Unmanned aerial vehicle autonomous inspection trajectory planning method and system based on channel environment real-time map

Abstract

The application provides a kind of unmanned aerial vehicle autonomous inspection trajectory planning method and system based on channel environment real-time map.The method first constructs real-time map containing semantic information by using multi-sensor fusion data online and generates Euclidean signed distance field synchronously;Then, based on the distance field, a safe flight corridor is generated, and a planning algorithm guided by a dynamic fuzzy logic controller that integrates energy state perception is used to plan a global spatiotemporal trajectory within the corridor that meets the dynamics and energy constraints;Next, taking the global trajectory as a reference, a local planner that integrates a velocity obstacle mechanism is used to perform real-time trajectory optimization and dynamic obstacle avoidance under the constraint of the safe corridor;Finally, the control instructions are executed and the image quality is evaluated, and based on the evaluation results and the state of the unmanned aerial vehicle, a dynamic decision is made on whether to adjust the trajectory or execute an emergency strategy.The application significantly improves the safety, efficiency and data acquisition quality of the unmanned aerial vehicle in the complex distribution network channel.

Description

Technical Field

[0001] This invention relates to the field of autonomous navigation and trajectory planning technology for unmanned aerial vehicles (UAVs), specifically to a method and system for autonomous inspection trajectory planning of UAVs applied to environments such as power distribution networks and transmission lines. Background Technology

[0002] With the rapid development of drone technology, its application in fields such as power grid inspection and facility monitoring is becoming increasingly widespread. Achieving safe, efficient, and high-quality autonomous inspection by drones in complex, narrow, and dynamically changing environments such as power distribution network corridors and transmission corridors is a key technological challenge. Existing drone inspection solutions mostly rely on pre-set fixed routes or path planning based on global static maps, making it difficult to cope with dynamic environmental changes, drone energy constraints, and real-time image quality assurance during the inspection process. Especially for corridor inspection tasks, the environment typically includes numerous static obstacles (such as poles, trees, and buildings) and dynamic interference (such as branches swaying in the wind). Furthermore, the quality of the inspection task highly depends on the resolution, clarity, and target coverage integrity of the drone-captured images. Traditional planning methods often separate path planning, obstacle avoidance control, and task quality assessment, lacking a closed-loop decision-making system that comprehensively considers real-time environmental perception, energy management, dynamic obstacle avoidance, and visual quality feedback. This leads to problems such as unsafe paths, insufficient battery life, untimely obstacle avoidance, and invalid images requiring rework and reflight when drones are inspecting complex passages, which seriously restricts inspection efficiency and automation level.

[0003] Therefore, there is an urgent need in this field for an autonomous inspection technology solution for unmanned aerial vehicles (UAVs) that can deeply integrate real-time environmental modeling, global and local adaptive trajectory planning, dynamic intelligent obstacle avoidance, and closed-loop feedback decision-making based on visual quality, so as to improve the operational safety, task completion efficiency, and data acquisition quality of UAVs in complex channel environments. Summary of the Invention

[0004] This invention aims to address the core problems of existing UAV inspection technologies, such as insufficient safety, low endurance, poor dynamic obstacle avoidance, and difficulty in guaranteeing the validity of inspection data in complex environments due to reliance on static paths, fragmented planning and control, and neglect of energy constraints and mission quality feedback. This invention provides a technical solution that enables intelligent operation throughout the entire process, from real-time environmental perception, global and local collaborative planning, dynamic intelligent obstacle avoidance to vision-quality-based closed-loop decision-making, significantly improving the overall safety, mission reliability, operational efficiency, and data acquisition quality of autonomous UAV inspections.

[0005] In a first aspect, embodiments of this application provide a method for autonomous inspection trajectory planning of unmanned aerial vehicles (UAVs) based on a real-time map of the channel environment, the method comprising:

[0006] Based on the real-time acquisition of channel environment data by UAV airborne fusion sensor, a real-time map containing semantic information is constructed online and a Euclidean symbolic distance field is generated simultaneously.

[0007] Based on the real-time map and distance field, a safe flight corridor connecting each inspection target point is generated online. A sampling planning algorithm guided by a dynamic fuzzy logic controller that integrates energy state perception is used to plan a global spatiotemporal trajectory that satisfies dynamic and energy constraints under the constraints of the safe flight corridor.

[0008] Using the global spatiotemporal trajectory as a reference, a local planner that integrates velocity obstacle mechanisms is employed for real-time trajectory optimization and dynamic obstacle avoidance under the constraints of a safe flight corridor.

[0009] Execute control commands to control the UAV to complete inspection flights and image acquisition, and perform real-time visual quality assessment of the acquired images;

[0010] Based on the visual quality assessment results and the real-time status information of the UAV, a dynamic decision is made on whether to trigger trajectory adjustment or execute emergency strategies.

[0011] Secondly, embodiments of this application provide an autonomous UAV inspection trajectory planning system based on a real-time map of the channel environment, applied to the autonomous UAV inspection trajectory planning method based on a real-time map of the channel environment as described in the first aspect, the system comprising:

[0012] The environmental perception and map building module is used to collect channel environmental data in real time based on UAV airborne fusion sensors, build a real-time map containing semantic information online and generate a Euclidean symbol distance field simultaneously.

[0013] The global trajectory planning module is used to generate safe flight corridors connecting each inspection target point online based on the real-time map and distance field, and to plan a global spatiotemporal trajectory that satisfies dynamic and energy constraints under the constraints of the safe flight corridor using a sampling planning algorithm guided by a dynamic fuzzy logic controller that integrates energy state perception.

[0014] The local trajectory optimization and obstacle avoidance module is used to perform real-time trajectory optimization and dynamic obstacle avoidance using a local planner that integrates a velocity obstacle mechanism under the constraint of a safe flight corridor, with the global spatiotemporal trajectory as a reference.

[0015] The task execution and quality assessment module is used to execute control commands to control the UAV to complete inspection flights and image acquisition, and to perform real-time visual quality assessment on the acquired images.

[0016] The closed-loop decision-making and adjustment module is used to dynamically decide whether to trigger trajectory adjustment or execute emergency strategies based on the visual quality assessment results and the real-time status information of the UAV.

[0017] Thirdly, embodiments of this application provide an electronic device, including:

[0018] processor;

[0019] Memory used to store processor-executable instructions;

[0020] The processor is configured to implement the UAV autonomous inspection trajectory planning method based on a real-time map of the channel environment as described in the first aspect when executing the instructions.

[0021] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the UAV autonomous inspection trajectory planning method based on a real-time map of the channel environment as described in the first aspect.

[0022] Compared with existing technologies, this invention has the following significant advantages: Through a safe flight corridor and speed obstacle mechanism, it achieves dual active obstacle avoidance against both static and dynamic obstacles, significantly reducing collision risks. Using energy state as a core planning parameter, it achieves a dynamic balance between mission path and endurance, extending effective operating time. Based on incremental updates of real-time semantic maps and dynamic fuzzy logic, the system can quickly adapt to dynamic changes and uncertainties in the environment. The introduction of real-time visual quality assessment and closed-loop feedback ensures the effectiveness of acquired images, reducing invalid operations and rework. It provides a complete methodology and system implementation from low-level perception to high-level decision-making, achieving truly autonomous and intelligent inspection. Attached Figure Description

[0023] Figure 1 A flowchart illustrating the method for autonomous UAV inspection trajectory planning based on real-time maps of the channel environment.

