction model is constructed based on the sparse three-dimensional U-Net model. Step S20: Based on the spatial features, perform foreground probability prediction on the original point cloud data to obtain a set of foreground points, perform multi-radius sphere neighborhood aggregation on the set of foreground points to obtain local features, and perform superpoint aggregation on the set of foreground points to obtain global features. Step S30: The local features and the global features are fused according to a preset feature fusion model to obtain a fused query vector. The feature fusion model is constructed based on a large language model. Step S40: Perform classification probability prediction on the fused query vector to obtain the category probability vector of the original point cloud data, and obtain the point cloud segmentation result of the power line based on the category probability vector.

[0019] This invention addresses the challenges of 3D point cloud data in power line inspection scenarios, including data sparsity, strong linearity, and significant background noise interference. It provides a rapid 3D point cloud segmentation method. Assume the original point cloud data of the power line is... , indicating that the original point cloud data contains N points, each point has d-dimensional attributes, R represents the set of real numbers, and the point cloud segmentation finally obtains the semantic category label of each point, the semantic categories include power lines, towers, vegetation and ground, etc.

[0020] In this embodiment, spatial features are first extracted from the raw point cloud data of the power line. Since point cloud data in power line inspection typically exhibits extremely high sparsity, direct computation on a dense 3D convolutional mesh not only consumes enormous resources but also easily leads to feature dilution. Therefore, this embodiment employs sparse convolution to construct a sparse 3D U-Net model as a multi-scale spatial feature extraction model. This model not only effectively extracts multi-scale spatial features but also significantly reduces computational complexity.

[0021] Specifically, the sparse 3D U-Net model contains a symmetric encoder-decoder structure. The encoder progressively downsamples the spatial resolution to extract high-level semantic features, while the decoder progressively recovers the resolution through sparse deconvolution, and simultaneously fuses shallow detail features from the encoder through skip connections. Assuming the initial input to the multi-scale spatial feature extraction model is denoted as F0, where F0 is the encoded feature obtained by voxelizing the original point cloud data, and the l-th layer feature of the encoder is represented as F... l The encoding process can then be represented as: In the formula, SparseConv l (*) represents the sparse convolution operation of the l-th layer of the encoder, and L represents the total number of layers of the encoder.

[0022] During the decoding stage, the spatial structure is recovered through sparse deconvolution, and the encoded features at the corresponding scale are fused, as shown in the following formula: In the formula, SparseDeconv represents the l-th layer features of the decoder. l (*) denotes the sparse deconvolution operation of the l-th layer of the decoder.

[0023] The features obtained by structural reconstruction and encoding feature fusion based on the decoder are spatial features. In this embodiment, the spatial features extracted by the multi-scale spatial feature extraction model are denoted as follows: This indicates that each point has a C-dimensional attribute, and this spatial feature will be used as input data for subsequent processing. It should be noted that the sparse 3D U-Net model in this embodiment can adopt the conventional architecture of the 3D U-Net model. The specific steps for extracting spatial features can refer to the conventional processing steps of the 3D U-Net model, and will not be repeated here.

[0024] The multi-scale spatial feature extraction model described above extracts local structural features from the point cloud data of power lines. To further improve feature quality, this embodiment, based on spatial features, models the local detailed structure of the point cloud data through a multi-scale semantic enhancement mechanism, highlighting small foreground regions such as power lines and insulators. Furthermore, it models the global topological relationships of the point cloud data through a superpoint aggregation mechanism, thereby encoding global contextual features. The steps for modeling the local detailed structure of the point cloud data and the global topological relationships are explained in detail below.

[0025] In the local feature extraction stage, this embodiment addresses the issue that targets of interest (such as power lines and towers) in power line detection scenarios typically exhibit strong linear structural characteristics, showing significant differences in geometric attributes from the background ground and vegetation point clouds. Traditional methods based on dense sampling or single-scale neighborhood feature aggregation struggle to fully capture these fine-grained and continuously changing linear features. To address this, a local feature extraction method based on multi-scale linear region semantic enhancement is proposed. By combining foreground point selection and multi-radius sphere query, local geometric semantic features are accurately extracted, enhancing the model's ability to perceive slender structures such as power lines and towers.

