Unmanned aerial vehicle autonomous inspection waypoint dynamic planning method for complex mountainous power grid

Abstract

The present application relates to the technical field of unmanned aerial vehicle inspection control, in particular to a kind of unmanned aerial vehicle autonomous inspection waypoint dynamic programming method for complex power grid in mountainous area.The present application obtains to-be-analyzed time stamp when triggering close-to-proximity complex mechanism, obtains spatial distance matrix and idle space topology diagram;Further obtain target three-dimensional coordinates;Further calculate the spatial channel extension form index of the node of idle space topology diagram, obtain the security topology diagram with traffic weight in combination with the numerical value of node in spatial distance matrix;Further, the traffic weight is used to build the search target node sequence of optimization cost, generate initial interpolation curve based on target node sequence, and execute anti-collision offset correction, generate continuous pose sequence;Finally, the body is driven to approach to the end point of the path and hovers, reduce the pose control instability caused by misentry semi-closed dead angle and trajectory discontinuity.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) inspection and control technology, specifically to a dynamic planning method for autonomous inspection waypoints of UAVs used in complex power grids in mountainous areas. Background Technology

[0002] With the development of automated inspection technology, power companies generally use drones to perform inspections according to preset offline global routes. However, when long-distance observation causes the onboard visual model to become blurry, it is necessary to guide the drone off the route and automatically approach the interior of the high-voltage transmission tower network to perform a second inspection.

[0003] High-voltage transmission towers are truss structures made of densely riveted angle steel. When flying close to such complex, non-convex, densely packed metal areas, existing technologies face serious control adaptation problems: First, obstacle avoidance algorithms based on conventional collision avoidance distances cannot distinguish between safe through corridors and semi-enclosed dead ends that are prone to signal multipath blockage from a geometric perspective, easily leading drones to blindly enter dangerous blind spots; Second, relying directly on the underlying coarse distance field data for path map search is prone to derivative singularities and trajectory discontinuities, making it impossible to provide smooth and continuous flight commands to the flight control system. This can lead to abrupt changes in the aircraft's attitude control during the shuttle process, easily causing instability or even collapse of the underlying flight control. Summary of the Invention

[0004] To address the technical problem of drone attitude control instability during close-range flight in complex power grid non-convex spaces due to discontinuous trajectory planning and easy entry into semi-enclosed blind spots, the present invention aims to provide a dynamic planning method for autonomous waypoints of drones for complex power grids in mountainous areas. The specific technical solution adopted is as follows:

[0005] Based on the fluctuation characteristics of the target recognition probability analyzed by a preset sliding window, a close-up reshooting mechanism is triggered to obtain the timestamp to be analyzed and the target bounding box; a three-dimensional occupancy grid matrix is ​​constructed based on a single frame three-dimensional point cloud, and distance transformation and skeleton extraction are performed to obtain a spatial distance matrix and a free space topology map; the geometric center of the target bounding box in the image with the timestamp to be analyzed is extracted as the target pixel point, and the three-dimensional coordinates of the target are obtained by combining the camera intrinsic parameter matrix and extrinsic parameter matrix;

[0006] Local voxel clusters are extracted from the nodes of the free space topology graph and subjected to three-dimensional Gaussian smoothing. The second-order partial derivative of the spatial distance Hessian matrix is ​​calculated and eigenvalue decomposition is performed. The spatial channel extension morphology index is calculated. Combined with the value of the node in the spatial distance matrix, a safe topology graph with passage weight is obtained.

[0007] Based on the security topology map, unobstructed detection is performed by combining the target's three-dimensional coordinates with the preset observation radius to obtain the pathfinding endpoint. The target node sequence is constructed using the passage weight, and an initial interpolation curve is generated based on the target node sequence. Collision avoidance offset correction is performed on the initial interpolation curve along the distance gradient direction of the spatial distance matrix to generate a continuous pose sequence.

[0008] The underlying flight control unit receives the continuous pose sequence, drives the aircraft to approach the pathfinding endpoint and hover.

[0009] Furthermore, the triggering determination of the close-up reshooting mechanism includes:

[0010] Calculate the mean and variance of the target recognition probability within the preset sliding window; when the mean is in a preset critical range, the variance is less than a preset fluctuation threshold, and the pixel area of ​​the target bounding box is less than a preset observation scale threshold, determine that the close-up re-shooting mechanism is triggered.

[0011] Furthermore, the method for obtaining the spatial channel extension morphology index includes:

[0012] The spatial distance second-order partial derivative Hessian matrix is ​​decomposed into eigenvalues ​​and its absolute value is taken to obtain the first, second, and third eigenvalues ​​arranged in ascending order; the first, second, and third eigenvalues ​​are fused to obtain the spatial channel extension morphology index.

[0013] Furthermore, the method for obtaining the security topology map includes:

[0014] The data values ​​of nodes extracted from the spatial distance matrix are combined with the spatial channel extension morphology index to obtain the passage weight of the nodes; nodes and connecting edges with passage weights lower than a preset truncation threshold in the free space topology map are deleted, and the safe topology map with passage weights is output.

[0015] Furthermore, the method for obtaining the pathfinding endpoint includes:

[0016] Using the origin of the local three-dimensional coordinate system as the starting point for pathfinding, the spatial straight line segment pointing from the target three-dimensional coordinates to the starting point of pathfinding is backed by a preset observation radius to obtain the coordinates of the pre-selected endpoint. The node that is closest to the Euclidean distance of the pre-selected endpoint coordinates in the safety topology map and whose spatial connection with the target three-dimensional coordinates does not intersect with the obstacle object element in the three-dimensional occupied grid matrix is ​​selected as the endpoint of pathfinding.

[0017] Furthermore, the method for obtaining the continuous pose sequence includes:

[0018] Based on the basic physical distance between adjacent nodes and the passage weight in the safety topology graph, the single-step optimization cost is obtained, and the target node sequence with the lowest comprehensive cost is searched. Spline interpolation is performed on the target node sequence to generate the initial interpolation curve. When the initial interpolation curve intersects with an obstacle voxel, the distance gradient vector of the center voxel of the intersection region is extracted from the spatial distance matrix. A positive offset operation is performed on the local control points of the initial interpolation curve along the rising direction of the distance gradient to reconstruct a collision-free continuous curve. Discrete sampling is performed on the collision-free continuous curve to generate the continuous pose sequence.

