An unmanned aerial vehicle inspection track automatic generation method based on power tower point cloud

Abstract

A kind of unmanned aerial vehicle inspection trajectory automatic generation method based on electric power tower point cloud, comprising: obtaining the space coordinates of the top end of target inspection tower;Extract the edge point cloud of the region within a certain range of space coordinates, use DBSCAN clustering algorithm to obtain point cloud cluster;Based on PointNet algorithm, each point cloud cluster is identified and classified, and point cloud with power line features is screened out;Apply the corrosion algorithm, with the outermost point away from the tower on the power line point cloud as the starting point, corrosion along the line direction, identify and extract insulator from line bifurcation;According to the shooting rule of unmanned aerial vehicle inspection and the inclination angle of tower connected power line, generate the point to be shot on the device, calculate the corresponding unmanned aerial vehicle flight path point;Generate and export unmanned aerial vehicle inspection route file.The present application reduces the labor cost of unmanned aerial vehicle route planning, and is widely used in point cloud processing of various electric power tower equipment, unmanned aerial vehicle inspection route planning scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of unmanned aerial vehicle power inspection, and particularly relates to a method for automatically generating an unmanned aerial vehicle inspection track based on power tower point clouds. BACKGROUND

[0002] With the rapid development of unmanned aerial vehicle technology and the increasing maturity of intelligent sensing systems, revolutionary changes have been brought to the equipment operation and safety protection of various industries. In the equipment inspection scene of the power industry, due to the complex and changeable inspection environment and the limitations of traditional ground inspection methods in geographical conditions and human resources, it is often difficult to achieve efficient and detailed inspection, and there are many safety hazards in climbing operations. As a new type of inspection mode constructed by a self-flying vehicle carrying multiple sensors and data transmission technology, unmanned aerial vehicle inspection can flexibly adapt to complex and changeable geographical environments, overcome environmental restrictions in special areas such as mountainous areas, line corridors, and achieve more efficient and safer three-dimensional inspection.

[0003] The existing work mainly has the following problems: first, in the power inspection operation, the processing of point cloud data still highly depends on manual operation, which needs to manually identify the equipment in the point cloud and set the aerial photography points and inspection track of the unmanned aerial vehicle according to the relevant technical specifications. Manual participation has a heavy workload, which not only affects the accuracy of the setting of the aerial photography points and the safety of the track, but also restricts the data processing efficiency and hinders the popularization and application of unmanned aerial vehicle inspection; second, the route matching function of the mainstream unmanned aerial vehicle route planning software is not intelligent enough, and the type of the tower needs to be judged manually when using it, and it often relies on the construction of a large-scale tower route database. After the matching is completed, a large amount of manual correction is still needed to obtain a suitable route; third, the generality and adaptability of the existing power tower inspection route planning method need to be improved. Due to the complex and diverse types of towers and equipment in the actual power lines, and the possibility of differential design or local structural differences, the effect of the conventional tower matching method is affected.

[0004] The existing technology needs to solve the problem that, under the condition that the collection task of the environment point cloud required by the unmanned aerial vehicle inspection has been completed, how to automatically process the point cloud data intelligently by the computer system to quickly and effectively plan the aerial photography points and the inspection track of the unmanned aerial vehicle, so as to support the unmanned aerial vehicle power inspection operation. SUMMARY

[0005] In view of the main problems existing in the traditional manual route planning method and the existing route planning software, the present application designs a method for automatically generating an unmanned aerial vehicle inspection track based on power tower point clouds, which reduces the labor cost of unmanned aerial vehicle inspection track planning operation and improves the efficiency of unmanned aerial vehicle inspection operation.

[0006] A method for automatically generating an unmanned aerial vehicle inspection track based on power tower point clouds, comprising the following steps:

[0007] Step (1), the point cloud data of the power tower and line collected by the laser radar device is visualized, the inspection tower coordinates or the manually marked tower top position to be inspected is input, and the spatial coordinates of the target inspection tower top are obtained; a region with a given length around the spatial coordinates is intercepted, and the edge point cloud of the region is obtained; the DBSCAN algorithm is used to cluster the edge point cloud, and the point cloud cluster is obtained. The PointNet classification algorithm is applied to determine whether the type of each point cloud cluster is a power line.

