Unmanned aerial vehicle route planning method and system based on lightweight artificial intelligence

Abstract

The invention relates to an unmanned aerial vehicle route planning method and system based on lightweight artificial intelligence, and the method comprises the following steps: S1, obtaining the environment data of a distribution line, and fusing the data of multiple sensors in real time; s2, constructing an environment map based on the fused multi-sensor data; s3, key points are extracted from the environment map to serve as path nodes, the number of the nodes is reduced through a clustering algorithm, and a three-dimensional environment map with the key path nodes is obtained; s4, optimizing path calculation by adopting an improved A * algorithm and utilizing a heuristic search strategy according to the electric power line inspection plan, and constructing a path plan; and S5, deploying an edge terminal at the key node, and triggering real-time data storage and transmission only when necessary by adopting an event-based detection algorithm. Through the lightweight artificial intelligence algorithm design, the method can efficiently operate on an unmanned aerial vehicle platform with limited computing resources, and can effectively execute the inspection task in a complex distribution line environment at the same time.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle (UAV) route planning, and in particular to a UAV route planning method and system based on lightweight artificial intelligence. Background Art

[0002] In contemporary society, with the increasing complexity and scale of power infrastructure, the inspection task of distribution lines has become more important and urgent. Traditional manual inspection faces many challenges such as low efficiency, high cost, and environmental hazards. Drones have been widely used in the inspection of power facilities due to their flexibility and efficiency. Their advantage is that they can quickly patrol a large area, and at the same time, they can achieve refined equipment inspection through high-definition cameras and sensor detection, helping to promptly discover and report line faults or potential safety hazards. However, in order to operate autonomously and safely in a complex distribution line environment, drones need to have highly intelligent and adaptive route planning technology. Summary of the invention

[0003] In order to solve the above problems, the purpose of the present invention is to provide a UAV route planning method based on lightweight artificial intelligence. Through the design of lightweight artificial intelligence algorithm, it can run efficiently on a UAV platform with limited computing resources, extend the UAV flight time, and effectively complete the inspection tasks in a complex distribution line environment.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] A method for UAV route planning based on lightweight artificial intelligence includes the following steps:

[0006] Step S1: Use the drone equipped with a camera and a lidar to obtain distribution line environmental data and fuse multi-sensor data in real time;

[0007] Step S2: constructing an environment map based on the fused multi-sensor data;

[0008] Step S3: extract key points from the environment map as path nodes, reduce the number of nodes by clustering algorithm, and obtain a three-dimensional environment map with key path nodes;

[0009] Step S4: According to the power line inspection plan, an improved version of the A* algorithm is used, based on the three-dimensional environment map with key path nodes obtained in step S2, and a heuristic search strategy is used to optimize the path calculation, reduce the amount of calculation, and construct a path plan;

[0010] Step S5: Edge terminals are deployed at key nodes, and event-based detection algorithms are used to accurately identify and respond to key environmental changes. UAVs trigger real-time data storage and transmission only when necessary.

[0011] Furthermore, multi-sensor data is fused in real time, based on the particle filter method, to continuously update the drone's estimate of the environmental state, as follows:

[0012] Assign N particles to the possible state space, each particle Have the same initial weights

[0013] For each particle, the state at the next moment is predicted based on the motion model:

[0014]

[0015] Where f is the state transition function, u k is the control input at time k; is the process noise;

[0016] Update the weight of each particle through sensor observation data:

[0017]

[0018] Among them, z k is the sensor observation value, including the fused observation data from the camera and lidar;

[0019] is the observation model, which represents the particle state Observe z k probability;

[0020] According to the updated weights, particles are resampled based on system resampling to reduce degradation;

[0021] Combine all updated and resampled particles to estimate the current environment state:

[0022] The current state estimate is obtained by calculating the weighted average of the particle states:

[0023]

[0024] Furthermore, an environmental map is constructed based on the fused multi-sensor data, specifically:

[0025] Use timestamps and synchronization algorithms to ensure that camera images and lidar point cloud data are aligned in time;

[0026] Based on the image data collected by the camera, the ORB algorithm is used to extract key points and generate descriptors;

[0027] Calculate curvature and plane geometry from LiDAR point cloud data and use normal estimation algorithms to determine the characteristics of the tangent plane at each point.