[0019] Furthermore, the method for obtaining the local voxel clusters includes:

[0020] Extract the three-dimensional coordinates of the nodes in the free space topology graph, and construct a three-dimensional data block by extracting adjacent voxels of a preset size from the spatial distance matrix using the three-dimensional coordinates as the center, as the local voxel cluster.

[0021] Furthermore, the method for obtaining the spatial distance matrix includes:

[0022] Identify the free voxels and obstacle voxels in the three-dimensional occupied grid matrix; calculate the three-dimensional Euclidean distance from each voxel to the nearest obstacle voxel, and construct the spatial distance matrix from all the three-dimensional Euclidean distance values.

[0023] Furthermore, the method for obtaining the free space topology map includes:

[0024] Based on the spatial distance matrix, a three-dimensional topology refinement algorithm is invoked to obtain a set of central voxels that maintain three-dimensional connectivity; the set of central voxels is then converted into an undirected graph network consisting of intersecting nodes and connecting edges, which serves as the free space topology graph.

[0025] Furthermore, the method for obtaining the target's three-dimensional coordinates includes:

[0026] The target pixel is mapped to a spatial projection ray in a local three-dimensional coordinate system using the camera intrinsic parameter matrix. Traversal calculations are performed along the extension direction of the spatial projection ray to obtain the first voxel in the three-dimensional occupied grid matrix that the spatial projection ray passes through, which is assigned the value of an obstacle. The center coordinates of this voxel are taken as the three-dimensional coordinates of the target.

[0027] The present invention has the following beneficial effects:

[0028] When the close-up re-shooting mechanism is triggered, this invention acquires the timestamp to be analyzed, obtains the spatial distance matrix and the topology map of the free space, and provides a decision-making basis for subsequent departure from the fixed flight path; further, it acquires the three-dimensional coordinates of the target, projects the defective target in the two-dimensional image onto the three-dimensional static grid space; further, it calculates the spatial channel extension morphology index of the nodes in the free space topology map, quantifies the underlying three-dimensional geometric attenuation features into macroscopic spatial anisotropic connectivity morphology criteria, and combines the node values ​​in the spatial distance matrix to obtain a safe topology map with passage weights, avoiding guiding the UAV into dead ends surrounded by metal; further, it constructs a target node sequence for optimization cost search based on the passage weights, generates an initial interpolation curve based on the target node sequence, and performs anti-collision offset correction to generate a continuous pose sequence, taking into account both trajectory smoothness and safety, and ensuring high control robustness in the complex grid shuttle process; finally, it drives the aircraft to approach to the pathfinding endpoint and hover. This invention calculates morphological indices by decomposing the eigenvalues ​​of the partial derivative matrix, constructs a safe topology graph and searches for target node sequences, performs offset correction on the interpolation curve to generate a continuous pose sequence, and reduces pose control instability caused by accidentally entering semi-enclosed dead angles and trajectory discontinuities. Attached Figure Description

[0029] Figure 1 A flowchart illustrating a dynamic planning method for autonomous waypoints of unmanned aerial vehicles (UAVs) used in complex power grids in mountainous areas, provided as an embodiment of the present invention.

[0030] Figure 2 This is a flowchart of a method for obtaining the endpoint of a pathfinding process, provided as an embodiment of the present invention. Detailed Implementation

[0031] The following description, in conjunction with the accompanying drawings, details a specific scheme for a dynamic planning method for autonomous unmanned aerial vehicle (UAV) inspection waypoints in complex mountainous power grids, provided by this invention.

[0032] Please see Figure 1 The diagram illustrates a flowchart of a dynamic planning method for autonomous waypoints of unmanned aerial vehicles (UAVs) used in complex power grids in mountainous areas, provided by an embodiment of the present invention. The method specifically includes:

[0033] Step S1: Analyze the fluctuation characteristics of target recognition probability based on a preset sliding window, trigger a close-up reshoot mechanism, and obtain the timestamp to be analyzed and the target bounding box; construct a three-dimensional occupancy grid matrix based on a single-frame three-dimensional point cloud, and perform distance transformation and skeleton extraction to obtain a spatial distance matrix and a free space topology map; extract the geometric center of the target bounding box in the image with the timestamp to be analyzed as the target pixel, and obtain the target's three-dimensional coordinates by combining the camera intrinsic and extrinsic parameter matrices.

[0034] In one embodiment of the present invention, before the UAV takes off, the system reads the calibration parameters of the onboard camera (camera), including the focal length component and the optical center component. Then, a fixed camera intrinsic parameter matrix is ​​constructed based on the focal length component and the optical center component. (Existing technology). At the same time, the system reads the electromechanical characteristic manual of the UAV power system, extracts the upper limit of the airframe flight speed (e.g., the preset value is 1.5 m / s) and the upper limit of the airframe acceleration (e.g., the preset value is 0.5 m / s²) from the manual, and saves the above two upper limit values ​​as the airframe flight limit thresholds, which are used to constrain the curvature of the planned trajectory in the future.

[0035] In addition, the system reads the offline global flight path sequence pre-set by the ground station from the internal storage medium, and the UAV performs automated flight according to the offline global flight path sequence.

[0036] During the drone's flight, the system continuously captures the real-time video stream output by the onboard camera. The system is set with a preset sliding window length. .

[0037] Considering the need to distinguish between instantaneous visual interference and resolution limitations caused by excessively long physical observation distances, the fluctuation characteristics of target recognition probability are analyzed based on a preset sliding time window. This triggers a close-up re-shooting mechanism to obtain the timestamp to be analyzed and the target bounding box, assessing the confidence level of the visual model and the target scale. This allows the re-shooting process to be activated at the optimal time and the time reference of the environmental data to provide a basis for decision-making when leaving the fixed flight path.

[0038] To isolate the risk of time drift caused by continuous updates of radar data, and to discretize the complex non-convex environment into a controllable structure, the system retrieves and extracts the single-frame 3D point cloud closest to the timestamp to be analyzed from the data cache queue of the airborne lidar. Then, a 3D occupancy grid matrix is ​​constructed based on the single-frame 3D point cloud, and the local spatial structure of the tower at the trigger time is frozen as static data in the form of voxel occupancy.

