Photovoltaic inspection waypoint dynamic planning and shooting attitude optimization system for unmanned aerial vehicle

Abstract

The invention discloses an unmanned aerial vehicle photovoltaic routing inspection waypoint dynamic planning and shooting attitude optimization system, relates to the technical field of photovoltaic power station routing inspection, and aims to solve the technical problems that traditional photovoltaic routing inspection is low in efficiency and high in labor cost, and a detection result is inaccurate due to unreasonable waypoint planning and poor shooting angle in unmanned aerial vehicle routing inspection. Comprising a scene three-dimensional reconstruction module which is used for generating a high-precision three-dimensional grid model containing component space coordinates, obstacle positions and topographic relief; and the local path planning module is used for carrying out unmanned aerial vehicle local target detection in a photovoltaic power station scene by combining a cluster center point anti-sun azimuth offset strategy and a genetic algorithm comprehensive cost optimization method, and dynamically generating local optimization waypoints adaptive to irregular component distribution and local obstacles. According to the invention, full-coverage inspection of the component in a complex scene is realized, tiny defect features can be clearly captured, and the core problems of incomplete coverage and low precision are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power plant inspection technology, and more specifically, to a dynamic planning and shooting attitude optimization system for unmanned aerial vehicle (UAV) photovoltaic inspection waypoints. Background Technology

[0002] In drone inspections of complex scenarios such as mountain photovoltaic power stations and rooftop distributed photovoltaic power stations, photovoltaic modules often face a triple challenge: irregular spatial distribution (such as being arranged in a stepped manner according to the terrain), dense local obstacles (such as rock protrusions in mountain power stations and pipe supports in rooftop power stations), and diverse types of surface defects (such as tiny scratches and local stains).

[0003] Current UAV inspection systems employ a "fixed waypoint + single attitude" model, which can only plan paths based on two-dimensional plane coordinates. This approach cannot accurately reconstruct the three-dimensional spatial position of components to adapt to irregular distributions, nor can it dynamically adjust waypoints according to local obstacles. Furthermore, it cannot capture minute defects through optimized shooting attitude. This leads to frequent problems during inspections, such as "waypoints deviating from the component area, shooting views being obstructed by obstacles, and minor defects being missed due to blurred imaging." Ultimately, this results in the core technical problem of "incomplete inspection coverage and low defect recognition accuracy in complex scenarios." In view of this, we propose a UAV photovoltaic inspection waypoint dynamic planning and shooting attitude optimization system. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic planning and shooting attitude optimization system for unmanned aerial vehicle (UAV) photovoltaic inspection waypoints, in order to solve the technical problems of low efficiency and high labor costs in traditional photovoltaic inspection, as well as inaccurate detection results caused by unreasonable waypoint planning and poor shooting angles in UAV inspection.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a dynamic planning and shooting attitude optimization system for unmanned aerial vehicle (UAV) photovoltaic inspection waypoints, comprising:

[0006] The scene 3D reconstruction module is used to generate a high-precision 3D mesh model that includes component spatial coordinates, obstacle locations, and terrain undulations.

[0007] The local path planning module is used to combine the cluster center point anti-solar azimuth offset strategy and the genetic algorithm comprehensive cost optimization method to perform UAV local target detection in photovoltaic power station scenarios, and dynamically generate local optimized waypoints that adapt to irregular component distribution and local obstacles.

[0008] The global path planning module is used to merge locally optimized waypoints to generate globally optimized waypoints by iteratively optimizing the roll angle, pitch angle, and yaw angle in the attitude parameters, as well as the comprehensive cost of generating globally optimized waypoints and shooting attitude.

[0009] An unmanned aerial vehicle cruise module for generating unmanned aerial vehicle cruise attitude data.

[0010] The application generates a high-precision three-dimensional grid model containing component space coordinates, obstacle positions and terrain undulations by using the full-process technology of aerial image stitching-image segmentation-multi-algorithm enhancement-principal component analysis reconstruction through the scene three-dimensional reconstruction module; the local optimization navigation points that adapt to irregular component distribution and local obstacles can be dynamically generated by combining the clustering cluster center point anti-solar azimuth deviation + genetic algorithm comprehensive cost optimization of the local path planning module; and the iterative optimization of the roll angle, pitch angle and yaw angle in the attitude parameters through the global path planning module ensures that the shooting angle avoids obstacles and fits the component surface, which can realize full-coverage inspection of components in complex scenes and clearly capture tiny defect features, effectively solving the core problems of incomplete coverage and low precision.