[0028] Use descriptors to match feature points between adjacent frames, connect images and point cloud data at different time points through feature descriptors, and combine RANSAC algorithm to remove outliers and improve matching accuracy;

[0029] Using the matched feature points, we solve the pose transformation through PnP, integrate the visual odometry calculation, and convert the feature matching results into robot motion estimation;

[0030] Use the ICP algorithm to align point cloud data and refine the LiDAR pose estimation;

[0031] The pose of each frame is cumulatively decomposed, the newly acquired point cloud data is converted to the global coordinate system, and merged to obtain the environment map.

[0032] Furthermore, by using the matched feature points, the pose transformation is solved through PnP, and the visual odometry calculation is integrated to convert the feature matching results into robot motion estimation, as follows:

[0033] 2D image feature points (x) extracted and matched from adjacent frames i′ ,y i′ ) and the corresponding 3D space point (X i′ ,Y i′ ,Z i′ );

[0034] Use the camera's intrinsic matrix K to convert 2D image coordinates into 3D rays;

[0035] Solve the posture transformation and projection relationship through PnP:

[0036]

[0037] Where R and t represent the rotation matrix and translation vector; s is the scaling factor;

[0038] Solve for the optimal R and t so that the reprojection error between the predicted 2D points and the actual 2D points is minimized:

[0039]

[0040] Among them, π is the projection function, which converts 3D coordinates into 2D coordinates, and p i′ is the i′th 2D point actually observed; a nonlinear optimization method is used to solve this problem;

[0041] Integrated visual odometry calculation, the pose transformation [Rk, tk] calculated for the current frame relative to the previous frame is added to the entire motion trajectory:

[0042] [R k+1 ,tk+1 ]=[R k ,t k ]·[R k-1 ,t k-1 ];

[0043] The pose change between consecutive frames can be simply described by solving PnP:

[0044] ΔT k =[R k ,t k ];

[0045] Use Kalman filtering to smooth the accumulated results, consider optimizing the estimation of position and attitude, and use bundle adjustment to optimize the global pose and map points when encountering loop closure.

[0046] Furthermore, when loop closure is encountered, bundle adjustment is used to optimize the global pose and map points, as follows:

[0047] The optimization objective is to optimize the camera pose and the 3D point coordinates it observes at the same time, and the error function is constructed to minimize the reprojection error:

[0048]

[0049] Where R i″ represents the rotation matrix of the i″th camera pose; t i″ represents the translation vector of the i″th camera pose; X j represents the coordinates of the jth 3D point in the world coordinate system; π projection function; ρ robustness loss function; q i″j represents the 2D image coordinates of the jth feature point observed in the i″th camera;

[0050] By minimizing the reprojection error, all camera poses and 3D point positions are optimized, and the Levenberg-Marquardt method is used to iteratively adjust the parameters to achieve the minimum error.

[0051] Furthermore, the Levenberg-Marquardt method is used to iteratively adjust the parameters to achieve the minimum error, as follows:

[0052] Starting with reasonable initial guesses for all existing poses and map points, evaluate all reprojection errors and the newly added loop closure constraint errors;

[0053] The updated value is obtained by solving the following linear equation:

[0054] (J T WJ+λ′I)Δ=-J T We;

[0055] Where J is the Jacobian matrix; W is the weight matrix; λ′ is the damping factor; e is the cumulative error vector; Δ is the change;

[0056] Update the state variables:

[0057] x new =x old +Δ;

[0058] Among them, x new is the new state variable; x old is the original state variable;

[0059] The damping factor λ′ is dynamically adjusted to adapt to the error reduction speed: when the error decreases to a threshold, the iteration is stopped.

[0060] Further, step S3 is specifically:

[0061] Based on Harris corner points, key feature points are extracted from the constructed environment map as potential path nodes, the constructed environment map is grayed out, and the gradient of the environment map in the x and y directions is calculated using the Sobel operator.

