Power transmission channel point cloud classification method and training method and system of point cloud classification model

Abstract

The invention provides a power transmission channel point cloud classification method and a training method and system of a point cloud classification model. And carrying out deep learning model training by adopting a multi-scale feature fusion and weighted loss method. Obtaining a first local feature, a second local feature and a third local feature corresponding to the original point cloud based on a multi-scale grouping mode; carrying out feature fusion on the three local features to obtain a comprehensive geometric feature; determining an original point cloud sample corresponding to the comprehensive geometric features; training a point cloud classification model through an equilibrium point cloud sample obtained after weighting processing is performed on the original point cloud sample; according to the method, feature extraction is carried out on original point clouds in a multi-scale grouping mode, higher-dimensional point cloud features are obtained, dynamic weighting among the features can be realized by adopting a multiplication feature fusion strategy, and local features have higher distinction degree and robustness; by performing weighting processing on an original point cloud sample, the problem of sample imbalance existing in the original point cloud is solved; accurate training of the point cloud classification model is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electric power engineering, and relates to a power transmission channel point cloud classification method, a point cloud classification model training method and system. BACKGROUND

[0002] In the field of electric power system engineering, the reliable operation of the power transmission line is a core element for ensuring the stability of the power grid and the quality of power supply. As a key infrastructure supporting power transmission, the power transmission tower is long-term exposed to complex and dynamically changing natural environment, and its structural integrity and operating state directly determine the safety margin and power supply continuity of the entire power transmission system. The traditional manual inspection and visual inspection mode faces severe challenges in terms of efficiency, cost and safety.

[0003] With the progress of science and technology, airborne laser scanning (LiDAR) technology has gradually become the mainstream technical solution for intelligent inspection of power transmission lines. The intelligent inspection system based on unmanned aerial vehicles usually includes two core operation stages: line channel inspection and detailed tower body inspection. In the channel inspection stage, the unmanned aerial vehicle needs to follow the preset line geometric structure information to realize autonomous path tracking and navigation, accurately maintain the flight trajectory and ensure a safe distance from the live body. After entering the detailed inspection stage, the unmanned aerial vehicle needs to accurately maneuver to the vicinity of the tower and hover at a series of preset or dynamically generated key detection points to collect high-resolution image data of key components such as power fittings and insulators for subsequent defect recognition and state analysis. However, the current technology applied to intelligent detection of power transmission towers still has several technical problems to be solved. SUMMARY

[0004] The application provides a power transmission channel point cloud classification method, a point cloud classification model training method and system, which are used to improve the accuracy of point cloud classification.

[0005] In a first aspect, the application provides a training method of a power transmission channel point cloud classification model, which includes: obtaining an original point cloud; performing feature extraction operation on the original point cloud based on a multi-scale grouping manner to obtain first, second and third local features corresponding to the original point cloud; fusing the first, second and third local features to obtain comprehensive geometric features; determining original point cloud samples corresponding to the comprehensive geometric features; performing weighted processing on the original point cloud samples to obtain balanced point cloud samples; and training the point cloud classification model based on the balanced point cloud samples.

[0006] In an implementation form of the first aspect, the performing feature extraction operation on the original point cloud based on the multi-scale grouping manner to obtain the first, second and third local features corresponding to the original point cloud includes:

[0007] The original point cloud is grouped according to three scales of small-scale grouping, medium-scale grouping and large-scale grouping, and first local features corresponding to the small-scale grouping, second local features corresponding to the medium-scale grouping and third local features corresponding to the large-scale grouping are obtained respectively, wherein the small-scale grouping focuses on fine changes and edges of the surface of the original point cloud, and is used to capture local details of the original point cloud; the medium-scale grouping understands the relationship between the parts and the surrounding environment corresponding to the original point cloud while seeing the details of the original point cloud; and the large-scale grouping focuses on the complete shape and spatial pose of the object corresponding to the original point cloud, and is used to capture information of macroscopic overall aspects.

[0008] In an implementation form of the first aspect, the obtaining the first local features corresponding to the small-scale grouping, the second local features corresponding to the medium-scale grouping and the third local features corresponding to the large-scale grouping comprises: determining reference information features and point-to-point relationship features corresponding to local point sets in the small-scale grouping, the medium-scale grouping and the large-scale grouping respectively, wherein the reference information features represent basic physical properties corresponding to the original point cloud, and the point-to-point relationship features represent geometric structures of local neighborhoods corresponding to the original point cloud; multiplying the reference information features and the point-to-point relationship features in the small-scale grouping to obtain the first local features; multiplying the reference information features and the point-to-point relationship features in the medium-scale grouping to obtain the second local features; and multiplying the reference information features and the point-to-point relationship features in the large-scale grouping to obtain the third local features.

[0009] In an implementation form of the first aspect, an expression of the multiplying the reference information features and the point-to-point relationship features in the small-scale grouping to obtain the first local features is:

[0010] A = mul (A1, A2)

[0011] wherein A represents the first local features, A1 represents the reference information features, and A2 represents the point-to-point relationship features.

