Power transmission line inspection point cloud data filtering method and system based on deep learning and feature fusion
Through the method of deep learning and feature fusion, combined with density clustering, local curvature weighting and multi-scale convolutional network, the accuracy and adaptability problems of traditional filtering methods in complex environments are solved, and efficient and accurate point cloud data filtering is achieved, which is suitable for transmission line inspection.
Patent Information
- Application Number
- CN202510668438.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-23
- Publication Date
- 2025-10-17
AI Technical Summary
Traditional filtering methods have insufficient accuracy and poor adaptability in complex environments, cannot effectively remove noise and retain key line information, and lack the robustness of multi-dimensional feature fusion.
A method based on deep learning and feature fusion is adopted. Through density clustering, local curvature weighting, Gaussian smoothing and multi-scale convolutional network, feature fusion and adaptive filtering are performed in combination with attention mechanism, and the filtering weights are dynamically adjusted to remove noise points.
It achieves high-precision and robust point cloud data filtering in complex environments, preserves the key structures of transmission lines, improves the quality and adaptability of inspection data, and is suitable for a variety of complex scenarios.
Smart Images
Figure CN120807943A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of three-dimensional computer vision and point cloud processing, and particularly relates to a power line inspection point cloud data filtering method and system based on deep learning and feature fusion. BACKGROUND
[0002] With the continuous development of power line inspection technology, devices such as laser radars can efficiently collect point cloud data of power lines and their surrounding environment, providing an important basis for line state monitoring and fault identification. However, the collected point cloud data often contains a large amount of noise and redundant information, such as background interference of vegetation, buildings, etc., which can seriously affect the accuracy of subsequent analysis and decision-making. Traditional filtering methods mainly rely on artificial design rules or semi-automatic algorithms based on geometric features, which perform well in simple scenarios but have low precision and poor adaptability in complex environments, and cannot effectively remove noise and retain key line information. In addition, traditional methods lack effective fusion of multi-source features when processing multi-dimensional point cloud data, resulting in a lack of robustness and practicality of the filtering results, making it difficult to meet the demand for efficient and accurate filtering of modern power line inspection.
[0003] Currently, some research has made certain progress in improving the precision and efficiency of point cloud filtering. Some research has proposed a filtering method based on statistical outlier removal, which effectively removes outliers by calculating the local density and distance features of the point cloud. This method performs well in simple scenarios, but it is difficult to handle non-linearly distributed noise points in complex environments, and its ability to capture local features of the point cloud is limited. Some research has proposed an adaptive filtering method based on geometric features, which uses curvature and normal vector information of the point cloud to distinguish noise, which can improve the filtering precision to some extent. However, this method is highly dependent on the resolution and quality of the point cloud data, and is prone to mistakenly deleting key points in complex scenarios, making it difficult to meet the needs of fine inspection. Some scholars have proposed a filtering method based on deep learning, which extracts point cloud features through a convolutional neural network and classifies them, significantly improving filtering performance. However, this method mainly focuses on global features and lacks fine modeling of local features, and does not fully utilize the multi-dimensional information of the point cloud, limiting its application value in practical scenarios. SUMMARY
[0004] In view of the above problems, the present application is proposed.
[0005] Therefore, the purpose is to solve the problem of insufficient precision and poor adaptability of traditional filtering algorithms in complex environments.
[0006] To solve the above technical problems, the present application provides the following technical solutions: a power line inspection point cloud data filtering method based on deep learning and feature fusion, which includes the following steps,
[0007] The collected power line inspection point cloud data is preprocessed; key features including local features and global features are extracted from the preprocessed multi-scale point cloud data, and feature fusion is performed through an attention mechanism; based on the attention mechanism, the filtering strategy is adaptively adjusted, the filtering weight is dynamically allocated according to the extracted point cloud local features, noise discrimination is performed, and the filtered point cloud data result is output.
[0008] As a preferred scheme of the power line inspection point cloud data filtering method based on deep learning and feature fusion, the preprocessing includes first processing, second processing and third processing.