[0062] Compute the structure tensor M around each pixel:

[0063]

[0064] in, Represents the sum of squared gradients of all pixels in the local neighborhood of the image in the x direction; Represents the sum of squared gradients of all pixels in the local neighborhood of the image in the y direction; ∑I x I y Represents the product of the gradients of pixels in the x and y directions within the local neighborhood of the image;

[0065] The Harris response function is used to calculate the corner response value R:

[0066] R = det(M) - k·trace(M) 2 ;

[0067] Where det(M) is the matrix determinant, trace(M) is the trace of the matrix, and k is an empirical constant;

[0068] The clustering algorithm DBSCAN is used to cluster potential path nodes, reduce the number of nodes, simplify the complexity of path calculation, and obtain a simplified environment map containing key path nodes.

[0069] Furthermore, step S4 is specifically as follows:

[0070] Set the start and end node sets, and mark the start and end locations according to the inspection plan;

[0071] Combined with the three-dimensional environment map with key path nodes constructed in step S3, the map includes key path nodes and topological structures, the three-dimensional environment map is allowed to expand to six directions from each node, the edges between nodes represent feasible direct paths, and the path cost function g(n) represents the accumulated actual cost from the starting node to the current node n;

[0072] The heuristic function h(n) estimates the minimum cost from the current node n to the nearest destination node:

[0073] h(n)=|x goal -x n ·+|y goal -y n |+|z goal -z n |;

[0074] Among them, (x goa l,y goal ,z goal ) is the coordinate of the target node; (x n ,y n ,z n ) is the coordinate of the current node n;

[0075] Use the minimum value of the heuristic function to select the path node and calculate f(n)=g(n)+h(n) to optimize the path;

[0076] Based on the optimized node selection, the path is gradually generated and adjusted to output the optimal path planning result.

[0077] Furthermore, S5 is specifically:

[0078] Deploy edge computing terminals at key locations of power lines, including substations and important line intersections, to handle preliminary data processing, event detection, and communication with drones;

[0079] Event detection sets multi-level thresholds according to different event types, comprehensively presets data at several moments to reduce noise, and is used to improve detection accuracy. It analyzes the rate of change within the sliding window based on a decision tree to identify events.

[0080] When the edge terminal detects that the event conditions are met, the drone begins to record and transmit real-time data; an optimized wireless transmission protocol is used to reduce data transmission delays; after the event is triggered, a priority queue is used to manage the transmission of different types of data.

[0081] A drone route planning system based on lightweight artificial intelligence includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically executes the steps in the drone route planning method based on lightweight artificial intelligence as described above.

[0082] The present invention has the following beneficial effects:

[0083] 1. The present invention uses a lightweight artificial intelligence algorithm design to efficiently run on a drone platform with limited computing resources, extending the drone's flight time, and can effectively complete inspection tasks in a complex distribution line environment;

[0084] 2. The ORB descriptor of the present invention is used in combination with RANSAC to make the feature matching process more robust, effectively remove false matches, and continuously track the motion path of the camera, converting the local matching results into global motion estimation for navigation and map construction. In addition, local dynamic filtering and smoothing are achieved through the Kalman filter, and global consistency optimization is achieved through bundling adjustment, which can effectively improve the accuracy and robustness of trajectory estimation and help solve real-time positioning and mapping problems in complex and dynamic environments.

[0085] 3. The present invention uses an improved A* algorithm combined with a heuristic strategy to significantly shorten the calculation time while ensuring the optimal path, respond to dynamic changes more quickly, adjust the inspection route in real time, ensure the smooth progress of the inspection task, and trigger data storage and transmission in real time when the environment changes or key events occur through edge terminal deployment and event detection algorithms. This reduces unnecessary data communication, saves communication bandwidth and energy consumption, and ensures system efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0087] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0088] refer to Figure 1 In this embodiment, a method for UAV route planning based on lightweight artificial intelligence is provided, comprising the following steps:

[0089] Step S1: Use the drone equipped with a camera and a lidar to obtain distribution line environmental data and fuse multi-sensor data in real time;

[0090] Step S2: constructing an environment map based on the fused multi-sensor data;

[0091] Step S3: extract key points from the environment map as path nodes, reduce the number of nodes by clustering algorithm, and obtain a three-dimensional environment map with key path nodes;

[0092] Step S4: According to the power line inspection plan, an improved version of the A* algorithm is used, based on the three-dimensional environment map with key path nodes obtained in step S2, and a heuristic search strategy is used to optimize the path calculation, reduce the amount of calculation, and construct a path plan;

[0093] Step S5: Edge terminals are deployed at key nodes, and event-based detection algorithms are used to accurately identify and respond to key environmental changes. UAVs trigger real-time data storage and transmission only when necessary.