[0039] Considering that high-voltage transmission towers are open truss structures with highly anisotropic interconnected forms in their internal non-convex free spaces, direct discrete point cloud data cannot intuitively reflect the connectivity of the channel. Therefore, distance transformation and skeleton extraction were performed on the three-dimensional occupancy grid matrix to obtain a spatial distance matrix and a free space topology map. This quantified the three-dimensional entity distribution and collision avoidance margin of the tower grid, and constructed a high-quality static road network base that is free from the influence of positioning drift, providing a safe geometric basis for subsequent topology feature extraction.

[0040] Simultaneously, the geometric center of the target bounding box in the image with the timestamp to be analyzed is extracted as the target pixel. The three-dimensional coordinates of the target are obtained by combining the camera intrinsic and extrinsic parameter matrices. The defect target in the two-dimensional image is projected and mapped to the three-dimensional static grid space, locking the accurate spatial orientation of the defect to be re-photographed in the local coordinate system, and providing a spatial anchor point for the determination of the destination of subsequent path finding.

[0041] Preferably, in one embodiment of the present invention, the frame rate of the video stream is 30 frames / second, the preset sliding window length is 1 second, the sliding step size is 1 frame, the system sequentially inputs each frame of image into the airborne target detection model for forward inference, obtains the detection result output by the model, the detection result includes the classification probability of a specific defect part and the target bounding box that defines the part; subsequently, the system extracts the pixel area occupied by the target bounding box output by the detection model in the image;

[0042] Considering that there are significant differences in the temporal distribution between instantaneous strong light interference and visual blur caused by excessive physical observation distance, the mean and variance of the target recognition probability within the preset sliding window are calculated.

[0043] When the mean is within the preset critical range, it indicates that the model's judgment of the target is in a state of ambiguity; at the same time, the variance is less than the preset fluctuation threshold, indicating that the model is not currently subjected to instantaneous strong interference, but has fallen into a continuous and stable state of fuzzy evaluation; at the same time, the pixel area of ​​the target bounding box is less than the preset observation scale threshold, indicating that the imaging size of the target in the image plane is lower than the lower limit of the pixel resolution required for effective identification. At this point, it is concluded that the current image blur is not caused by optical interference such as backlight, but is actually limited by the resolution due to the excessive physical distance of the observation, and the mechanism of close-up re-shooting is triggered.

[0044] As an example, the closer the mean of the target recognition probability is to 0.5, the more ambiguous the judgment. A preset critical interval is set as follows: The theoretical standard deviation of the binomial distribution centered at 0.5 is approximately 0.09 (with a sample size of 30 frames). This is narrowed to 0.05 as a strict judgment line to ensure that the detected confidence fluctuations belong to a continuous and stable low-discretion state rather than random jitter. Therefore, the preset fluctuation threshold is set to 0.05. The preset observation scale threshold is set to 1024 pixels. This threshold corresponds to a target imaging size of 32×32 pixels. Under the typical field of view of airborne inspection, this scale represents that the imaging projection of the target in the image plane is close to the lower limit of the pixel area for effective feature extraction by the convolutional neural network. Below this area, the spatial texture details in the deep feature map are severely attenuated.

[0045] The acquisition timestamp of the frame corresponding to the determination trigger time is used as the timestamp to be analyzed. The target bounding box corresponding to the timestamp to be analyzed is extracted, and the geometric center of the target bounding box in the image of the timestamp to be analyzed is used as the target pixel.

[0046] Furthermore, the system reads the absolute spatial position of the geometric center of the body at the time stamp to be analyzed and sets this position as the origin of the coordinate system. A local 3D coordinate system is constructed using the direction the aircraft's nose is facing at the time of analysis as the positive X-axis, the direction perpendicular to the X-axis and pointing to the left of the aircraft as the positive Y-axis, and the direction perpendicular to the XOY plane and pointing towards the zenith as the positive Z-axis (satisfying the right-hand rule). Then, existing coordinate translation and rotation matrix methods are used to transform the extracted single-frame 3D point cloud from the original coordinate system to this local 3D coordinate system.

[0047] For the converted single-frame 3D point cloud, the system first runs a reflection intensity filtering algorithm. This algorithm reads the laser reflectivity data attached to the point cloud data and identifies data with reflectivity below a preset threshold as free noise caused by dust or weak reflections from angle steel, and removes them. In this example, considering that man-made metal structures such as metal angle steel and buildings have high laser reflectivity due to their high surface conductivity (typically in the range of 80-250), the preset threshold is set to 50. This effectively filters out low reflectivity interference such as suspended dust particles and weak reflective noise while completely preserving the point cloud data of the iron tower's metal structure.

[0048] After the filtering operation is completed, the system divides the space where the local three-dimensional coordinate system is located into continuous discrete solid cubes, i.e., voxels, according to a preset fixed side length (e.g., 0.1 meters in this embodiment).

[0049] Specifically, with the origin of the coordinate system Starting from the vertex, and using a preset fixed side length as the step size, the system advances layer by layer along the X-axis, Y-axis, and Z-axis of the local three-dimensional coordinate system, so that the entire truncated space is completely covered by a continuous, non-overlapping, and gapless voxel array.

[0050] The system iterates through all discrete cubes, determining whether each cube contains filtered point cloud data points. For a cube containing at least one point cloud data point, the system assigns it a value... Marked as an obstacle element, indicating the presence of a metallic obstacle boundary at that location. For cubes that do not contain any point cloud data points, the system assigns them a value. The data is marked as an empty voxel, indicating that the location belongs to a safe, empty region. After traversing and assigning values ​​to all voxels, the data is packaged and encapsulated, and finally output as a 3D occupancy grid matrix representing the distribution of obstacles.

[0051] The spatial boundary of the 3D-occupied raster matrix is ​​determined by the minimum bounding box of a single-frame 3D point cloud in the local 3D coordinate system. In this example, to avoid the raster matrix size from getting out of control due to distant mountains and irrelevant background point clouds, the system uses the trigger position of the machine as the center and a cutoff radius of three times the preset observation radius (for example, when the preset observation radius is 2.0 meters, the cutoff radius is 6.0 meters). Only the point clouds within this spherical cutoff range are retained for raster construction, and point clouds outside the cutoff range are considered invalid backgrounds and are removed.

