A point cloud automatic recognition method for power UAV inspection of pole towers

Abstract

The present invention provides a method for automatically identifying point clouds of electric power unmanned aerial vehicle inspection towers, including: pre-processing the point cloud data of electric power lines collected by laser radar equipment, screening and clustering the pre-processed point cloud data to obtain point cloud clusters, and extracting the point cloud clusters of the towers from the point cloud clusters based on the difference characteristics between the target inspection tower and the environment; calculating the key points and characteristic values ​​in the tower point cloud, combining the tower types in the power inspection scene, and screening out the typical towers of each tower type based on the key point distribution and noise content of the tower point cloud, and constructing a device feature type library through the characteristic values ​​of the typical towers; calculating the key points and characteristic values ​​in the point cloud to be processed, matching with the typical towers in the device feature type library, and selecting the type with the highest matching degree as the recognition result. The present invention reduces the labor cost of point cloud processing and equipment identification, and improves the efficiency of unmanned aerial vehicle inspection operations.

Description

Technical Field

[0001] The invention relates to the field of point cloud data recognition, and in particular to an automatic point cloud recognition method for power unmanned aerial vehicle inspection of pole towers. Background Art

[0002] With the advancement of drone technology and the maturity of intelligent perception systems, revolutionary changes have been brought to equipment operation and maintenance and safety assurance in all walks of life. In the equipment inspection scene of the power industry, due to the complexity and changeability of the inspection environment and the limitation of traditional ground inspection methods on geographical conditions and human resources, it is often not efficient and meticulous, and there are many safety hazards in high-altitude operations. As a new inspection mode constructed by autonomous aircraft equipped with multiple sensors and data transmission technology, drone inspection can flexibly adapt to the complex and changeable geographical environment, overcome the environmental restrictions of special areas such as mountainous areas and line corridors, and realize more efficient and safer three-dimensional inspection. For the task of drone power inspection, laser point cloud data is an important basis for spatial information perception and inspection trajectory planning. However, in addition to the target equipment of the inspection task, the original point cloud data also contains a large amount of environmental information and noise. The traditional method of manually identifying inspection equipment in the point cloud space also faces the problem of low processing efficiency as the amount of point cloud data increases. Therefore, designing a point cloud automatic recognition method that can realize the inspection of power drone towers is of great significance to enhance the digitalization and intelligence level of power inspection.

[0003] The existing work mainly has the following problems: First, the existing point cloud preprocessing methods are often designed based on general scenarios, and the particularity of the tower structure in the power inspection scenario is not considered enough. For example, there may be a large number of interference factors such as vegetation and buildings around the tower, resulting in a large number of noisy point clouds. Conventional filtering methods may not be able to effectively remove these interferences, affecting the subsequent tower recognition accuracy; Second, the existing tower extraction or tower segmentation methods are highly dependent on label data. When training the model, it is often necessary to manually annotate the point cloud data and assign the category label of the point to all points, which consumes too much labor cost; Third, the versatility and adaptability of the existing tower type recognition methods need to be improved. Due to the complexity and diversity of tower types in actual power lines, including straight towers, corner towers, tension towers and other forms, and there may be differentiated designs or local structural differences, the recognition effect of conventional tower recognition methods is easily affected; Finally, the existing point cloud processing and tower type recognition methods lack the effective use of the tower position information obtained during the point cloud acquisition process to achieve more accurate point cloud segmentation in the point cloud preprocessing stage to assist in the subsequent equipment type recognition.

[0004] The problem that the existing technology urgently needs to solve is, under the condition that the point cloud collection task of the UAV inspection environment has been completed, how to quickly and effectively process the point cloud data and identify the power equipment that needs to be inspected to support the UAV power inspection operation. Summary of the invention

[0005] In view of the main defects of the traditional manual methods and existing point cloud recognition methods in the power inspection environment, the present invention designs an automatic point cloud recognition method for power tower inspection by unmanned aerial vehicles, which reduces the labor cost of point cloud processing and equipment identification and improves the efficiency of unmanned aerial vehicle inspection operations.