[0012] In an implementation form of the first aspect, a weighted cross-entropy loss function corresponding to the balanced point cloud sample obtained by performing the weighted processing on the original point cloud sample is:

[0013]

[0014] wherein K represents a total number of categories, y i represents a true label of a point, p i represents a probability that the model predicts that the point belongs to a category i, and w i represents a weight of the category i.

[0015] In a possible implementation manner of the first aspect, the expression corresponding to the weight is:

[0016]

[0017] where t represents a hyperparameter greater than 1, per i represents the proportion of the category i in the total categories.

[0018] In the method for training the power transmission channel point cloud classification model provided in the embodiments of the present application, the original point cloud is subjected to feature extraction in a multi-scale grouping manner, so that the geometric information of the original point cloud in all directions from fine local details to macro overall morphology can be obtained; the original point cloud samples are subjected to loss optimization to perform sample weighting processing, so that balanced point cloud samples are obtained, and the technical problem of unbalanced samples in which the ground and vegetation point cloud proportion is high and the power facility point cloud proportion is extremely low in the original point cloud in the power transmission channel data is solved; by combining the multi-scale grouping manner with the weighting processing, accurate point cloud data is provided for training the point cloud classification model, the accurate training of the point cloud classification model is realized, and the classification accuracy can reach more than 90%, thereby providing an accurate point cloud classification model for the efficient and accurate classification of the five types of ground objects, i.e., the tower, the grounding wire, the insulator, the ground channel, and the building in the power transmission channel.

[0019] In a second aspect, the present application provides a method for classifying point clouds of a power transmission channel, the method comprising: obtaining an original point cloud; and classifying the original point cloud based on the trained point cloud classification model of any one of the first aspect to obtain ground object samples corresponding to the original point cloud.

[0020] The embodiments of the present application provide a method for classifying point clouds of a power transmission channel, in which the original point cloud is classified based on the trained point cloud classification model of any one of the first aspect to obtain ground object samples corresponding to the original point cloud, so that the efficient and accurate classification of the five types of ground objects, i.e., the tower, the grounding wire, the insulator, the ground channel, and the building in the power transmission channel is realized, and the accuracy of the classification of the five types of ground objects in the power transmission channel is greatly improved.

[0021] Thirdly, embodiments of this application provide a training apparatus for a point cloud classification model. The apparatus includes: an original point cloud acquisition module for acquiring an original point cloud; a feature extraction module for performing feature extraction operations on the original point cloud based on a multi-scale grouping method to obtain a first local feature, a second local feature, and a third local feature corresponding to the original point cloud; a comprehensive geometric feature determination module for fusing the first local feature, the second local feature, and the third local feature to obtain a comprehensive geometric feature; an original point cloud sample determination module for determining the original point cloud sample corresponding to the comprehensive geometric feature; a balanced point cloud sample determination module for performing weighted processing on the original point cloud sample to obtain a balanced point cloud sample; and a training module for training the point cloud classification model based on the balanced point cloud sample.

[0022] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the training method for the power transmission channel point cloud classification model as described in any one of the first aspects of this application and the power transmission channel point cloud classification method as described in any one of the second aspects.

[0023] Fifthly, embodiments of this application provide an electronic device, the electronic device comprising: a memory storing a computer program; and a processor communicatively connected to the memory, wherein when the computer program is invoked, the processor executes the training method for the power transmission channel point cloud classification model described in any one of the first aspects of the embodiments of this application and the power transmission channel point cloud classification method described in any one of the second aspects.

[0024] As described above, the power transmission channel point cloud classification method, the point cloud classification model training method, and the system described in this application have the following beneficial effects:

[0025] 1) By extracting features from the original point cloud through multi-scale grouping, we can capture the geometric information of the original point cloud from fine local details to macroscopic overall shape. By multiplying the baseline information features and inter-point relationship features in each scale group, we obtain the high-dimensional features corresponding to the original point cloud. The feature vector has a dimension far exceeding the original dimension and contains rich geometric and semantic information, which improves the training accuracy of the point cloud classification model.

[0026] 2) By applying a weighted cross-entropy loss function to the original point cloud samples, the technical problem of sample imbalance, which is common in the original point clouds of power transmission channel data, is solved. This makes the point cloud classification model pay more attention to each category of samples during training, thereby significantly improving the stability and convergence quality of the point cloud classification model training.

[0027] 3) By combining multi-scale grouping with weighted processing, accurate point cloud data is provided for training the point cloud classification model, enabling accurate training of the point cloud classification model with a classification accuracy of over 90%. This provides an accurate point cloud classification model for the efficient and precise classification of various ground features in the power transmission channel.