[0052] Next, to avoid the risk of absolute positioning drift, the target's three-dimensional coordinates are obtained by combining the camera's intrinsic and extrinsic parameter matrices. Specifically, considering that simple discrete point clouds are prone to penetration and missed detection by the camera's optical projection rays due to their sparse density, a spatial ray in the camera's local coordinate system is generated using the camera's intrinsic parameter matrix. Then, the spatial ray is transformed and mapped into a spatial projection ray in the local three-dimensional coordinate system by combining the extrinsic parameter matrix from the camera to the local three-dimensional coordinate system. The extrinsic parameter matrix is ​​obtained by pre-calibrating the airborne camera and the UAV body. The intrinsic and extrinsic parameter calibration and coordinate transformation applications of this type of camera are well-known technologies in the field.

[0053] Traverse the spatial projection ray along its extension direction to obtain the first voxel in the 3D occupied grid matrix through which the spatial projection ray passes, which is assigned the value of an obstacle. Take the center coordinates of this voxel as the target's 3D coordinates.

[0054] It should be noted that when traversing along the spatial projection ray to the spatial truncation boundary of the 3D occupancy grid matrix and still not acquiring a voxel assigned as an obstacle, the truncation radius is increased by a preset observation radius as the step size, and the point cloud is re-extracted to construct an expanded 3D occupancy grid matrix. Within the expanded 3D occupancy grid matrix, traversal calculations continue along the projection ray from the boundary point until the first voxel assigned as an obstacle is captured. If the truncation radius is increased to the maximum detection limit of the airborne radar and still no obstacle voxel is captured, coordinate acquisition is stopped and a safety fallback strategy of waypoint planning exit is triggered.

[0055] Finally, we obtain the spatial distance matrix and the free space topology graph, as an example:

[0056] Identify free voxels and obstacle voxels in the 3D occupancy grid matrix; calculate the 3D Euclidean distance from each voxel to the nearest obstacle voxel, and construct a spatial distance matrix from all the 3D Euclidean distance values. The spatial distance matrix has the same spatial dimension as the 3D occupancy grid matrix, the element value at the corresponding position of the obstacle voxel is 0, and each value within the spatial distance matrix represents the safety collision margin of the corresponding spatial point relative to the nearest angle steel structure.

[0057] Based on the spatial distance matrix, a three-dimensional topology refinement algorithm (i.e., a three-dimensional skeleton extraction algorithm) is called to obtain a set of central voxels that maintain three-dimensional connectivity. These retained sets of central voxels are connected end to end in space to form a generalized skeleton of the width of a single voxel. The set of central voxels is then converted into an undirected graph network composed of intersection nodes and connecting edges, which serves as the topology graph of the free space.

[0058] It should be noted that the 3D topology refinement algorithm, based on the principle of 26-neighborhood connectivity, peels away the boundary voxels with the smallest distance values ​​in the spatial distance matrix layer by layer. The peeling process starts from the edge near the obstacle and proceeds towards the interior of space until the set of central voxels necessary to maintain 3D connectivity is retained, which is a well-known technique in the art. In other embodiments of the present invention, the implementer can adjust multiple thresholds corresponding to the determination of triggering the proximity repeating mechanism according to the needs of the scene, and can also adjust the size of the 3D occupancy grid matrix.

[0059] It should be noted that during the period between triggering the analysis and generating and sending the new planned trajectory to the flight control system, the underlying flight control module of the system continues to read the real-time obstacle avoidance radar data to maintain the basic collision avoidance response mechanism and prevent the aircraft from flying blindly.

[0060] Step S2: Extract local voxel clusters from the nodes of the free space topology graph and perform 3D Gaussian smoothing. Calculate the second-order partial derivative Hessian matrix of the spatial distance and perform eigenvalue decomposition. Calculate the spatial channel extension morphology index. Combine the node values ​​in the spatial distance matrix to obtain a safe topology graph with passage weights.

[0061] Given that existing algorithms only use the shortest Euclidean distance as the path cost, they are highly susceptible to guiding drones blindly into dead ends that are surrounded by metal and are closed in one direction, despite having ample collision avoidance margins. Within a dead end, the drone's direction-finding module will experience severe signal abrupt changes due to signal reflections from the angle steel on both sides, leading to flight control malfunctions.

[0062] Therefore, this embodiment of the invention calculates the second-order partial derivative Hessian matrix of spatial distance and performs eigenvalue decomposition, introduces the second-order partial derivative of the three-dimensional distance field to evaluate the geometric shape, calculates the spatial channel extension shape index, quantifies the underlying three-dimensional geometric attenuation features into a macroscopic spatial anisotropic connectivity criterion, and combines the node values ​​in the spatial distance matrix to integrate the collision avoidance margin and geometric shape index into a node passage weight, thereby obtaining a safe topology map with passage weights, fundamentally avoiding the traditional obstacle avoidance algorithm that only relies on scalar distance to guide the drone into a dead end surrounded by metal.

[0063] Furthermore, considering that calculating the difference directly on the skeleton of the single-voxel width can easily produce severe singular noise, we first extract local voxel clusters from the nodes of the free space topology graph and perform three-dimensional Gaussian smoothing before analysis. This ensures that the eigenvalue decomposition of the second-order partial derivative matrix is ​​stable and convergent in numerical calculation, and avoids drastic changes in morphological indicators caused by small fluctuations in the underlying distance field data.

[0064] As a preferred approach, nodes in the free space topology graph are extracted one by one as nodes to be analyzed. The 3D coordinates of each node are extracted, and adjacent voxels of a preset size are extracted from the spatial distance matrix centered on these coordinates to construct a 3D data block, which serves as a local voxel cluster for the node to be analyzed. Specifically, when the extracted range exceeds the boundary of the spatial distance matrix, the voxel positions exceeding the boundary are filled with a value of 0 or by copying the value of their nearest boundary to prevent matrix index out-of-bounds errors and to obtain a complete 3D data block.

[0065] In this example, the default size is The system constructs a three-dimensional Gaussian filter kernel function (with a preset standard deviation of 0.5). The system uses this three-dimensional Gaussian filter kernel function to perform convolution smoothing calculations on the distance values ​​within local adjacent voxel clusters.

[0066] Through this smoothing calculation, the distance value of the central voxel absorbs the smoothing transition characteristics of the surrounding voxels, effectively eliminating the inherent numerical mutations and derivative singularities near the topological skeleton ridges. Then, the three-dimensional data block after smoothing calculation is denoted as the smoothed distance field voxel cluster.