[0008] Step (2), using the point cloud cluster determined as a power line category in step (1), applying an erosion algorithm, selecting the endpoint of the power line away from the tower in the point cloud cluster as the starting point, and eroding each line along the line towards the tower direction, and determining whether there is a line bifurcation generated by the connection of the insulator during the erosion process. Based on the physical structure difference between the insulator and the line and whether it is directly connected to the tower, the point cloud cluster is classified to distinguish the insulator and the line.

[0009] Step (3), for each point cloud cluster identified as an insulator in step (2), generating camera shooting points on the device point cloud cluster and corresponding path points in the unmanned aerial vehicle flight trajectory according to the device photo shooting rules of the power unmanned aerial vehicle inspection, and adjusting the path point position after calculating the tower pull line inclination angle; according to the device photo shooting rules of the power unmanned aerial vehicle inspection, supplementing the remaining camera shooting points and corresponding path points in the unmanned aerial vehicle flight trajectory in the unmanned aerial vehicle inspection process except for the insulator device, sequentially connecting each path point by using ACO and RRT* path planning algorithm, and obtaining the optimal unmanned aerial vehicle inspection trajectory.

[0010] Further, step (1) obtains the spatial coordinates of the tower top in the point cloud data, which specifically includes:

[0011] 1) Import tower top coordinate data: import the longitude and latitude coordinates of the power tower to be inspected and the tower top height data from the point cloud data;

[0012] 2) Manually mark the tower top: visualize the point cloud data, and manually click the tower top position in the visualized point cloud when the imported data is not accurate or lacks coordinate data, to correct or obtain the spatial coordinates of the tower top.

[0013] Further, step (1) uses the DBSCAN algorithm to cluster the edge point cloud to obtain the point cloud cluster, and applies the PointNet classification algorithm to determine whether the type of each point cloud cluster is a power line, which specifically includes:

[0014] The DBSCAN clustering algorithm is used to group the collected blue point set, to ensure that the points in each cluster have similar spatial distribution characteristics. The DBSCAN algorithm assigns a unique label to each point in the point cloud data set. Initially, all points are not visited. For each point P in the point cloud data set, the number of points in its neighborhood ε is calculated:

[0015] N ε (P)={Q∈D|dist(P,Q)≤ε}

[0016] If |N∈(P)|≥MinPts, then P is a core point, and it is added to the core point set. An unvisited core point is selected as a seed point to start a new clustering process. For the current seed point P, all density-reachable points are found, that is, there is a point sequence P1,P2,…,P n , where P1=P, P n =Q. All density-reachable points are added to the current cluster, and the labels of these points are updated to the identifier of the current cluster. For each point in the current cluster, if it has not been visited and is a core point, it is used as a new seed point to continue expanding the cluster. The above process is repeated until all points in the current cluster are visited.

[0017] In each cluster, three representative points are randomly selected, which will serve as the basis for subsequent processing. Then, an erosion algorithm is used to process these points, further generating three data sets each containing 1024 points.

[0018] Finally, the three data sets each containing 1024 points are input into the PointNet algorithm for classification. The PointNet algorithm is used to make three independent judgments for each cluster. If two or more of the judgment results are power lines, the cluster is finally determined to be a power line.

[0019] When the classification result continues to output as a tower, the classification of the upper half of the tower and all power lines has been completed, and the PointNet algorithm is terminated.

[0020] Further, in step (2), the erosion algorithm is applied, and the end point of the power line far from the tower in the point cloud cluster is selected as the starting point. Each line is eroded in the direction of the tower along the line. Specifically, it includes:

[0021] 1) Along the power line corrosion: using corrosion algorithm along the power line point cloud cluster to the tower center step by step corrosion, the stop condition of corrosion is set as follows: one is to encounter power line bifurcation; two is to corrode to the tower area, that is, the area with high density, when the corrosion reaches the bifurcation position, the two groups are divided from the bifurcation point for corrosion treatment, and when similar bifurcation situation or high density is encountered again, the corrosion process stops;

[0022] 2) Power line bifurcation corrosion: corrode the two bifurcation point cloud clusters at the power line bifurcation position, and set the corrosion stop condition again when the high-density area of the tower structure is corroded; the points on the two bifurcation paths after corrosion are fitted by straight lines respectively, and the average distances obtained after fitting the two groups of points are compared, and the categories of the two bifurcation path point cloud clusters are judged by comparing the average distances, wherein the average distance of the line point cloud cluster is greater than that of the insulator string point cloud cluster.

