Unmanned aerial vehicle inspection dynamic path planning method and system based on channel feature points

Abstract

The application discloses a dynamic path planning method and system for unmanned aerial vehicle inspection based on channel feature points, and belongs to the technical field of unmanned aerial vehicle path planning. The method comprises the following steps: a directed weighted graph is constructed as a feature point map; dynamic environment information is acquired in real time, and is spatiotemporally associated with the feature point map, so as to give dynamic risk coefficients and dynamic flight costs to nodes and edges; global path planning is performed on the graph with updated weights, and an inspection path composed of a feature point sequence is obtained; the risk is continuously monitored during flight, when the risk exceeds a threshold, a spatiotemporal forbidden area is generated based on a dynamic obstacle prediction trajectory, local dynamic re-planning is performed on a subgraph with the corresponding nodes and edges removed, and a safe path is obtained. The application realizes the leap of unmanned aerial vehicle inspection from'static open-loop execution' to 'dynamic closed-loop intelligent response', and improves the safety, reliability and overall efficiency of the inspection task in a complex channel environment.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) path planning technology, and in particular to a dynamic path planning method and system for UAV inspection based on airway feature points. Background Technology

[0002] With the deepening development of smart waterways and digital transportation, drones, with their advantages of high mobility, wide coverage, and low operating costs, have been widely applied in waterway inspection, port monitoring, maritime supervision, and emergency response. Drone inspection systems can be equipped with high-definition cameras, infrared thermal imaging, and multispectral sensors to effectively collect key data such as ship traffic flow, navigation mark status, shoreline utilization, and navigation environment along the waterway. This provides crucial technical support for building digitalized waterways and intelligent operational monitoring.

[0003] Inland waterways and coastal waterways are generally characterized by complex navigation environments, high vessel density, and variable hydrological and meteorological conditions. Traditional manual inspection methods, such as boat or vehicle patrols, suffer from inherent drawbacks such as low efficiency, limited field of vision, and delayed response, making it difficult to meet the demands of high-frequency, large-scale, and detailed inspections. Therefore, developing unmanned aerial vehicle (UAV)-based autonomous inspection technology to shift from "manual remote control" to "automatic task execution" has become an urgent need to improve the public service capabilities and emergency response speed of waterway management departments.

[0004] In practical applications, existing UAV (Unmanned Aerial Vehicle) airway inspection path planning technology mainly relies on preset fixed flight path patterns. Technicians manually mark a series of waypoints on a two-dimensional map based on offline geographic information system data or historical remote sensing imagery, and connect them sequentially to form a closed or open flight path covering the entire section to be inspected. This flight path typically employs a "bow" or "serpentine" scanning strategy to ensure coverage of the waterway surface and features on both banks. The generated flight path file is then pre-uploaded to the UAV's onboard flight control system or cloud management platform. During the inspection mission, the UAV strictly follows the preset flight path, and its onboard camera triggers images at a fixed frequency or equal interval, storing the collected images or video data on an onboard memory card or transmitting it back to the ground control station in real time.

[0005] However, the aforementioned existing solutions have at least the following technical drawbacks: First, the static environment assumption leads to a lack of adaptability to dynamic waterway elements. The core of existing solutions is route pre-planning based on static or quasi-static geographic information, without effectively integrating real-time dynamic data sources within the waterway into its decision-making closed loop. The system cannot dynamically integrate real-time hydrological sensor data, vessel automatic identification system signals, and meteorological monitoring data during route generation and execution. This "static planning, open-loop execution" model leaves UAVs lacking the ability for autonomous perception, real-time decision-making, and dynamic replanning when facing sudden severe weather, uncontrollable vessels temporarily appearing in the waterway, or large dynamic obstacle avoidance scenarios, posing flight safety risks and the possibility of missing key targets.

[0006] Secondly, local optimization strategies are prone to generating global task conflicts and resource imbalances. Existing solutions, when faced with complex inspection tasks involving multiple UAVs and long ranges, generally employ local or greedy optimization strategies. This planning approach, lacking overall coordination, easily leads to system-level conflicts during multi-UAV collaborative operations. For example, multiple UAVs independently responding to local dynamic events can cause high-frequency task overlap, airspace resource competition, and inter-UAV avoidance risks, resulting in an imbalance in the allocation of overall inspection resources and a systemic decline in inspection efficiency.