[0011] Preferably, the three-dimensional grid model of the photovoltaic power station scene is generated in the following manner: an aerial image sequence is collected, the aerial image sequence is spliced to obtain a scene view through scale-invariant feature conversion points and rotation and translation technology, the scene view is converted into a single-channel view through image segmentation technology, and the single-channel view is converted into a three-dimensional grid model through image filtering technology.

[0012] The step of converting the scene view into a single-channel view includes: setting a binary threshold T, comparing the gray value I(x, y) of each pixel point in the scene view with the binary threshold T, extracting all structure edges of the scene view, and obtaining a contour image of the scene model; using a median filter and a morphological closing operation to process the contour image to generate a single-channel view.

[0013] The step of converting the single-channel view into a three-dimensional grid model through image filtering technology includes: respectively performing image enhancement on the single-channel view through a bilateral filter, a high-pass filter, a gradient algorithm, histogram equalization and a wavelet transform algorithm to extract three-dimensional point cloud data; and performing image decomposition and image reconstruction through a principal component analysis method to obtain a two-dimensional grid model and a three-dimensional grid model.

[0014] Preferably, the step of extracting three-dimensional point cloud data comprises: smoothing the single-channel view by a Gaussian filter, sharpening the single-channel view processed by the Gaussian filter by a Laplace operator, extracting image structure edge point cloud by pixel value of the image, clustering the image structure edge point cloud to obtain a plurality of pixel sets, extracting principal component direction of the image structure edge point cloud by a principal component analysis algorithm, and sequentially taking adjacent pixel points in each pixel set to project vectors to the principal component direction of the respective pixel set, and retaining the principal component direction of the pixel set with a projection distance exceeding a preset distance to generate a principal component vector set of the image structure edge point cloud; and the principal component direction of each pixel set is connected to the midpoint of the adjacent pixel points selected thereby, as three-dimensional point cloud data.

[0015] Preferably, the step of generating local optimization waypoints by local target detection comprises: processing the current three-dimensional grid model by a target detection algorithm to generate a photovoltaic component detection frame, processing the photovoltaic component detection frame by a clustering algorithm to obtain a clustering cluster, obtaining a clustering cluster center point of each clustering cluster, taking a plurality of clustering cluster center points to move unit coordinate points in the opposite direction of the solar direction to determine local optimization points, and obtaining local optimization waypoints by a genetic algorithm with the local optimization points as the center to calculate the comprehensive cost of all local optimization point subsets and obtaining the local optimization waypoints according to the comprehensive cost.

[0016] Preferably, the step of calculating the comprehensive cost of all local optimization point subsets comprises: calculating distance cost, obstacle cost and view cost, correcting the shooting distance cost, obstacle cost and view cost, setting the relationship between the comprehensive cost of the local optimization point subset and the shooting distance cost, obstacle cost and view cost, and the sum of the three types of corrected costs constitutes the final total comprehensive cost.

[0017] Preferably, the step of generating local optimization waypoints comprises: setting fitness function F(X) and fitness threshold F th , calculating the predicted comprehensive cost of each local optimization point subset within the optimization step by the fitness function, setting local iterative search step size loal and local iterative termination distance d end,local , taking the local optimization point subset as the initial search point to perform iterative search, taking the search point with a comprehensive cost less than a preset limit value C lim,capture as the best image capture point to obtain a plurality of best image capture points, and performing Delaunay triangulation on the best image capture points to obtain a plurality of local optimization waypoints.

[0018] Preferably, the step of generating the comprehensive cost and the shooting pose of the global optimization waypoint comprises: taking the local optimization waypoint as the initial optimization point, setting the global iterative search step size globaland a global iteration termination distance d end,global , the global optimization flight point is calculated by iteratively processing each local optimization flight point through a genetic algorithm, and the local optimization flight point with the minimum comprehensive cost is selected as the global optimization flight point; the attitude parameter of the global optimization flight point is iteratively optimized and calculated through the genetic algorithm, and the optimal attitude parameter of the global optimization flight point is obtained.

[0019] The attitude parameter includes a roll angle size , a pitch angle size θ, and a yaw angle size ψ.

[0020] Preferably, the process of generating the UAV cruising attitude data is as follows:

[0021] An initial speed v0 and a camera field of view angle ω0 of the global optimization flight point are initialized.

[0022] The comprehensive cost is calculated in the roll, pitch, and yaw angle directions using the Gauss-Newton method.

[0023] The angle variable step and the optimization step are determined based on the directional derivative, and the larger the optimization step, the larger the angle adjustment amplitude.