[0023] Further, the camera shooting point positions on the device point cloud cluster and the path points in the unmanned aerial vehicle flight trajectory generated by step (3) specifically include:

[0024] 1) Camera shooting point position generation: according to the different voltage levels of the power scene to which the point cloud data belongs, combining the power professional device operation related regulations and the needs of unmanned aerial vehicle inspection work, generating camera shooting point positions at the end points and intermediate points in the device point cloud cluster according to the type of equipment to be inspected on the tower to be inspected and the equipment shooting rules;

[0025] 2) Path point generation in unmanned aerial vehicle flight trajectory: based on the determined camera shooting point positions, based on the inclination angle formed by the stay wire and the tower body on the power tower, adjust the extension direction to the outside of the tower on both sides, and extend according to the length of the pre-set shooting distance, to obtain the unmanned aerial vehicle flight path points, and calculate the gimbal pitch angle in the unmanned aerial vehicle shooting process and the yaw angle information in the unmanned aerial vehicle flight attitude.

[0026] Further, the step (3) of sequentially connecting each path point by using ACO and RRT* path planning algorithm to obtain the optimal unmanned aerial vehicle inspection trajectory specifically includes:

[0027] 1) ACO and RRT* path planning algorithm:

[0028] First, use ACO algorithm to connect each path point to generate a preliminary inspection path, and optimize the path based on the distance between points and points; then, use RRT* algorithm to further optimize the preliminary inspection path to ensure that the unmanned aerial vehicle will not collide with the environment during flight; this process is described by the following formula:

[0029]

[0030] where τ(i,j) represents the pheromone concentration on edge (i,j); η(i,j) is the heuristic information defined according to distance; and α and β reflect the relative importance of pheromone and heuristic information;

[0031] By continuously adjusting the pheromone concentration, the ACO algorithm gradually finds the optimal path from the starting point to the ending point;

[0032] The RRT* algorithm is applied to path optimization: a path tree is maintained, starting from the starting point, randomly sampling in the space, and finding the point on the path tree closest to the sampling point and capable of connecting it without obstacles, connecting the point and the sampling point, and adding the sampling point to the path tree until the area near the ending point is explored;

[0033] 2) Route safety check: according to the safety distance of the unmanned aerial vehicle in the flight operation process and the obstacle, the route planned by the algorithm is verified, and if the requirement is not met, the path planning is corrected again

[0034] The unmanned aerial vehicle inspection track of the tower in the generated power line point cloud is taken as a total target, clustering analysis and category judgment of the point cloud, power line corrosion and insulator device category judgment, and generation of the unmanned aerial vehicle inspection track are realized, the labor cost of unmanned aerial vehicle route planning is reduced, and the method can be widely applied to point cloud processing and route planning scenes of unmanned aerial vehicle inspection of various power tower devices. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a flowchart of the power tower unmanned aerial vehicle inspection track automatic generation method of the embodiment of the present application;

[0036] Figure 2 is a power line feature recognition schematic diagram of the embodiment of the present application;

[0037] Figure 3 is an insulator recognition algorithm flowchart of the embodiment of the present application;

[0038] Figure 4 is a power line bifurcation recognition schematic diagram of the embodiment of the present application;

[0039] Figure 5 is a bifurcation corrosion and category judgment schematic diagram of the embodiment of the present application.

[0040] Figure 6 is a route planning result schematic diagram of the embodiment of the present application. DETAILED DESCRIPTION

[0041] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0042] Referring to Figure 1 The embodiments of the present application provide a point cloud automatic identification method for power unmanned aerial vehicle inspection of towers, comprising the following steps:

[0043] Step (1), visualizing the point cloud data of power towers and lines collected by a laser radar device, inputting inspection tower coordinates or manually marking the top position of the tower to be inspected, and obtaining the spatial coordinates of the top end of the target inspection tower; taking the spatial coordinates as the center to intercept a region with a given length, obtaining the edge point cloud of the region; using a DBSCAN algorithm to cluster the edge point cloud to obtain point cloud clusters, and using a PointNet classification algorithm to determine whether the type of each point cloud cluster is a power line.