[0026] In this embodiment, local feature extraction includes two stages: foreground point selection and multi-scale local feature aggregation. In the foreground point selection stage, the probability of each point belonging to the foreground is obtained by predicting the foreground probability of the spatial features of each point output by the multi-scale spatial feature extraction model. Specific steps include: The spatial features are input into the foreground probability prediction head to obtain the foreground probability of each point. The foreground probability prediction head consists of a linear layer and an activation function. The foreground probability is compared with a foreground probability threshold, and the points corresponding to the foreground probabilities that are greater than the foreground probability threshold are taken as foreground points, thus obtaining a set of foreground points.

[0027] In this embodiment, the spatial features of each point output by the multi-scale spatial feature extraction model are first input into a pre-established foreground probability prediction head to calculate the probability that each point belongs to the foreground, that is, the probability that it belongs to a target such as a power line or insulator. Preferably, the foreground probability prediction head consists of a linear layer plus a sigmoid activation function. The linear layer is used to perform a linear transformation on the spatial features of each point, and the activation function is used to predict the probability of foreground points on the linearly transformed spatial features. Therefore, the foreground probability of each point can be expressed as: in, This represents the foreground probability of the i-th point. and All represent linear transformation parameters. This represents the Sigmoid function.

[0028] For each point, the foreground probability m i Compare it with the set foreground probability threshold. Compare them, and then select the ones that satisfy the criteria. The point is used as the foreground point p i Thus forming a collection of foreground sites. : In the formula, The total number of foreground points is represented by P, which represents the original point cloud data, and m is the number of points in the foreground. i Let β represent the foreground probability of the i-th point, and let β represent the foreground probability threshold. Preferably, ... Set it to 0.5 to ensure a high prospect recall rate.

[0029] Because power cables extend across space and have a coherent structure, they are suitable for modeling contextual coherence in a larger neighborhood. However, small targets such as insulators require a smaller neighborhood to capture their detailed features. Single-radius aggregation often cannot simultaneously accommodate both types of targets, easily leading to issues such as cable break detection or blurred boundaries of small targets. Therefore, this embodiment introduces multi-scale neighborhood aggregation. Small radii enhance geometric detail representation, while large radii model structural continuity. Multi-scale complementarity is achieved during the fusion phase, thereby improving structural preservation and segmentation accuracy. The specific multi-scale neighborhood aggregation steps include: Based on multiple preset scale radii, a spherical neighborhood search is performed on each foreground point in the foreground point set within each scale radius to obtain multiple first neighborhoods. Each neighborhood point in the first neighborhood belongs to a foreground point in the foreground point set. Average pooling is used to aggregate the spatial features of all neighboring points in each first neighborhood to obtain the scale local features corresponding to each scale radius. The local features corresponding to each scale radius are spliced ​​and linearly fused to obtain the local features corresponding to each foreground point.

[0030] In this embodiment, at least two search radii of different scales are first set. Then, for each foreground point, a spherical neighborhood point search is performed within the search radii of different scales to extract local context information at different scales. Taking the pre-set two scale radii as an example, assuming that the radius values ​​of these two scale radii are respectively and ,and For each foreground point , respectively at the radius value and radius value By performing a spherical neighborhood search within the foreground point, we can obtain the neighborhood set of the foreground point at different radii, expressed as follows: In the formula, Indicates the foreground point p i In the neighborhood at scale s, p j Indicates in The j-th neighboring point P within the range f Indicates the meeting point of the previous site, r s This represents the radius value corresponding to scale s. This represents the L2 norm. Here, the scale 's' can be understood as different scale radii.

[0031] Then, average pooling is used to aggregate the spatial features of all neighboring points within the neighborhood at each scale, resulting in the local scale features at scale s, expressed as: In the formula, Indicates the foreground point p i Local scale features at scale s p represents the j-th neighboring point. j Spatial characteristics, Indicates the foreground point p i The number of neighborhood points at scale s.