[0028] 4) The multiplicative feature fusion strategy is adopted. Compared with the additive fusion or splicing fusion, the multiplicative feature fusion strategy can retain more information of the original features during the fusion process and reduce information loss. Moreover, by multiplying the reference information features and the inter-point relationship features, the multiplicative feature fusion strategy can realize dynamic weighting between features, so that the inter-point relationship features (A2) can modulate the expression intensity of the reference information features (A1), realizing the fine fusion of the reference information features and the inter-point relationship features, thereby making the final fused local features have stronger discriminativeness and robustness.

[0029] 5) Based on the trained point cloud classification model, the point cloud of the power transmission channel is classified, realizing the automated classification of power transmission channels (including five types of ground features: towers, grounding wires, insulators, ground channels, and buildings). Attached Figure Description

[0030] Figure 1A The diagram shown is a hardware scene diagram corresponding to the power transmission channel point cloud classification method provided in an embodiment of this application.

[0031] Figure 1B The flowchart shown is a training method for a power transmission channel point cloud classification model provided in an embodiment of this application.

[0032] Figure 2A The flowchart shown is for determining a first local feature, a second local feature, and a third local feature according to an embodiment of this application.

[0033] Figure 2B The flowchart shown is for determining comprehensive geometric features according to an embodiment of this application.

[0034] Figure 3 The flowchart shown is another method for determining a first local feature, a second local feature, and a third local feature, as provided in an embodiment of this application.

[0035] Figure 4A The flowchart shown is a power transmission channel point cloud classification method provided in an embodiment of this application.

[0036] Figure 4B The flowchart shown is another power transmission channel point cloud classification method provided in an embodiment of this application.

[0037] Figure 5 The diagram shown is a schematic diagram of a training apparatus for a point cloud classification model provided in an embodiment of this application.

[0038] Figure 6 The diagram shown is a structural diagram of an electronic device provided in an embodiment of this application.

[0039] Component designation explanation

[0040] Steps S11-S16, Step 55: Equalization Point Cloud Sample Determination Module

[0041] Steps S21-S24, 56. Training Module

[0042] Steps S31-S32, 60. Electronic equipment.

[0043] Steps 61 of S41-S42: Processor

[0044] 50 Training device for point cloud classification model; 62 Non-volatile storage medium

[0045] 51 Original point cloud acquisition module 63 System bus

[0046] 52 Feature extraction module 64 Internal memory

[0047] 53 Comprehensive Geometric Feature Determination Module 65 Network Interface

[0048] 54. Original Point Cloud Sample Determination Module Detailed Implementation

[0049] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0050] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0051] The following embodiments of this application provide a method for classifying point clouds of power transmission channels, a method for training point cloud classification models, and a system, including but not limited to the hardware application scenarios listed in these embodiments. The following description will take the hardware application scenario corresponding to the training method for point cloud classification models of power transmission channels as an example.

[0052] likeFigure 1A As shown in the figure, this application embodiment provides a hardware application scenario diagram of a power transmission channel point cloud classification method, specifically including: a power transmission tower and a drone. The drone is equipped with sensors used to collect the original point cloud of the power transmission tower and wirelessly transmit the collected original point cloud to a point cloud classification model in a ground-based electronic device for classification.

[0053] The point cloud classification model in this application is based on the classic PointNet++ algorithm. To improve the model's ability to perceive complex geometric structures, key improvements were made in the feature extraction stage by adopting a multi-scale grouping (MSG) design. Specifically, by constructing three different scale grouping levels (small, medium, and large), comprehensive geometric information from fine local details to macroscopic overall morphology is captured. During the training of the point cloud classification model, a weighted cross-entropy loss function is introduced to balance the samples and improve the classification accuracy. In the subsequent point cloud classification method for power transmission channels, the original point cloud is directly input into the trained point cloud classification model, realizing the automated classification of end-to-end power transmission channels (including five types of ground features: towers, grounding wires, insulators, ground channels, and buildings).

[0054] The technical solutions in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0055] like Figure 1B As shown in the figure, this embodiment provides a flowchart of a training method for a point cloud classification model of a power transmission channel, as follows: Figure 1B As shown, the training method for the power transmission channel point cloud classification model provided in this application embodiment includes the following steps S11 to S16.

[0056] S11, obtain the original point cloud.

[0057] The original point cloud represents a set of points in three-dimensional space.

[0058] For example, it can be based on a set of three-dimensional spatial points directly collected by lidar, drones or other sensors, i.e., the original point cloud.

[0059] In addition, the original point cloud also includes information such as reflection intensity and color.

[0060] It should be noted that the methods for obtaining the original point cloud listed above are merely illustrative examples. In practical applications, the original point cloud can be obtained through any other suitable method, and this application does not impose any restrictions on this.

[0061] S12, based on a multi-scale grouping method, perform feature extraction on the original point cloud to obtain the first local feature, the second local feature and the third local feature corresponding to the original point cloud.

[0062] Among them, the first local feature represents the fine local features of the ground objects corresponding to the original point cloud; the second local feature represents the feature information between the microscopic details and macroscopic structure of the ground objects corresponding to the original point cloud; and the third local feature represents the overall structure and spatial distribution of the ground objects corresponding to the original point cloud.