[0009] The first processing adopts a clustering algorithm to perform density clustering on the signal, and dynamically identifies and preliminarily removes noise points according to the local density distribution of the point cloud.
[0010] As a preferred scheme of the power line inspection point cloud data filtering method based on deep learning and feature fusion, the second processing includes normalizing the preliminarily denoised data, and calculating the local region weight by considering the adjustment density and curvature feature in the normalization process to obtain the preliminarily cleaned features.
[0011] As a preferred scheme of the power line inspection point cloud data filtering method based on deep learning and feature fusion, the third processing includes smoothing the power line inspection point cloud data, and dynamically adjusting the weight by the distance between the point and the center of the neighborhood using an improved Gaussian weighted smoothing filtering method.
[0012] As a preferred scheme of the power line inspection point cloud data filtering method based on deep learning and feature fusion, the extraction of key features includes using a point cloud convolution network as a basic network, constructing a graph structure through a local neighborhood, and extracting local features and global features.
[0013] For each point, the local features are extracted through convolution operation, and the global features are extracted through maximum pooling.
[0014] By adjusting the convolution kernel and the neighborhood range, different scale key features are obtained.
[0015] As a preferred scheme of the power line inspection point cloud data filtering method based on deep learning and feature fusion, the feature fusion includes dynamically adjusting the weight of the key features of different scales through an attention mechanism.
[0016] Two or more scales of key features are extracted through a multi-scale convolution network.
[0017] Attention weights are calculated for each scale feature;
[0018] The attention weights of all scale features are weighted and summed.
[0019] As a preferred scheme of the power line inspection point cloud data filtering method based on deep learning and feature fusion, the adaptive adjustment filtering strategy includes using a point cloud convolution network to extract local features of the point cloud.
[0020] An attention mechanism is constructed, and the local features are mapped to a query matrix and a key matrix through linear transformation.
[0021] A multi-head attention mechanism is added to adaptively adjust the filtering strategy, and filtering weights are calculated for each point according to the attention weight matrix.
[0022] The point cloud data is filtered using the filtering weights to remove noise points.
[0023] Another object of the present application is to provide a power line inspection point cloud data filtering system based on deep learning and feature fusion.
[0024] To solve the above technical problems, the present application provides the following technical scheme: a power line inspection point cloud data filtering system based on deep learning and feature fusion, comprising: a point cloud data processing module, a feature extraction module, and a filtering module.
[0025] The point cloud data processing module pre-processes the collected power line inspection point cloud data.
[0026] The feature extraction module extracts key features including local features and global features from the pre-processed multi-scale point cloud data, and performs feature fusion through an attention mechanism.
[0027] The filtering module dynamically allocates filtering weights based on the attention mechanism and the extracted local features of the point cloud, discriminates noise, and outputs the filtered point cloud data result.
[0028] The present application provides a computer device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the power line inspection point cloud data filtering method based on deep learning and feature fusion when executing the computer program.
[0029] The present application provides a computer readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of the power line inspection point cloud data filtering method based on deep learning and feature fusion.
[0030] The beneficial effects of the present application: the density-aware feature optimization processing of the present application: through the improved density-based clustering denoising algorithm, combined with local curvature weighting and Gaussian smoothing, the high-fidelity retention of the key components of the power transmission line (conductor, insulator, etc.) is realized.
[0031] The hierarchical feature extraction system is constructed by using the DGCNN architecture, and the geometric features of different neighborhood ranges are aggregated through the differentiable graph convolution. The cross-scale feature association is established combined with the multi-head attention mechanism, and the feature breaking problem of small components such as insulator strings is effectively solved.
[0032] An end-to-end filtering framework is constructed based on a lightweight attention network. The system has dynamic scene adaptation capability, can automatically adjust the filtering strategy for complex environments such as urban dense vegetation and overhead lines, does not need to rely on scene prior knowledge, and effectively meets the engineering requirements of real-time and stability for unmanned aerial vehicle inspection. BRIEF DESCRIPTION OF DRAWINGS
[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.
[0034] Figure 1 The overall flowchart of the power transmission line inspection point cloud data filtering method based on deep learning and feature fusion provided by an embodiment of the present application is shown.