[0007] Third, the lack of an energy management model makes it difficult for the actual effective endurance to support long-range missions. Existing solutions mostly use simplified energy consumption estimation models, which typically assume that the UAV flies at a constant speed in a standard atmospheric environment and simply treats power consumption as a linear relationship proportional to the path length. This model does not establish a dynamic correlation between power consumption and real-time flight attitude, payload power consumption, ambient wind field, and flight speed. This lack of energy management means that the system can only rely on a preset fixed low power threshold to trigger a return to base. In long-distance, strong wind scenarios, the UAV may either underestimate energy consumption and land prematurely, or overestimate its endurance and fail to complete the intended mission, severely limiting the effective operating radius and mission execution reliability of the UAV. Summary of the Invention

[0008] This invention overcomes the shortcomings of the prior art and provides a dynamic path planning method and system for UAV inspection based on airway feature points.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention discloses a dynamic path planning method for UAV inspection based on airway feature points, comprising the following steps: S1. Construct a channel feature point map of the target channel. The channel feature point map is a directed weighted graph composed of multiple channel feature points extracted from the channel geographic information in advance as nodes and the navigable paths between nodes as edges. S2. Real-time acquisition of dynamic environmental data within the waterway, spatiotemporal correlation of the dynamic environmental data with nodes and edges in the waterway feature point map, generation of dynamic risk coefficient for each node and dynamic flight cost for each edge, and updating the weights of the directed weighted graph. S3. Based on the task feature points that need to be covered as determined by the inspection task, perform global path planning on the directed weighted graph after weight update to obtain a global inspection path composed of a sequence of feature points. S4. Control the drone to fly along the global inspection path, and during the flight, monitor the dynamic risk of the feature points ahead based on the real-time updated dynamic risk coefficient and dynamic flight cost; when the dynamic risk coefficient of any feature point or the dynamic flight cost of any edge exceeds the preset safety threshold, trigger local dynamic replanning. S5. The local dynamic replanning includes: generating a spatiotemporal restricted area that is prohibited from entering in the future based on the predicted motion trajectory of the dynamic obstacle; after removing the nodes and edges involved in the spatiotemporal restricted area from the current directed weighted graph, re-searching on the remaining subgraph to obtain a local safe path to avoid the dynamic obstacle, and controlling the UAV to fly along the local safe path.

[0010] Furthermore, the waterway feature points include geometric feature points, navigation feature points, facility feature points, and risk feature points. Among them, the geometric feature points include at least the apex of a bend and a junction; the navigation feature points include at least the convergence area of ​​ship flows and the anchorage; the facility feature points include at least navigation marks and hydrological stations; and the risk feature points include at least accident-prone areas and the edge of shoals.

[0011] Furthermore, the dynamic environmental data includes real-time AIS ship data, hydrological and meteorological data, and UAV video stream detection results; The spatiotemporal correlation specifically refers to: For each feature point, calculate the dynamic set of ships within its radius of influence, generate a dynamic risk label, and use it as the basis for the dynamic risk coefficient; For each edge, a wind resistance correction factor calculated based on real-time wind speed and direction is superimposed, and combined with the physical distance to form the dynamic flight cost.

[0012] Furthermore, the global path planning employs an improved A* algorithm, with the following cost function: f(n) = g(n) + h(n) + λ·Risk(n) Where g(n) is the actual cost from the starting point to node n; h(n) is the heuristic function; Risk(n) is the dynamic risk coefficient of node n; and λ is the risk weight factor.

[0013] Furthermore, in local dynamic replanning, spatiotemporal restricted areas that are prohibited from entering in the future are generated based on the predicted motion trajectories of dynamic obstacles, specifically: Based on the current position, speed and heading of the dynamic vessel, the trajectory within a future time window is predicted, and the feature points and their associated edges traversed by the trajectory are marked as prohibited elements within the corresponding time window, forming a spatiotemporal prohibited region. On the subgraph after removing the spatiotemporally forbidden regions, an improved A* algorithm is used to re-search for local safe paths.