[0006] A method for automatically identifying point clouds of power towers inspected by electric drones comprises the following steps:

[0007] Step (1), preprocessing the power line point cloud data collected by the laser radar equipment, including data clipping, point cloud denoising, and point cloud downsampling, to obtain preprocessed point cloud data;

[0008] Step (2), extracting the target inspection equipment from the preprocessed point cloud data, obtaining the target point cloud of the target inspection equipment, comparing the height value of each point in the point cloud with the corresponding threshold value set based on the height range of the actual tower and filtering, clustering the target point cloud based on the spatial distance, calculating the discretely distributed point cloud clusters in the point cloud space, and then calculating the average value of the color in the point cloud cluster and comparing it with the color threshold value set based on the color range of the actual tower and filtering, finally calculating the point cloud density and the area and length of the horizontal projection of each point cloud cluster to obtain the calculation results of the above three types of features, taking the weighted value of the calculation result as a measurement indicator, and extracting the point cloud cluster of the tower;

[0009] Step (3), from the point cloud cluster of the tower extracted in step (2), select the significant points whose local structure meets the feature requirements as the key points of the tower point cloud, and analyze the geometric characteristics of the local surface of the point cloud as the feature value of the tower point cloud based on the relative position and surface normal direction between the key points and their neighboring points;

[0010] Step (4), using the tower point cloud feature values ​​calculated in step (3), combined with the tower types in the power inspection scene, and based on the key point distribution and noise content of the tower point cloud, screen out typical towers of each tower type, and construct a device feature type library through the feature values ​​of the typical towers;

[0011] Step (5), calculate the key points and feature values ​​in the point cloud to be processed, match them with the typical towers in the equipment feature type library, and select the type with the highest matching degree as the recognition result.

[0012] Furthermore, the step (1) of preprocessing the point cloud data specifically includes:

[0013] 1) Data clipping: Using the coordinate data of the power tower collected during the point cloud scanning of the power line, the original point cloud data of the drone inspection line is clipped to obtain segmented inspection routes or single equipment inspection routes, and the clipping method is selected according to the route application scenario;

[0014] 2) Point cloud denoising: Using the spatial distribution of the point cloud, outliers in the spatial dimension are considered as noise points and filtered out;

[0015] 3) Point cloud downsampling: The point cloud data is gridded. Each grid after gridding is called a voxel. In each voxel, a point is selected to replace all the points in the original grid to achieve downsampling of the point cloud data.

[0016] Furthermore, the step (2) specifically includes:

[0017] 1) Height filtering: Using the height attribute of the tower relative to the environment, the minimum Z-axis coordinate value of all points in the pre-processed point cloud data plus 5, that is, the lowest height in the environment plus 5 meters is used as the height restriction condition, and the point cloud clusters and point clouds below the height restriction are filtered;

[0018] 2) Clustering based on point cloud density: Taking advantage of the fact that there is a certain spatial distance between different objects in space, the points in the point cloud data are clustered according to the spatial distance and divided into point cloud clusters of different unknown types and discrete distributions;

[0019] 3) Color filtering: Based on the color attributes of the tower, the RGB color filtering threshold is fixed to [180, 180, 180], the RGB color average of each point cloud cluster is calculated, and the point cloud clusters with any color value in the color average value lower than the corresponding tower color threshold are filtered out;

[0020] 4) Extract tower point cloud: Extract the tower point cloud based on the differential features between the tower and other objects, including point cloud density, area and length of horizontal plane projection. Calculate the above three types of features for all point cloud clusters and normalize the results to obtain a ternary array consisting of three numbers in the [0,1] interval. Then perform weighted calculation, square the point cloud density and add the length of the horizontal plane projection divided by the area of ​​the horizontal plane projection to obtain the weighted value. The point cloud cluster with the largest weighted value is retained as the tower extraction result.