[0067] Next, the system performs second-order central difference calculations along the three orthogonal directions of the X, Y, and Z axes, calculating the second-order independent partial derivatives of the smoothing distance in the three independent directions, as well as the second-order mixed partial derivatives in the three planes; the obtained partial derivative values ​​are then arranged sequentially and assembled into a single... The size matrix is ​​used as the Hessian matrix, which represents the second-order partial derivative of the spatial distance between the nodes to be analyzed.

[0068] The system calls the eigenvalue decomposition function to perform eigenvalue decomposition on the Hessian matrix of the second-order partial derivative of spatial distance, and obtains three real eigenvalues ​​representing the principal eigendirection. Then, the absolute values ​​of these three real eigenvalues ​​are taken to obtain the first, second, and third eigenvalues ​​in ascending order.

[0069] Further integrate the first, second, and third eigenvalues ​​to obtain spatial channel extension morphology indicators.

[0070] As an example, considering that when in a standard tubular or corridor-shaped space, the distance attenuation of the surrounding angle steel is symmetrical, the second characteristic value and the third characteristic value are very close, and the ratio is close to 1. Conversely, the ratio drops rapidly. Therefore, the second characteristic value is used as the numerator, the sum of the third characteristic value and the preset positive parameter divided by zero is used as the denominator, and the ratio of the fraction is used as the first index to characterize the symmetry of the channel cross section.

[0071] Considering that the longitudinal distance changes very gradually in a continuous channel open at both ends, the first eigenvalue approaches 0; considering that the geometric mean of the two eigenvalues ​​of the cross section can be used as the denominator to provide a physical benchmark scale representing the overall entity approximation rate of the channel cross section, a dimensionless relative rate of change can be constructed by comparing the longitudinal attenuation feature with it, so that the system can effectively measure the significance of longitudinal continuity compared with lateral closure.

[0072] Based on this, the first eigenvalue is used as the numerator, the arithmetic square root of the product of the second and third eigenvalues ​​plus the sum of the preset positive parameter divided by zero is used as the denominator, the ratio of the fraction is used as the independent variable x, and the value is substituted into the exponential function exp(-x) with the natural constant e as the base. The mapping value is used as the second index to characterize the longitudinal extension and connectivity of the channel.

[0073] Finally, the product of the first and second indicators is used as the spatial channel extension morphology index of the node to be analyzed. Only when the topological node is exactly located inside a through corridor with open ends and approximately symmetrical cross-section will the value of this index approach 1 to the maximum extent.

[0074] The addition of a preset division-to-zero positive parameter at two denominator positions ensures that even when the distance change eigenvalues ​​in all directions of space simultaneously approach zero in an open and flat airspace with no obstructions, the underlying program will not crash or deadlock due to a zero arithmetic denominator. The preset division-to-zero positive parameter is a very small constant with the same dimension as the eigenvalue, such as 0.01. This numerical magnitude can effectively mask the inherent numerical noise of discrete voxel-based calculations, preventing the noise in open areas from being amplified abnormally, and will not cause numerical interference to the high-significance distance gradient values ​​generated when approaching the metal grid.

[0075] In a preferred embodiment of the present invention, the data values ​​of nodes extracted from the spatial distance matrix and the spatial channel extension morphology index are combined to obtain the passage weight of the nodes.

[0076] As an example, the product of the data value of the spatial distance matrix of the node to be analyzed and the spatial channel extension morphology index is used as the passage weight. Through the multiplicative coupling relationship, the comprehensive passage weight of any node with a very small collision avoidance distance (with collision risk) or a very small morphology index (belonging to a semi-closed dead end) will be rapidly reduced.

[0077] After traversing all nodes to be analyzed and obtaining the passage weight, the network dead end truncation and cleanup mechanism is activated: the system reads the preset truncation threshold. If the passage weight is lower than the preset truncation threshold, it means that the node is either adjacent to the angle steel entity (insufficient anti-collision margin) or in a one-way closed metal-enclosed space (the geometry does not have the characteristics of a through corridor). Either one or both of these conditions do not meet the conditions for safe close-range flight of the UAV.

[0078] Therefore, nodes and (directly connected) edges with passage weights lower than the preset truncation threshold are deleted from the free space topology graph, and a safe topology graph with passage weights is output, or simply a safe topology graph.

[0079] The preset cutoff threshold is limited by the product of the equivalent collision radius of the body and the morphological safety threshold. In this example, the equivalent collision radius of the body is set to 0.3 meters and the morphological safety threshold is set to 0.5. In this example, the passage weight is obtained by multiplying the anti-collision distance (in meters) and the spatial passage extension morphological index (dimensionless). Therefore, the dimension of the weight value is consistent with the distance, and the value is 0.15.

[0080] Step S3: Based on the safety topology map, perform unobstructed detection by combining the target's three-dimensional coordinates with the preset observation radius to obtain the pathfinding endpoint. Construct a search target node sequence with the access weight, generate an initial interpolation curve based on the target node sequence, and perform anti-collision offset correction on the initial interpolation curve along the distance gradient direction of the spatial distance matrix to generate a continuous pose sequence.

[0081] Considering that the underlying dynamics control needs to execute instructions smoothly, coherently, and with absolute safety to avoid cost computation deadlock and sudden changes in fuselage attitude, it is necessary to generate continuous pose sequences.

[0082] Considering that the nodes in the safety topology map are skeleton nodes with a single element width, which only represent the theoretical unobstructed centerline, directly using them as flight trajectories cannot meet the physical size constraints and dynamic smoothness requirements of the fuselage. Therefore, based on the safety topology map, we combine the target's three-dimensional coordinates with the preset observation radius to perform unobstructed detection, obtain the pathfinding endpoint, establish a clear boundary for subsequent trajectory search, and ensure the spatial accessibility of the endpoint through unobstructed detection.

[0083] Next, a sequence of target nodes for optimization cost search is constructed using passage weights. This allows the path search algorithm to prioritize connected corridors with optimal through-flow morphology and the largest collision avoidance space as candidate paths. An initial interpolation curve is generated based on the target node sequence to eliminate sharp turns between waypoints. Collision avoidance offset correction is then performed on the initial interpolation curve along the distance gradient direction of the spatial distance matrix. This ensures that the trajectory is always constrained within a space with sufficient collision avoidance margin, while maintaining the continuity of the entire curve. A continuous pose sequence is generated, balancing trajectory smoothness and safety, and guaranteeing high control robustness during the complex network traversal process.