[0063] S13, the first local feature, the second local feature and the third local feature are fused to obtain the comprehensive geometric feature.

[0064] Specifically, in the classic PointNet++ framework, the most common and direct method for fusing first, second, and third local features is concatenation.

[0065] For example, the first local feature is represented as: V_small = [s1, s2, s3, ...], the second local feature is represented as: V_medium = [m1, m2, m3, ...], and the third local feature is represented as: V_large = [l1, l2, l3, ...]. Concatenation involves joining the three strings of numbers end to end to form a longer string of numbers, which is the comprehensive geometric feature: V_final = [s1, s2, ..., m1, m2, ..., l1, l2, ...]. The purpose is to completely preserve all information, ensuring that information at all scales is retained intact.

[0066] The stitched composite geometric features are fed into the subsequent neural network. The network automatically learns which scale of the first, second, or third local features to rely on more when making the final judgment. For example, when judging fine texture, it may pay more attention to the first local features; when judging the overall shape, it may pay more attention to the third local features. S14, determine the original point cloud sample corresponding to the composite geometric features.

[0067] For example, in the point feature propagation stage, the comprehensive geometric features are propagated layer by layer and restored to the original point cloud scale through upsampling operations such as inverse distance weighted interpolation, so as to obtain the original point cloud sample corresponding to the comprehensive geometric features.

[0068] S15, perform weighted processing on the original point cloud sample to obtain a balanced point cloud sample.

[0069] In some embodiments, the weighted processing of the original point cloud samples to obtain the weighted cross-entropy loss function corresponding to the balanced point cloud samples is as follows: Where K represents the total number of categories, yi p represents the actual label of a point. i w represents the probability that the model predicts the point belongs to class i. i Let L represent the weight of category i, and L represent the incorrect rating.

[0070] During training, the model calculates an L score based on each prediction it makes. A higher score indicates a more serious error. The goal of the entire training process is to keep the L score as low as possible. Furthermore, based on specially designed rules, if the model misclassifies an important target like an insulator, the L score will rise particularly quickly, thus "punishing" the model and forcing it to learn to recognize these important minority targets.

[0071] Based on the aforementioned weighted entropy loss function, the sample imbalance problem commonly found in power transmission channel data, where the original point cloud corresponding to the ground and vegetation accounts for nearly 90%, while the original point cloud corresponding to power facilities accounts for a very low percentage, can be solved.

[0072] Specifically, by assigning higher weight values ​​to rarer categories such as poles and towers, the penalty for misclassifying these key categories by the point cloud classification model is significantly increased, thereby guiding the point cloud classification model to pay more attention to learning from minority class samples during training.

[0073] In some embodiments, the expression corresponding to the weight is: Where t represents a hyperparameter greater than 1, per i w represents the proportion of category i to the total number of categories. i This represents the weight of category i.

[0074] It should be noted that the value of t can be determined based on the accuracy of the validation set in the point cloud classification model, and this application does not impose any restrictions on the specific value of t.

[0075] S16, Train the point cloud classification model based on the balanced point cloud samples.

[0076] Specifically, multiple balanced point cloud samples are sequentially input into the point cloud classification model to train the point cloud classification model.

[0077] This application provides a training method for a point cloud classification model for power transmission channels. The method includes acquiring an original point cloud; performing feature extraction on the original point cloud based on a multi-scale grouping method to obtain comprehensive geometric features corresponding to the original point cloud; determining the original point cloud samples corresponding to the comprehensive geometric features; performing weighted processing on the original point cloud samples to obtain balanced point cloud samples; and training the point cloud classification model based on the balanced point cloud samples. By using a multi-scale grouping method to extract features from the original point cloud, it can capture comprehensive geometric information of the original point cloud, from fine local details to macroscopic overall morphology, providing a richer and more robust feature foundation. This approach improves the training accuracy of point cloud classification models. By performing weighted processing on the original point cloud samples, it solves the technical problem of sample imbalance commonly found in original point clouds in power transmission channel data. This makes the point cloud classification model pay more attention to samples of each category during training, thus significantly improving the stability and convergence quality of point cloud classification model training. By combining multi-scale grouping with weighted processing, accurate point cloud data is provided for training the point cloud classification model, achieving accurate training of the point cloud classification model with a classification accuracy of over 90%. This provides an accurate point cloud classification model for the subsequent efficient and precise classification of various ground features in power transmission channels.

[0078] like Figure 2A As shown, this embodiment provides a flowchart for determining a first local feature, a second local feature, and a third local feature, as follows: Figure 2A As shown, the method for determining the first local feature, the second local feature and the third local feature provided in this application embodiment includes the following steps S21 to S22.

[0079] S21, the original point cloud is grouped into three scales: small scale grouping, medium scale grouping, and large scale grouping.