[0094] In this embodiment, multi-sensor data is fused in real time, and the sensor data is fused based on the particle filter method to continuously update the drone's estimate of the environmental state, as follows:

[0095] Assign N particles to the possible state space, each particle Have the same initial weights

[0096] For each particle, the state at the next moment is predicted based on the motion model:

[0097]

[0098] Where, f is the state transfer function (based on the UAV's dynamic model), u k is the control input at time k (the speed and direction of the drone); is the process noise;

[0099] Update the weight of each particle through sensor observation data:

[0100]

[0101] Among them, z k is the sensor observation value, including the fused observation data from the camera and lidar;

[0102] is the observation model, which represents the particle state Observe z k probability;

[0103] According to the updated weights, the particles are resampled based on the system resampling to reduce degradation (i.e., the particle weights differ too much);

[0104] Combine all updated and resampled particles to estimate the current environment state:

[0105] The current state estimate is obtained by calculating the weighted average of the particle states:

[0106]

[0107] In this embodiment, an environment map is constructed based on the fused multi-sensor data, specifically:

[0108] Use timestamps and synchronization algorithms to ensure that camera images and lidar point cloud data are aligned in time;

[0109] Based on the image data collected by the camera, the ORB algorithm is used to extract key points and generate descriptors;

[0110] Calculate curvature and plane geometry from LiDAR point cloud data and use normal estimation algorithms to determine the characteristics of the tangent plane at each point.

[0111] Use descriptors (such as BRIEF) to match feature points between adjacent frames, connect images and point cloud data at different time points through feature descriptors, and combine RANSAC algorithm to remove outliers and improve matching accuracy;

[0112] Using the matched feature points, we solve the pose transformation through PnP, integrate the visual odometry calculation, and convert the feature matching results into robot motion estimation;

[0113] Use the ICP algorithm to align point cloud data and refine the LiDAR pose estimation;

[0114] The pose of each frame is cumulatively decomposed, the newly acquired point cloud data is converted to the global coordinate system, and merged to obtain the environment map.

[0115] In this embodiment, the matched feature points are used to solve the pose transformation through PnP, integrate the visual odometry calculation, and convert the feature matching results into robot motion estimation, as follows:

[0116] 2D image feature points (x) extracted and matched from adjacent frames i′ ,y i′ ) and the corresponding 3D space point (X i′ ,Y i′ ,Z i′ );

[0117] Use the camera's intrinsic matrix K to convert 2D image coordinates into 3D rays;

[0118] Solve the posture transformation and projection relationship through PnP:

[0119]

[0120] Where R and t represent the rotation matrix and translation vector; s is the scaling factor;

[0121] Solve for the optimal R and t so that the reprojection error between the predicted 2D points and the actual 2D points is minimized:

[0122]

[0123] Among them, π is the projection function, which converts 3D coordinates into 2D coordinates, and p i′ is the i′th 2D point actually observed; a nonlinear optimization method (such as Levenberg-Marquardt) is used to solve this problem;

[0124] Integrated visual odometry calculation, the pose transformation [Rk, tk] calculated for the current frame relative to the previous frame is added to the entire motion trajectory:

[0125] [R k+1 ,t k+1 ]=[R k ,t k ]·[R k-1 ,t k-1 ];

[0126] The pose change (relative pose) between consecutive frames can be simply described by solving PnP:

[0127] ΔT k =[R k ,t k ];

[0128] Use Kalman filtering to smooth the accumulated results, consider optimizing the estimation of position and attitude, and use bundle adjustment to optimize the global pose and map points when encountering loop closure.