[0084] Preferably, in one embodiment of the present invention, please refer to Figure 2 The flowchart illustrates a method for obtaining the endpoint of a pathfinding process according to an embodiment of the present invention, specifically including:

[0085] Step S301: Using the origin of the local three-dimensional coordinate system as the starting point for pathfinding, follow the spatial straight line segment pointing from the target three-dimensional coordinates to the starting point for pathfinding, and backtrack by a preset observation radius to obtain the coordinates of the pre-selected endpoint.

[0086] In this example, a pathfinding reference anchor point is established using the target's three-dimensional coordinates, and the origin of the local three-dimensional coordinate system is used as the starting point for the UAV's pathfinding flight. The two are connected to form a spatial straight line segment. Then, the target's three-dimensional coordinates are used as the starting point for the backtracking operation. The UAV backtracks along this line segment from the target's three-dimensional coordinates towards the origin by a distance of a preset observation radius to obtain the coordinates of the pre-selected endpoint. This ensures that the center of the airborne camera's field of view remains aligned with the target and does not deviate from the observation optical axis, while also ensuring that the UAV maintains sufficient physical collision avoidance space with the metal entity when hovering.

[0087] The preset observation radius is set based on the outer dimensions of the aircraft rotor and the minimum clear focusing distance of the airborne camera. In this example, the value is 2.0 meters. This distance is greater than the collision envelope of a typical aircraft of about 1 meter and is within the optimal focal length range for high-resolution imaging by the visual sensor.

[0088] Step S302: In the safety topology map, find the node that is closest to the pre-selected endpoint coordinates in Euclidean distance and whose spatial connection with the target's three-dimensional coordinates does not intersect with the obstacle object in the three-dimensional grid matrix, and use it as the pathfinding endpoint.

[0089] Considering that in a dense array of angle steel trusses, the node with the closest Euclidean distance to the pre-selected endpoint coordinates may be located behind a certain angle steel (on the side facing away from the target), if the underlying flight control guides the drone to fly to that node and hover for reshooting, the drone's lens will be completely blocked by the angle steel, making it impossible to capture the defective target, thus causing the entire reshooting mission to fail completely. Therefore, instead of directly using the pre-selected endpoint coordinates as the pathfinding endpoint, a point is found nearby that has the closest Euclidean distance to the pre-selected endpoint coordinates, and the spatial connection between the point and the target's three-dimensional coordinates does not intersect with the obstacle object pixels in the three-dimensional occupied grid matrix.

[0090] As an example, calculate the three-dimensional Euclidean distance between all nodes in the security topology graph and the coordinates of the pre-selected endpoint, and generate a queue of candidate nodes by sorting them in ascending order of distance.

[0091] Nodes in the candidate node queue are extracted sequentially, and a 3D line segment from the node to the target's 3D coordinates is constructed. All spatial voxels traversed by the 3D line segment in the local 3D coordinate system are extracted. If all voxels traversed have a value of 0 in the 3D occupancy grid matrix, it is determined that there is no occlusion, the node is established as the pathfinding endpoint, and the search is terminated; if there is an obstacle voxel with a value of 1 among the traversed voxels, the node is removed, and the next node in the candidate node queue is extracted to continue the detection.

[0092] It should be noted that if the traversal distance exceeds the preset maximum deviation radius (e.g., set to 1.5 meters) and no node that meets the conditions is found, the target is determined to be a visually unreachable dead zone, a pathfinding failure flag is output, and the safety fallback strategy of the aircraft exiting along the original offline route is triggered.

[0093] Furthermore, the system utilizes the nearest neighbor algorithm to find the node closest to the pathfinding start point in the safe topology graph and use it as the graph start point, transforming continuous coordinates into nodes. The pathfinding end point is already a node during the search, so it is directly used as the graph end point.

[0094] After the transformation is completed, the existing Dijkstra shortest path search algorithm is enabled on the safe topology graph. During the step traversal of the search algorithm, considering that the closer the basic physical distance between adjacent nodes is and the greater the passage weight of the node relative to the destination, the lower the basic passage cost is. At the same time, the nodes are in a corridor structure that is open in space and connected at both ends. Therefore, based on the basic physical distance and passage weight between adjacent nodes, the single-step optimization cost is obtained, and the target node sequence with the lowest comprehensive cost is searched.

[0095] For example, for adjacent nodes A and B, B is the relative endpoint from A to B; the Euclidean distance between the two voxel centers of the adjacent nodes is used as the basic physical distance, the basic physical distance is used as the numerator, the passage weight of the relative endpoint node is used as the denominator, and the ratio of the fractions is used as the corresponding single-step optimization cost; since the passage weight of the node is always positive (obtained by multiplying the collision avoidance distance and the morphological index, both of which are positive), the denominator is not 0.

[0096] The system uses Dijkstra's shortest path search algorithm to perform a traversal search. If a valid connected path exists from the start point to the end point of the graph, the node path sequence with the lowest overall cost (sum of single-step optimization costs) is selected as the target node sequence. If no path exists, the system outputs a pathfinding failure flag and triggers a safe return strategy where the UAV hovers in place and exits along the original offline flight path, serving as a bottom-line safety protection.

[0097] After receiving the target node sequence, the system performs spline interpolation on the target node sequence to generate an initial interpolation curve, and fits the discrete piecewise line sequence to generate an initial interpolation curve that eliminates sharp corner breaks;

[0098] As an example, the system reads the aircraft's flight limit thresholds and inputs them, along with the target node sequence, into a preset B-spline curve interpolation module. The interpolation module generates an initial interpolation curve while satisfying the upper limits of the aircraft's dynamic velocity and acceleration.

[0099] The B-spline curve interpolation module employs a cubic uniform B-spline curve fitting algorithm. Using each node in the target node sequence as a control point, it generates a third-order continuous parametric curve through recursive calculation of the basis function. Each segment of the curve is determined by four adjacent control points, ensuring that the curve passes through the first and last control points and that the overall curvature is continuous. The system uses the upper limits of the airframe's velocity and acceleration as constraints. By adjusting the node vectors and time allocation of the B-spline curve, it ensures that the velocity and acceleration at any point on the curve do not exceed the preset limit thresholds, thereby meeting the dynamic feasibility requirements of the underlying flight control system.