[0014] Furthermore, when deploying multiple drones, the method also includes a multi-drone collaborative task allocation step: The waterway to be inspected is divided into multiple task sections, each of which contains a set of continuous feature points, forming a task pool. The system acquires the real-time status of each drone and performs global optimal allocation based on the contract network protocol, so as to minimize the total task completion time and balance the energy consumption of each drone, while avoiding multiple drones from entering the same high-risk feature point at the same time. When multiple drones are detected to be approaching the same feature point at the same time, the relevant drones are controlled to hover or detour according to the preset task priority instructions to resolve the conflict between drones.

[0015] Furthermore, it also includes dynamic energy management steps: A basic energy consumption model is established, which determines the basic energy consumption based on flight distance, flight speed, and payload. This basic energy consumption is denoted as E. base ; Introducing environmental correction factor δ wind =1+k w ·(v wind ·cosθ) 2 , where v wind Let θ be the wind speed, θ be the angle between the wind direction and the drone's flight direction, and k be the wind speed. w These are the coefficients obtained by fitting historical flight data; During flight, for the waiting segment, the segment energy consumption E is predicted based on the aforementioned environmental correction factor. segment =E base ·δ wind ; The deviation between actual current consumption and predicted energy consumption is compared in real time. If the deviation continues to exceed the preset deviation threshold, the energy consumption coefficient of the remaining flight range is dynamically adjusted, and the energy consumption of subsequent routes is verified. If the verification result indicates that the remaining battery power is insufficient to continue the mission, the local dynamic replanning is triggered to adjust the path or control the drone to return to home.

[0016] Furthermore, it also includes: After each inspection mission is completed, the waterway feature point map is incrementally updated based on the dynamic events actually perceived during the mission. Flight data is uploaded to the cloud, and machine learning algorithms are used to periodically update the risk weight factor λ in the improved A* algorithm of the global path planning, and / or update the environmental correction factor coefficient k in the dynamic energy management step. w .

[0017] The second aspect of the present invention discloses a dynamic path planning system for UAV inspection based on airway feature points. The system includes a memory and a processor. The memory stores a program for a dynamic path planning method for UAV inspection based on airway feature points. When the program for a dynamic path planning method for UAV inspection based on airway feature points is executed by the processor, the steps of the dynamic path planning method for UAV inspection based on airway feature points as described in any one of the present invention are implemented.

[0018] This invention addresses the technical deficiencies in the prior art and has the following beneficial effects: This invention replaces continuous coordinate points with discrete waterway feature points as the basic unit of path planning, and integrates real-time AIS vessel dynamics, hydrological and meteorological information, and other multi-source information with the feature point network in a spatiotemporal manner. This constructs a two-layer adaptive architecture of "global pre-planning and local dynamic re-planning," fundamentally overcoming the bottleneck of traditional solutions' insufficient adaptability to dynamic navigation environments. Simultaneously, it introduces a dynamic energy consumption prediction model coupled with environmental wind fields and a multi-aircraft collaborative global scheduling mechanism with feature point segments as the granularity. While ensuring flight safety, this effectively solves the problems of unreliable endurance under complex weather conditions and multi-aircraft task conflicts, realizing a leap from "static open-loop execution" to "dynamic closed-loop intelligent response" for UAV inspection, improving the safety, reliability, and overall system efficiency of inspection tasks in complex waterway environments. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.

[0020] Figure 1 Flowchart of a dynamic path planning method for UAV inspection based on airway feature points; Figure 2 This is a diagram of the dynamic path planning system architecture for UAV inspection based on airway feature points. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0022] Example 1

[0023] like Figure 1 As shown, this embodiment provides a dynamic path planning method for UAV inspection based on airway feature points. This method can be run in the system provided in subsequent embodiments of the present invention, and is implemented by the system's processor executing program instructions stored in the memory.