[0035] Figure 2 The Dynamic Graph CNN network structure diagram of the power transmission line inspection point cloud data filtering method based on deep learning and feature fusion provided by an embodiment of the present application is shown. DETAILED DESCRIPTION
[0036] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.
[0037] Embodiment 1, refer to Figure 1 For the first embodiment of the present application, the embodiment provides a power transmission line inspection point cloud data filtering method based on deep learning and feature fusion, which comprises:
[0038] The method can effectively distinguish noise points and key points by introducing a deep learning model and combining multi-level extraction of local features and global features. Meanwhile, the method integrates the geometric properties, spatial distribution and multi-source feature information of the point cloud, and further enhances the robustness and practicality of filtering through feature fusion. In addition, the filtered point cloud data can seamlessly connect to the subsequent analysis and decision system, providing high-quality data support for power line inspection, and the specific implementation process is as shown in Figure 1 .
[0039] S1: Preprocessing the collected power line inspection point cloud data.
[0040] It should be noted that the main purpose of preprocessing the collected data is to eliminate noise and redundant information in the point cloud data and provide high-quality data input for subsequent processing.
[0041] The preprocessing is divided into first processing, second processing and third processing.
[0042] In the first processing, a clustering algorithm is used to perform density clustering on the signal, and noise points are dynamically identified and removed according to the local density distribution of the point cloud.
[0043] Specifically, in the embodiments of the present application, the point cloud data is easily disturbed by the environment (such as dust, vegetation, etc.) during the collection process, resulting in noise points. Traditional methods usually use statistical filtering or radius filtering, but these methods have limited effect when dealing with complex scenes. To solve this problem, the present application uses the DBSCAN algorithm (density clustering algorithm) to perform density clustering on the signal, sets the neighborhood radius to 1.2-1.5 times the average distance of the point cloud, and the minimum clustering point number MinPts is a dynamic adaptive value. This method can dynamically identify and remove noise points according to the local density distribution of the point cloud. The improved denoising formula is as follows:
[0044]
[0045] where P i is the target point, P j is the neighborhood point, k is the number of neighborhood points, and sigma is the weight coefficient. By setting the density threshold, low-density points are identified as noise and removed.
[0046] In an optional embodiment, the preliminary removal of noise points can be based on a random sample consensus algorithm (RANSAC), a certain number of points are randomly selected from the point cloud data to establish a preliminary model, the model fitting is performed using these randomly selected points, the distance between the model and all points is calculated, and the points within the preset threshold are marked as inliers, and the remaining points are marked as noise points, the random sampling and model fitting process is repeated for a maximum of N iterations, and the number of inliers of each model is recorded for selecting the best model, and data cleaning is performed according to the identified noise points, and filtered point cloud data is output.
[0047] In another optional embodiment, the preliminary removal of noise points can also be based on a bounding box constraint filtering method, based on the local density distribution of the point cloud, a bounding box is established according to a preset power transmission line spatial model to limit the processing range of the point cloud data, points falling outside the bounding box are directly determined as noise and removed, and for points falling within the bounding box, the local density of each point is calculated, a local density threshold is set, low-density points are marked as noise points and removed, and point cloud data after bounding box pruning and local density evaluation is output.
[0048] The second processing includes normalizing the preliminary denoised data, in order to avoid weakening of local features by global normalization, local region weights are calculated by considering adjusting density and curvature features during normalization to overcome the problem of weakening of local features caused by traditional global normalization method, and preliminary cleaned features are obtained, specifically:
[0049]
[0050] wherein P norm is the normalized feature value, P min and P max are the minimum and maximum feature values of the entire data set for global normalization; w(P) is a local weight function defined as the density or curvature feature of the points in the neighborhood:
[0051] w(P) = 1 + a · Density(P) + β · Curvature(P)
[0052] wherein a and β are weight coefficients for adjusting the influence of density and curvature on normalization; Density(P) is the density feature of the neighborhood of point P, and Curvature(P) is the curvature feature of the neighborhood of point P.