[0044] Step (2), using the point cloud clusters determined as power line categories in step (1), applying an erosion algorithm, selecting the end point of the power line away from the tower in the point cloud cluster as the starting point, and eroding each line along the line towards the tower direction, and determining whether there is a line bifurcation generated by the connection of the insulator in the erosion process, and classifying the point cloud clusters based on the physical structure difference between the insulator and the line and whether they are directly connected to the tower to distinguish the insulator and the line.

[0045] Step (3), for each point cloud cluster identified as an insulator in step (2), generating camera shooting points on the device point cloud cluster and corresponding path points in the unmanned aerial vehicle flight trajectory according to the device photo shooting rules of the power unmanned aerial vehicle inspection, and adjusting the path point positions after calculating the tower pull wire inclination angle; according to the device photo shooting rules of the power unmanned aerial vehicle inspection, supplementing the remaining camera shooting points and unmanned aerial vehicle flight path points in the unmanned aerial vehicle inspection process except for the insulator device, sequentially connecting each path point by using ACO and RRT* path planning algorithms, and obtaining the optimal unmanned aerial vehicle inspection trajectory.

[0046] Step (1) applies the PointNet algorithm to point cloud to identify and extract towers and power lines, which specifically includes:

[0047] 1) Edge clustering:

[0048] First, a specific range area around the top of the tower is selected, and the edge point cloud in this area is collected, aiming to accurately capture the spatial structure information. Then, the DBSCAN clustering algorithm is used to group the collected blue point set, ensuring that the points in each cluster have similar spatial distribution characteristics.

[0049] The DBSCAN algorithm assigns a unique label to each point in the point cloud data set, and initially all points are not visited. For each point P in the point cloud data set, the number of points in its ε-neighborhood is calculated.

[0050] N ε (P)={Q∈D|dist(P,Q)≤ε}

[0051] If |N ∈ (P)|≥MinPts, P is a core point and is added to the core point set. An unvisited core point is selected as a seed point to start a new clustering process. For the current seed point P, find all density-reachable points (density-reachable: there is a point sequence P1, P2, …, P n , where P1=P, P n =Q) of P and add them to the current clustering cluster. Update the labels of these points to the identifier of the current clustering cluster. For each point in the current clustering cluster, if it has not been visited and is a core point, it is used as a new seed point to continue expanding the clustering cluster. Repeat the above process until all points in the current clustering cluster are visited.

[0052] In each cluster, the algorithm randomly selects three representative points, which will serve as the basis for subsequent processing. Then, an erosion algorithm is used to process these points, further generating three data sets each containing 1024 points. This step aims to provide more abundant input information for the subsequent deep learning model through data augmentation.

[0053] 2) Category judgment:

[0054] The three 1024-point data sets are sent to the PointNet network for intelligent judgment, which can distinguish whether the input data belongs to the power line or the environment. To improve the accuracy of the judgment, the algorithm makes three independent judgments for each cluster, and if two or more of the judgment results are power lines, the cluster is finally determined to be a power line.

[0055] Since point cloud data is arranged in random order, when an N×D point cloud data is randomly shuffled in the N dimension, it still represents the same object and does not affect the overall expression of the object, which is called the permutation invariance of point cloud.

[0056] To preserve the invariance of point cloud permutation, the PointNet network is designed as a symmetric function. Therefore, a general function defined on a point set is approximated by applying a symmetric function on the transformed elements in the set:

[0057] f({x1,x2,…,x n})≈g(h(x1),h(x2),…,h(x n )),x i ∈R N

[0058] where h is a mapping function from R N to R K , and the function has symmetry. This module uses a multi-layer perceptron network to approximate the function h, and uses a combination of a single-variable function and a max-pooling function to approximate the function g, which works well. Through a series of h, multiple f can be learned to capture different properties of the point set.