[0024] Figure 2 Architecture diagram of UAV autonomous inspection trajectory planning system based on real-time map of channel environment.

[0025] Figure 3 Schematic diagram of the system's electronic equipment. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0027] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0028] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] Example 1

[0030] Figure 1 This is a schematic flowchart illustrating a method for autonomous UAV inspection trajectory planning based on a real-time map of the channel environment, provided in one embodiment of this application. Figure 1 As shown, a method for autonomous UAV inspection trajectory planning based on a real-time map of the channel environment includes:

[0031] S101. Based on the UAV-borne fusion sensor, real-time acquisition of channel environment data is carried out, and a real-time map containing semantic information is constructed online and Euclidean symbolic distance field (ESDF) is generated simultaneously.

[0032] Based on real-time acquisition of channel environment data by a UAV-borne fusion sensor combining vision and lidar, a real-time map incorporating dense point clouds and semantic segmentation information is constructed online. Simultaneously, a Euclidean symbolic distance field describing the distances from spatial points to obstacles is generated. Here, the "real-time map with dense point clouds and semantic segmentation information" refers to a 3D environment model, and the "Euclidean symbolic distance field" refers to the distance map from each point in space to the nearest obstacle. By fusing sensor data to construct a channel map with semantic information in real time and simultaneously generating an Euclidean symbolic distance field describing obstacle distances, a precise environmental safety representation foundation is provided for path planning.

[0033] Specifically, in this embodiment, the process of constructing a real-time map online and generating a Euclidean symbolic distance field includes:

[0034] A dense point cloud map is generated by fusing LiDAR point clouds acquired from multiple sensors with visual images. The LiDAR point cloud and visual images refer to 3D point cloud data (such as XYZ coordinate information) generated by LiDAR scanning and 2D image data acquired by a visual camera, respectively. Data fusion algorithms (such as coordinate transformation based on calibration parameters and feature matching) unify the data from different sensors into the same coordinate system, generating a more complete and accurate dense point cloud map. For example, fusing the precise geometric information from LiDAR with the rich texture information from vision generates a 3D point cloud model with color attributes.

[0035] A deep learning model is used to perform real-time semantic segmentation on the dense point cloud map, identifying key inspection components such as conductors, insulators, towers, vegetation, and buildings. The deep learning model refers to neural networks such as PointNet++ and RandLA-Net, specifically designed for 3D point cloud processing; key inspection components refer to specific target objects in power line inspections. The model classifies each point in the point cloud and assigns it a semantic label. For example, a set of points is classified as "conductor" (label 1), "insulator" (label 2), and "vegetation" (label 3). The output is a point cloud with each point labeled.

[0036] Based on the semantic segmentation results, the point cloud is voxelized to generate a multi-scale semantic occupancy map containing semantic information and occupancy probability. Voxelization refers to dividing a continuous three-dimensional space into a regular grid (voxels), with each voxel summarizing the information of its internal points; occupancy probability refers to the likelihood that a voxel is occupied by an object. This transforms dense, unstructured semantic point clouds into regular, storage-efficient multi-scale semantic occupancy maps. For example, a 1 cubic meter space is divided into voxels with 10 cm sides. If a voxel contains multiple "vegetation" points, then that voxel is labeled as "vegetation," and its occupancy probability is set to 1.0.

[0037] An incremental update mechanism is used to maintain the multi-scale semantic occupancy map, where only the affected areas are locally updated when environmental changes are detected. This incremental update mechanism means that only the differences between newly acquired data and historical maps are updated, rather than a global reconstruction. This significantly reduces computational load and ensures the map's real-time performance. For example, when a drone flies to a new area or detects local environmental changes (such as swaying tree branches), only the voxels and their attributes in the changed area are updated.

[0038] Based on the multi-scale semantic occupancy map, Euclidean signed distance fields in both positive and negative directions are calculated in parallel. The positive direction distance field represents the distance from a spatial point to the nearest obstacle, and the negative direction distance field represents the distance from a spatial point to the nearest free space boundary. The updates of these distance fields are synchronized with the updates of the semantic occupancy map. The Euclidean signed distance field is a three-dimensional scalar field, where the value of each point represents the signed distance from that point to the nearest object surface; the positive direction distance field value represents the distance to the nearest obstacle; and the negative direction distance field value represents the distance to the nearest free space boundary. This provides crucial gradient information for path planning. The planning algorithm can utilize the gradient information of the distance field to efficiently guide the path away from obstacles (along the descent direction of the positive distance field gradient) or into the center of free space (along the ascent direction of the negative distance field gradient).

[0039] For a point in space Its Euclidean distance field (ESDF) value It can be defined as:

[0040] ,

[0041] in, To represent any point in three-dimensional space, coordinates are usually used. Point The ESDF value (signed distance) at that location. It is a point The Euclidean distance to the nearest obstacle surface (boundary). This represents the set of all obstacle surfaces. It is the boundary between "occupied" and "free" space in the environment map. When the occupancy status (semantic or probabilistic) of a voxel in the map is updated, the ESDF value of its neighboring area is recalculated using a fast algorithm (such as BFS or FastMarching Method) to ensure the timeliness of planning information.

[0042] S102. Based on the real-time map and distance field, a safe flight corridor connecting each inspection target point is generated online. A sampling planning algorithm guided by a dynamic fuzzy logic controller that integrates energy state perception is used to plan a global spatiotemporal trajectory that satisfies dynamic and energy constraints under the constraints of the safe flight corridor.

[0043] Based on the real-time map and Euclidean symbolic distance field, a safe flight corridor connecting each inspection target point is generated online. A sampling planning algorithm guided by a dynamic fuzzy logic controller is then used to plan a global spatiotemporal trajectory that satisfies the UAV's dynamic constraints, energy constraints, and inspection coverage requirements under the constraints of the safe flight corridor, outputting its spatiotemporal keyframe sequence. The input of the dynamic fuzzy logic controller includes parameters representing the energy state. Here, the safe flight corridor refers to an unobstructed, flyable space conduit; the dynamic fuzzy logic controller refers to an intelligent planning module that can adjust its strategy based on states such as battery power; and the spatiotemporal keyframe sequence refers to a sequence of trajectory points containing time, position, and attitude. Within the safe space, a globally optimal flight path and timetable that balances flight performance, power consumption, and mission requirements are planned.

[0044] Specifically, in this embodiment, the sampling planning algorithm guided by a dynamic fuzzy logic controller includes:

[0045] S1021: A safe flight corridor connecting various inspection waypoints is generated online based on the Euclidean symbolic distance field. This safe flight corridor is a sequence of polygonal convex hulls. Specifically, the UAV needs to inspect insulators B, towers C, and conductor joints D sequentially from inspection point A in the power corridor. First, based on a real-time ESDF map, the feasible space between point A and target B is modeled: a collision-free pipe region is generated by expanding the surface of obstacles (such as conductors and trees). Using a convex decomposition algorithm, this pipe is approximated as a series of interconnected three-dimensional convex polyhedra (convex hulls). For example, the first convex hull extends from near point A to the bend in the channel, and the second convex hull covers the straight channel after the bend. Each convex hull is defined by its vertex coordinates, forming a spatial sequence that ensures geometric safety, i.e., the safe flight corridor. When planning, the UAV's path only needs to be constrained within the union of these convex hulls to ensure physical obstacle avoidance.