[0021] Furthermore, the step (3) specifically includes:

[0022] 1) Key point extraction: Based on the internal morphological descriptor, the local geometric features of the point cloud are analyzed. The local geometric features of the point cloud include the curvature and neighborhood structure of the point. The shape features and directionality are used to identify the salient points in the local structure:

[0023] First, a neighborhood analysis is performed. For each point p in the point cloud P i Establish a local coordinate system, define the neighborhood by setting the search radius r, and calculate the neighborhood midpoint p j The weight w ij , whose expression is:

[0024] w ij =1 / |p i -p j |,|p i -p j | <r

[0025] Next, calculate each point p i The covariance matrix cov(p i ), whose expression is:

[0026]

[0027] Among them, T represents the transpose operation;

[0028] Then calculate each point p i The covariance matrix cov(p i ) The three eigenvalues ​​correspond to the three axes of the local coordinate system of the point cloud. The size of the eigenvalues ​​can reflect the degree of anisotropy and diffusion of the distribution of neighborhood points.

[0029] Finally, the eigenvalue ratio thresholds ε1 and ε2 are set to screen out the key points of the tower point cloud that meet the conditions, which serve as the basis for subsequent eigenvalue calculation and tower type identification:

[0030]

[0031] 3) Eigenvalue calculation: Based on the selected key points of the tower point cloud, the eigenvalue of the point feature histogram FPFH is calculated for the neighborhood of each key point through the local feature descriptor, and a histogram for the eigenvalue is established:

[0032] First, for each key point P in the point cloud s Set the neighborhood with radius r, point P t is a point in the field, key point P s The expressions of the three axes of the local coordinate system are as follows:

[0033] u=n s

[0034] v=(P t -P s )×u

[0035] Among them, n s P sThe normal vector of a point represents the surface normal direction of the point, P t -P s is the vector between two points, representing the relative position of the two points in the neighborhood;

[0036] Then based on point P s The local coordinate system of P t Point and P s The triplet feature {α, φ, θ} between points, also known as the simplified point feature histogram SPFH, is calculated as follows:

[0037]

[0038] Where d is P t Point and P s Euclidean distance between points;

[0039] After calculating the key point P s After the triple features of the total number of neighboring points in the neighborhood of k, calculate each neighboring point P k The triplet features in its neighborhood are represented by point P s and point P k The Euclidean distance between them is used as the weight w k , and finally combine the triplets to get the point feature histogram FPFH:

[0040]

[0041] Furthermore, the step (4) specifically includes:

[0042] 1) Tower type definition: According to the power scenarios of different voltage levels to which the point cloud data belongs, combined with the relevant regulations for the operation of professional power equipment and the needs of drone inspection work, all the inspection tower types that need to be identified are determined;

[0043] 2) Construct a pole tower type feature library: Based on the determined pole tower type, calculate the average spatial coordinates of all the key points of the pole tower point cloud, and calculate the distance from the center point of the pole tower point cloud on the horizontal plane projection; perform noise analysis on the pole tower point cloud, screen out drift points, isolated points, and redundant points as noise points based on the distance between adjacent points, and calculate the noise content in the point cloud. Based on the above calculation results, the distribution of key points and the characteristic values ​​of the noise content of the pole tower point cloud are normalized and sorted from small to large. The pole tower with uniform key point distribution and low noise content is selected as a typical example of this type, and the FPFH eigenvalue is calculated to construct a pole tower type feature library.

[0044] Furthermore, the step (5) specifically includes:

[0045] 1) Similarity calculation: extract the key points of the unknown type of tower point cloud and calculate the eigenvalues, and perform similarity calculations with the eigenvalues ​​of each type of typical tower in the type library one by one;

[0046] 2) Type matching and result output: Based on the similarity calculation results, the tower type with the highest matching degree is output as the final result of equipment type identification. When the similarity calculation result is lower than the threshold, a matching error message is output.