[0032] Finally, the local features at each scale are concatenated and further fused using a feedforward mapping module to form the final local features of the foreground point. Taking the two scale radii mentioned above as an example, the foreground point p... i The local features can be represented as: In the formula, Indicates the foreground point p i Local features, The FM() denotes the feature concatenation operation, and the FM() denotes the feedforward mapping module. This module consists of two linear transformations and aims to unify, compress, and weight the multi-scale feature channels to achieve a learnable fusion effect.

[0033] Through the above steps, new enhanced local features corresponding to each foreground point can be obtained. However, in the point cloud segmentation task of power lines, while local features can effectively characterize the fine-grained structural information of the point cloud, relying solely on the local receptive field easily overlooks the broader contextual structural relationships, especially when power lines span large spatial distances, where local point features tend to lack overall coherence. Therefore, this invention introduces a superpoint aggregation mechanism, integrating spatially adjacent points with similar features into a higher-order unit, extracting superpoint features with global structure awareness capabilities to enhance subsequent global feature modeling and fusion. Specific steps include: Within a preset superpoint radius, a spherical neighborhood search is performed on each foreground point in the set of foreground points to obtain a second neighborhood. The superpoint radius is greater than the scale radius, and each neighborhood point in the second neighborhood belongs to a point in the original point cloud data. The second neighborhood is segmented using a region growing method to obtain multiple super-point regions; Global pooling is used to aggregate the spatial features of all neighboring points within each superpoint region to obtain a superpoint feature set. The superpoint feature set contains the superpoint features corresponding to each superpoint region and is used as the global feature set.

[0034] In this embodiment, the superpoint radius is first set. This radius is typically larger than the radius of the multi-radius sphere neighborhood aggregation to ensure the formation of a more connected region. Assume the superpoint radius is... For each foreground point A spherical neighborhood search is performed within the radius of the superpoint, and the resulting neighborhood can be represented as: In the formula, P represents the original point cloud data, and N(p i () indicates the foreground location In the radius of the superpoint The neighborhood within, Indicates that in N(p) i The j-th neighboring point within the superpoint radius. It should be noted that the neighboring points within the superpoint radius are selected from all points in the original point cloud data, while the neighborhood points within the multi-scale radius are selected from all foreground points in the foreground point set.

[0035] Then, a region growing method is used, starting from a randomly selected seed point, to recursively expand the neighborhood of points with similar features, ultimately forming M super-point regions, denoted as M0. The specific region partitioning steps are the standard steps of the region growing method, and will not be elaborated here. Each resulting super-point region... Contains several origin points By analyzing each super-point region The spatial features of all original points are globally pooled to obtain a high-level representation of the superpoint region. Specifically, max pooling is used to obtain the superpoint features of the superpoint region. In the formula, Indicates the super-point area The superpoint feature, max( () indicates taking the maximum value element by element. Represents the original point in the superpoint region. Spatial characteristics.

[0036] Finally, all the superpoint features are aggregated to obtain the global feature F. G : Where R represents the set of real numbers, M represents the number of superpoint regions, and C represents that each region has C-dimensional attributes.

[0037] In this embodiment, superpoint features can comprehensively encode feature differences within the region while retaining the most representative response, thus obtaining representative features of the region, which is beneficial for subsequent global context modeling.

[0038] In the 3D point cloud of power lines, the spatial structure of power lines exhibits high linear extensibility and weak semantic variation. This makes traditional Transformer somewhat redundant when modeling long-distance dependencies. In contrast, power poles and towers demonstrate clear local geometric boundaries, requiring more refined local perception capabilities. To effectively integrate local fine-grained features with global structural features and improve segmentation performance, this embodiment employs a feature fusion model based on a large language model to fuse the extracted local and global features. Preferably, a RWKV (Receptance-Weighted Key-Value) model is used to construct the feature fusion model.

[0039] The RWKV model is an architecture that combines long-range feature representation capabilities with computational orderliness. Its advantages include the ability to model long-range dependencies, making it suitable for handling structurally strong power lines; avoiding the secondary computational complexity of self-attention, resulting in higher computational efficiency; a built-in sequential update mechanism, adapting to linear path data; and strong extensibility, making it naturally applicable to point sequence or region sequence modeling.