[0024] According to the angle change, the camera shooting attitude is adjusted based on the principle that the angle is negatively correlated with the attitude.

[0025] The field of view center is determined through the camera shooting attitude, the UAV flight speed is adjusted based on the principle that the attitude is positively correlated with the speed, the UAV height and the remaining cruising distance are obtained.

[0026] If the remaining cruising distance exceeds a threshold d th,remain , the derivative calculation is repeated; otherwise, the relative speed and angle of the UAV and the target component are calculated based on the camera field of view angle, and the cruising attitude data is output.

[0027] Preferably, the formula for determining the angle variable step based on the directional derivative is as follows:

[0028]

[0029] In the formula, Δθ k , and Δψ k are the variable steps of the roll angle size, the pitch angle size, and the yaw angle size in the kth iteration, respectively. Ck, Ck, and Ck are the directional derivatives of the comprehensive cost in the roll angle size, the pitch angle size, and the yaw angle size in the kth iteration, respectively. k-1 Ck-1 is the comprehensive cost in the (k-1)th iteration.

[0030] The calculation formula of the optimization step is as follows: In the formula, δθ k, δψ k respectively are the optimization step of the roll angle size, the pitch angle size, and the yaw angle size in the kth iteration; θ k , ψ k respectively are the optimization step of the roll angle size, the pitch angle size, and the yaw angle size in the kth iteration.

[0031] Preferably, the calculation formula of the UAV cruising attitude data is as follows: In the formula, φ is the heading angle of the UAV; ω k is the optimal camera field of view angle in the kth iteration; is the maximum roll angle; θ max is the maximum pitch angle; ψ max is the maximum yaw angle; ω0 is the initial camera field of view angle;

[0032] wherein the pitch angle limit when the UAV cruises is set as: In the formula, θ is the current pitch angle; v max is the maximum flight speed of the UAV; ω max is the maximum value of the camera field of view angle; a is the equivalent aperture of the camera. g and h are different calculation functions.

[0033] Compared with the prior art, the present application has the following beneficial effects:

[0034] 1. The scene three-dimensional reconstruction module is adopted in the present application, the full-process technology of aerial image stitching-image segmentation-multi-algorithm enhancement-principal component analysis reconstruction is adopted, a high-precision three-dimensional grid model containing component space coordinates, obstacle positions and terrain undulations is generated, the clustering cluster center point anti-solar azimuth offset of the local path planning module is combined with the genetic algorithm comprehensive cost optimization, the local optimization navigation point which is adaptive to irregular component distribution and local obstacles can be dynamically generated, the iteration optimization of the roll angle, the pitch angle and the yaw angle in the attitude parameter is carried out through the global path planning module, the shooting angle is ensured to avoid the obstacles and to be close to the component surface, the full-coverage inspection of the components in the complex scene can be realized, the small defect features can be clearly captured, and the core problems of incomplete coverage and low precision are effectively solved.

[0035] 2. In the three-dimensional point cloud data extraction link, the preprocessing process of Gaussian filter smoothing-Laplacian operator sharpening-pixel value edge point cloud extraction is adopted in the present application, the gray gradient difference between the small defect area and the normal component is strengthened, the pixel gray threshold compensation is set for the defect sensitive area in the clustering stage, the defect pixels are avoided to be misclassified into the normal pixel set, a smaller preset distance threshold is separately set for the defect area pixel set in the subsequent vector projection screening, the three-dimensional point cloud data of the defect area is independently generated, the spatial coordinates and the morphology of the defect can be accurately restored, the bottleneck of improving the identification precision of the small defect is further solved, and the millimeter-level data support is provided for the defect positioning.

[0036] 3、The application also establishes a dynamic mapping relationship with the scene characteristics through the weight coefficient for the comprehensive cost calculation - automatically increasing the obstacle cost weight and reducing the distance cost weight in the mountain power station obstacle dense area; increasing the view cost weight and reducing the obstacle cost weight in the roof power station component dense area, to ensure that the comprehensive cost calculation fits the scene requirements; and constructing a field of view angle-speed real-time feedback mechanism in the cruise control, when the camera field of view angle is reduced, the direction derivative of the speed adjustment is calculated synchronously through the Gauss-Newton method, so that the flight speed is reduced synchronously with the field of view angle, avoiding the missed shooting or overlap caused by the speed-field of view angle linkage delay, and further solving the problem of balancing the inspection efficiency and accuracy in complex scenes. DETAILED DESCRIPTION

[0037] The application relates to a UAV photovoltaic inspection flight point dynamic planning and shooting posture optimization system, which comprises:

[0038] A scene three-dimensional reconstruction module is used for generating a high-precision three-dimensional grid model containing component space coordinates, obstacle positions and terrain undulations.