[0053] The third processing includes smoothing the power transmission line inspection point cloud data to effectively eliminate local fluctuations and extract cleaner geometric features.
[0054] Specifically, a modified Gaussian weighted smoothing filtering method is adopted, in which the weight is dynamically adjusted by the distance between the point and the center of the neighborhood:
[0055]
[0056] wherein q i is a Gaussian weight:
[0057]
[0058] wherein P center is the center point of the neighborhood, and sigma1 is a parameter for controlling the smoothing strength.
[0059] S2: key features including local features and global features are extracted from the preprocessed multi-scale point cloud data, and feature fusion is performed through an attention mechanism.
[0060] Feature extraction and fusion are performed on the preprocessed point cloud information. Feature extraction and fusion are the core steps of point cloud data processing, and the purpose is to extract key features from multi-scale data and improve the expression ability of features through fusion mechanism.
[0061] Specifically, features are extracted based on a multi-scale convolutional neural network. Point cloud data has unstructured characteristics, and traditional CNN cannot be directly processed. The present application constructs multi-scale convolution operation based on the spatial distribution of point cloud.
[0062] A point cloud convolution network DynamicGraphCNN (DGCNN) is used as the basic network, a graph structure is constructed through a local neighborhood, and key features including local features and global features are extracted;
[0063] In the embodiments of the present application, for each point P i , a local feature F local is extracted through a convolution operation:
[0064]
[0065] wherein N(P i ) represents the neighborhood of point P i , w j is a convolution weight, and Conv(·) is a convolution operation.
[0066] A global feature F global is extracted through maximum pooling (MaxPooling):
[0067]
[0068] By adjusting the size of the convolution kernel or the neighborhood range, features of different scales are obtained, for example, using a small-scale convolution kernel to extract detailed features and a large-scale convolution kernel to extract overall structural features.
[0069] Attention mechanism is introduced into feature fusion. The traditional local weighted summation feature fusion method ignores the nonlinear relationship between features. To solve this problem, the attention mechanism is used to dynamically adjust the weights of key features of different scales;
[0070] N-scale key features {F1, F2,..., F N} are extracted by a multi-scale convolutional network.
[0071] For each scale of feature F i , the attention weight is calculated as:
[0072] Attention(F i )=softmax(W a ·F i +b a )
[0073] The attention weights of all scale features are weighted and summed:
[0074]
[0075] Where W a and b a are learnable parameters.
[0076] In an optional embodiment, the key feature extraction can be a random forest-based feature extraction method. The collected point cloud data is preprocessed, and the local neighborhood features (such as normal vector, curvature, etc.) of each point are calculated using a neighborhood search method (such as K-Nearest Neighbor algorithm) as local feature descriptors. The neighborhood size is set to 1.2-1.5 times the average distance of the point cloud; the model is trained by a random forest classifier, and the local features of the sample points are used as input to learn the global features (such as overall shape, distribution characteristics), and the entire point cloud is used as a sample to generate a global feature vector using the label.
[0077] The local feature vector and the global feature vector are fused by simple concatenation or weighted summation to obtain the final feature vector.
[0078] In another optional embodiment, key features can be extracted using an enhanced feature extraction method based on AdaBoost. This method preprocesses the collected point cloud data, uses local point cloud features, and calculates local geometric properties of the points (e.g., point density, normal vector, etc.) in combination with the convolution operation of the present invention. Furthermore, the AdaBoost algorithm is used to train the model on the extracted local features, enhancing the ability to learn global features. The model is updated using multiple rounds of learning, with each round focusing on misclassified sample points.
[0079] The local features and global features are fused by weighted averaging, and the weights can be dynamically adjusted according to the performance of the local features in the classification.
[0080] S3: Based on the attention mechanism, the filtering strategy is adaptively adjusted and the filtering weights are dynamically allocated according to the extracted local features of the point cloud, thereby effectively removing noise without losing key information, performing noise discrimination, and outputting the filtered point cloud data results.
[0081] Furthermore, the point cloud data is represented as P = {p1, p2, ..., p N}, where P i ∈R 3 Contains 3D coordinate information.