[0040] This embodiment utilizes the aforementioned characteristics of the RWKV model to establish a parallel RWKV feature fusion structure feature fusion model. This model includes two parallel feature fusion modules: a local feature fusion module and a global feature fusion module. A query vector fusion module is connected in series after these two parallel modules. Both the local and global feature fusion modules are built based on the RWKV model, while the query vector fusion module is built based on a gated fusion mechanism. Under this model architecture, local features and global superpoint features are fed into two separate RWKV paths for independent modeling. Finally, query vector fusion is used to generate an instance-aware representation for subsequent point-level category prediction. The specific feature fusion steps include: The local features are input into the local feature fusion module to obtain the local query vector; The global features are input into the global feature fusion module to obtain the global query vector; The local query vector and the global query vector are input into the query vector fusion module. The local query vector and the global query vector are weighted and fused according to the gating fusion mechanism to obtain the fused query vector.

[0041] In this embodiment, the local feature is set as F. L The global feature is F GThe local features are represented as follows: In the formula, This represents the total number of foreground points.

[0042] Local features are input into the local feature fusion module, and global features are input into the global feature fusion module. The local features are then processed through two RWKV branches. With global features To perform modeling, specifically, K learnable query vectors are introduced into each RWKV model, denoted as: In the formula, This represents the k-th query vector at level 0. This represents the set of query vectors at level 0, where the k-th query vector has C-dimensional attributes, and K represents the total number of query vectors.

[0043] The query vector Q output by the l-th layer of the RWKV model l The update is obtained through the following methods: In the formula, X represents the features of the input module, including local features and global features, and RWKV(*) represents the modeling process of the RWKV model.

[0044] In the structure of the RWKV model, sequence modeling (abbreviated as single vector dimension) is based on the following formula: In the formula, The model output represents time t. For Receptivity Gating, Transformation for the key vector. For the transformation of the value vector, This represents a weighted memory of historical states. For element-wise multiplication, For the Sigmoid function, This represents the input sequence at time t. This mechanism maintains the representational power of the Transformer model while avoiding complex attention matrix calculations. It should be noted that the RWKV model in this embodiment can refer to the structure and modeling process of the conventional RWKV model, and is not specifically limited here.

[0045] Assume the local query vector Q output by the local feature fusion module and the global feature fusion module L and global query vector Q G They are represented as follows: In the formula, RWKV l RWKV represents the local feature fusion module. g This indicates the global feature fusion module.

[0046] Finally, the query vectors output by the two modules are input into the query vector fusion module. A gated fusion mechanism is used to weight and fuse the local and global query vectors to obtain the fused query vector. Preferably, in this embodiment, a learnable function is used to control the dynamic weights of the two query vectors, and the weights can be expressed as: In the formula, α represents the gating weight. This is the Sigmoid activation function, and [.|.] indicates feature concatenation. and All of these are gating parameters, and gating parameters are learnable parameters.

[0047] The final fused query vector Q fused for: The gating fusion mechanism in this embodiment can adaptively adjust the degree of attention to local or global branches according to the semantic context of the current query, and has good interpretability and expressiveness.

[0048] In another preferred embodiment, the present invention also provides a feature fusion mechanism based on uncertainty gating, the specific steps of which include: Based on the multilayer perceptron, uncertainty prediction is performed on the local query vector and the global query vector respectively to obtain branch uncertainty values, which include local branch uncertainty values ​​and global branch uncertainty values; Based on the branch uncertainty value, the feature weights are calculated, and the local query vector and the global query vector are weighted and fused according to the feature weights to obtain the fused query vector.

[0049] In this embodiment, point cloud data in actual power line inspection scenarios often exhibits sparse local areas, severe occlusion, or high noise levels, leading to unreliable features for a particular path (local or global). If reliable and unreliable features cannot be distinguished during fusion, missegmentation or missed segmentation can easily occur. Therefore, this embodiment predicts an uncertainty value for each branch before fusion; the larger the value, the less reliable the features of that branch at the current point. During fusion, an inverse weighting is applied based on the uncertainty, giving reliable branches a higher weight.