[0047] The present invention aims to identify the pole tower types in the power line point cloud, realizes the preprocessing of the point cloud, pole tower feature analysis and type identification, reduces the manpower cost of UAV route planning, and can be widely used in point cloud processing and route planning scenarios in various power UAV inspection environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flow chart of a method for automatically identifying point clouds of power towers inspected by an electric power drone according to an embodiment of the present invention;

[0049] Figure 2 is the pseudo code of the algorithm of the embodiment of the present invention;

[0050] Figure 3 is a schematic diagram of point cloud clustering visualized in an embodiment of the present invention;

[0051] Figure 4 It is a schematic diagram of a point cloud of an inspection device extracted by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] See also Figure 1 The embodiment of the present invention provides a method for automatically identifying point clouds of power towers inspected by a power drone, comprising the following steps:

[0054] Step (1), point cloud data preprocessing: read the power line point cloud data and equipment location information collected by the laser radar equipment in the specified location file, the equipment location information mainly includes the equipment number, equipment name, longitude, latitude and altitude; then carry out point cloud data cropping, and divide the point cloud into single-level towers according to the tower location information, so that the point cloud only contains the tower and the environmental information in the close range of the tower; then perform noise reduction on the point cloud data to remove the discrete points distributed in the point cloud space; finally, perform down-sampling on the point cloud, and use the voxelization method to reduce the number of points, reduce the scale of the point cloud data, and reduce the computational complexity of subsequent processing.

[0055] Step (2), target device extraction: the pre-processed point cloud data is filtered based on the tower attributes. Based on the height attribute range of the tower, a height threshold of the minimum Z-axis coordinate value in the point cloud plus five is set to filter out the point clouds below the height threshold. For the remaining point clouds that may contain devices after filtering, the spatial distance characteristics between different objects in space are used to cluster the points in the point cloud data based on the distance between the points, and the points are divided into different unknown types and discretely distributed point cloud clusters, such as Figure 3 As shown in the figure; based on the color attribute range of the tower, the RGB color threshold of [180,180,180] is set to filter out any point cloud cluster whose color value of the average color in the point cloud cluster is lower than the color threshold; then the point cloud cluster is further screened, firstly, the point cloud density, the area and length of the horizontal plane projection are calculated for all point cloud clusters, and the calculation results of the three types of features are normalized to obtain a ternary array consisting of three numbers in the interval [0,1], and then a weighted calculation is performed, the square of the point cloud density is added to the length of the horizontal plane projection divided by the area of ​​the horizontal plane projection to obtain the weighted value, and the point cloud cluster with the largest weighted value is retained as the tower extraction result, and the following is obtained: Figure 4 Point cloud cluster of the target tower is shown.

[0056] Step (3), filter the key points of the point cloud and calculate the characteristic values ​​of the tower point cloud

[0057] Screening point cloud key points: the implementation steps are as follows Figure 2 As shown in the pseudo code of point cloud key point extraction, firstly, neighborhood analysis is performed. For each point p in the target device point cloud P i Establish a local coordinate system, define the neighborhood by setting the search radius r, and calculate the neighborhood midpoint p j The weight w ij , whose expression is:

[0058] w ij =1 / |p i -p j |,|p i -p j | <r

[0059] Next, calculate each point p i The covariance matrix cov(p i ), whose expression is:

[0060]

[0061] Where T represents the transpose operation. Then calculate each point p i The covariance matrix cov(p i ) The three eigenvalues ​​correspond to the three axes of the local coordinate system of the point cloud. The size of the eigenvalue can reflect the degree of anisotropy and diffusion of the distribution of neighborhood points. Finally, the eigenvalue ratio thresholds ε1 and ε2 are set to filter out key points that meet the conditions. The expression is:

[0062]

[0063] Eigenvalue calculation: Based on the key points of the point cloud selected in the previous step, the eigenvalues ​​of the neighborhood of each key point are calculated through the local feature descriptor, and a histogram of the eigenvalues ​​is established. First, for each key point P in the point cloud s Set the neighborhood with radius r, point P t is a point in the field, key point P s The expressions of the three axes of the local coordinate system are as follows:

[0064] u=n s

[0065] v=(P t -P s )×u

[0066] Among them, n s P s The normal vector of a point represents the surface normal direction of that point, P t -P s is the vector between two points, representing the relative position of the two points in the neighborhood. s The local coordinate system of P t Point and P s The triplet feature {α, φ, θ} between points, also known as the simplified point feature histogram SPFH, is calculated as follows:

[0067]

[0068] Where d is P t Point and P s The Euclidean distance between points. After calculating the key point P s After the triple features of the total number of neighboring points in the neighborhood of k, calculate each neighboring point P kThe triplet features in its neighborhood are represented by point P s and point P k The Euclidean distance between them is used as the weight w k ,Finally, the triplets are combined to obtain the point feature histogram FPFH.

[0069]

[0070] Step (4), constructing a device type feature library: The goal of the present invention is to realize device type identification on the basis of completing the extraction of device point cloud clusters, calculation of pole tower key points and characteristic values. The pole tower types in the power inspection scene are known. According to the power scenes of different voltage levels to which the point cloud data belongs, combined with the relevant regulations for the operation of power professional equipment and the needs of drone inspection work, all the types of inspection pole towers that need to be identified are determined. At the same time, a typical type of pole tower feature library can be constructed for matching and type judgment. For different types of overhead line pole towers in the power inspection scene, such as cat head towers, wine glass towers, dry word towers and other tower types in the transmission scene, typical pole towers are screened and the pole tower key points are calculated. Then, the average spatial coordinates of all the tower point cloud key points are calculated, and the distance from the center point of the tower point cloud on the horizontal plane projection is calculated; the tower point cloud is subjected to noise analysis, and based on the distance between adjacent points, drift points, isolated points, and redundant points are screened as noise points, and the noise content in the point cloud is calculated. Based on the above calculation results, the key point distribution and the characteristic values ​​of the noise content of the tower point cloud are normalized as the measurement standard, and the towers with uniform key point distribution and low noise content are selected as typical examples of this type. Then, the characteristic values ​​of the typical sample towers are calculated and added to the characteristic library of this type of towers.

[0071] Step (5), equipment type identification: perform the calculation of key points and eigenvalues ​​in step (3) on the point cloud data of the unknown pole tower to be identified, extract the key points and calculate the eigenvalues, then match the calculation results with the pole tower feature library, and select the pole tower type with the largest eigenvalue similarity as the type identification result among all the typical pole tower types. If the similarity of all typical pole tower types is less than 50%, it is determined that the pole tower is a pole tower with a special structural design and cannot be accurately identified. Relevant error information is prompted and further manual judgment is required.

[0072] In the task of power inspection, the refined inspection of equipment by drones has gradually been promoted and applied, but drone patrols rely on planned and designed routes. At present, route planning relies on relevant software and pre-collected line point cloud data to manually identify inspection equipment in the point cloud and manually mark the drone route. Once the amount of point cloud data to be processed increases to a certain extent, the cost and complexity of manual point cloud identification will increase significantly. At this time, the manual method is obviously not a suitable method. At the same time, point cloud recognition is a repetitive process of continuously screening point cloud clusters. This process conforms to the working characteristics of computers, so designing a point cloud automatic recognition method is a better choice.

[0073] The present invention first performs a preprocessing operation on the point cloud, removes the noise information in the point cloud data, and performs voxel downsampling on the point cloud data, while ensuring that the structural features are not destroyed, the density of the point cloud is reduced to reduce the subsequent recognition burden and speed up the calculation. Then, we designed a tower point cloud extraction algorithm based on spatial distance clustering and tower attribute analysis, and used a feature analysis method based on point cloud key points and eigenvalues ​​to calculate the tower point cloud features without manual labeling. Finally, by screening typical tower samples for each tower type in the power inspection scene, and using the eigenvalues ​​of typical towers to build a tower type feature library, a tower type recognition method based on type feature library matching is implemented, which is universal between power inspection scenes within a similar voltage range and adaptable to some special structure towers.