[0059] The output of the above function forms a vector [f1,f2,…,f K ], which is the global feature representation of the input point set. Then it is spliced with the feature of each point, and then a new per-point feature is extracted based on the combined point feature. The feature of each point contains both local and global information. To ensure that the semantic label is invariant after the point cloud undergoes geometric transformation, compared with the method of aligning two-dimensional images through customized layers, the PointNet network simplifies the alignment process by using the point cloud input form, and directly transforms the coordinates by predicting the affine transformation matrix through the T-net. This idea can be extended to feature space alignment, but a regularization term is needed to constrain the feature transformation matrix to be close to an orthogonal matrix to reduce the optimization difficulty:

[0060]

[0061] where A is a feature alignment matrix predicted by a small network, and I represents the identity matrix. An orthogonal transformation does not lose information in the input, so this is ideal. By adding this regularization term, the optimization process becomes more stable.

[0062] Finally, as shown in Figure 2 , the PointNet algorithm records all the clustered points that are judged as power lines and marks them in red, clearly showing the spatial distribution of the power lines.

[0063] 3) End extraction: when the category judgment result continues to output as a tower, the extraction of the upper part of the tower and all power lines is completed, and the point cloud analysis algorithm is terminated.

[0064] The flowchart of the power line corrosion and insulator identification algorithm in step (2) is shown in Figure 3 , which specifically includes:

[0065] 1) Line erosion: As shown in FIG. 2, when processing each power line category, an erosion algorithm is used to gradually erode from the power line to the tower center. The stopping condition of erosion is set as follows: one is to encounter a power line bifurcation; the other is to erode to the tower area, i.e. the area with high density. When eroding to the bifurcation position, the bifurcation point is divided into two groups for erosion processing. When encountering similar bifurcation conditions again or the density is too large, the erosion process will stop. Figure 4

[0066] 2) Power line bifurcation erosion: As shown in FIG. 3, at the bifurcation position of the line, the point cloud clusters of each bifurcation are eroded, and a straight line is used to fit the points on each bifurcation path, and the average distance of the two groups of point fitting is compared. Since the insulator string mostly presents a straight line form, and the power line may appear curved due to the need to connect with the insulator. At the same time, the erosion stopping condition is set, and the erosion is stopped when it reaches the high-density area of the tower structure. Figure 5

[0067] Step (3) camera shooting point and UAV flight path point generation specifically includes:

[0068] 1) Camera shooting point generation: According to the different voltage levels of the power scene to which the point cloud data belongs, combined with the relevant regulations of power professional equipment operation and the needs of unmanned aerial vehicle inspection work, according to the type of equipment to be inspected on the tower to be inspected and the equipment shooting rules, the camera shooting points are accurately generated at the endpoints and intermediate positions of the equipment point cloud clusters. This process ensures that the UAV can strictly meet the preset safety distance requirements when hovering to take pictures, thereby ensuring the safety of the inspection operation while ensuring the quality of the shooting.

[0069] 2) UAV flight path planning:

[0070] After generating a set of points for UAV photo shooting, the flight trajectory of the UAV needs to be further planned to effectively avoid collision with the environment. For this purpose, the strategy of combining Ant Colony Optimization (ACO) algorithm and Rapidly-exploring Random Tree Star (RRT*) algorithm is adopted.

[0071] ACO algorithm has excellent ability in solving the trajectory planning problem of given access targets. However, it mainly focuses on the shortest distance between points and points, and does not have the ability to detect whether collision will occur during flight between target points. In order to make up for this deficiency, RRT algorithm is introduced. RRT algorithm has high search ability and adaptability to the environment, and can optimize the flight path between the given target points, effectively avoiding potential collision problems.​​

[0072] In a specific implementation, a preliminary trajectory planning is first generated using the ACO algorithm, which is mainly optimized based on the distance between points. Subsequently, the RRT* algorithm is used to further optimize the preliminary inspection path to ensure that the UAV does not collide with the environment during flight. This process can be described by the following formula:

[0073]

[0074] where τ(i,j) represents the pheromone concentration on edge (i,j); η(i,j) is the heuristic information defined according to distance; and α and β reflect the relative importance of pheromone and heuristic information. This formula reflects how ants make decisions based on pheromone concentration and heuristic information when choosing the next location. The higher the pheromone concentration and the better the heuristic information, the greater the probability of the ant choosing that path. By continuously adjusting the pheromone concentration, the ACO algorithm can gradually find the optimal path from the starting point to the end point.