[0082] Use DGCNN to extract local features of point cloud F = {f1,f2,...,f N}, where f i ∈ò d is the feature vector of the i-th point, and d is the feature dimension.
[0083] Construct an attention mechanism to map the local features F to the query matrix Q and key matrix K through linear transformation:
[0084] Q=W Q ·F+b Q K=W K ·F+b K
[0085] Q=W Q ·F+b Q K=W K ·F+b K
[0086] Among them, W Q ,W K is the learnable weight matrix, b Q ,b K is the bias vector.
[0087] Traditional single-head attention may not be able to fully capture the complex relationship of the point cloud, so a multi-head attention mechanism is added:
[0088] A multi =Concat(head1,head2,...,head h )·W O
[0089] where each attention head head h independently calculates the weight, W O is the output weight matrix, and h is the number of heads.
[0090] The adaptive adjustment filtering strategy is performed, and according to the attention weight matrix A, the filtering weight w i of each point p i is calculated:
[0091]
[0092] The point cloud data is filtered using the filtering weight to remove the noise points. The filtering formula is as follows:
[0093] P filtered ={p i ∣w i >T}
[0094] where T is a filtering threshold, which can be adjusted according to the task requirements.
[0095] It should also be noted that the point cloud data result after intelligent filtering is output. After completing the feature fusion enhancement and noise discrimination, the system will output the high-precision point cloud data after intelligent filtering. This data result retains the key structures of the power transmission line, such as the conductor, insulator, and tower components. At the same time, it effectively eliminates background noise, such as vegetation, buildings, and irregular point cloud interference, thereby significantly improving the data quality. The strategy based on deep learning and feature fusion ensures the adaptability of the filtering process, so that this method can maintain stable performance in different scenarios, such as mountainous areas, cities, dense vegetation areas, and different resolution point cloud data.
[0096] The application provides a power transmission line inspection point cloud data intelligent filtering method based on deep learning and feature fusion. By introducing multi-scale feature extraction, adaptive feature fusion enhancement and noise discrimination mechanism based on deep learning, the technical problems of poor adaptability and insufficient precision of traditional filtering methods in complex environments are solved. The method provides a more intelligent and robust solution for power transmission line inspection point cloud data denoising. By dynamically optimizing the feature extraction strategy, enhancing the interaction between multi-scale features, and accurately discriminating noise points, the finally output intelligent filtering point cloud data efficiently removes background noise while maximally preserving the fine structure of key components such as ground wires and insulators, providing high-quality and reliable data support for three-dimensional modeling, defect detection and intelligent inspection of power transmission lines, significantly improving the automation level and precision of inspection. Compared with other previous technologies, the application has the advantages shown in Table 1.
[0097] Table 1 Comparison of the application and previous technologies
[0098]
[0099] Embodiment 2, which is different from the previous embodiment, is a second embodiment of the application.
[0100] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the application or the parts that essentially contribute to the prior art or parts of the technical solutions can be embodied in the form of software products, which are stored in a storage medium and include instructions to make a computer device (which can be a personal computer, a server, or a network device) execute all or part of the steps of the methods described in the embodiments of the application. The aforementioned storage medium includes U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, and various program code storage media.
[0101] The logic and / or steps represented in flow diagrams or otherwise described herein, for example, can be considered as a sequence of executable instructions, and can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions, or a combination of the above. For the purposes of this specification, a "computer-readable medium" can be any apparatus that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.
[0102] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection having one or more wires (electrical connections), a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can also be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example, via optical scanning of the paper or other medium, then compiled, interpreted, or otherwise processed in a suitable manner, if necessary, and stored in a computer memory.
[0103] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the above embodiments, the various steps or methods can be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, and in another embodiment, any of the following technologies can be used in the implementation: discrete logic circuitry having logic gates for implementing logic functions upon data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and so forth, as well as technologies and techniques that are well known and appreciated in the art.