[0024] To facilitate understanding, the core concept of this invention will first be briefly described. This invention breaks away from the traditional approach of using continuous geographic coordinates as the basic unit of path planning, proposing instead that "channel feature points" be used as the basic atomic unit of path planning. Channel feature points refer to discrete location points with clear navigational or safety significance, pre-extracted from channel geographic information, such as bend apex, channel intersections, ship flow convergence areas, navigation mark locations, and accident-prone areas. By constructing a directed weighted graph with these feature points as nodes and flyable paths between nodes as edges, the path planning problem in complex continuous space is reduced to a graph search problem on a discrete feature point network, thus reducing planning complexity while ensuring the effectiveness of inspection coverage. Based on this, dynamic environmental information such as real-time AIS ship data and hydrological and meteorological data are further introduced and spatiotemporally correlated and fused with feature point networks. Dynamic weights are assigned to nodes and edges in the graph, thus forming a two-layer path planning architecture of "global pre-planning and local dynamic replanning". With the help of dynamic energy consumption management model and multi-machine collaboration mechanism, autonomous, safe and efficient dynamic planning of UAV inspection paths in complex general aviation environments is realized.

[0025] The steps of this method will be explained in detail below.

[0026] Step S1: Construct a map of the target waterway's feature points.

[0027] The core of this step is to replace traditional continuous path points with "channel feature points" as the basic atomic units for path planning. The channel feature point map is a directed weighted graph composed of multiple channel feature points extracted from channel geographic information in advance as nodes and traversable paths between nodes as edges.

[0028] Specifically, the extraction of waterway feature points is completed automatically or semi-automatically based on high-precision electronic waterway charts (ENC) and survey point cloud data. In this embodiment, waterway feature points are divided into four types: geometric feature points, navigation feature points, facility feature points, and risk feature points.

[0029] Among these, geometric feature points include at least bend apex and channel intersections; these feature points represent key turning or bifurcation points in the channel's spatial morphology and are nodes that must be passed through or turned at in route planning. Navigation feature points include at least convergence areas of vessel traffic and anchorage points; these feature points represent areas with high vessel traffic density requiring focused inspection. Facility feature points include at least navigation marks and hydrological stations; these feature points are channel infrastructure that requires inspection or monitoring. Risk feature points include at least accident-prone areas and shoal edges; these feature points are high-risk areas requiring close monitoring or detour.

[0030] After extracting feature points, a feature point network is constructed. Specifically, each feature point is treated as a node in the graph, and the existence of connected edges between nodes is determined based on whether a flyable, unobstructed path exists between them. The criteria for determining a flyable path include factors such as straight-line visibility, no-fly zone avoidance, and safe distance from obstacles. If a flyable path exists between two nodes, a directed edge is created between them. Each edge is assigned an initial weight, which can be determined based on the physical distance between nodes or further superimposed with a static risk coefficient. For example, an edge passing through a shoal edge, even if the physical distance is short, can have a higher initial weight to reflect its potential flight risk. This results in a directed weighted graph covering the entire target flight path, composed of feature point nodes and directed weighted edges. This graph structure is stored on edge computing nodes and a cloud platform, supporting incremental updates.

[0031] As a specific example, the target waterway is the "Y"-shaped intersection area of ​​the 200,000-tonnage channel in Guangzhou Port and its surrounding connecting waters. This area contains the intersection of the main channel and branch channels (geometric feature points), areas of dense vessel flow (navigation feature points), multiple navigation marks and hydrological stations (facility feature points), and the edge of a historical shoal (risk feature point). The system extracts dozens of these feature points from the high-precision electronic waterway map of the area and constructs connecting edges based on flight path determination, forming a waterway feature point map covering the target waterway.

[0032] Step S2: Acquire dynamic environmental data within the waterway in real time, and spatiotemporally correlate the dynamic environmental data with nodes and edges in the waterway feature point map. Generate a dynamic risk coefficient for each node and a dynamic flight cost for each edge, and update the weights of the directed weighted graph.

[0033] The core objective of this step is to deeply integrate multi-source real-time environmental data with a static waterway feature point map to generate a dynamic weighted map that reflects the current navigation dynamics.