[0039] In the embodiment of the application, the three-dimensional grid model of the photovoltaic power station scene is generated in the following manner: aerial image sequences are collected, scene views are obtained by splicing the aerial image sequences through scale-invariant feature transformation points and rotation and translation technology, a single-channel view is obtained by converting the scene views through image segmentation technology, and a three-dimensional grid model is obtained by converting the single-channel view through image filtering technology.

[0040] The step of converting the scene view into the single-channel view comprises the following steps: setting a binary threshold T, comparing the gray value I(x, y) of each pixel point in the scene view with the binary threshold T, extracting all structure edges of the scene view, and obtaining a contour image of the scene model; using a median filter and a morphological closing operation to process the contour image, and generating the single-channel view.

[0041] The step of converting the single-channel view into the three-dimensional grid model through the image filtering technology comprises the following steps: respectively performing image enhancement on the single-channel view through a bilateral filter, a high-pass filter, a gradient algorithm, histogram equalization and a wavelet transform algorithm, and extracting three-dimensional point cloud data; performing image decomposition and image reconstruction through a principal component analysis method, and obtaining a two-dimensional grid model and a three-dimensional grid model.

[0042] The step of extracting three-dimensional point cloud data comprises: smoothing a single-channel view through a Gaussian filter, sharpening the single-channel view processed by the Gaussian filter through a Laplace operator, extracting image structure edge point cloud through pixel values of the image, clustering the image structure edge point cloud to obtain a plurality of pixel sets, extracting a principal component direction of the image structure edge point cloud through a principal component analysis algorithm, sequentially taking adjacent pixel points in each pixel set and respectively projecting the adjacent pixel points to the principal component direction of the respective pixel set, and retaining a principal component direction of a pixel set whose projection distance exceeds a preset distance to generate a principal component vector set of the image structure edge point cloud; and a midpoint of a line connecting the principal component direction of each pixel set to the adjacent pixel point selected by the pixel set is taken as the three-dimensional point cloud data.

[0043] The local path planning module is used for combining the cluster center point anti-solar direction offset strategy and the genetic algorithm comprehensive cost optimization method to perform local target detection of the UAV in the photovoltaic power station scene, and dynamically generate local optimization waypoints adaptive to irregular component distribution and local obstacles.

[0044] In the embodiment of the application, the step of performing local target detection to generate local optimization waypoints comprises: processing a current three-dimensional grid model through a target detection algorithm to generate a photovoltaic component detection frame, processing the photovoltaic component detection frame through a clustering algorithm to obtain a cluster, obtaining a cluster center point of each cluster, moving unit coordinate points in the opposite direction of the solar direction from a plurality of cluster center points respectively to determine local optimization points, obtaining a local optimization point subset with the local optimization points as the center through a genetic algorithm, calculating comprehensive costs of all the local optimization point subsets, and obtaining local optimization waypoints according to the comprehensive costs.

[0045] The step of calculating the comprehensive costs of all the local optimization point subsets comprises: calculating distance costs, obstacle costs and view costs.

[0046]

[0047] The formula is used for comprehensively evaluating the total cost of the jth local optimization point subset. In the formula, C total,j,caw represents the comprehensive cost of the local optimization point subset (j=1, 2, …, n, n is the total number of the local optimization point subsets). The three terms on the right side of the formula correspond to different types of costs respectively:

[0048] The first term represents a shooting distance cost; l j is the total path length of the jth local optimization point subset, l max is a maximum value of the total path length allowed by the local optimization point subset, is a soft threshold function, when , otherwise This function ensures that only paths with a length exceeding a certain proportion are taken into account in the cost.

[0049] Second term represents the obstacle cost; v j is the obstacle volume of the jth local optimization point subset, v max is the maximum allowed obstacle volume, and similarly, The cost is calculated according to the proportion of the obstacle volume to the maximum allowed value.

[0050] Third term represents the viewshed cost, s j is the viewshed coverage of the jth local optimization point subset, s min is the minimum value of the viewshed coverage, and this formula measures the proportion of the part of the viewshed coverage that does not meet the minimum requirement in the total cost.

[0051] α, β, γ are cost weight coefficients, with a value range of [0, 1] and α + β + γ = 1;

[0052] The shooting distance cost, obstacle cost, and viewshed cost are corrected as follows:

[0053]

[0054] where λ d , λ o , and λ s represent the shooting distance correction factor, obstacle volume correction factor, and viewshed coverage correction factor, respectively, which are used to adjust the original cost according to actual conditions; C' d,j , C' o,j , and C' s,j are the corrected shooting distance cost, obstacle cost, and viewshed cost of the jth local optimization point subset, respectively.