[0050] This embodiment uses a lightweight multilayer perceptron (MLP) to perform uncertainty prediction on local and global query vectors respectively. The activation function of the MLP model is the Softplus function, which is a smooth activation function to ensure that the uncertainty value output by the MLP model is positive. It should be noted that the MLP for uncertainty prediction of local and global query vectors is trained on different types of datasets.

[0051] By inputting the local and global query vectors into the corresponding MLP models for uncertainty prediction, we can obtain the local branch uncertainty value and the global branch uncertainty value. A larger uncertainty value indicates that the branch is less reliable. Then, the reciprocal of the uncertainty value is used as the basis for setting the gating weights. In the formula, Represents the local branch weight, u L u represents the uncertainty value of a local branch. G This represents the uncertainty value of the global branch.

[0052] Then, the local query vector and the global query vector are weighted and fused to obtain the fused query vector: Using the above fusion formula, if the uncertainty of a certain path is large, its weight in the fusion result will be automatically reduced, thereby effectively suppressing the interference of unreliable features on the final prediction.

[0053] In this embodiment, the fusion query vector integrates local and global information queries. Based on the fusion query vector, a lightweight point-level prediction head is designed to directly output the semantic segmentation result of each point. Preferably, the point-level prediction head consists of a linear layer and a softmax activation function. The softmax function ensures that the output is an effective probability distribution. The point-level prediction head performs a linear transformation on the input fusion query vector, and then uses the softmax activation function to perform classification probability prediction on the linearly transformed fusion query vector, thereby obtaining the category probability vector of each foreground point in the original point cloud data. Based on the category probability vector of each point, the point cloud segmentation result of the power line is obtained.

[0054] To verify the performance improvement of this invention in segmenting slender linear structures (such as power cables) and small target objects (such as insulators) in power scenarios, this embodiment conducted comparative experiments on a real power line point cloud dataset. This dataset originates from a ±500kV transmission line UAV inspection mission in a certain region, covering various complex environments including mountainous areas, plains, and suburban areas, and provides manually labeled cable and insulator categories. The conventional comparison methods used were the PointNet++ point cloud network model and the PointTransformer point cloud transformer network model. The semantic segmentation evaluation metric used was the average intersection-over-union ratio (AOR). The comparison results are shown below. Figure 2 As shown, according to Figure 2 As can be seen, the method proposed in this invention achieves the best result of 0.833 in the average intersection-union ratio (IU / R), which is about 7.8% and 4.6% higher than PointNet++ (0.762) and PointTransformer (0.788), respectively. This result fully demonstrates the superiority and robustness of the proposed method in the segmentation of slender cables and small targets, and shows good application prospects.

[0055] This embodiment provides a segmentation method for 3D point clouds of power lines. It efficiently extracts multi-scale local structural information from the point cloud through sparse convolution, reducing redundant computation and ensuring effective perception of low-density areas. Through foreground point filtering and multi-scale sphere querying, it accurately extracts local structural features with power line morphology, effectively improving the separability of slender targets. Superpoint partitioning based on spatial proximity and feature similarity compresses the point cloud size while improving global semantic consistency, facilitating the modeling of large-scale structural relationships. A lightweight RWKV architecture is used to model local detail paths and global structural paths separately, avoiding the computational bottlenecks of traditional Transformers and providing stronger long-distance dependency modeling capabilities, thus improving prediction accuracy at semantic boundaries. This embodiment achieves high-precision, high-efficiency, and highly scalable semantic segmentation results in large-scale, sparse, and structurally complex power scene point cloud data, demonstrating significant practicality and generalizability.

[0056] Please see Figure 3 Based on the same inventive concept, the second embodiment of this invention proposes a segmentation system for three-dimensional point clouds of power lines, comprising: The spatial feature extraction module 10 is used to acquire the original point cloud data of the power line, input the original point cloud data into a preset multi-scale spatial feature extraction model, and obtain the spatial features of the original point cloud data. The multi-scale spatial feature extraction model is constructed based on the sparse three-dimensional U-Net model. The feature aggregation module 20 is used to perform foreground probability prediction on the original point cloud data based on the spatial features to obtain a set of foreground points, perform multi-radius sphere neighborhood aggregation on the set of foreground points to obtain local features, and perform superpoint aggregation on the set of foreground points to obtain global features. The parallel perceptual fusion module 30 is used to fuse the local features and the global features according to a preset feature fusion model to obtain a fused query vector. The feature fusion model is constructed based on a large language model. The classification prediction module 40 is used to perform classification probability prediction on the fusion query vector to obtain the category probability vector of the original point cloud data, and obtain the point cloud segmentation result of the power line based on the category probability vector.