[0046] S1022: Configure a dynamic fuzzy logic controller whose input variable set includes: the width of the safe flight corridor section where the current sampling point is located, the obstacle density within the corridor, the normalized remaining distance from the current point to the target waypoint, and the percentage of the drone's real-time remaining battery power. Specifically, during the flight from convex hull 1 (6 meters wide) to convex hull 2 ​​(3 meters wide), the controller collects four real-time inputs: Corridor width: Queryed from the ESDF map, the effective width of the narrowest section of convex hull 2 ​​is 3.0 meters. Obstacle density: The percentage of voxels marked as "obstacles" within convex hull 2 ​​is statistically analyzed, and the density is calculated to be 0.15 (15% of the space is occupied). Normalized remaining distance: The straight-line distance from the current sampling point to the target waypoint D is 120 meters, the total planned distance is 600 meters, and the normalized value is 0.2. Remaining battery power percentage: The drone's current remaining battery power is 35%. These parameters are fuzzified (e.g., mapping the width "3 meters" to "narrow" in the membership function) and transformed into semantic variables that the controller can process.

[0047] S1023: The path optimization weights are dynamically adjusted through the rule base of the dynamic fuzzy logic controller. When the remaining battery power is below a first safety threshold, the path length penalty coefficient is significantly increased. When the remaining battery power is above a second safety threshold and the obstacle density is below a set density threshold, the reward weight for the integrity of the task area coverage is increased. The first safety threshold is lower than the second safety threshold. Specifically, the preset first battery safety threshold is 20%, the second battery safety threshold is 50%, and the obstacle density threshold is 0.1. The current battery power is 35%, which is between the two, and the obstacle density of 0.15 is higher than the density threshold, so the controller executes the default rule. Assuming that the battery power drops to 18% (below the first threshold) during subsequent flights, the controller triggers an emergency rule: the path length penalty coefficient is dynamically adjusted from 1.0 to 5.0, and the reward weight for coverage integrity indicators such as "optimal image acquisition angle" is significantly reduced in the evaluation function. This makes sampling planning algorithms (such as RRT*) more inclined to select short path nodes directly facing the target when expanding the random tree, even if the path may sacrifice some inspection views, thus prioritizing safe return. Conversely, if the battery is sufficient (e.g., 60%) and the area is open (density 0.05), the coverage integrity weight is increased, and the algorithm will explore more lateral nodes to ensure complete imaging of all sides of the tower.

[0048] Furthermore, the online generation of safe flight corridors connecting each inspection target point, and the use of a sampling planning algorithm guided by a dynamic fuzzy logic controller to plan a global spatiotemporal trajectory that satisfies the UAV's dynamic constraints, energy constraints, and inspection coverage requirements under the constraints of the safe flight corridors, specifically also includes:

[0049] An improved fast expanding random tree algorithm is used for path search within a safe flight corridor, where the target node sampling probability during random tree expansion is dynamically adjusted by a bias strategy output by the dynamic fuzzy logic controller. For example, the UAV needs to plan an initial path from its starting point to insulator B. The improved RRT* algorithm (an optimized version of the fast expanding random tree algorithm, which builds a search tree through random sampling to find feasible paths) is activated, with its search space strictly limited to a pre-calculated safe flight corridor (a series of convex hulls). The dynamic fuzzy logic controller outputs a bias strategy based on the current battery level (60%) and the environment (low obstacle density): the target-oriented sampling probability is set to 70%, and the random exploration probability is set to 30%. This means that when the algorithm grows the random tree, 70% of the attempts will directly generate sampling points in the direction of target point B, and 30% of the attempts will randomly sample within the corridor to discover potentially better paths. If the battery level drops below 30%, the controller will increase the target-oriented probability to 90%, forcing the algorithm to quickly converge to the shortest path. The algorithm iterates continuously, eventually generating a tree in the corridor that connects the starting point to the target point B, and backtracking from it to find a preliminary, collision-free polyline path.

[0050] The sequence of discrete waypoints obtained from the search is input into the minimum control trajectory optimization framework. Under the spatial constraints, UAV dynamic constraints, and energy consumption constraints of the safe flight corridor, a B-spline trajectory that satisfies multi-objective optimization and whose control points are located within the convex hull of the safe flight corridor is generated. Specifically, RRT* searches for a polyline composed of 10 discrete waypoints. This sequence is then input into the MINCO framework (an efficient trajectory optimization framework specifically designed for differentially flat systems, whose objective function minimizes the time integral of the control quantity, such as jerk, to generate the most energy-efficient or smoothest trajectory). The optimization problem is modeled as follows: under the premise of satisfying spatial constraints (the trajectory must be located within each convex hull of the safe corridor), dynamic constraints (maximum UAV speed 10 m / s, maximum acceleration 5 m / s²), and energy constraints (associating the total trajectory time with a predicted energy consumption model to penalize excessively long flight times), minimize the control quantity of the trajectory (such as the sum of squares of jerk), and simultaneously optimize a multi-objective weighted function that considers path length and target observation angle. The MINCO framework, through numerical optimization, outputs a trajectory represented by a B-spline curve (a smooth parametric curve defined by a small number of control points, the shape of which is entirely determined by the positions of these control points). After optimization, all control points of the B-spline curve are strictly constrained within their corresponding convex hulls, thus ensuring the safety of the entire trajectory.

[0051] The optimized B-spline trajectory is then time-parameterized, extracting a spatiotemporal keyframe sequence containing spatial coordinates, velocity, acceleration, and the desired gimbal attitude as a reference for subsequent trajectory tracking. Specifically, the optimized B-spline trajectory is a geometric curve with respect to parameter u. Time parameterization is performed based on the UAV's motion performance: a timestamp is assigned to each point on the curve, ensuring that the velocity and acceleration along the trajectory do not exceed the UAV's dynamic limits. For example, a higher velocity (8 m / s) can be assigned to straight sections, while a lower velocity (3 m / s) is required for turning sections. Then, samples are taken along this time-parameterized trajectory at fixed time intervals (e.g., 0.1 seconds) or fixed distance intervals. Each sample point constitutes a "spatiotemporal keyframe," containing: spatial coordinates: (x, y, z), such as (102.3, 55.7, 50.1) meters; velocity: (v...). x ,v y ,v z For example, (0, 5.0, 0) m / s. Acceleration: (a x ,a y ,a z For example, (0, 0.5, 0) m / s². Desired gimbal attitude: To ensure the camera is directly facing insulator B at this moment, calculate the required gimbal pitch and yaw angles, such as (pitch -10 degrees, yaw 45 degrees). Finally, this series of time-ordered keyframes is sent to the underlying flight controller as precise reference instructions for trajectory tracking. Simultaneously, it provides a spatiotemporal reference for local planning and image acquisition triggering in step three.

[0052] S103. Using the global spatiotemporal trajectory as a reference, a local planner employing a fusion velocity obstacle mechanism is used to perform real-time trajectory optimization and dynamic obstacle avoidance under the constraints of a safe flight corridor.