[0129] In this embodiment, when a loop closure is encountered, bundle adjustment is used to optimize the global pose and map points, as follows:

[0130] The optimization objective is to optimize the camera pose and the 3D point coordinates it observes at the same time, and the error function is constructed to minimize the reprojection error:

[0131]

[0132] Where R i″ represents the rotation matrix of the i″th camera pose; t i″ represents the translation vector of the i″th camera pose; X j represents the coordinates of the jth 3D point in the world coordinate system; π projection function; ρ robustness loss function; q i″jrepresents the 2D image coordinates of the jth feature point observed in the i″th camera;

[0133] By minimizing the reprojection error, all camera poses and 3D point positions are optimized, and the Levenberg-Marquardt method is used to iteratively adjust the parameters to achieve the minimum error.

[0134] In this embodiment, the Levenberg-Marquardt method is used to iteratively adjust the parameters to achieve the minimum error, as follows:

[0135] Starting with reasonable initial guesses for all existing poses and map points, evaluate all reprojection errors and the newly added loop closure constraint errors;

[0136] The updated value is obtained by solving the following linear equation:

[0137] (J T WJ+λ′I)Δ=-J T We;

[0138] Where J is the Jacobian matrix; W is the weight matrix; λ′ is the damping factor; e is the cumulative error vector; Δ is the change;

[0139] Update the state variables:

[0140] x new =x old +Δ;

[0141] Among them, x new is the new state variable; x old is the original state variable;

[0142] The damping factor λ′ is dynamically adjusted to adapt to the error reduction speed: when the error decreases to a threshold, the iteration is stopped.

[0143] In this embodiment, step S3 is specifically:

[0144] Based on Harris corner points, key feature points are extracted from the constructed environment map as potential path nodes, the constructed environment map is grayed out, and the gradient of the environment map in the x and y directions is calculated using the Sobel operator.

[0145] Compute the structure tensor M around each pixel:

[0146]

[0147] in, Represents the sum of squared gradients of all pixels in the local neighborhood of the image in the x direction; surface

[0148] ∑I x I y

[0149] represents the sum of squared gradients of all pixels in the local neighborhood of the image in the y direction; represents the sum of the products of the gradients of pixels in the x and y directions in the local neighborhood of the image;

[0150] The Harris response function is used to calculate the corner response value R:

[0151] R = det(M) - k·trace(M) 2 ;

[0152] Where det(M) is the matrix determinant, trace(M) is the trace of the matrix, and k is an empirical constant;

[0153] The clustering algorithm DBSCAN is used to cluster potential path nodes, reduce the number of nodes, simplify the complexity of path calculation, and obtain a simplified environment map containing key path nodes.

[0154] In this embodiment, step S4 is specifically as follows:

[0155] Set the start and end node sets, and mark the start and end locations according to the inspection plan;

[0156] Combined with the three-dimensional environment map with key path nodes constructed in step S3, the map includes key path nodes and topological structures, and the three-dimensional environment map allows expansion from each node to six directions (up and down, front and back, left and right), the edges between nodes represent feasible direct paths, and the path cost function g(n) represents the accumulated actual cost from the starting node to the current node n;

[0157] The heuristic function h(n) estimates the minimum cost from the current node n to the nearest destination node:

[0158] h(n)=|x goal -x n |+|y goal -y n |+|z goal -z n |;

[0159] Among them, (x goal ,y goal ,z goal ) is the coordinate of the target node; (x n ,y n ,z n ) is the coordinate of the current node n;

[0160] Use the minimum value of the heuristic function to select the path node and calculate f(n)=g(n)+h(n) to optimize the path;

[0161] Based on the optimized node selection, the path is gradually generated and adjusted to output the optimal path planning result.

[0162] In this embodiment, S5 is specifically:

[0163] Deploy edge computing terminals at key locations of power lines, including substations and important line intersections, to handle preliminary data processing, event detection, and communication with drones;

[0164] Event detection sets multi-level thresholds according to different event types, comprehensively presets data at several moments to reduce noise, and is used to improve detection accuracy. It analyzes the rate of change within the sliding window based on a decision tree to identify events.

[0165] When the edge terminal detects that the event conditions are met, the drone starts real-time data recording and transmission; uses optimized wireless transmission protocols (such as LoRa, WiFi Direct) to reduce data transmission delays; after the event is triggered, uses priority queues to manage the transmission of different types of data, such as fault images and low-priority regular data.