[0055] The relationship between the comprehensive cost of the local optimization point subset and the shooting distance cost, obstacle cost, and viewshed cost is set as follows:

[0056] C total,j,corr = C' d,j + C' o,j + C' s,j That is, the sum of the three types of corrected costs constitutes the final total comprehensive cost.

[0057] The step of generating the local optimization waypoints includes setting a fitness function F(X) and a fitness threshold F th , calculating the predicted comprehensive cost of each local optimization point subset within the optimization step length through the fitness function, setting a local iterative search step length Δs local (Δs localThe iteration search step length of the local path planning module is set according to the photovoltaic module density, 0.5 m in a module dense area and 1 m in a sparse area end,local The iteration search is performed with the local optimization point subset as initial search points, and the comprehensive cost less than a preset limit value C is obtained lim,capture The search points are used as the best image capture points, and a plurality of best image capture points are obtained; the best image capture points are subjected to Delaunay triangulation to obtain a plurality of local optimization waypoints.

[0058] The global path planning module is used for merging the local optimization waypoints to generate global optimization waypoints by iteration optimization of the roll angle, the pitch angle and the yaw angle in the attitude parameter, and generating the comprehensive cost and the shooting attitude of the global optimization waypoints;

[0059] In the embodiment of the application, the step of generating the comprehensive cost and the shooting attitude of the global optimization waypoints comprises: taking the local optimization waypoints as initial optimization points, setting a global iteration search step length Δs global (Δs global The iteration search step length of the global path planning module is set according to the power station scale, which is 5-10 times of Δs local , and a global iteration termination distance d end,global (d end,local The difference between d end,global and d local is that the former is used for iteration termination judgment of the local path planning, the iteration search step length is Δs global The genetic algorithm is used for iteration processing of each local optimization waypoint, the local optimization waypoint with the minimum comprehensive cost is selected as the global optimization waypoint, and the comprehensive cost of each global optimization waypoint is calculated; the genetic algorithm is used for iteration optimization calculation of the attitude parameter of the global optimization waypoint, the optimal attitude parameter of the global optimization waypoint is obtained, and the attitude parameter comprises the roll angle size of the global optimization waypoint, the pitch angle size θ and the yaw angle size ψ.

[0060] The UAV cruise module is used for generating UAV cruise attitude data;

[0061] In the embodiments of the present application, the process of generating the unmanned aerial vehicle cruise attitude data is as follows: setting the initial speed v0 and the camera field of view angle ω0 of the global optimization waypoint; calculating the direction derivative of the comprehensive cost in the roll angle size, the pitch angle size and the yaw angle size by the Gauss-Newton method; calculating the variable step size of the roll angle size, the pitch angle size and the yaw angle size according to the direction derivative of the comprehensive cost in the roll angle size, the pitch angle size and the yaw angle size respectively; determining the optimization step size according to the variable step size; optimizing the roll angle size, the pitch angle size and the yaw angle size of the global optimization waypoint according to the optimization step size, and the larger the optimization step size is, the larger the optimization range of the roll angle size, the pitch angle size and the yaw angle size is; adjusting the camera shooting attitude according to the changes of the roll angle size, the pitch angle size and the yaw angle size, and the larger the roll angle size, the pitch angle size and the yaw angle size are, the smaller the camera shooting attitude is; obtaining the camera shooting field center according to the camera shooting attitude, adjusting the unmanned aerial vehicle flight speed through the distance between the camera optical axis and the camera shooting field center, and the larger the camera shooting attitude is, the larger the unmanned aerial vehicle flight speed is; obtaining the unmanned aerial vehicle height through the camera shooting attitude, and obtaining the remaining cruise distance of the unmanned aerial vehicle through the unmanned aerial vehicle height; when the remaining cruise distance of the unmanned aerial vehicle exceeds the preset distance threshold d th,remain The difference between the two is that the former is based on the remaining cruise distance to determine whether to continue to iteratively adjust the attitude parameters, and the latter is based on the comprehensive cost to select the best image capture point. lim,capture When the remaining cruise distance of the unmanned aerial vehicle does not exceed the preset distance threshold, the distance between the unmanned aerial vehicle and the target component is less than the preset distance, the relative speed and the relative angle between the unmanned aerial vehicle and the target component are calculated according to the camera field of view angle, and the unmanned aerial vehicle cruise attitude data is obtained.