[0053] Using the global spatiotemporal trajectory and its keyframe sequence as a reference, and combining real-time updated map information and dynamic obstacle detection results, under the constraints of the safe flight corridor, a local planner integrating a velocity obstacle collision prediction mechanism is employed for real-time trajectory optimization and dynamic obstacle avoidance, generating local control commands. The velocity obstacle collision prediction mechanism refers to the algorithm used to predict and avoid collisions with moving obstacles, while the local control commands refer to the specific motion commands that drive the UAV to execute. The global path is fine-tuned according to real-time environmental changes to ensure safe and accurate flight in dynamic environments.

[0054] Specifically, in this embodiment, the local planner employing a fusion velocity obstacle collision prediction mechanism for real-time trajectory optimization and dynamic obstacle avoidance includes:

[0055] S1031. Using the global spatiotemporal trajectory and its spatiotemporal keyframe sequence as a reference path, and combining it with a real-time updated multi-scale semantic occupancy map, a local trajectory replanning is performed. Specifically, the UAV is tracking keyframe 45 (position P45, time t=45s) towards keyframe 46. At this time, the local planner receives an update from the real-time semantic map: 5 meters ahead, at the edge of a static area previously marked as "vegetation," a small patch of moving point cloud has been newly detected and temporarily marked as "unclassified dynamic object." The planner does not completely discard the global trajectory, but instead uses the global trajectory from P45 to P50 (as a reference path) as a basis, combined with the real-time map containing this new dynamic information, to replan a locally optimal trajectory within the next few seconds (e.g., from t=45s to t=48s) to cope with unexpected situations.

[0056] S1032. Under the hard spatial constraint of the safe flight corridor, a local optimizer based on Model Predictive Control (MPC) is used for trajectory generation. To generate a new local trajectory, the system initiates an MPC-based local optimizer. MPC is a rolling time-domain control algorithm that, in each control cycle, solves an optimization problem based on the current state and a system model prediction of the future finite time domain (e.g., the next 2 seconds, called the prediction time domain) to obtain the current optimal control input. In this example, the constraint is set as follows: the optimizer first treats the safe flight corridor generated in the global planning stage as an inviolable hard spatial constraint. This means that any candidate trajectory point must fall within the corresponding three-dimensional convex hull. The optimization process is as follows: the optimizer solves an optimization problem in the current state (position, velocity) with the next 2 seconds as the prediction time domain. The decision variables of this problem are a series of future control commands (e.g., acceleration), and the objective is to minimize an objective function that includes trajectory deviation, energy consumption, etc., while adhering to all constraints. After solving, only the first control command is applied to the UAV, and rolling optimization is performed again in the next cycle.

[0057] S1033. For detected dynamic obstacles, a collision prediction model based on velocity-velocity obstacles (VO) is constructed to calculate the relative velocity between each dynamic obstacle and the UAV, and to predict potential future collision time windows. For the unclassified dynamic object, the system estimates its centroid movement velocity using two frames of point cloud data. =(0,1,0)m / s (assuming slow movement along the Y-axis). Current speed of the drone. =(5,0,0) m / s. Calculate the relative velocity. = - =(5,-1,0)m / s. The core idea of ​​the Velocity Object (VO) model is to define the set of velocity vectors from the UAV's current position to the area that a dynamic obstacle might occupy in the future as the conflict velocity set in velocity space (rather than position space). In this example, the VO model calculates the conflict velocity set based on the relative velocity and the estimated geometry of the obstacle, if the UAV maintains its current velocity. Therefore, within a time window of approximately 0.8 to 1.2 seconds, the two will collide. The construction of VO is a geometric operation that defines a region in velocity space that needs to be avoided.

[0058] S1034. The predicted collision-free time interval is used as a hard constraint for trajectory optimization to ensure that the generated local trajectory avoids collisions with dynamic obstacles in the time dimension. In solving the problem, the MPC optimizer adds a hard time constraint from the VO (Voice of Occurrence) in addition to the spatial corridor constraint. For each future time t within the prediction time domain (0-2 seconds), the optimizer requires that the distance between the predicted position of the UAV and the predicted position of the dynamic obstacle must be greater than a safety threshold at that time. Specifically, in this example, the optimizer knows that there is a collision risk between 0.8s and 1.2s, therefore it will enforce that at discrete times such as t=0.8s, 0.9s, 1.0s, 1.1s, and 1.2s, the predicted trajectory point of the UAV and the predicted position point of the dynamic obstacle maintain a safe distance of at least 2 meters. This is equivalent to adding a series of inequality constraints to the optimization problem.

[0059] S1035. The evaluation function of the local optimizer simultaneously considers trajectory tracking accuracy, safe distance from static obstacles, dynamic obstacle avoidance requirements, and motion smoothness indicators to complete the local trajectory optimization process. Specifically, the evaluation function of the MPC optimizer is a weighted sum, calculated as follows: 1. Trajectory tracking accuracy term: penalizes the deviation between the predicted trajectory point and the corresponding point (P45, P46...) of the reference global trajectory. For example, the sum of squared position errors. 2. Static safe distance term: penalizes the predicted trajectory point for being too close to all static obstacles (such as the nearest tower) in the real-time semantic map. Using ESDF, the closer the distance, the greater the penalty. 3. Dynamic obstacle avoidance term: has been transformed into the above-mentioned VO hard constraints. A reward term for the minimum interval distance can be added to the evaluation function to encourage a larger safety margin. 4. Motion smoothness term: penalizes drastic changes in the control input (acceleration), i.e., minimizing the derivative of the acceleration (jerk). The optimizer minimizes this evaluation function through numerical methods (such as quadratic programming), and finally outputs a local trajectory control sequence that satisfies all hard constraints and is comprehensively optimal. The drone executes the first control command of the sequence, thereby achieving smooth flight while maintaining the general direction of tracking, moving away from static towers, and actively avoiding dynamic objects (which may manifest as slight deceleration or minor detours). This process is repeated in the next control cycle (e.g., after 0.1 seconds), continuously optimizing the trajectory in dynamic environments.

[0060] The evaluation function of the local optimizer adopts a multi-objective optimization function under the model predictive control (MPC) framework, specifically in the form of:

[0061] ,

[0062] in, Indicates the first The predicted state of the step, including position and speed , This represents the global reference trajectory state at the corresponding moment. This is a static obstacle penalty term based on a real-time Euclidean symbolic distance field; the closer the distance, the greater the penalty. For the velocity obstacle region (or composite VO region), if the velocity is predicted... Falling into this area will result in a heavy penalty. For control input (such as acceleration). These are the weighting coefficients for trajectory tracking, static safety, dynamic obstacle avoidance, and control smoothness, which can be dynamically adjusted according to the mission stage. Let be the square of the Euclidean norm. This is a penalty for dynamic obstacle avoidance.

[0063] This optimization problem is solved under the premise of satisfying the dynamic constraints of the UAV, the spatial constraints of the safe flight corridor, and the velocity obstacle avoidance constraints, and finally outputs the locally optimal control command sequence.