[0166] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0167] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0168] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0169] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0170] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A UAV route planning method based on lightweight artificial intelligence, characterized in that: The following steps are involved: Step S1: Use the drone equipped with a camera and a lidar to obtain distribution line environmental data and fuse multi-sensor data in real time; Step S2: constructing an environment map based on the fused multi-sensor data; Step S3: extract key points from the environment map as path nodes, reduce the number of nodes by clustering algorithm, and obtain a three-dimensional environment map with key path nodes; Step S4: According to the power line inspection plan, an improved version of the A* algorithm is used, based on the three-dimensional environment map with key path nodes obtained in step S2, and a heuristic search strategy is used to optimize the path calculation, reduce the amount of calculation, and construct a path plan; Step S5: Edge terminals are deployed at key nodes, and event-based detection algorithms are used to accurately identify and respond to key environmental changes. UAVs trigger real-time data storage and transmission only when necessary.

2. The method for UAV route planning based on lightweight artificial intelligence according to claim 1 is characterized in that: The real-time fusion of multi-sensor data is based on the particle filter method to fuse sensor data and continuously update the drone's estimate of the environmental state, as follows: Assign N particles to the possible state space, each particle Have the same initial weights For each particle, the state at the next moment is predicted based on the motion model: Where f is the state transition function, u k is the control input at time k; is the process noise; Update the weight of each particle through sensor observation data: Among them, z k is the sensor observation value, including the fused observation data from the camera and lidar; is the observation model, which means that in the particle state Observe z k The probability of According to the updated weights, particles are resampled based on system resampling to reduce degradation; Combine all updated and resampled particles to estimate the current environment state: The current state estimate is obtained by calculating the weighted average of the particle states:

3. The method for UAV route planning based on lightweight artificial intelligence according to claim 2 is characterized in that: The construction of the environment map based on the fused multi-sensor data is specifically as follows: Use timestamps and synchronization algorithms to ensure that camera images and lidar point cloud data are aligned in time; Based on the image data collected by the camera, the ORB algorithm is used to extract key points and generate descriptors; Calculate curvature and plane geometry from LiDAR point cloud data and use normal estimation algorithms to determine the characteristics of the tangent plane at each point. Use descriptors to match feature points between adjacent frames, connect images and point cloud data at different time points through feature descriptors, and combine RANSAC algorithm to remove outliers and improve matching accuracy; Using the matched feature points, we solve the pose transformation through PnP, integrate the visual odometry calculation, and convert the feature matching results into robot motion estimation; Use the ICP algorithm to align point cloud data and refine the LiDAR pose estimation; The pose of each frame is cumulatively decomposed, the newly acquired point cloud data is converted to the global coordinate system, and merged to obtain the environment map.

4. The method for UAV route planning based on lightweight artificial intelligence according to claim 3 is characterized in that: The method utilizes the matched feature points, solves the pose transformation through PnP, integrates the visual odometer calculation, and converts the feature matching results into robot motion estimation, as follows: 2D image feature points (x) extracted and matched from adjacent frames i′ ,y i′ ) and the corresponding 3D space point (X i′ ,Y i′ ,Z i′ ); Use the camera's intrinsic matrix K to convert 2D image coordinates into 3D rays; Solve the posture transformation and projection relationship through PnP: Where R and t represent the rotation matrix and translation vector; s is the scaling factor; Solve for the optimal R and t so that the reprojection error between the predicted 2D points and the actual 2D points is minimized: Among them, π is the projection function, which converts 3D coordinates into 2D coordinates, and p i′ is the i′th 2D point actually observed; a nonlinear optimization method is used to solve this problem; Integrated visual odometry calculation, the pose transformation [Rk, tk] calculated for the current frame relative to the previous frame is added to the entire motion trajectory: [R k+1 ,t k+1 ]=[R k ,t k ]·[R k-1 ,t k-1 ]; The pose change between consecutive frames can be simply described by solving PnP: ΔT k =[R k ,t k ]; Use Kalman filtering to smooth the accumulated results, consider optimizing the estimation of position and attitude, and use bundle adjustment to optimize the global pose and map points when encountering loop closure.