[0057] The technical features and effects of the power line 3D point cloud segmentation system proposed in this embodiment are the same as those of the method proposed in this embodiment, and will not be repeated here. Each module in the above-mentioned power line 3D point cloud segmentation system can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0058] Furthermore, embodiments of the present invention also propose a computer 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 steps of the above-described method.

[0059] Please see Figure 4 The diagram illustrates the internal structure of a computer device in one embodiment. This computer device can specifically be a terminal or a server. The computer device includes a processor, memory, network interface, display, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for segmenting three-dimensional point clouds of power lines. The display screen of the computer device can be a liquid crystal display (LCD) or an e-ink display. The input devices of the computer device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse, etc.

[0060] Those skilled in the art will understand that Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computing devices may include more or fewer components than those shown in the figure, or combine certain components, or have the same component arrangement.

[0061] Furthermore, embodiments of the present invention also propose a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0062] In summary, the present invention provides a method, system, device, and storage medium for segmenting three-dimensional point clouds of power lines. The method acquires raw point cloud data of the power line, inputs the raw point cloud data into a preset multi-scale spatial feature extraction model to obtain spatial features of the raw point cloud data. The multi-scale spatial feature extraction model is constructed based on a sparse three-dimensional U-Net model. Based on the spatial features, foreground probability prediction is performed on the raw point cloud data to obtain a set of foreground points. Multi-radius sphere neighborhood aggregation is performed on the set of foreground points to obtain local features, and superpoint aggregation is performed on the set of foreground points to obtain global features. The local features and global features are fused according to a preset feature fusion model to obtain a fusion query vector. The feature fusion model is constructed based on a large language model. Classification probability prediction is performed on the fusion query vector to obtain a category probability vector of the raw point cloud data. Based on the category probability vector, the point cloud segmentation result of the power line is obtained. This invention efficiently extracts multi-scale local structural information from point clouds through sparse convolution, reducing redundant computation while ensuring effective perception of low-density regions. It effectively improves the separability of slender targets through foreground point selection and multi-scale sphere query. By using superpoint partitioning based on spatial proximity and feature similarity, it compresses the point cloud size while enhancing global semantic consistency. A lightweight RWKV architecture is used to model local detail paths and global structural paths separately, effectively integrating local fine-grained features with global structural features, significantly improving prediction accuracy at semantic boundaries. This invention effectively enhances the semantic recognition capability of sparsely structured targets while maintaining low computational cost and low model complexity, achieving high-precision and high-efficiency semantic segmentation results.

[0063] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0064] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this invention, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A method of segmenting a power line three-dimensional point cloud, the method comprising: The method comprises the following steps: obtaining original point cloud data of a power line, inputting the original point cloud data into a preset multi-scale spatial feature extraction model to obtain spatial features of the original point cloud data, and the multi-scale spatial feature extraction model is constructed based on a sparse three-dimensional U-Net model; performing foreground probability prediction on the original point cloud data according to the spatial features to obtain a foreground point set, performing multi-radius spherical neighborhood aggregation on the foreground point set to obtain local features, and performing superpoint aggregation on the foreground point set to obtain global features; fusing the local features and the global features according to a preset feature fusion model to obtain a fusion query vector, and the feature fusion model is constructed based on a large language model; performing classification probability prediction on the fusion query vector to obtain a category probability vector of the original point cloud data, and obtaining a point cloud segmentation result of the power line according to the category probability vector.

2. The segmentation method of a power line three-dimensional point cloud according to claim 1, characterized in that, The step of performing foreground probability prediction on the original point cloud data according to the spatial features to obtain a foreground point set comprises the following steps: inputting the spatial features into a foreground probability prediction head to obtain the foreground probability of each point, and the foreground probability prediction head is composed of a linear layer and an activation function; comparing the foreground probability with a foreground probability threshold value, and regarding the points corresponding to the foreground probability greater than the foreground probability threshold value as foreground points to obtain a foreground point set.