[0104] Embodiment 3, as a third embodiment of the present application, provides a power line inspection point cloud data filtering system based on deep learning and feature fusion, including a point cloud data processing module, a feature extraction module, and a filtering module;
[0105] The point cloud data processing module pre-processes the collected power line inspection point cloud data;
[0106] The feature extraction module extracts key features including local features and global features from the pre-processed multi-scale point cloud data, and performs feature fusion through an attention mechanism;
[0107] The filtering module dynamically allocates filtering weights according to the extracted local features of the point cloud by adaptively adjusting the filtering strategy based on the attention mechanism, discriminates noise, and outputs filtered point cloud data results.
[0108] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and they should be covered in the scope of the claims of the present application.
Claims
1. A transmission line inspection point cloud data filtering method based on deep learning and feature fusion, characterized by: include, Pre-process the collected transmission line inspection point cloud data; Extract key features including local features and global features from preprocessed multi-scale point cloud data, and perform feature fusion through the attention mechanism; Based on the attention mechanism, the filtering strategy is adaptively adjusted, the filtering weights are dynamically allocated according to the extracted local features of the point cloud, noise discrimination is performed, and the filtered point cloud data results are output.
2. The method for filtering transmission line inspection point cloud data based on deep learning and feature fusion according to claim 1, characterized in that: The pretreatment includes a first treatment, a second treatment and a third treatment; The first processing uses a clustering algorithm to perform density clustering on the signal, and dynamically identifies and preliminarily removes noise points based on the local density distribution of the point cloud.
3. The method for filtering transmission line inspection point cloud data based on deep learning and feature fusion according to claim 2, characterized in that: The second processing includes normalizing the data subjected to the preliminary denoising, and calculating the weight of the local area by considering the adjustment density and curvature features during the normalization process to obtain the features after preliminary cleaning.
4. The method for filtering transmission line inspection point cloud data based on deep learning and feature fusion according to claim 3, characterized in that: The third processing includes smoothing the transmission line inspection point cloud data, using an improved Gaussian weighted smoothing filter method, and dynamically adjusting the weight according to the distance between the point and the neighborhood center.
5. The method for filtering transmission line inspection point cloud data based on deep learning and feature fusion according to claim 4, characterized in that: Extracting key features includes using a point cloud convolutional network as a basic network, constructing a graph structure through local neighborhoods, and extracting local features and global features; For each point, local features are extracted through convolution operations, and global features are extracted through maximum pooling; By adjusting the convolution kernel and neighborhood range, key features of different scales are obtained.
6. The method for filtering transmission line inspection point cloud data based on deep learning and feature fusion according to claim 5, characterized in that: The feature fusion includes dynamically adjusting the weights of key features of different scales through an attention mechanism; Extract key features of two or more scales through multi-scale convolutional networks; Calculate attention weights for features at each scale; The attention weights of all scale features are weighted summed.
7. The method for filtering transmission line inspection point cloud data based on deep learning and feature fusion according to claim 6, characterized in that: The adaptive adjustment filtering strategy includes extracting local features of the point cloud using a point cloud convolutional network; Construct an attention mechanism to map local features to query and key matrices through linear transformations; Add a multi-head attention mechanism to adaptively adjust the filtering strategy and calculate the filtering weight for each point based on the attention weight matrix; Use filter weights to filter the point cloud data and remove noise points.
8. A power transmission line inspection point cloud data filtering system based on deep learning and feature fusion, applying the power transmission line inspection point cloud data filtering method based on deep learning and feature fusion as described in any one of claims 1 to 7, characterized in that: include: Point cloud data processing module, feature extraction module, filtering module; The point cloud data processing module pre-processes the collected transmission line inspection point cloud data; The feature extraction module extracts key features including local features and global features from the preprocessed multi-scale point cloud data, and performs feature fusion through the attention mechanism; The filtering module, based on the attention mechanism, adaptively adjusts the filtering strategy, dynamically allocates filtering weights according to the extracted local features of the point cloud, performs noise discrimination, and outputs the filtered point cloud data results.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the transmission line inspection point cloud data filtering method based on deep learning and feature fusion according to any one of claims 1 to 7 are implemented.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the transmission line inspection point cloud data filtering method based on deep learning and feature fusion according to any one of claims 1 to 7 are implemented.