[0075] The optimization theory of the RRT* algorithm is actually to maintain a path tree: starting from the starting point, randomly sampling in space, and finding the point on the path tree closest to the sampling point and able to connect it without obstacles, connecting this point with the sampling point, and adding the sampling point to the path tree until the end point vicinity is explored. Finally Figure 6 The final results of the algorithm are shown, which can accurately identify insulator strings, towers, and power line features.

[0076] 3) Route safety verification: According to the safety distance between the UAV and obstacles during flight, the algorithm-planned route is verified, and if it does not meet the requirements, the path planning is revised.

[0077] This example process starts with point cloud data reading, first obtaining the coordinates of the tower top, and using the PointNet network to classify the point cloud to identify the environment and power lines. Then, through erosion detection, the possible corrosion areas in the power line are identified to determine the key equipment to be inspected. Based on the equipment location and power equipment inspection specifications, the best shooting point is generated and further UAV flight path is generated to ensure that all target areas are covered during inspection. On this basis, the route planning algorithm generates the complete inspection route of the UAV, ensuring efficient and safe completion of the inspection task. Finally, the UAV inspection route file is output, providing a basis for subsequent equipment maintenance and decision-making. This example realizes the complete inspection route planning task chain from data collection, analysis to path planning and result output.

[0078] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any changes or replacements within the technical scope disclosed by the present application, which can be easily thought by any person skilled in the art, should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for automatically generating a UAV inspection trajectory based on a power tower point cloud, characterized in that, The method comprises the following steps: Step (1), visualizing the point cloud data of the power tower and line collected by the laser radar device, inputting the inspection tower coordinates or manually marking the position of the top of the tower to be inspected, and obtaining the spatial coordinates of the top of the target inspection tower; taking the spatial coordinates as the center to intercept a region with a given length, obtaining the edge point cloud of the region; using the DBSCAN algorithm to cluster the edge point cloud to obtain a point cloud cluster, and using the PointNet classification algorithm to determine whether the type of each point cloud cluster is a power line; Step (2), using the point cloud cluster determined as the power line category in step (1), applying the erosion algorithm, selecting the endpoint of the power line away from the tower in the point cloud cluster as the starting point, and eroding each line along the line towards the tower direction, and determining whether there is a line bifurcation generated by the connection of the insulator during the erosion process, and classifying the point cloud cluster based on the physical structure difference between the insulator and the line and whether it is directly connected to the tower to distinguish the insulator and the line; Step (3), for each point cloud cluster identified as an insulator in step (2), generating camera shooting points on the device point cloud cluster and corresponding path points in the unmanned aerial vehicle flight trajectory according to the device photo shooting rules of the power unmanned aerial vehicle inspection, and adjusting the path point position after calculating the tower pull line inclination angle; according to the device photo shooting rules of the power unmanned aerial vehicle inspection, supplementing the remaining camera shooting points and corresponding path points in the unmanned aerial vehicle flight trajectory in the unmanned aerial vehicle inspection process except for the insulator device, sequentially connecting each path point by using the ACO and RRT* path planning algorithm, and obtaining the optimal unmanned aerial vehicle inspection trajectory. 2.The method of claim 1, wherein, Step (1) obtains the spatial coordinates of the top of the tower in the point cloud data, specifically including: 1) Import tower top coordinate data: import the longitude and latitude coordinates of the power tower to be inspected and the height data of the tower top from the point cloud data; 2) Manually mark the tower top: visualize the point cloud data, and manually click the tower top position in the visualized point cloud when the imported data is not accurate or lacks coordinate data to correct or obtain the spatial coordinates of the tower top. 3.The method of claim 1, wherein, Step (1) uses the DBSCAN algorithm to cluster the edge point cloud to obtain a point cloud cluster, and uses the PointNet classification algorithm to determine whether the type of each point cloud cluster is a power line, specifically including: The DBSCAN clustering algorithm is used to group the collected blue point set, ensuring that the points in each cluster have similar spatial distribution characteristics. The DBSCAN algorithm assigns a unique label to each point in the point cloud data set. Initially, all points are not visited. For each point P in the point cloud data set, calculate the number of points in its neighborhood ε: N ε (P) = { Q E D | dist(P, Q) < e} If |N∈(P)|≥MinPts, P is a core point and is added to the core point set; an unvisited core point is selected as a seed point to start a new clustering process; for the current seed point P, find all density reachable points of P, i.e. there is a point sequence P1, P2,..., P n , wherein P1=P, P n =Q; all the density reachable points are added to the current clustering cluster, and the labels of these points are updated to the identifier of the current clustering cluster; for each point in the current clustering cluster, if it has not been visited and is a core point, it is taken as a new seed point to continue expanding the clustering cluster; the above process is repeated until all points in the current clustering cluster are visited. In each cluster, three representative points are randomly selected, which will be the basis for subsequent processing. Then, the erosion algorithm is used to process these points to further generate three data sets each containing 1024 points; Finally, three data sets each containing 1024 points are input into the PointNet algorithm for category judgment, and the PointNet algorithm is used to make three independent judgments for each cluster. If two or more of the judgment results are power lines, the cluster is finally determined to be a power line. When the category judgment result continues to output as a tower, the category judgment of the upper half of the tower and all power lines has been completed, and the PointNet algorithm is terminated.