[0062] The formula for calculating the variable step size of the roll angle size, the pitch angle size and the yaw angle size is as follows:

[0063]

[0064] The above formula is used to calculate the variable step size of the roll angle, the pitch angle and the yaw angle at the kth iteration. Wherein, Δθ k , Δψ k are the variable step sizes of the roll angle size, the pitch angle size and the yaw angle size at the kth iteration, respectively; are the direction derivatives of the comprehensive cost in the roll angle size, the pitch angle size and the yaw angle size at the kth iteration, respectively, reflecting the sensitivity of the comprehensive cost to the angle change; C k-1 is the comprehensive cost of the k-1th iteration, which is used to normalize the direction derivative, so that the calculation of the variable step size can consider the cost of the historical iteration.

[0065] The formula for calculating the optimization step size is as follows:

[0066]

[0067] here, δθ k ,δψ k These are the optimization step sizes for the roll angle, pitch angle, and yaw angle in the k-th iteration, respectively. θ k ψ k These are the roll angle, pitch angle, and yaw angle for the kth iteration, respectively. Δθ k , Δψ k These represent the variable step sizes for the roll, pitch, and yaw angles in the k-th iteration. This formula dynamically adjusts the magnitude of the optimized angle for each iteration by multiplying the variable step size by the current angle value, thus achieving fine-tuning of the attitude parameters.

[0068] The formula for calculating the drone's cruise attitude data is as follows:

[0069]

[0070] This formula indicates that the UAV heading angle φ is related to the optimal camera field of view ω in the k-th iteration. k Maximum roll angle Maximum pitch angle θ max Maximum yaw angle ψ max And a function of the initial camera field of view ω0. By comprehensively considering these parameters, the heading angle of the UAV during the cruise process is finally determined to achieve accurate shooting and inspection.

[0071] Among them, the pitch angle limit is set during drone cruise:

[0072]

[0073] Where, θ max The maximum pitch angle is θ, and the current pitch angle is θ; when the UAV cruises, θ ≤ θ max At that time, the maximum flight speed of the drone is v max =g(ω) max ,a); otherwise, v max =h(ω) max ,a);ω max denoted by , where is the maximum field of view of the camera; and 'a' is the equivalent aperture of the camera. This conditional expression uses different functions 'g' and 'h' to calculate the maximum flight speed of the UAV depending on whether the pitch angle exceeds the maximum limit, ensuring the flight safety and shooting effect of the UAV in different attitudes.

[0074] Logical relationship between each formula and derivation process

[0075] (I) Overall logical framework

[0076] Local optimization point subset total cost evaluation stage: First, the total comprehensive cost is calculated by the initial calculation formula, which integrates the costs of shooting distance, obstacles and field of view in three dimensions to obtain the preliminary total comprehensive cost. However, since the initial cost does not consider the special circumstances of the actual scene, the individual costs need to be adjusted by the cost correction formula, and then the corrected individual costs are added together using the modified total comprehensive cost formula to obtain the accurate total comprehensive cost, which provides a reliable basis for the selection of local optimization points.

[0077] Attitude parameter optimization stage: After determining the local optimization points, the attitude parameter optimization of the UAV is entered. First, the variable step size of the attitude parameter is obtained by the attitude parameter variable step size calculation formula combined with the derivative information of the comprehensive cost and the historical iteration cost, and then the final optimization step size is determined by combining the variable step size with the current attitude parameter value according to the attitude parameter optimization step size calculation formula, which realizes the fine adjustment of the roll angle, pitch angle and yaw angle.

[0078] UAV cruise control stage: On the basis of attitude parameter optimization, the heading angle of the UAV is determined by the UAV heading angle calculation formula, which takes into account parameters such as camera field of view angle and maximum attitude angle. At the same time, according to the relationship between the current pitch angle of the UAV and the maximum pitch angle, the maximum flight speed under different attitudes is determined using the UAV maximum flight speed calculation formula, thereby realizing the safe and precise cruise of the UAV.

[0079] (II) Specific derivation logic chain

[0080] Total cost calculation logic chain: Total comprehensive cost initial calculation formula (multi-objective weighted summation) → cost correction formula (linear correction) → modified total comprehensive cost formula (addition operation). This logic chain first constructs a preliminary comprehensive evaluation of multi-dimensional cost, then corrects the cost in combination with the actual scene, and finally obtains the accurate total cost, which provides a quantitative standard for subsequent optimization.