[0074] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by technicians in this technical field within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A point cloud automatic recognition method for electric power drone inspection tower, characterized in that: The following steps are involved: Step (1), preprocessing the power line point cloud data collected by the laser radar equipment, including data clipping, point cloud denoising, and point cloud downsampling, to obtain preprocessed point cloud data; Step (2), extracting the target inspection equipment from the preprocessed point cloud data, obtaining the target point cloud of the target inspection equipment, comparing the height value of each point in the point cloud with the corresponding threshold value set based on the height range of the actual tower and filtering, clustering the target point cloud based on the spatial distance, calculating the discretely distributed point cloud clusters in the point cloud space, and then calculating the average value of the color in the point cloud cluster and comparing it with the color threshold value set based on the color range of the actual tower and filtering, finally calculating the point cloud density and the area and length of the horizontal projection of each point cloud cluster to obtain the calculation results of the above three types of features, taking the weighted value of the calculation result as a measurement indicator, and extracting the point cloud cluster of the tower; Step (3), from the point cloud cluster of the tower extracted in step (2), select the significant points whose local structure meets the feature requirements as the key points of the tower point cloud, and analyze the geometric characteristics of the local surface of the point cloud as the feature value of the tower point cloud based on the relative position and surface normal direction between the key points and their neighboring points; Step (4), using the tower point cloud feature values ​​calculated in step (3), combined with the tower types in the power inspection scene, and based on the key point distribution and noise content of the tower point cloud, screen out typical towers of each tower type, and construct a device feature type library through the feature values ​​of the typical towers; Step (5), calculating the key points and feature values ​​in the point cloud to be processed, matching them with the typical pole towers in the equipment feature type library, and selecting the type with the highest matching degree as the recognition result; The step (3) specifically comprises: 1) Key point extraction: Based on the internal morphological descriptor, the local geometric features of the point cloud are analyzed. The local geometric features of the point cloud include the curvature and neighborhood structure of the point. The shape features and directionality are used to identify the salient points in the local structure: First, a neighborhood analysis is performed. For each point p in the point cloud P i Establish a local coordinate system, define the neighborhood by setting the search radius r, and calculate the neighborhood midpoint p j The weight w ij , whose expression is: w ij =1 / |p i -p j |,|p i -p j |<r Next, calculate each point p i The covariance matrix cov(p i ), whose expression is: Among them, T represents the transpose operation; Then calculate each point p i The covariance matrix cov(p i ) The three eigenvalues ​​correspond to the three axes of the local coordinate system of the point cloud. The size of the eigenvalues ​​can reflect the degree of anisotropy and diffusion of the distribution of neighborhood points. Finally, the eigenvalue ratio thresholds ε1 and ε2 are set to screen out the key points of the tower point cloud that meet the conditions, which serve as the basis for subsequent eigenvalue calculation and tower type identification: 2) Eigenvalue calculation: Based on the selected key points of the tower point cloud, the eigenvalue of the point feature histogram FPFH is calculated for the neighborhood of each key point through the local feature descriptor, and a histogram for the eigenvalue is established: First, for each key point P in the point cloud s Set the neighborhood with radius r, point P t is a point in the field, key point P s The expressions of the three axes of the local coordinate system are as follows: u=n s v=(P t -P s )×u Among them, n s P s The normal vector of a point represents the surface normal direction of that point, P t -P s is the vector between two points, representing the relative position of the two points in the neighborhood; Then based on point P s The local coordinate system of P t Point and P s The triplet feature {α, φ, θ} between points, also known as the simplified point feature histogram SPFH, is calculated as follows: Where d is P t Point and P s Euclidean distance between points; After calculating the key point P s After the triple features of the total number of neighboring points in the neighborhood of k, calculate each neighboring point P k The triplet features in its neighborhood are represented by point P s and point P k The Euclidean distance between them is used as the weight w k , and finally combine the triplets to get the point feature histogram FPFH:

2. According to the method for automatic point cloud recognition of power tower inspection by electric drone in claim 1, it is characterized in that: The step (1) of preprocessing the point cloud data specifically includes: 1) Data clipping: Using the coordinate data of the power tower collected during the point cloud scanning of the power line, the original point cloud data of the drone inspection line is clipped to obtain segmented inspection routes or single equipment inspection routes, and the clipping method is selected according to the route application scenario; 2) Point cloud denoising: Using the spatial distribution of the point cloud, outliers in the spatial dimension are considered as noise points and filtered out; 3) Point cloud downsampling: The point cloud data is gridded. Each grid after gridding is called a voxel. In each voxel, a point is selected to replace all the points in the original grid to achieve downsampling of the point cloud data.

3. According to the method for automatic point cloud recognition of power tower inspection by electric drone in claim 1, it is characterized in that: The step (2) specifically comprises: 1) Height filtering: Using the height attribute of the tower relative to the environment, the minimum Z-axis coordinate value of all points in the pre-processed point cloud data plus 5, that is, the lowest height in the environment plus 5 meters is used as the height restriction condition, and the point cloud below the height limit is filtered; 2) Clustering based on point cloud density: Taking advantage of the fact that there is a certain spatial distance between different objects in space, the points in the point cloud data are clustered according to the spatial distance and divided into point cloud clusters of different unknown types and discrete distributions; 3) Color filtering: Based on the color attributes of the tower, the RGB color filtering threshold is fixed to [180, 180, 180], the RGB color average of each point cloud cluster is calculated, and the point cloud clusters with any color value in the color average value lower than the color threshold are filtered out; 4) Extract tower point cloud: Extract the tower point cloud based on the differential features between the tower and other objects, including point cloud density, area and length of horizontal plane projection. Calculate the above three types of features for all point cloud clusters and normalize the results to obtain a ternary array consisting of three numbers in the [0,1] interval. Then perform weighted calculation, square the point cloud density and add the length of the horizontal plane projection divided by the area of ​​the horizontal plane projection to obtain the weighted value. The point cloud cluster with the largest weighted value is retained as the tower extraction result.

4. According to the method for automatic point cloud recognition of power tower inspection by electric drone in claim 1, it is characterized in that: The step (4) specifically comprises: 1) Tower type definition: According to the power scenarios of different voltage levels to which the point cloud data belongs, combined with the relevant regulations for the operation of professional power equipment and the needs of drone inspection work, all the inspection tower types that need to be identified are determined; 2) Construct a pole tower type feature library: Based on the determined pole tower type, calculate the average spatial coordinates of all the key points of the pole tower point cloud, and calculate the distance from the center point of the pole tower point cloud on the horizontal plane projection; perform noise analysis on the pole tower point cloud, screen out drift points, isolated points, and redundant points as noise points based on the distance between adjacent points, and calculate the noise content in the point cloud. Based on the above calculation results, the distribution of key points and the characteristic values ​​of the noise content of the pole tower point cloud are normalized and sorted from small to large. The pole tower with uniform key point distribution and low noise content is selected as a typical example of this type, and the FPFH eigenvalue is calculated to construct a pole tower type feature library.

5. According to the method for automatic point cloud recognition of power tower inspection by electric drone in claim 1, it is characterized in that: The step (5) specifically comprises: 1) Similarity calculation: extract the key points of the unknown type of tower point cloud and calculate the eigenvalues, and perform similarity calculations with the eigenvalues ​​of each type of typical tower in the type library one by one; 2) Type matching and result output: Based on the similarity calculation results, the tower type with the highest matching degree is output as the final result of equipment type identification. When the similarity calculation result is lower than the threshold, a matching error message is output.

Images