4. The method of claim 1, wherein, In step (2), the corrosion algorithm is applied, and the end point of the power line away from the tower is selected as the starting point in the point cloud cluster. Each line is corroded along the line towards the tower. Specifically, it includes: 1) Corrosion along the power line: the corrosion algorithm is used to gradually corrode the power line point cloud cluster towards the tower center. The stopping condition of corrosion is set as follows: one is to encounter a power line bifurcation; two is to corrode to the tower area, i.e. a region with high density. When corrosion reaches the bifurcation position, two groups are formed from the bifurcation point for corrosion processing. When similar bifurcation conditions or high density are encountered again, the corrosion process stops. 2) Power line bifurcation corrosion: the two bifurcation point cloud clusters at the power line bifurcation position are corroded, and the corrosion stopping condition is set again. When the high-density area of the tower structure is corroded, it is stopped. Straight lines are used to fit the points on the two bifurcation paths after corrosion, and the average distances obtained after fitting the two groups of points are compared. The category of the two bifurcation path point cloud clusters is determined by comparing the average distances, wherein the average distance of the line point cloud cluster is greater than the average distance of the insulator string point cloud cluster.

5. The method of claim 1, wherein, The path points in the camera shooting point position on the device point cloud cluster and the corresponding UAV flight trajectory generated in step (3) specifically include: 1) Camera shooting point position generation: according to the different voltage levels of the point cloud data belonging to the power scene, combined with the relevant regulations of power professional equipment operation and the needs of UAV inspection work, according to the type of equipment to be inspected on the tower to be inspected and the equipment shooting rules, camera shooting point positions are generated at the end points and intermediate points in the device point cloud cluster. 2) Path point generation in UAV flight trajectory: based on the camera shooting point positions determined, based on the inclination angle formed by the tower body and the tower body, the direction of extension to the outside of the tower is adjusted, and the length of the pre-set shooting distance is extended to obtain the UAV flight path points. At the same time, the gimbal pitch angle and the yaw angle information in the UAV flight attitude during the UAV shooting process are calculated.

6. The method of claim 1, wherein, The ordered connection of each path point using ACO and RRT* path planning algorithms in step (3) to obtain the optimal UAV inspection trajectory specifically includes: 1) ACO and RRT* path planning algorithm: First, the ACO algorithm is used to connect each path point to generate a preliminary inspection path, and the path is optimized based on the distance between points. Then, the RRT* algorithm is used to further optimize the preliminary inspection path to ensure that the UAV does not collide with the environment during flight. This process is described by the following formula: Where τ(i, j) represents the pheromone concentration on edge (i, j); η(i, j) is the heuristic information defined according to distance; α and β reflect the relative importance of pheromone and heuristic information; By constantly adjusting the pheromone concentration, the ACO algorithm gradually finds the optimal path from the starting point to the end point; Path optimization is applied by RRT* algorithm: maintain a path tree, start from the starting point, randomly sample in space, find the point on the path tree closest to the sampling point and can be connected with it without obstacles, connect this point with the sampling point, add the sampling point to the path tree, until the end point vicinity area is explored; 2) Route safety check: according to the safety distance between the unmanned aerial vehicle and the obstacle in the flight operation process, the route planned by the algorithm is verified, and if it does not meet the requirements, the path planning correction is re-performed.

Images