[0080] The small-scale group focuses on the fine changes and edges of the original point cloud surface to capture local details of the original point cloud; the medium-scale group, while clearly seeing the details of the original point cloud, understands the relationship between the corresponding components and the surrounding environment; and the large-scale group focuses on the complete shape and spatial posture of the objects corresponding to the original point cloud to capture macroscopic overall information.

[0081] For example, the neighborhood radius corresponding to the small-scale group is between 0.5m and 1m.

[0082] It should be noted that the specific values ​​of the neighborhood radius corresponding to the small-scale groupings listed above are only for illustrative purposes. In practical applications, other suitable neighborhood radii of small-scale groupings can be selected according to specific application requirements, and this application does not impose any restrictions on this.

[0083] It should be noted that small-scale grouping, by setting a small neighborhood radius, accurately extracts surface texture variations and subtle structural features. This level can identify subtle changes and edge features on the surface of ground objects, providing local detail information for subsequent classification. It is particularly crucial when distinguishing highly similar object components, effectively capturing structural features that differ only at the microscopic level.

[0084] For example, the neighborhood radius corresponding to the mesoscale group is between 1m and 5m.

[0085] It should be noted that the specific values ​​of the neighborhood radius corresponding to the mesoscale groupings listed above are only for illustrative purposes. In practical applications, other suitable neighborhood radii of mesoscale groupings can be selected according to specific application requirements, and this application does not impose any restrictions on this.

[0086] It should be noted that mesoscale grouping establishes a connection between microscopic details and macroscopic structure, using an appropriate neighborhood range. It preserves local feature information while integrating the context of the surrounding environment. It is an important bridge for hierarchical expression of features and plays a decisive role in understanding the components of an object and their connection methods. It can effectively identify the intermediate semantic information of the structure.

[0087] For example, the neighborhood radius corresponding to large-scale grouping is between 5m and 10m or larger.

[0088] It should be noted that the specific values ​​of the neighborhood radius corresponding to the large-scale groupings listed above are only for illustrative purposes. In practical applications, other suitable neighborhood radii of large-scale groupings can be selected according to specific application requirements, and this application does not impose any restrictions on this.

[0089] It should be noted that large-scale grouping can help the system quickly identify the main structure and large components, providing a key basis for the recognition of the whole object.

[0090] S22, respectively obtain the first local features corresponding to the small-scale group, the second local features corresponding to the medium-scale group, and the third local features corresponding to the large-scale group.

[0091] Please see Figure 2B , Figure 2B The flowchart shown is a method for determining comprehensive geometric features according to an embodiment of this application. Figure 2B Each step is similar to the method / step in step S13 of Figure 1 above, and will not be described again in this application.

[0092] This application provides a method for determining comprehensive geometric features. In this method, the original point cloud is grouped to obtain the first local features corresponding to the original point cloud in the small-scale group, the second local features corresponding to the original point cloud in the medium-scale group, and the third local features corresponding to the original point cloud in the large-scale group. This method can capture the local features of the original point cloud at different scales, providing a richer and more robust local feature foundation for subsequent training of the point cloud classification model. Furthermore, the first local features, the second local features, and the third local features are fused to obtain comprehensive geometric features containing multi-scale information, making the comprehensive geometric features more accurately reflect the feature information of the original point cloud.

[0093] like Figure 3 As shown, this embodiment provides another flowchart for determining the first local feature, the second local feature, and the third local feature, as follows: Figure 3 As shown, another method for determining the first local feature, the second local feature, and the third local feature provided in this application embodiment includes the following steps S31 to S34.

[0094] S31, determine the reference information features and inter-point relationship features corresponding to the local point sets within the small-scale group, the medium-scale group, and the large-scale group, respectively.

[0095] The reference information features represent the basic physical properties corresponding to the original point cloud, and the point relationship features represent the geometric structure of the local neighborhood corresponding to the original point cloud.

[0096] In this model, the baseline information feature represents the basic physical properties of a point, while the inter-point relationship feature A2 represents the geometric structure of the local neighborhood. By multiplying the two, dynamic weighting between the features can be achieved, allowing the inter-point relationship feature (A2) to modulate the expressive strength of the baseline information feature (A1). For example, in structurally flat regions, the value of A2 is smaller, which can suppress the influence of noise on the baseline features; while in structurally complex edge regions, the value of A2 is larger, which can amplify key geometric change features, thereby giving the final comprehensive geometric features stronger discriminative power and robustness.

[0097] It should be noted that in multi-scale grouped feature learning, each point initially has only three-dimensional spatial coordinates (X, Y, Z), possibly with a few other dimensions such as the normal vector. 'High-dimensional features' refer to feature vectors with dimensions far exceeding the original dimensions, containing rich geometric and semantic information, thus achieving an increase in the dimensionality of the feature vectors.

[0098] Specifically, the reference information features are determined based on the spatial coordinates and normal vectors of points within the local point sets in the small-scale group, medium-scale group, and large-scale group, respectively.