[0034] Dynamic environmental data includes real-time AIS vessel data, hydro-meteorological data, and UAV video stream detection results. Among them, AIS vessel data provides information such as the position, speed, and heading of surrounding vessels; hydro-meteorological data includes at least wind speed, wind direction, visibility, current speed, and water level; and UAV video stream detection results can identify water surface targets in real time through an airborne visual perception system.

[0035] Data access is achieved through standardized interfaces. For example, AIS data can be obtained through shore-based AIS base stations or maritime VTS system interfaces, while hydrological and meteorological data can be obtained through hydrological and meteorological stations deployed along the waterway.

[0036] The specific process of spatiotemporal correlation is as follows: For each feature point, the dynamic set of vessels within its radius of influence is calculated, generating a dynamic risk label, which serves as the basis for the dynamic risk coefficient. The radius of influence can be set according to the type of feature point and safety requirements; for example, a smaller radius can be set for the apex of a curve, while a larger radius can be set for intersections or convergence zones of vessel flows. The dynamic risk coefficient can be graded; for example, a medium risk value is assigned when there are stationary or low-speed vessels within the radius of influence, and a high risk value is assigned when there are high-speed approaching vessels. As one implementation method, a risk level mapping table can be predefined, mapping the dynamic risk coefficient based on the nearest encounter distance (CPA) and nearest encounter time (TCPA) between the vessel and the feature point.

[0037] For each edge, a wind resistance correction factor calculated based on real-time wind speed and direction is superimposed, and combined with the physical distance to form the dynamic flight cost. Specifically, for an edge connecting node A and node B, its physical distance d is first obtained, and then the dynamic flight cost is calculated based on the real-time wind speed v. wind The drag correction factor is calculated based on the angle between the wind direction and the edge direction. Flight cost decreases with the wind and increases with the wind. The dynamic flight cost can be expressed as the product or weighted sum of the physical distance d and the drag correction factor.

[0038] Therefore, each node in the route feature point map is assigned a dynamic risk coefficient, and each edge is assigned a dynamic flight cost, forming a dynamic weighted map. This dynamic weighted map will be used in subsequent path planning steps.

[0039] Step S3: Based on the task feature points that need to be covered as determined by the inspection task, perform global path planning on the directed weighted graph after weight update to obtain a global inspection path composed of the feature point sequence.

[0040] Inspection tasks are issued by operators through a shore-based visualization platform. Operators can define the waterway area to be inspected on the interactive interface, and the system automatically extracts relevant waterway feature points in and around that area as task feature points, i.e., the target nodes that the inspection task needs to cover or pass through. At the same time, the system's takeoff point (such as the location of the UAV's nest) is also mapped as a node on the feature point map.

[0041] Global path planning employs an improved A* algorithm. The cost function of the improved A* algorithm is: f(n) = g(n) + h(n) + λ·Risk(n) Where g(n) is the actual cost consumed from the starting point to the current node n, which can be obtained by accumulating the dynamic flight costs along each edge; h(n) is a heuristic function representing the estimated remaining cost from node n to the target node, which can usually be obtained by using the physical distance between nodes as a heuristic value; Risk(n) is the current dynamic risk coefficient of node n; λ is a risk weighting factor used to adjust the proportion of risk items in the total cost. λ can be set by the operator according to the actual application scenario, or optimized from historical flight data through machine learning.

[0042] By improving the A* algorithm, an optimal or suboptimal path is searched on the dynamic weight map, passing through all task feature points in sequence, and an ordered sequence of feature points is output, for example: takeoff point → feature point A → feature point C → feature point F → ... → takeoff point. This ordered sequence of feature points is the global inspection path.

[0043] Step S4: Control the UAV to fly along the global inspection path, and during the flight, monitor the dynamic risk of the feature points ahead based on the real-time updated dynamic risk coefficient and dynamic flight cost; when the dynamic risk coefficient of any feature point or the dynamic flight cost of any edge exceeds the preset safety threshold, trigger local dynamic replanning.