5. The method for UAV route planning based on lightweight artificial intelligence according to claim 4 is characterized in that: When encountering loop closure, bundle adjustment is used to optimize the global pose and map points, as follows: The optimization objective is to optimize the camera pose and the 3D point coordinates it observes at the same time, and the error function is constructed to minimize the reprojection error: Where R i″ represents the rotation matrix of the i″th camera pose; t i″ represents the translation vector of the i″th camera pose; X j represents the coordinates of the jth 3D point in the world coordinate system; π projection function; ρ robustness loss function; q i″j represents the 2D image coordinates of the jth feature point observed in the i″th camera; By minimizing the reprojection error, all camera poses and 3D point positions are optimized, and the Levenberg-Marquardt method is used to iteratively adjust the parameters to achieve the minimum error.

6. The method for UAV route planning based on lightweight artificial intelligence according to claim 5 is characterized in that: The Levenberg-Marquardt method is used to iteratively adjust the parameters to achieve the minimum error, as follows: Starting with reasonable initial guesses for all existing poses and map points, evaluate all reprojection errors and the newly added loop closure constraint errors; The updated value is obtained by solving the following linear equation: (J T WJ+λ′I)Δ=-J T We; Where J is the Jacobian matrix; W is the weight matrix; λ′ is the damping factor; e is the cumulative error vector; Δ is the change; Update the state variables: x new =x old +Δ; Among them, x new is the new state variable; x old is the original state variable; The damping factor λ′ is dynamically adjusted to adapt to the error reduction speed: when the error decreases to a threshold, the iteration is stopped.

7. The method for UAV route planning based on lightweight artificial intelligence according to claim 2 is characterized in that: The step S3 is specifically: Based on Harris corner points, key feature points are extracted from the constructed environment map as potential path nodes, the constructed environment map is grayed out, and the gradient of the environment map in the x and y directions is calculated using the Sobel operator. Compute the structure tensor M around each pixel: in, Represents the sum of squared gradients of all pixels in the local neighborhood of the image in the x direction; ∑I x I y It represents the sum of squared gradients of all pixels in the local neighborhood of the image in the y direction; it represents the sum of the products of the gradients of pixels in the x and y directions in the local neighborhood of the image; The Harris response function is used to calculate the corner response value R: R=det(M)-k·trace(M) 2 ; Where det(M) is the matrix determinant, trace(M) is the trace of the matrix, and k is an empirical constant; The clustering algorithm DBSCAN is used to cluster potential path nodes, reduce the number of nodes, simplify the complexity of path calculation, and obtain a simplified environment map containing key path nodes.

8. The method for UAV route planning based on lightweight artificial intelligence according to claim 1 is characterized in that: The step S4 is specifically as follows: Set the start and end node sets, and mark the start and end locations according to the inspection plan; Combined with the three-dimensional environment map with key path nodes constructed in step S3, the map includes key path nodes and topological structures, the three-dimensional environment map is allowed to expand to six directions from each node, the edges between nodes represent feasible direct paths, and the path cost function g(n) represents the accumulated actual cost from the starting node to the current node n; The heuristic function h(n) estimates the minimum cost from the current node n to the nearest destination node: h(n)=|x goal -x n |+|y goal -y n |+|z goal -z n |; Among them, (x goal ,y goal ,z goal ) is the coordinate of the target node; (x n ,y n ,z n ) is the coordinate of the current node n; Use the minimum value of the heuristic function to select the path node and calculate f(n)=g(n)+h(n) to optimize the path; Based on the optimized node selection, the path is gradually generated and adjusted to output the optimal path planning result.

9. The method for UAV route planning based on lightweight artificial intelligence according to claim 1, characterized in that: The S5 is specifically: Deploy edge computing terminals at key locations of power lines, including substations and important line intersections, to handle preliminary data processing, event detection, and communication with drones; Event detection sets multi-level thresholds according to different event types, comprehensively presets data at several moments to reduce noise, and is used to improve detection accuracy. It analyzes the rate of change within the sliding window based on a decision tree to identify events. When the edge terminal detects that the event conditions are met, the drone begins to record and transmit real-time data; an optimized wireless transmission protocol is used to reduce data transmission delays; after the event is triggered, a priority queue is used to manage the transmission of different types of data.

10. A UAV route planning system based on lightweight artificial intelligence, characterized in that: It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the lightweight artificial intelligence-based unmanned aerial vehicle route planning method as described in any one of claims 1 to 9.

Images