[0064] Furthermore, the local trajectory optimization process also includes:

[0065] S1036. When multiple dynamic obstacles are detected, a composite velocity obstacle region is constructed, which is the union of the velocity obstacle regions of each dynamic obstacle. The UAV is flying in a passage and simultaneously detects two dynamic obstacles: obstacle A (possibly a fluttering plastic sheet) traveling at a speed V... A Moving northeast, obstacle B (swaying tree branch) moves at a speed of V B The system first constructs a velocity obstacle (VO) region for each obstacle based on its own velocity (or velocity range). This is a geometric region defined in the UAV's velocity space, indicating that if the UAV's velocity vector falls within this region, it will collide with the corresponding obstacle at some future time. Then, the system calculates the union of these two VO regions, forming a composite velocity obstacle region. This composite region represents the set of all UAV velocities that would lead to a collision with obstacle A or obstacle B. The optimizer must avoid selecting velocities located within this composite region during planning.

[0066] S1037. In each optimization iteration of model predictive control, the intersection of the UAV's predicted trajectory and the composite velocity obstacle region is used as an infeasible region constraint. In each iteration step of the MPC optimizer to solve for the trajectory in the next 2 seconds, it is necessary to evaluate the predicted trajectories corresponding to a set of candidate future control sequences. For this predicted trajectory, the optimizer checks its instantaneous velocity vector at each future sampling time (e.g., every 0.1 seconds). If the instantaneous velocity vector at a certain time falls into the composite velocity obstacle region calculated in step S1036, it means that flying at this speed will result in a collision with obstacle A or B in the future. Therefore, MPC will add the constraint that the instantaneous velocity at any point on the predicted trajectory must not fall into the composite VO region as an infeasible region to the optimization problem. This forces the optimizer to actively avoid all speed choices that would lead to a collision when searching for the optimal solution.

[0067] S1038. An adaptive safety margin mechanism is introduced to dynamically adjust the expansion radius of the velocity obstacle region based on the motion uncertainty and perception error of dynamic obstacles. The system not only uses the basic geometric model of the VO (Vehicle Object) but also introduces a safety margin around the VO boundary. This margin is not fixed but dynamically adjusted based on the perception uncertainty and motion prediction uncertainty of the obstacle. For example, for a high-speed, stable obstacle A (such as a bird), the velocity estimation is relatively accurate with low uncertainty, and the safety margin expansion radius might be set to R1 = 0.5 meters. For an obstacle B with complex motion patterns and inaccurate velocity estimation (such as a swaying tree branch), its velocity direction and magnitude have significant uncertainty, so the system increases the safety margin, and the expansion radius might be set to R2 = 1.5 meters. When constructing the VO region for each obstacle, its geometry (usually based on the obstacle's conservative outer envelope size) is "expanded" outward by this dynamically calculated expansion radius. This makes the composite VO region larger and more conservative, and the planned trajectory is further away from the actual dynamic obstacle, thus providing a buffer for perception and prediction errors and improving obstacle avoidance robustness.

[0068] S1039. When a feasible solution cannot be found under the constraints of a composite velocity obstacle, a local trajectory pause mechanism is initiated, controlling the UAV to hover at the current safe position until the dynamic obstacle passes or the preset waiting timeout is reached. Assume the passage ahead is temporarily blocked by two slowly moving large obstacles (such as temporary construction vehicles), causing any speed selected in any direction to fail to meet the constraint of avoiding the composite velocity obstacle region (i.e., all possible speeds fall within the composite velocity obstacle region). At this point, the MPC optimizer reports no feasible solution after multiple iterations. The system then triggers the local trajectory pause mechanism: 1. Safe hover: The UAV immediately executes an emergency stop procedure, using its position controller to hover stably at the current position (which is within the global safe corridor and far from static obstacles). 2. Continuous monitoring: During hovering, the system continuously updates the composite velocity obstacle region, monitoring the movement of the two dynamic obstacles. 3. Exit conditions include: Obstacle passage: When the obstacle moves to a point sufficient to open a gap in the composite velocity obstacle region, allowing the UAV to pass at a certain safe speed (i.e., the optimizer finds a feasible solution again), the system automatically releases the hover, continues planning, and executes the flight. Waiting timeout: If the obstacle is not moved after hovering for more than the preset time (e.g., 30 seconds), in order to prevent the task from being stalled for a long time or the battery from being consumed too much, the system determines that this path is temporarily impassable and reports this situation to the higher-level task decision module, which may trigger global replanning or task adjustment.

[0069] S104. Execute control commands to control the UAV to complete the inspection flight and image acquisition, and perform real-time visual quality assessment on the acquired images.

[0070] The system executes local control commands to control the UAV to complete inspection flights and image acquisition, and performs real-time visual quality assessment on the acquired images. Real-time visual quality assessment refers to the instantaneous analysis of indicators such as sharpness and contrast of the captured images. The technical quality of the images is evaluated simultaneously during acquisition to ensure the acquisition of effective and usable inspection data.

[0071] Specifically, in this embodiment, the real-time visual quality assessment of the acquired images includes:

[0072] S1041. A deep learning-based image quality assessment model is used to score the quality of each frame of the inspection image in multiple dimensions. The scoring dimensions include at least image sharpness, target contrast, illumination uniformity, and image noise level. Specifically, the drone hovers near the tower, and the gimbal camera captures a frame of insulator image with a resolution of 1920x1080. This image is immediately input into a pre-trained deep learning-based image quality assessment model (e.g., using a convolutional neural network structure, with the image as input and multiple quality scores as output). The model outputs scores in parallel for four dimensions (all normalized to 0-1, with 1 being optimal): Sharpness: 0.92 (sharp image details, visible insulator ceramic texture). Target contrast: 0.85 (clear distinction between the insulator and the background sky). Illumination uniformity: 0.65 (one side of the image is slightly dark due to backlighting). Noise level: 0.88 (clean image, no obvious noise). These four scores constitute the preliminary quality vector of this frame of image.

[0073] S1042. The key inspection components of the image, including conductors, insulators, and towers, are designated as priority areas for quality assessment. Local quality scores are calculated separately for each priority area. The system calls a lightweight object detection model (such as YOLOv5) to quickly locate the bounding box of the insulator in the image. Then, the algorithm crops this insulator region from the entire image. For this priority area, an evaluation algorithm optimized specifically for small regions is used again to calculate its local sharpness, contrast, and other scores. For example, the overall image illumination uniformity is 0.65, but the insulator region may be completely in shadow, resulting in a lower local illumination uniformity score (0.5). Finally, a comprehensive quality score is obtained by combining the overall image score and the priority area score (e.g., weighted average or taking the lowest score). ).

[0074] S1043. Construct a dynamic quality threshold adjustment mechanism to adaptively adjust the pass thresholds for each quality dimension based on the type of inspection target, shooting distance, and ambient lighting conditions. The system presets a set of baseline quality thresholds, such as sharpness > 0.8 and contrast > 0.7. However, these thresholds are not fixed. In this inspection task: Target type: The current target is an insulator, whose defect detection requires high sharpness. Therefore, the system automatically increases the sharpness pass threshold from 0.8 to 0.85. Shooting distance: Since the drone is approximately 15 meters away from the target (medium distance), the system allows the contrast threshold to be slightly lowered to 0.68. Ambient lighting: The current time is dusk, and the ambient lighting is weak. The system automatically lowers the pass thresholds for lighting uniformity and noise level (e.g., from 0.7 and 0.8 to 0.6 and 0.75) to adapt to reasonable performance under low-light conditions. In this way, the quality assessment standard changes dynamically according to the actual operating conditions, making it more reasonable.