[0099] For example, if the spatial coordinates of a point within a local point set in a small-scale group are (x, y, z), and the normal vector coordinates are (n... x n y n z If the reference information feature of this point is A1=∑F1(x,y,z,n), then the reference information feature of this point is represented as A1=∑F1(x,y,z,n). x ,n y ,n z ). Where F1 represents the feature extraction function used to encode the baseline information.

[0100] Specifically, the point-to-point relationship feature encodes in detail the geometric relationship between the center point selected by the sampling layer, the mean coordinates of the local region (i.e., the centroid), and any neighboring point within the region. Specifically, this feature encodes the Euclidean distance (d1) connecting the neighboring point to the center point, the Euclidean distance (d2) connecting the neighboring point to the centroid, and the Euclidean distance (d3) connecting the center point to the centroid. Simultaneously, to supplement directional information, cosine similarities (θ1, θ2, θ3) corresponding to these three sets of distances are also encoded, describing the directional relationship of each pair of points relative to the third point.

[0101] For example, the feature of the relationship between points within a small-scale group is represented as A2 = ∑F2(d1,d2,d3,θ1,θ2,θ3). Where F2 represents the feature extraction function used for the relationship information between points.

[0102] It should be noted that the method for determining the reference information features and inter-point relationship features corresponding to the local point sets within the mesoscale group and the large-scale group is similar to the method for determining the reference information features and inter-point relationship features corresponding to the local point sets within the small and large-scale groups, and will not be repeated here.

[0103] S32, the reference information features and the inter-point relationship features within the small-scale group are multiplied to obtain the first local feature.

[0104] In some embodiments, the expression for multiplying the reference information features and the inter-point relationship features within the small-scale group to obtain the first local feature is:

[0105] A = mul(A1, A2)

[0106] Where A represents the first local feature, A1 represents the baseline information feature, and A2 represents the inter-point relationship feature.

[0107] By multiplying A1 and A2 to obtain "high-dimensional features", dynamic weighting between features is achieved, enabling the geometric structural feature A2 to modulate the expression intensity of the basic physical feature A1.

[0108] S33, Multiply the reference information features and the inter-point relationship features within the mesoscale group to obtain the second local feature.

[0109] Specifically, the expression corresponding to the second local feature is obtained by multiplying the reference information feature and the inter-point relationship feature within the mesoscale group. This expression is similar to the expression for determining the first local feature, and will not be repeated here.

[0110] S34, Multiply the reference information features and the inter-point relationship features within the large-scale group to obtain the third local feature.

[0111] Specifically, the expression corresponding to the third local feature is obtained by multiplying the reference information feature and the inter-point relationship feature within the large-scale group. This expression is similar to the expression for determining the first local feature, and will not be repeated here.

[0112] In another method for determining the first local feature, the second local feature, and the third local feature provided in this application embodiment, the reference information features and inter-point relationship features corresponding to the local point sets within the small-scale group, the medium-scale group, and the large-scale group are determined respectively; the reference information features and the inter-point relationship features are fused to obtain the first local feature, the second local feature, and the third local feature respectively; wherein, in the process of feature fusion of the reference information features and the inter-point relationship features, a multiplicative feature fusion strategy is adopted. Compared with the traditional feature fusion method, the multiplicative feature fusion strategy can retain more information of the original features during the fusion process, reduce information loss, and achieve dynamic weighting between features by multiplying the reference information features and the inter-point relationship features, so that the inter-point relationship features (A2) can modulate the expression intensity of the reference information features (A1), thereby achieving fine fusion of the reference information features and the inter-point relationship features, laying the foundation for subsequent point cloud classification.

[0113] like Figure 4A As shown in the figure, this embodiment provides a flowchart of a point cloud classification method for power transmission channels, as follows: Figure 4A As shown, the power transmission channel point cloud classification method provided in this application embodiment includes the following steps S41 to S42.

[0114] S41, obtain the original point cloud.

[0115] S42, classify the original point cloud based on the point cloud classification model trained according to any one of claims 1 to 6 to obtain the ground feature samples corresponding to the original point cloud.

[0116] For example, ground feature samples include poles, grounding wires, insulators, ground passages, and buildings.

[0117] It should be noted that the specific types of land cover samples listed above are only for illustrative purposes. In practical applications, other land cover samples corresponding to the original point cloud can be obtained according to actual needs, and this application does not impose any restrictions on this.

[0118] This application provides a point cloud classification method for power transmission channels. In this method, the original point cloud is input into a point cloud classification model that has been trained based on any one of the first aspects. The original point cloud is classified in the point cloud classification model, and the point cloud classification model outputs the ground feature samples corresponding to the original point cloud. This reduces the possibility of misclassification of the original point cloud and greatly improves the accuracy of the classification of the original point cloud.

[0119] The protection scope of the training method and the classification method of the power transmission channel point cloud described in the embodiments of this application is not limited to the execution order of the steps listed in this embodiment. Any solution implemented by adding, subtracting or replacing steps in the prior art based on the principles of this application is included within the protection scope of this application.