3. The method of claim 1, wherein, The step of performing multi-radius spherical neighborhood aggregation on the foreground point set to obtain local features comprises the following steps: performing spherical neighborhood point search on each foreground point in the foreground point set within each scale radius according to a plurality of preset scale radii to obtain a plurality of first neighborhoods, and each neighborhood point in the first neighborhood belongs to the foreground point set; performing feature aggregation on the spatial features of all neighborhood points in each first neighborhood by using an average pooling operation to obtain a scale local feature corresponding to each scale radius; performing feature splicing and linear fusion on the scale local features corresponding to each scale radius to obtain a local feature corresponding to each foreground point.

4. The segmentation method of a power line three-dimensional point cloud according to claim 3, characterized in that, The step of performing superpoint aggregation on the foreground point set to obtain global features comprises the following steps: performing spherical neighborhood point search on each foreground point in the foreground point set within a preset superpoint radius to obtain a second neighborhood, and the superpoint radius is greater than the scale radius, and each neighborhood point in the second neighborhood belongs to the points in the original point cloud data; performing region segmentation on the second neighborhood by using a region growing method to obtain a plurality of superpoint regions; performing feature aggregation on the spatial features of all neighborhood points in each superpoint region by using a global pooling operation to obtain a superpoint feature set, the superpoint feature set contains a superpoint feature corresponding to each superpoint region, and the superpoint feature set is taken as a global feature.

5. The method of claim 1, wherein, The feature fusion model comprises a local feature fusion module, a global feature fusion module and a query vector fusion module, and the local feature fusion module and the global feature fusion module are connected with the query vector fusion module. The local feature fusion module and the global feature fusion module are constructed by using an RWKV model, and the query vector fusion module is constructed based on a gated fusion mechanism.

6. The segmentation method of a power line three-dimensional point cloud according to claim 5, characterized in that, The step of aggregating the local feature and the global feature according to a preset feature fusion model to obtain a fusion query vector includes: inputting the local feature into the local feature fusion module to obtain a local query vector; inputting the global feature into the global feature fusion module to obtain a global query vector; inputting the local query vector and the global query vector into the query vector fusion module, and performing weighted fusion on the local query vector and the global query vector according to a gated fusion mechanism to obtain a fusion query vector.

7. The segmentation method of a power line three-dimensional point cloud according to claim 6, characterized in that, The step of performing weighted fusion on the local query vector and the global query vector according to a gated fusion mechanism to obtain a fusion query vector includes: performing uncertainty prediction on the local query vector and the global query vector based on a multi-layer perception respectively to obtain branch uncertainty values, the branch uncertainty values including local branch uncertainty values and global branch uncertainty values; calculating feature weights according to the branch uncertainty values, and performing weighted fusion on the local query vector and the global query vector according to the feature weights to obtain a fusion query vector.

8. A system for segmentation of a power line 3D point cloud, the system being applied to the method of any one of claims 1 to 7, characterized in that, It includes: a spatial feature extraction module configured to obtain original point cloud data of a power line, input the original point cloud data into a preset multi-scale spatial feature extraction model, and obtain spatial features of the original point cloud data, the multi-scale spatial feature extraction model being constructed based on a sparse three-dimensional U-Net model; a feature aggregation module configured to perform foreground probability prediction on the original point cloud data according to the spatial features, obtain a foreground point set, perform multi-radius spherical neighborhood aggregation on the foreground point set to obtain a local feature, and perform superpoint aggregation on the foreground point set to obtain a global feature; a parallel perception fusion module configured to fuse the local feature and the global feature according to a preset feature fusion model to obtain a fusion query vector, the feature fusion model being constructed based on a large language model; a classification prediction module configured to perform classification probability prediction on the fusion query vector to obtain a category probability vector of the original point cloud data, and obtain a point cloud segmentation result of the power line according to the category probability vector.

9. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the steps of the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 7.

Images