[0044] After the drone automatically takes off from its nest, the flight control system uses the first feature point as the current waypoint based on the sequence of feature points defined in the global inspection path, and flies along the edges between those feature points. During flight, onboard sensors continuously transmit video streams, and edge computing nodes or cloud platforms continuously receive the latest AIS and hydrological and meteorological data, constantly updating the dynamic weights of the feature point map.

[0045] The system continuously monitors the dynamic risks of feature points ahead of the flight path. Specifically, before the drone approaches the next feature point to be visited, the edge computing node checks the real-time dynamic risk coefficient of that feature point. Simultaneously, it checks the dynamic flight cost of the edges corresponding to the preceding flight segment. If the dynamic risk coefficient of a feature point exceeds a preset safety threshold (e.g., an out-of-control vessel suddenly appears within the feature point's influence radius), or if the dynamic flight cost of an edge exceeds a safety threshold due to a sudden increase in wind speed, the current flight segment is determined to have unacceptable risks, and local dynamic replanning is automatically triggered.

[0046] Step S5: Local Dynamic Replanning. This step includes: generating a spatiotemporally prohibited region that is forbidden to enter in the future based on the predicted motion trajectory of the dynamic obstacle; after removing the nodes and edges involved in the spatiotemporally prohibited region from the current directed weighted graph, re-searching on the remaining subgraph to obtain a local safe path to avoid the dynamic obstacle, and controlling the UAV to fly along the local safe path.

[0047] Local dynamic replanning is completed by edge computing nodes within milliseconds. Specifically, dynamic obstacles that trigger replanning are first identified. Taking a dynamic ship as an example, based on the ship's current position, speed (SOG), and heading (COG), and assuming it maintains its current motion state, the ship's trajectory is predicted within a preset time window (e.g., 1 to 3 minutes in the future). A uniform linear motion prediction model is used as the basis here, but a more complex ship trajectory prediction model based on channel constraints can also be employed.

[0048] The feature points traversed by the predicted trajectory and their associated edges are marked as prohibited elements within the corresponding time window. In other words, these nodes and edges are not permanently deleted, but rather marked as unavailable for a specific future time period. This creates a spatiotemporally prohibited region.

[0049] In the current directed weighted graph, nodes and edges involved in the spatiotemporally forbidden region are removed. That is, on the subgraph with the remaining nodes and edges, an improved A* algorithm is used to re-search for a local safe path from the UAV's current position to the subsequent task feature point. The re-planning still uses the cost function f(n) = g(n) + h(n) + λ·Risk(n), at which point the candidate nodes and edges have excluded those affected by dynamic obstacles. The re-searched local safe path will replace the corresponding flight segment in the original global inspection path.

[0050] Edge computing nodes send the replanned local safe path to the UAV flight control system through a dedicated 5G communication network. The UAV then adjusts its course and flies along the new safe path, enabling it to actively avoid dynamic obstacles.

[0051] Implementation of multi-machine collaborative task allocation: When multiple drones are deployed to perform inspection tasks collaboratively, this method also includes a multi-drone collaborative task allocation step: First, the waterway to be inspected is divided into multiple task segments, each containing a set of continuous feature points. This division can be based on the geographical distribution of the feature points and the workload. The divided task segments are then placed into a task pool, forming a set of tasks to be assigned.

[0052] The system acquires the real-time status of each drone, including its current location, remaining battery power, and current payload status. Global optimal allocation is performed based on the Contract Network Protocol. The basic process of the Contract Network Protocol is as follows: the cloud platform, acting as the bidding party, publishes task information from the task pool to each drone (bidder); each drone calculates the cost of executing each task based on its own status and task requirements, and submits a bid; the cloud platform aggregates all bids and allocates tasks based on the principle of optimizing the system's global objective. The global objective can be set as minimizing the total task completion time and balancing the energy consumption of each drone, while constraining the allocation scheme to avoid multiple drones being assigned to the same high-risk feature point area simultaneously, reducing the possibility of inter-drone conflicts from the allocation source.

[0053] When multiple drones are detected approaching the same feature point simultaneously during flight (i.e., potential inter-drone conflict exists), the relevant drones are controlled to hover or detour according to preset task priority instructions to resolve the conflict. Task priority can be dynamically determined based on factors such as task type (e.g., emergency response takes precedence over routine inspection) and drone status.