[0100] Considering that the initial interpolation curve may intersect with obstacle objects in the 3D occupied grid matrix, it cannot be directly used as the final flight command. Discrete point collision detection is also required, specifically including:

[0101] When the initial interpolation curve intersects with an obstacle voxel, the distance gradient vector of the central voxel of the intersection region is extracted from the spatial distance matrix. Specifically, the distance values ​​of the adjacent voxels of the central voxel along the X-axis, Y-axis, and Z-axis in the spatial distance matrix are retrieved, and the distance differences in the three directions are calculated respectively. These are then combined and normalized into a three-dimensional gradient vector. Subsequently, a positive offset operation is performed on the local control points of the initial interpolation curve along the ascending direction of the distance gradient to reconstruct a collision-free continuous curve. The collision-free continuous curve is then discretely sampled to generate a continuous pose sequence.

[0102] The forward offset operation includes iteratively translating local control points intersecting with obstacles on the initial interpolation curve along the ascending direction of the distance gradient with a preset safety step size. After each translation, spline interpolation and discrete point collision detection are re-executed until all sampling points on the reconstructed curve correspond to free voxels in the 3D occupied grid matrix. In this example, the preset safety step size is set based on the voxel side length of the 3D occupied grid matrix, and is less than or equal to half of the voxel side length to prevent the offset step from being too large and exceeding the limit.

[0103] Discrete sampling includes: combining the set upper limit of the aircraft's flight speed, and according to the preset control frequency of the underlying flight control unit, extracting a set of three-dimensional coordinate nodes with time-series timestamps along a collision-free continuous curve at equal time intervals, as a continuous pose sequence. In this example, the underlying flight control frequency is 50 times per second.

[0104] If the initial interpolation curve does not intersect with any obstacle object after discrete point collision detection, there is no need to perform offset correction. The initial interpolation curve can be directly used as a collision-free continuous curve, and a continuous pose sequence can be generated after discrete sampling.

[0105] In step S4, the underlying flight control unit receives a continuous pose sequence, drives the aircraft to approach the pathfinding endpoint and hover.

[0106] In order to realize the closed-loop transformation of dynamic programming strategy into execution driven by underlying physics, the underlying flight control unit receives a continuous pose sequence, drives the aircraft to approach the pathfinding endpoint and hover, and promotes the automated approach and reshooting task.

[0107] In a preferred embodiment of the present invention, the underlying flight control unit is an airborne flight controller (such as PX4, ArduPilot, or other open-source flight controllers or customized autopilot systems). It can analyze the position, attitude, and timing information in the continuous pose sequence according to a preset internal control law (such as a PID control law or a model predictive control algorithm) to generate corresponding motor speed control quantities to control the UAV's flight; this is already a prior art method. After the underlying flight control unit completes execution according to the continuous pose sequence, the UAV reaches the pathfinding endpoint (i.e., the spatial hovering point corresponding to the pre-selected endpoint coordinates). After hovering, the system controls the airborne camera to acquire a clear real-time image of the current viewpoint and reactivates the airborne target detection model for target inference. The system determines whether the newly output classification probability is greater than the system's preset target recognition probability release threshold (e.g., preset to 0.85, which can be adjusted according to the scenario requirements).

[0108] If the determination result is true, it indicates that the resolution bottleneck has been resolved after the distance is reduced, and the reshoot defect has been diagnosed. The system sends a release command to the memory to clear the previously fixed local 3D coordinate system and 3D occupancy grid matrix and other static cached data. The system reloads the previously paused offline global flight path sequence, directs the drone to leave the tower area and fly to the subsequent site, completing the exit loop of the close-up reshoot planning mechanism.

[0109] If the judgment result is false, it indicates that the current approach position has not yet reached the ideal observation conditions. The system will re-trigger the approach re-photographing mechanism, re-execute the planning process of steps S1 to S4, and guide the UAV to further adjust its hovering position until the classification probability meets the release conditions, or if it is repeatedly triggered multiple times (e.g., 5 times), it will be determined that the defective target cannot be effectively re-photographed under the current mission conditions due to objective physical limitations (e.g., the defect size is too small to exceed the physical resolution limit of the airborne optical system, or the defect is located in a blind spot area blocked by the angle steel structure, etc.). The system will terminate the re-photographing attempt at that point, upload a "re-photographing unconfirmed" status flag with timestamp and spatial coordinates to the ground station, release the static cache data, reload the offline global flight path sequence, and command the UAV to leave the tower area and continue to perform routine inspection tasks at subsequent stations. The ground station will then conduct manual analysis or arrange special approach flight missions based on the returned data.

[0110] In summary, to address the technical problem of UAV attitude control instability during close-range flight in complex, non-convex power grids due to discontinuous trajectory planning and easy entry into semi-enclosed blind spots, this invention aims to provide a dynamic planning method for autonomous waypoints of UAVs in complex mountainous power grids. When the close-range re-shooting mechanism is triggered, this invention acquires the timestamp to be analyzed, obtains the spatial distance matrix and the topology map of the available space; further, it acquires the three-dimensional coordinates of the target; further, it calculates the spatial channel extension morphology index of the nodes in the available space topology map, and combines the node values ​​in the spatial distance matrix to obtain a safe topology map with passage weights; further, it constructs an optimization cost search for the target node sequence using the passage weights, generates an initial interpolation curve based on the target node sequence, and performs anti-collision offset correction to generate a continuous pose sequence; finally, it drives the UAV to approach the pathfinding endpoint and hovers. This invention calculates morphology indices through partial derivative matrix eigenvalue decomposition, constructs a safe topology map and searches for the target node sequence, performs offset correction on the interpolation curve to generate a continuous pose sequence, reducing attitude control instability caused by entering semi-enclosed blind spots and trajectory discontinuities.