[0075] S1044. When the overall quality score obtained from the evaluation is lower than the preset first quality threshold, a local shooting parameter adjustment command is triggered, automatically adjusting the gimbal's shooting angle, focal length, or exposure parameters. The calculated overall quality score... =0.78, while the first quality threshold after dynamic adjustment is 0.80. A score below the threshold triggers the adjustment mechanism. The system analyzes the scores of each dimension and finds that "lighting uniformity" (0.65) is the main weakness. Therefore, a local shooting parameter adjustment instruction is generated and executed: fine-tune the gimbal angle: deflect the gimbal 5 degrees towards the light source to improve the lighting in the target area. Adjust the exposure parameters: increase exposure compensation to brighten shadow areas. After adjustment, the drone immediately re-captures a frame while almost hovering and re-evaluates. This process may iterate rapidly several times until the score reaches the target or the maximum number of fine-tuning attempts is reached.

[0076] S1045. When the overall quality score is lower than a preset second quality threshold (which is lower than the first quality threshold) for multiple consecutive frames, it is determined that the current shooting point cannot obtain a qualified image, triggering a trajectory replanning command to replan the flight trajectory including the backup shooting point. At a certain shooting point, due to severe backlighting and a dusty environment, even after multiple parameter adjustments, the overall quality scores of three consecutive frames of images are 0.72, 0.70, and 0.68, respectively, all lower than the more stringent second quality threshold (e.g., 0.75). The system determines that the current shooting point can no longer obtain a qualified image. Therefore, a trajectory replanning command is sent to the trajectory planning module, requesting to abandon the current point and carrying a new constraint: a backup shooting point needs to be planned for this target (e.g., "the insulator of tower number #5"). Based on the real-time map, the planning module may choose to replan a trajectory to the new shooting point from the other side of the tower, or a closer location with better lighting conditions, and update the subsequent waypoint sequence to ensure that the critical target can still be effectively inspected in the end.

[0077] S105. Based on the visual quality assessment results and the real-time status information of the UAV, make a dynamic decision on whether to trigger trajectory adjustment or execute an emergency strategy.

[0078] Based on the visual quality assessment results and the UAV's real-time status information, a dynamic decision is made regarding whether to trigger local fine-tuning of the trajectory, global replanning, or to execute an emergency return-to-home strategy. Dynamic decision-making refers to the process of automatically selecting a response strategy based on real-time data, while the emergency return-to-home strategy refers to a contingency plan for safe return in emergency situations. The operational plan is intelligently adjusted according to mission execution quality and system status, forming an autonomous, adaptive, and fault-tolerant closed-loop control system.

[0079] Specifically, in this embodiment, the dynamic decision-making based on the visual quality assessment results and the real-time status information of the UAV to trigger local fine-tuning of the trajectory, global replanning, or execution of an emergency return-to-home strategy includes at least the following steps:

[0080] S1051. Establish a multi-dimensional monitoring system to collect real-time data on the UAV's remaining battery power, communication link quality, mission area coverage, and image acquisition progress. During the inspection mission, a background monitoring system collects the following data at a fixed frequency (e.g., 1Hz): Remaining battery power: Read from the flight controller, currently 45%. Communication link quality: The data transmission module reports a signal strength of -70dBm, indicating a good link status. Mission area coverage: Comparing the inspected path with the planned total path, 60% of the area coverage has been completed. Image acquisition progress: Comparing the number of tower targets already photographed with the planned total, the progress is 55% (due to reshooting some points). This data is packaged into a real-time system state vector.

[0081] S1052. Set tiered trigger thresholds, including image quality pass threshold, battery safety threshold, communication interruption threshold, and task progress lag threshold. The decision-making system presets a series of key thresholds as trigger lines for decisions, including: Image quality pass threshold: a dynamically adjusted first quality threshold, for example, 0.8. Battery safety threshold: set to 20%. When the battery level is below this value, returning to base must be considered first. Communication interruption threshold: determined by a complete link interruption lasting more than 5 seconds. Task progress lag threshold: defined as the current progress lagging behind the planned schedule by more than 10% (e.g., planned to be 70% complete, actually 55% complete).

[0082] S1053. When an image quality score is detected to be continuously below the acceptable threshold, a local adjustment instruction is generated. This instruction includes gimbal angle fine-tuning parameters, flight speed adjustment parameters, and shooting distance optimization parameters. The monitoring system found that in a recent shot of a wire connector, the quality scores for three consecutive frames were 0.75, 0.78, and 0.76, all below the set acceptable threshold of 0.8. The decision module determines this to be a local quality issue and generates a specific instruction: the instruction content includes: {Instruction type: Local adjustment, Target: Wire connector X, Action: [Gimbal deflects 10 degrees to the right, flight speed reduced to 0.5 m / s, advances 2 meters closer to the target]. The drone slowly moves forward and adjusts the gimbal, attempting to reshoot from the new position and angle. This process closely coordinates the quality assessment and parameter adjustment closed loop.

[0083] S1054. When the remaining battery power is detected to be lower than the aforementioned safety threshold, an emergency task compression mechanism is activated to skip non-critical inspection points and generate a safe return trajectory directly to the return point. The monitoring system reads that the battery power has dropped to 18%, below the 20% safety threshold. The decision module immediately activates the emergency task compression mechanism. 1. Task Analysis: The module reviews the task list and identifies critical points (such as specified defect points that must be inspected) and non-critical points (such as routine inspection points) among the remaining uninspected points. 2. Compression and Replanning: The module commands the global planner to skip all non-critical points. Based on the current location and real-time map, the planner replans a safe and shortest path, which may only contain a few unavoidable critical points and ultimately leads directly to the charging station or data center (return point). The new trajectory is immediately issued and executed to ensure the drone's safe return.

[0084] S1055. When it is detected that the mission progress is behind the planned progress and the remaining battery power is higher than the battery power safety threshold, the accelerated mission mode is activated to increase the flight speed and reduce the shooting time at each inspection point. The decision module analyzes the state vector: the mission progress is lagging (actual 55% vs. planned 70%), but the remaining battery power is still 45%, far higher than the safety threshold of 20%. The module determines that the progress can be caught up, so the accelerated mission mode is activated. 1. Parameter adjustment: Send instructions to the flight control and mission execution modules to increase the cruise flight speed from the default 5m / s to 7m / s. The hovering shooting time at each inspection point is shortened from 10 seconds to 6 seconds (only ensuring the minimum number of shots required for quality). 2. Risk control: This mode will correspondingly increase the obstacle avoidance sensitivity of the local planner and may temporarily relax the secondary requirements for image quality, while making every effort to catch up on the progress under the premise of ensuring safety and the validity of core data.

[0085] Furthermore, the dynamic decision-making process also includes:

[0086] S1056. When a communication link interruption is detected, the system switches to autonomous operation mode and continues to execute the current task, while caching all collected data. While inspecting poles in a valley, the drone flew into a signal blind spot, causing a 4G / data transmission link interruption with the ground control station. After losing five consecutive heartbeat packets, the system monitoring module determined that the communication link was interrupted. The decision module then executed the contingency plan: Mode Switching: The drone immediately switches from remote control mode to autonomous operation mode. In this mode, it will no longer attempt to receive new task instructions, but will continue to perform tasks such as flight, obstacle avoidance, and photography based on the loaded global task plan, its own perception and decision-making capabilities. Data Caching: All collected images, videos, status logs, and other data will no longer attempt to be uploaded, but will be encrypted and stored in the local cache area of ​​the onboard solid-state drive. Recovery Strategy: The drone flies to the next waypoint according to the original plan. The decision module will periodically (e.g., every second) attempt to re-establish the connection. Once communication is restored, the cached data will be uploaded in batches immediately, the latest status will be synchronized, and possible instruction updates will be received.