[0054] Regarding the implementation of dynamic energy management: This method also includes dynamic energy management steps, which are implemented throughout the entire flight process.

[0055] First, a basic energy consumption model is established. This model determines the basic energy consumption based on flight distance, flight speed, and payload, and this basic energy consumption is denoted as E. base As a specific implementation method, the basic energy consumption can be obtained by fitting flight test data, for example, by establishing a multinomial regression model with flight distance d, flight speed v, and payload m as independent variables.

[0056] An environmental correction factor is introduced to compensate for the impact of environmental wind fields on energy consumption. The environmental correction factor is defined as follows: δ wind =1+k w ·(v wind ·cosθ) 2 Where v_wind is the wind speed, θ is the angle between the wind direction and the drone's flight direction, θ=0° indicates a completely headwind, and θ=180° indicates a completely tailwind; k wThese are coefficients obtained by fitting historical flight data, reflecting the aerodynamic shape and propulsion system characteristics of the UAV. When there is a tailwind, cosθ is negative, and δ... wind A value less than 1 indicates reduced energy consumption; when facing headwinds, δ wind A value greater than 1 indicates an increase in energy consumption.

[0057] During flight, for the waiting segment, the segment energy consumption E is predicted based on the aforementioned environmental correction factor. segment Satisfying E segment = E base ·δ wind .

[0058] The system compares the actual current consumption of the drone with the predicted energy consumption in real time. This is done by monitoring the real-time discharge current of the drone's battery, calculating the actual energy consumption, and comparing it with the predicted energy consumption for the corresponding flight segment. If the deviation consistently exceeds a preset deviation threshold (e.g., a deviation exceeding 15% for 30 consecutive seconds), it is determined that the current energy consumption model deviates significantly from actual operating conditions. The system then dynamically adjusts the energy consumption coefficient for the remaining flight distance, for example, by introducing an online correction factor to compensate for subsequent predictions. Finally, the energy consumption of the remaining inspection path is verified based on the adjusted prediction model.

[0059] If the verification results indicate that the remaining battery power is insufficient to support the continued execution of all remaining tasks, the local dynamic replanning is triggered. Possible adjustment strategies include: shortening the inspection path, prioritizing the inspection of important feature points before returning to base; or directly aborting the task, planning the optimal path back to the nest from the current location, and controlling the drone to return safely. This ensures that the drone will not make an early forced landing due to insufficient energy consumption estimation.

[0060] Implementation of incremental updates and model optimization for waterway feature point maps: This method also includes a closed-loop optimization mechanism.

[0061] After each inspection task is completed, the system incrementally updates the waterway feature point map based on the dynamic events actually detected during the task. For example, if a new construction vessel is found to be permanently stationed in a certain area during the inspection, that area can be added as a temporary feature point and marked; if the risk level of an accident-prone area changes, the attributes of the corresponding risk feature points can be updated. The results of the incremental updates are stored on edge nodes and in the cloud for use in subsequent tasks.

[0062] Simultaneously, flight data from each mission (including actual flight trajectory, energy consumption data, obstacle avoidance event records, AIS encounter data, etc.) is uploaded to a cloud-based machine learning platform. The cloud platform utilizes the accumulated dataset to periodically optimize and train key parameters in the algorithm model. For example, using labeled data from successful obstacle avoidance and risk encounters in historical missions, the risk weight factor λ in the A* algorithm's cost function is optimized and improved; using actual energy consumption data under different wind field conditions during historical flights, the environmental correction factor coefficient k in the dynamic energy consumption model is fitted and updated. w The updated optimization parameters are distributed from the cloud to the edge nodes, enabling continuous iterative improvement of algorithm performance.