[0120] This application also provides a training device for a point cloud classification model and a system corresponding to a point cloud classification method for power transmission channels. The system includes: a training device for a point cloud classification model, a computer-readable storage medium, and an electronic device.

[0121] This application also provides a point cloud classification model training device, which can implement the power transmission channel point cloud classification model training method described in this application. However, the implementation device of the power transmission channel point cloud classification model training method described in this application includes, but is not limited to, the structure of the point cloud classification model training device listed in this embodiment. All structural modifications and substitutions of the prior art made in accordance with the principles of this application are included within the protection scope of this application.

[0122] Please see Figure 4B , Figure 4B The flowchart shown is another point cloud classification method provided in an embodiment of this application. The steps in Figure B are described above. Figures 1B to 4A The above has already been explained in detail, and this application will not repeat it here.

[0123] like Figure 5 As shown, in one embodiment, the training device 50 of the point cloud classification model of this application includes an original point cloud acquisition module 51, a feature extraction module 52, a comprehensive geometric feature determination module 53, an original point cloud sample determination module 54, an equalized point cloud sample determination module 55, and a training module 56.

[0124] The original point cloud acquisition module 51 is used to acquire the original point cloud.

[0125] The feature extraction module 52 is used to perform feature extraction operations on the original point cloud based on a multi-scale grouping method to obtain the first local feature, the second local feature and the third local feature corresponding to the original point cloud.

[0126] The comprehensive geometric feature determination module 53 is used to fuse the first local feature, the second local feature and the third local feature to obtain comprehensive geometric features.

[0127] The original point cloud sample determination module 54 is used to determine the original point cloud sample corresponding to the comprehensive geometric features.

[0128] The balanced point cloud sample determination module 55 is used to perform weighted processing on the original point cloud sample to obtain a balanced point cloud sample.

[0129] Training module 56 is used to train the point cloud classification model based on the balanced point cloud samples.

[0130] The structure and principle of the original point cloud acquisition module 51, feature extraction module 52, comprehensive geometric feature determination module 53, original point cloud sample determination module 54, balanced point cloud sample determination module 55, and training module 56 correspond one-to-one with the steps in the above feature vector extraction method, so they will not be described in detail here.

[0131] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, or methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules / units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or units may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or modules or units may be electrical, mechanical, or other forms.

[0132] The modules / units described as separate components may or may not be physically separate. The components shown as modules / units may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules / units can be selected to achieve the objectives of the embodiments of this application, depending on actual needs. For example, the functional modules / units in the various embodiments of this application may be integrated into one processing module, or each module / unit may exist physically separately, or two or more modules / units may be integrated into one module / unit.

[0133] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0134] This application also provides an electronic device. Figure 6 The diagram shown is a structural schematic of an electronic device 60 in one embodiment of this application. The process action determination method provided in this embodiment can be applied to... Figure 6 The electronic device shown is 60, but it is not limited to this. For example... Figure 6 As shown, the electronic device 60 includes a processor 61, a memory, a system bus 63, and a network interface 65. The memory may include a non-volatile storage medium 62 and internal memory 64.

[0135] The non-volatile storage medium 62 can store an operating system and a computer program. The computer program includes program instructions that, when executed, cause the processor to perform any of the training methods and classification methods for power transmission channel point cloud classification models provided in the embodiments of this application.

[0136] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0137] The internal memory 64 provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to execute any of the training methods and classification methods for power transmission channel point cloud models provided in the embodiments of this application.

[0138] This network interface 65 is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 6 The 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 computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0139] It should be understood that processor 61 can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, the general-purpose processor can be a microprocessor or any conventional processor.

[0140] The electronic device 60 in this application embodiment may include terminal devices such as tablet computers, laptop computers, mobile phones, supercomputers, and smart wearable devices. It can also be applied to databases, servers, and service response systems based on terminal artificial intelligence. This application embodiment does not impose any restrictions on the specific type of electronic device.

[0141] For example, electronic devices can be stations (STAION, ST) in WLANs, cellular phones, cordless phones, Session Initiation Protocol (SIP) phones, Wireless Local Loop (WLL) stations, handheld devices with wireless communication capabilities, computing devices or other processing devices connected to a wireless modem, computers, laptops, handheld communication devices, handheld computing devices, and / or other devices for communicating over wireless systems, as well as next-generation communication systems, such as mobile terminals in 5G networks, mobile terminals in future evolved Public Land Mobile Networks (PLMNs), or mobile terminals in future evolved Non-terrestrial Networks (NTNs).

[0142] This application also provides a computer-readable storage medium. Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing a processor. The program can be stored in a computer-readable storage medium, which is a non-transitory medium, such as random access memory, read-only memory, flash memory, hard disk, solid-state drive, magnetic tape, floppy disk, optical disk, and any combination thereof. The storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital video disc (DVD)), or a semiconductor medium (e.g., solid-state drive (SSD)).