[0087] S1057. When encountering unavoidable obstacles or hazardous environmental conditions, a hovering and waiting mechanism is activated to continuously monitor environmental changes and continue the mission after safe conditions are restored. Specifically, when the UAV plans to fly to a certain insulator shooting point, its local planner (MPC), combined with real-time perception, discovers that the only passage ahead is completely and continuously blocked by a large, unpredictable dynamic obstacle (such as an agricultural UAV that has temporarily entered the area). All avoidance attempts are unsolvable due to the constraints of the composite speed obstacle, and it is predicted that this state will last for a short time (e.g., 2 seconds). The decision module determines that this is an unavoidable temporary blockage. Therefore: Hovering is activated: The decision module sends a "safe hovering" command to the flight controller, and the UAV immediately hovers stably in the current safe position. Continuous monitoring: During hovering, all sensors remain active, continuously assessing the state of the dynamic obstacle and environmental changes. Condition restoration judgment: When the obstacle is detected to have moved out of the critical area, or after waiting for more than a preset timeout period (e.g., 30 seconds), the system re-evaluates. If the blockage is cleared, exit the hover and continue the task; if the timeout still prevents passage, report the event, which may trigger a higher-level task replanning.

[0088] S1058. During mission execution, based on the image quality distribution and energy efficiency data of completed tasks, the system dynamically updates the rule base weight parameters of the dynamic fuzzy logic controller to optimize subsequent trajectory planning strategies. Specifically, after half of the mission is completed, the system performs mid-term data analysis and strategy optimization: Data collection: The system summarizes the data of the completed portion: Statistics show that targets photographed under backlight conditions have an average image quality score that is about 15% lower than those photographed under frontlight conditions; at the same time, analysis of flight logs shows that when using a higher flight speed (7m / s) to traverse dense vegetation areas, energy consumption is 20% higher than expected. Strategy update: Based on this analysis, the system dynamically updates the rule base of the dynamic fuzzy logic controller. Weight adjustment: Add a rule: If the next flight segment is predicted to be backlight and is not a critical target, appropriately reduce the image quality coverage integrity reward weight to reduce redundant maneuvers performed in pursuit of perfect angles. Or parameter adjustment: Adjust the weight parameters in the rule base that relate energy consumption and speed so that when planning to enter high-density obstacle areas in subsequent missions, the system tends to choose a slightly slower but more energy-efficient smooth path.

[0089] The updated fuzzy logic controller will leverage this online learning experience in subsequent global trajectory planning to generate trajectories that better reflect actual operational efficiency (balancing quality and energy consumption). This demonstrates the system's self-learning and adaptive capabilities, forming a complete closed loop from execution to optimization.

[0090] Example 2

[0091] like Figure 2 As shown, this application provides an architecture diagram of an autonomous UAV inspection trajectory planning system based on a real-time map of the channel environment, which is applied to the autonomous UAV inspection trajectory planning system based on a real-time map of the channel environment as described in Embodiment 1. It includes: an environmental perception and map construction module 210, a global trajectory planning module 220, a local trajectory optimization and obstacle avoidance module 230, a task execution and quality evaluation module 240, and a closed-loop decision and adjustment module 250.

[0092] The environmental perception and map building module 210 is used to collect channel environmental data in real time based on the UAV's airborne fusion sensor, build a real-time map containing semantic information online, and simultaneously generate a Euclidean symbolic distance field.

[0093] The global trajectory planning module 220 is used to generate a safe flight corridor connecting each inspection target point online based on the real-time map and distance field, and to plan a global spatiotemporal trajectory that satisfies dynamic and energy constraints under the constraints of the safe flight corridor by using a sampling planning algorithm guided by a dynamic fuzzy logic controller that integrates energy state perception.

[0094] The local trajectory optimization and obstacle avoidance module 230 is used to perform real-time trajectory optimization and dynamic obstacle avoidance using a local planner that integrates velocity obstacle mechanism under the constraint of safe flight corridor, with the global spatiotemporal trajectory as a reference.

[0095] The mission execution and quality assessment module 240 is used to execute control commands to control the UAV to complete the inspection flight and image acquisition, and to perform real-time visual quality assessment on the acquired images.

[0096] The closed-loop decision-making and adjustment module 250 is used to dynamically decide whether to trigger trajectory adjustment or execute emergency strategies based on the visual quality assessment results and the real-time status information of the UAV.

[0097] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 301 and memory 300, communication interface 303, and bus 302.

[0098] In this embodiment of the application, memory 300 is used to store executable instructions of processor 301, which, when configured to execute instructions, implements the method as described in the first aspect.

[0099] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0100] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0101] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0102] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0103] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0104] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0105] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for autonomous inspection trajectory planning of unmanned aerial vehicles (UAVs) based on real-time maps of the channel environment, characterized in that, Includes the following steps: Based on the real-time acquisition of channel environment data by UAV airborne fusion sensor, a real-time map containing semantic information is constructed online and a Euclidean symbolic distance field is generated simultaneously. Based on the real-time map and distance field, a safe flight corridor connecting each inspection target point is generated online. A sampling planning algorithm guided by a dynamic fuzzy logic controller that integrates energy state perception is used to plan a global spatiotemporal trajectory that satisfies dynamic and energy constraints under the constraints of the safe flight corridor. Using the global spatiotemporal trajectory as a reference, a local planner that integrates velocity obstacle mechanisms is employed for real-time trajectory optimization and dynamic obstacle avoidance under the constraints of a safe flight corridor. Execute control commands to control the UAV to complete inspection flights and image acquisition, and perform real-time visual quality assessment of the acquired images; Based on the visual quality assessment results and the real-time status information of the UAV, a dynamic decision is made on whether to trigger trajectory adjustment or execute emergency strategies.

2. The method according to claim 1, characterized in that, The online construction of a real-time map containing semantic information and the synchronous generation of an Euclidean symbolic distance field include: Dense point cloud maps are generated by fusing lidar point clouds collected by multiple sensors with visual images. A deep learning model is used to perform real-time semantic segmentation on the dense point cloud map to identify key inspection components such as conductors, insulators, towers, vegetation, and buildings. Based on the semantic segmentation results, the point cloud is voxelized to generate a multi-scale semantic occupancy map containing semantic information and occupancy probability. An incremental update mechanism is used to maintain the multi-scale semantic occupancy map, and when environmental changes are detected, only the affected areas are locally updated; Based on the multi-scale semantic occupancy map, Euclidean symbolic distance fields in both positive and negative directions are calculated in parallel. The positive direction distance field represents the distance from a spatial point to the nearest obstacle, and the negative direction distance field represents the distance from a spatial point to the nearest free space boundary. The update of the distance fields is synchronized with the update of the semantic occupancy map.