Claims

1. A dynamic planning method for autonomous waypoints of unmanned aerial vehicles (UAVs) used in complex power grids in mountainous areas, characterized in that, The method includes: Based on the fluctuation characteristics of the target recognition probability analyzed by a preset sliding window, a close-up reshooting mechanism is triggered to obtain the timestamp to be analyzed and the target bounding box; a three-dimensional occupancy grid matrix is ​​constructed based on a single frame three-dimensional point cloud, and distance transformation and skeleton extraction are performed to obtain a spatial distance matrix and a free space topology map; the geometric center of the target bounding box in the image with the timestamp to be analyzed is extracted as the target pixel point, and the three-dimensional coordinates of the target are obtained by combining the camera intrinsic parameter matrix and extrinsic parameter matrix; Local voxel clusters are extracted from the nodes of the free space topology graph and subjected to three-dimensional Gaussian smoothing. The second-order partial derivative of the spatial distance Hessian matrix is ​​calculated and eigenvalue decomposition is performed. The spatial channel extension morphology index is calculated. Combined with the value of the node in the spatial distance matrix, a safe topology graph with passage weight is obtained. Based on the security topology map, unobstructed detection is performed by combining the target's three-dimensional coordinates with the preset observation radius to obtain the pathfinding endpoint. The target node sequence is constructed using the passage weight, and an initial interpolation curve is generated based on the target node sequence. Collision avoidance offset correction is performed on the initial interpolation curve along the distance gradient direction of the spatial distance matrix to generate a continuous pose sequence. The underlying flight control unit receives the continuous pose sequence, drives the aircraft to approach the pathfinding endpoint and hover.

2. The method for dynamic planning of waypoints for autonomous inspection by unmanned aerial vehicles (UAVs) in complex power grids in mountainous areas, as described in claim 1, is characterized in that... The triggering determination of the close-up reshooting mechanism includes: Calculate the mean and variance of the target recognition probability within the preset sliding window; when the mean is in a preset critical range, the variance is less than a preset fluctuation threshold, and the pixel area of ​​the target bounding box is less than a preset observation scale threshold, determine that the close-up re-shooting mechanism is triggered.

3. The method for dynamic planning of waypoints for autonomous inspection by unmanned aerial vehicles (UAVs) in complex power grids in mountainous areas, as described in claim 1, is characterized in that... The method for obtaining the spatial channel extension morphology index includes: The spatial distance second-order partial derivative Hessian matrix is ​​decomposed into eigenvalues ​​and its absolute value is taken to obtain the first, second, and third eigenvalues ​​arranged in ascending order; the first, second, and third eigenvalues ​​are fused to obtain the spatial channel extension morphology index.

4. The method for dynamic planning of waypoints for autonomous inspection by unmanned aerial vehicles (UAVs) in complex power grids in mountainous areas, as described in claim 1, is characterized in that... The method for obtaining the security topology map includes: The data values ​​of nodes extracted from the spatial distance matrix are combined with the spatial channel extension morphology index to obtain the passage weight of the nodes; nodes and connecting edges with passage weights lower than a preset truncation threshold in the free space topology map are deleted, and the safe topology map with passage weights is output.

5. The method for dynamic planning of waypoints for autonomous inspection by unmanned aerial vehicles (UAVs) in complex power grids in mountainous areas, as described in claim 1, is characterized in that... The method for obtaining the pathfinding endpoint includes: Using the origin of the local three-dimensional coordinate system as the starting point for pathfinding, the spatial straight line segment pointing from the target three-dimensional coordinates to the starting point of pathfinding is backed by a preset observation radius to obtain the coordinates of the pre-selected endpoint. The node that is closest to the Euclidean distance of the pre-selected endpoint coordinates in the safety topology map and whose spatial connection with the target three-dimensional coordinates does not intersect with the obstacle object element in the three-dimensional occupied grid matrix is ​​selected as the endpoint of pathfinding.

6. The method for dynamic planning of waypoints for autonomous inspection by unmanned aerial vehicles (UAVs) in complex power grids in mountainous areas, as described in claim 1, is characterized in that... The method for obtaining the continuous pose sequence includes: Based on the basic physical distance between adjacent nodes and the passage weight in the safety topology graph, the single-step optimization cost is obtained, and the target node sequence with the lowest comprehensive cost is searched. Spline interpolation is performed on the target node sequence to generate the initial interpolation curve. When the initial interpolation curve intersects with an obstacle voxel, the distance gradient vector of the center voxel of the intersection region is extracted from the spatial distance matrix. A positive offset operation is performed on the local control points of the initial interpolation curve along the rising direction of the distance gradient to reconstruct a collision-free continuous curve. Discrete sampling is performed on the collision-free continuous curve to generate the continuous pose sequence.

7. The method for dynamic planning of waypoints for autonomous inspection by unmanned aerial vehicles (UAVs) in complex power grids in mountainous areas, as described in claim 1, is characterized in that... The method for obtaining the local voxel clusters includes: Extract the three-dimensional coordinates of the nodes in the free space topology graph, and construct a three-dimensional data block by extracting adjacent voxels of a preset size from the spatial distance matrix using the three-dimensional coordinates as the center, as the local voxel cluster.

8. The method for dynamic planning of waypoints for autonomous inspection by unmanned aerial vehicles (UAVs) in complex power grids in mountainous areas, as described in claim 1, is characterized in that... The method for obtaining the spatial distance matrix includes: Identify the free voxels and obstacle voxels in the three-dimensional occupied grid matrix; calculate the three-dimensional Euclidean distance from each voxel to the nearest obstacle voxel, and construct the spatial distance matrix from all the three-dimensional Euclidean distance values.

9. A dynamic planning method for autonomous waypoints of unmanned aerial vehicles (UAVs) for complex power grids in mountainous areas, as described in claim 1, is characterized in that... The method for obtaining the free space topology map includes: Based on the spatial distance matrix, a three-dimensional topology refinement algorithm is invoked to obtain a set of central voxels that maintain three-dimensional connectivity; the set of central voxels is then converted into an undirected graph network consisting of intersecting nodes and connecting edges, which serves as the free space topology graph.

10. A dynamic planning method for autonomous waypoints of unmanned aerial vehicles (UAVs) for complex power grids in mountainous areas, as described in claim 1, is characterized in that... The method for obtaining the target's three-dimensional coordinates includes: The target pixel is mapped to a spatial projection ray in a local three-dimensional coordinate system using the camera intrinsic parameter matrix. Traversal calculations are performed along the extension direction of the spatial projection ray to obtain the first voxel in the three-dimensional occupied grid matrix that the spatial projection ray passes through, which is assigned the value of an obstacle. The center coordinates of this voxel are taken as the three-dimensional coordinates of the target.

Images