[0081] Attitude parameter optimization logic chain: Attitude parameter variable step size calculation formula (derivative and normalization) → attitude parameter optimization step size calculation formula (multiplication operation). This logic chain first determines the variable step size based on the sensitivity of cost change and historical cost, and then obtains the optimization step size in combination with the current attitude parameter value, which realizes the dynamic and fine adjustment of the attitude parameter.

[0082] Cruise control logic chain: UAV heading angle calculation formula (function modeling); UAV maximum flight speed calculation formula (piecewise function modeling). The two formulas build control models from two key aspects of heading and speed respectively, which together guarantee the shooting accuracy and flight safety during the UAV cruise process, and both depend on the results of the previous attitude parameter optimization, such as the maximum roll angle, maximum pitch angle and other parameters.

[0083] The embodiments of the present application are disclosed, but not limited to the preferred embodiments, and the ordinary skilled in the art can easily understand the spirit of the present application and make different inferences and changes according to the above embodiments, as long as they do not deviate from the spirit of the present application, which are within the protection scope of the present application.

Claims

1. A dynamic planning and shooting attitude optimization system for unmanned aerial vehicle (UAV) photovoltaic inspection waypoints, characterized in that, include: The scene 3D reconstruction module is used to generate a high-precision 3D mesh model that includes component spatial coordinates, obstacle locations, and terrain undulations. The local path planning module is used to combine the cluster center point anti-solar azimuth offset strategy and the genetic algorithm comprehensive cost optimization method to perform UAV local target detection in photovoltaic power station scenarios, and dynamically generate local optimized waypoints that adapt to irregular component distribution and local obstacles. The global path planning module is used to merge locally optimized waypoints to generate globally optimized waypoints by iteratively optimizing the roll angle, pitch angle, and yaw angle in the attitude parameters, as well as the comprehensive cost of generating globally optimized waypoints and shooting attitude. The drone cruise module is used to generate drone cruise attitude data.

2. The UAV photovoltaic inspection waypoint dynamic planning and shooting attitude optimization system according to claim 1, characterized in that, The method for generating the three-dimensional mesh model of the photovoltaic power station scene is as follows: acquiring aerial image sequences, stitching the aerial image sequences together using scale-invariant feature transformation points and rotation and translation techniques to obtain a scene view, converting the scene view into a single-channel view using image segmentation techniques, and converting the single-channel view into a three-dimensional mesh model using image filtering techniques. The step of converting the scene view into a single-channel view includes: setting a binary threshold T, comparing the gray value I(x,y) of each pixel in the scene view with the binary threshold T, extracting all structural edges of the scene view to obtain the contour image of the scene model; and processing the contour image using a median filter and morphological closing operation to generate a single-channel view. The step of converting a single-channel view into a three-dimensional mesh model using image filtering technology includes: enhancing the single-channel view using bilateral filters, high-pass filters, gradient algorithms, histogram equalization, and wavelet transform algorithms to extract three-dimensional point cloud data; and performing image decomposition and reconstruction using principal component analysis to obtain a two-dimensional mesh model and a three-dimensional mesh model.

3. The UAV photovoltaic inspection waypoint dynamic planning and shooting attitude optimization system according to claim 2, characterized in that, The steps for extracting 3D point cloud data include: smoothing a single-channel view using a Gaussian filter, sharpening the single-channel view after Gaussian filter processing using a Laplacian operator, extracting image structure edge point clouds from the image pixel values, clustering the image structure edge point clouds to obtain several pixel sets, extracting the principal component directions of the image structure edge point clouds using a principal component analysis algorithm, sequentially selecting adjacent pixels within each pixel set and projecting them onto the principal component directions of their respective pixel sets, retaining the principal component directions of pixel sets whose projection distance exceeds a preset distance, and generating a principal component vector set of the image structure edge point cloud; and using the midpoint of the line connecting the principal component direction of each pixel set to its selected adjacent pixels as the 3D point cloud data.

4. The UAV photovoltaic inspection waypoint dynamic planning and shooting attitude optimization system according to claim 1, characterized in that, The steps for generating locally optimized waypoints through local target detection include: processing the current 3D mesh model using a target detection algorithm to generate photovoltaic module detection boxes; processing the photovoltaic module detection boxes using a clustering algorithm to obtain clusters; obtaining the cluster center point of each cluster; moving several cluster center points by unit coordinates in the opposite direction of the sun's azimuth to determine local search advantages; obtaining subsets of local search advantages centered on these local search advantages using a genetic algorithm; calculating the comprehensive cost of all subsets of local search advantages; and obtaining locally optimized waypoints based on the comprehensive cost. The locally optimized waypoints include their coordinates.