3. The method according to claim 1, characterized in that, The sampling planning algorithm, guided by a dynamic fuzzy logic controller that integrates energy state awareness, specifically includes: Online generation of safe flight corridors connecting various inspection waypoints is based on Euclidean symbolic distance field; the safe flight corridor is a sequence of polygonal convex hulls. Configure a dynamic fuzzy logic controller, whose input variable set includes: the cross-sectional width of the safe flight corridor where the current sampling point is located, the obstacle density in the corridor, the normalized remaining distance from the current point to the target waypoint, and the percentage of the UAV's real-time remaining battery power; The path optimization weights are dynamically adjusted through the rule base of the dynamic fuzzy logic controller. When the remaining battery power is lower than the first safety threshold, the path length penalty coefficient is significantly increased. When the remaining battery power is higher than the second safety threshold and the obstacle density is lower than the set density threshold, the reward weight for the integrity of the task area coverage is increased. The first safety threshold is lower than the second safety threshold.

4. The method according to claim 3, characterized in that, Also includes: An improved fast-expanding random tree algorithm is used to perform path search within the safe flight corridor, wherein the sampling probability of the target node during the random tree expansion process is dynamically adjusted by the bias strategy output by the dynamic fuzzy logic controller. The sequence of discrete path points obtained by the search is input into the minimum control quantity trajectory optimization framework. Under the spatial constraints, UAV dynamics constraints and energy consumption constraints of the safe flight corridor, a B-spline trajectory that satisfies multi-objective optimization and whose control points are located within the convex hull of the safe flight corridor is generated. The optimized B-spline trajectory is subjected to time parameterization, and a spatiotemporal keyframe sequence containing spatial coordinates, velocity, acceleration, and gimbal desired attitude is extracted as a reference for subsequent trajectory tracking.

5. The method according to claim 1, characterized in that, The local planner employing a fusion velocity obstacle mechanism for real-time trajectory optimization and dynamic obstacle avoidance under the constraints of a safe flight corridor specifically includes: Using the global spatiotemporal trajectory and its spatiotemporal keyframe sequence as reference paths, local trajectory replanning is performed in conjunction with a real-time updated multi-scale semantic occupancy map. Under the rigid spatial constraints of the safe flight corridor, a local optimizer based on model predictive control is used for trajectory generation; For detected dynamic obstacles, a collision prediction model based on velocity obstacles is constructed to calculate the relative velocity between each dynamic obstacle and the drone and predict the potential collision time window in the future. The predicted collision-free time interval is used as a hard constraint for trajectory optimization to ensure that the generated local trajectory avoids collisions with dynamic obstacles in the time dimension. The evaluation function of the local optimizer simultaneously considers trajectory tracking accuracy, safe distance from static obstacles, dynamic obstacle avoidance requirements, and motion smoothness index to complete the local trajectory optimization process.

6. The method according to claim 5, characterized in that, The local trajectory optimization process also includes: When multiple dynamic obstacles are detected, a composite velocity obstacle region is constructed, which is the union of the velocity obstacle regions of each dynamic obstacle. In each optimization iteration of model predictive control, the intersection of the UAV's predicted trajectory and the composite velocity obstacle region is used as an infeasible region constraint. An adaptive safety margin mechanism is introduced to dynamically adjust the expansion radius of the speed obstacle region based on the motion uncertainty and perception error of the dynamic obstacle. When a feasible solution cannot be found under the constraint of a composite velocity obstacle, a local trajectory pause mechanism is activated to control the drone to hover in the current safe position until the dynamic obstacle passes or the preset waiting timeout period is reached.

7. The method according to claim 1, characterized in that, The real-time visual quality assessment of the acquired images specifically includes: A deep learning-based image quality assessment model is used to score the quality of each frame of the inspection image in multiple dimensions. The scoring dimensions include at least image sharpness, target contrast, illumination uniformity and image noise level. The key inspection components of the image, such as conductors, insulators, and towers, are identified as key areas for quality assessment, and local quality scores are calculated separately for these key areas. A dynamic quality threshold adjustment mechanism is constructed to adaptively adjust the pass thresholds for each quality dimension based on the type of inspection target, shooting distance, and ambient lighting conditions. When the overall quality score obtained from the evaluation is lower than the preset first quality threshold, a local shooting parameter adjustment command is triggered, which automatically adjusts the shooting angle, focal length, or exposure parameters of the gimbal. When the overall quality score is lower than the preset second quality threshold, which is lower than the first quality threshold, for multiple consecutive frames, it is determined that the current shooting point cannot obtain a qualified image, triggering a trajectory replanning instruction to replan the flight trajectory including the backup shooting point.

8. The method according to claim 1, characterized in that, The dynamic decision-making process based on the visual quality assessment results and the real-time status information of the UAV, including whether to trigger trajectory adjustment or execute emergency strategies, includes: Establish a multi-dimensional monitoring system to collect data on the drone's remaining battery power, communication link quality, mission area coverage, and image acquisition progress in real time. Set tiered trigger thresholds, including image quality pass threshold, power safety threshold, communication interruption threshold, and task progress lag threshold; When the image quality score is detected to be continuously lower than the qualified threshold, a local adjustment instruction is generated. The instruction includes gimbal angle fine-tuning parameters, flight speed adjustment parameters, and shooting distance optimization parameters. When the remaining battery power is detected to be lower than the battery safety threshold, the emergency task compression mechanism is activated to skip non-critical inspection points and generate a safe return trajectory that leads directly to the return point. When the mission progress is detected to be lagging behind the planned progress and the remaining battery power is higher than the battery safety threshold, the accelerated mission mode is activated to increase the flight speed and reduce the shooting time at each inspection point.

9. The method according to claim 8, characterized in that, The dynamic decision-making also includes: When a communication link interruption is detected, it switches to autonomous operation mode and continues to execute the current task, while caching all collected data. When encountering unavoidable obstacles or dangerous environmental conditions, the hovering waiting mechanism is activated to continuously monitor environmental changes and continue to perform the mission after the safety conditions are restored; During task execution, the rule base weight parameters of the dynamic fuzzy logic controller are dynamically updated based on the image quality distribution and energy efficiency data of the completed tasks to optimize the subsequent trajectory planning strategy.

10. A UAV autonomous inspection trajectory planning system based on a real-time map of the channel environment, applied to the method described in any one of claims 1 to 9, characterized in that, The system includes: The environmental perception and map building module is used to collect channel environmental data in real time based on UAV airborne fusion sensors, build a real-time map containing semantic information online and generate a Euclidean symbol distance field simultaneously. The global trajectory planning module is used to generate safe flight corridors connecting each inspection target point online based on the real-time map and distance field, and to plan a global spatiotemporal trajectory that satisfies dynamic and energy constraints under the constraints of the safe flight corridors using a sampling planning algorithm guided by a dynamic fuzzy logic controller that integrates energy state perception. The local trajectory optimization and obstacle avoidance module is used to perform real-time trajectory optimization and dynamic obstacle avoidance using a local planner that integrates a velocity obstacle mechanism under the constraint of a safe flight corridor, with the global spatiotemporal trajectory as a reference. The task execution and quality assessment module is used to execute control commands to control the UAV to complete inspection flights and image acquisition, and to perform real-time visual quality assessment on the acquired images. The closed-loop decision-making and adjustment module is used to dynamically decide whether to trigger trajectory adjustment or execute emergency strategies based on the visual quality assessment results and the real-time status information of the UAV.

Images