[0143] This application embodiment may also provide a computer program product comprising one or more computer instructions. When the computer instructions are loaded and executed on a computing device, all or part of the processes or functions described in this application embodiment are generated. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0144] When the computer program product is executed by a computer, the computer performs the method described in the foregoing method embodiments. The computer program product can be a software installation package; when the foregoing method is required, the computer program product can be downloaded and executed on the computer.

[0145] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0146] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A training method for a point cloud classification model of a power transmission channel, characterized in that, The method includes: Obtain the original point cloud; Based on a multi-scale grouping method, feature extraction is performed on the original point cloud to obtain the first local feature, the second local feature and the third local feature corresponding to the original point cloud. The first local feature, the second local feature, and the third local feature are fused to obtain a comprehensive geometric feature. Determine the original point cloud sample corresponding to the comprehensive geometric features; The original point cloud samples are weighted to obtain balanced point cloud samples; The point cloud classification model is trained based on the balanced point cloud samples.

2. The training method for the point cloud classification model of power transmission channels according to claim 1, characterized in that, The feature extraction operation based on multi-scale grouping of the original point cloud yields a first local feature, a second local feature, and a third local feature corresponding to the original point cloud, including: The original point cloud is grouped into three scales: small-scale, medium-scale, and large-scale. First local features corresponding to the small-scale group, second local features corresponding to the medium-scale group, and third local features corresponding to the large-scale group are obtained respectively. Specifically, the small-scale group focuses on the fine variations and edges of the original point cloud surface to capture local details; the medium-scale group, while revealing the details of the original point cloud, helps to understand the relationship between the corresponding components and the surrounding environment; and the large-scale group focuses on the complete shape and spatial orientation of the objects corresponding to the original point cloud to capture macroscopic overall information.

3. The training method for the point cloud classification model of power transmission channels according to claim 2, characterized in that, The step of obtaining the first local features corresponding to the small-scale group, the second local features corresponding to the medium-scale group, and the third local features corresponding to the large-scale group includes: The baseline information features and inter-point relationship features corresponding to the local point sets within the small-scale group, the medium-scale group, and the large-scale group are determined respectively. The baseline information features represent the basic physical properties corresponding to the original point cloud, and the inter-point relationship features represent the geometric structure of the local neighborhood corresponding to the original point cloud. The first local feature is obtained by multiplying the baseline information feature and the inter-point relationship feature within the small-scale group; The second local feature is obtained by multiplying the baseline information feature and the inter-point relationship feature within the mesoscale group; The third local feature is obtained by multiplying the baseline information feature and the inter-point relationship feature within the large-scale group.

4. The training method for the point cloud classification model of the power transmission channel according to claim 3, characterized in that: The expression for the first local feature obtained by multiplying the baseline information feature and the inter-point relationship feature within the small-scale group is as follows: A = mul(A1, A2) Where A represents the first local feature, A1 represents the baseline information feature, and A2 represents the inter-point relationship feature.

5. The training method for the point cloud classification model of power transmission channels according to claim 1, characterized in that: The weighted processing performed on the original point cloud samples yields the weighted cross-entropy loss function corresponding to the balanced point cloud samples: Where K represents the total number of categories, y i p represents the actual label of a point. i w represents the probability that the model predicts the point belongs to class i. i This represents the weight of category i.

6. The training method for the point cloud classification model of the power transmission channel according to claim 5, characterized in that: The expression corresponding to the weight is: Where t represents a hyperparameter greater than 1, per i This indicates the proportion of category i out of the total number of categories.

7. A point cloud classification method for power transmission channels, characterized in that: The method includes: Obtain the original point cloud; The original point cloud is classified based on the point cloud classification model trained according to any one of claims 1 to 6 to obtain the ground feature samples corresponding to the original point cloud.

8. A training device for a point cloud classification model, characterized in that: The device includes: The raw point cloud acquisition module is used to acquire raw point clouds; The feature extraction module is used to perform feature extraction operations on the original point cloud based on a multi-scale grouping method to obtain the first local feature, the second local feature and the third local feature corresponding to the original point cloud; The comprehensive geometric feature determination module is used to fuse the first local feature, the second local feature, and the third local feature to obtain the comprehensive geometric feature; The original point cloud sample determination module is used to determine the original point cloud sample corresponding to the comprehensive geometric features; The balanced point cloud sample determination module is used to perform weighted processing on the original point cloud sample to obtain a balanced point cloud sample. The training module is used to train the point cloud classification model based on the balanced point cloud samples.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the training method for the power transmission channel point cloud classification model according to any one of claims 1 to 6 and the power transmission channel point cloud classification method according to any one of claims 7.

10. An electronic device, characterized in that, The electronic device includes: A memory that stores a computer program; The processor, which is communicatively connected to the memory, executes the training method for the power transmission channel point cloud classification model according to any one of claims 1 to 6 and the power transmission channel point cloud classification method according to any one of claims 7 when calling the computer program.

Images