5. The UAV photovoltaic inspection waypoint dynamic planning and shooting attitude optimization system according to claim 4, characterized in that, The step of calculating the comprehensive cost of all the local search advantage subsets includes: calculating the distance cost, obstacle cost, and view cost; correcting the shooting distance cost, obstacle cost, and view cost; setting the relationship between the comprehensive cost of the local search advantage subset and the shooting distance cost, obstacle cost, and view cost; and the sum of the three types of costs after correction constitutes the final total comprehensive cost.

6. The UAV photovoltaic inspection waypoint dynamic planning and shooting attitude optimization system according to claim 4, characterized in that, The step of generating locally optimized waypoints includes: setting the fitness function F(X) and the fitness threshold F. th The predicted comprehensive cost of each local search advantage subset within the search step size is calculated using the fitness function, and the local iterative search step size Δs is set. local and the local iteration termination distance d end,local Iterative search is performed using a subset of locally advantageous locations as initial search points to obtain a result where the overall cost is less than a preset limit value C. lim,capture The search point is used as the optimal image capture point to obtain several optimal image capture points; Delaunay triangulation is performed on the optimal image capture points to obtain several local optimized waypoints.

7. The UAV photovoltaic inspection waypoint dynamic planning and shooting attitude optimization system according to claim 1, characterized in that, The steps for generating the comprehensive cost and shooting attitude of the globally optimized waypoints include: using the locally optimized waypoints as initial search points, and setting the global iterative search step size Δs. global and global iteration termination distance d end,global The genetic algorithm iteratively processes each local optimization waypoint to select the local optimization waypoint with the minimum comprehensive cost as the global optimization waypoint, and calculates the comprehensive cost of each global optimization waypoint. The attitude parameters of the global optimization waypoint are then iteratively optimized using the genetic algorithm to obtain the optimal attitude parameters of the global optimization waypoint. The attitude parameters include the roll angle of the globally optimized waypoint. Pitch angle θ, yaw angle ψ.

8. The UAV photovoltaic inspection waypoint dynamic planning and shooting attitude optimization system according to claim 1, characterized in that, The process of generating UAV cruise attitude data is as follows: Initialize the initial velocity v0 and camera field of view ω0 of the global optimization waypoint; The derivatives of the overall cost in the roll, pitch, and yaw directions were calculated using the Gauss-Newton method. The step size and optimization step size for each angle are determined based on the directional derivative. The larger the optimization step size, the greater the angle adjustment range. Based on the change in angle, the camera's shooting posture is adjusted according to the principle of negative correlation between angle and posture; The center of the field of view is determined by capturing the attitude with a camera, and the flight speed of the drone is adjusted based on the principle of positive correlation between attitude and speed to obtain the drone's altitude and remaining cruising distance. If the remaining cruising distance exceeds the threshold d th,remain Repeat the derivative calculation; otherwise, calculate the relative speed and angle between the UAV and the target component based on the camera's field of view, and output the cruise attitude data.

9. The UAV photovoltaic inspection waypoint dynamic planning and shooting attitude optimization system according to claim 8, characterized in that, The formula for determining the variable step size for each angle based on the directional derivative is as follows: In the formula, Δθ k , Δψ k These are the variable step sizes for the roll angle, pitch angle, and yaw angle in the kth iteration, respectively. These are the directional derivatives of the comprehensive cost at the k-th iteration with respect to the magnitudes of the roll, pitch, and yaw angles, respectively; C k-1 This represents the total cost of the (k-1)th iteration. The formula for calculating the optimization step size is as follows: In the formula, δθ k ,δψ k These are the optimization step sizes for the roll angle, pitch angle, and yaw angle in the k-th iteration, respectively. θ k ψ k These represent the roll angle, pitch angle, and yaw angle for the kth iteration, respectively.

10. The UAV photovoltaic inspection waypoint dynamic planning and shooting attitude optimization system according to claim 8, characterized in that, The formula for calculating the UAV cruise attitude data is as follows: In the formula, φ is the heading angle of the UAV; ω k The optimal camera field of view for the k-th iteration; Maximum roll angle; θ max Maximum pitch angle; ψ max Maximum yaw angle; ω0 is the initial camera field of view angle; Among them, the pitch angle limit is set during drone cruise: In the formula, θ is the current pitch angle; v max ω is the maximum flight speed of the drone; max denoted as the maximum field of view of the camera; denoted as a; denoted as the equivalent aperture of the camera; and denoted as g and h, which are different calculation functions.

Images