[0063] In this embodiment, it also includes: For each feature point in the waterway feature point map, multi-source dynamic environmental data composed of real-time AIS ship data, hydrological and meteorological data, and video stream detection results are collected within its influence radius. Ship density, average speed variance, wind speed and wind direction combination vector, visibility index, and frequency of moving target intrusion are extracted from it. The rate of change of each statistic within a preset time window is calculated. The statistic and the rate of change are spliced ​​and normalized to form a multi-dimensional dynamic risk feature vector of the corresponding feature point at the current moment. Configure a Dirichlet process mixture model, select a multivariate Gaussian distribution that matches the dimension of the risk feature vector as the base distribution, and set a concentration parameter to control the growth tendency of the number of clusters; regard the set of risk feature vectors of all feature points as the dataset to be clustered, and randomly assign an initial risk pattern atom label to each feature point, with each risk pattern atom representing a potential risk pattern; Sampling iteration: In each iteration, for each feature point, its current risk mode atom label is temporarily removed. Based on the atom allocation, basis distribution and concentration parameters of other feature points, the probability of the feature point belonging to each existing atom and the probability of it belonging to a new atom are calculated. New atom labels are obtained by sampling according to the probability distribution, and the statistics of the corresponding atoms are updated. Iteration continues until the number of atoms and the allocation are stable. At this time, the number of atoms automatically inferred is the number of risk mode categories, and each feature point obtains a risk mode atom hard allocation. Using the converged sampling state, the posterior probability of each feature point belonging to each risk pattern atom is calculated to form a probabilistic risk label. This label is a soft assignment vector that indicates that the feature point belongs to each risk pattern with different probabilities. At the same time, the multivariate Gaussian distribution parameters corresponding to each risk pattern atom are extracted and persisted to form a risk pattern base that describes the current environmental risk pattern morphology. When the dynamic environmental data of the waterway is updated, the risk feature vector of the feature point changes. Using the stored risk pattern base as a priori, the updated risk feature vector is sampled and iterated, and a probabilistic risk label is assigned to it based on the fixed existing risk pattern atoms. Continuously monitor the cumulative likelihood of new data on existing risk pattern bases. When the cumulative likelihood drops below a preset threshold, a new risk pattern is identified. New risk pattern atoms are automatically introduced and the probabilistic risk labels of all feature points are readjusted, so that risk perception capabilities can adaptively grow with environmental evolution.

[0064] It should be noted that in actual waterway inspections, the risk patterns of ship movements near the same feature point may differ fundamentally under different seasons, time periods, and traffic densities. Furthermore, new risk patterns may emerge with changes in port operations or waterway conditions. Relying solely on static risk level mapping rules makes it difficult to capture structural changes in risk patterns. Therefore, this embodiment collects multi-source dynamic environmental data within the influence radius of each feature point on the waterway feature point map, composed of real-time AIS ship data, hydrological and meteorological data, and video stream detection results. Several statistical measures are extracted from this data. Among them, ship density reflects the congestion level of the waters surrounding the feature point; the average speed variance reflects the consistency of ship movement states, with a larger variance indicating more unpredictable ship behavior; the wind speed and direction combination vector characterizes the impact of current wind force on flight and ship maneuvering; the visibility index reflects the limitations of perception capabilities imposed by meteorological conditions such as fog and haze; and the frequency of moving target intrusion directly counts the number of ships entering the influence radius of the feature point within a specified time window. Calculate the rate of change of each statistic within a preset time window, concatenate the statistic and its rate of change and normalize them to form a multidimensional dynamic risk feature vector corresponding to the current moment of each feature point.

[0065] Next, the Dirichlet process mixture model is configured. The Dirichlet process mixture model is a non-parametric Bayesian model whose advantage lies in that it does not require pre-specifying the number of clusters, but can automatically infer a reasonable number of categories based on the distribution characteristics of the data itself. In this embodiment, a multivariate Gaussian distribution matching the dimension of the risk feature vector is selected as the base distribution to describe the form of each risk pattern. Simultaneously, a concentration parameter is set to control the tendency for the number of clusters to grow. The larger the concentration parameter, the more likely the model is to generate new risk pattern categories; the smaller the concentration parameter, the more likely it is to use fewer categories to explain the data. The specific value of the concentration parameter can be set empirically or determined through prior knowledge or cross-validation.

[0066] The set of risk feature vectors for all feature points is considered as the dataset to be clustered. During initialization, each feature point is randomly assigned an initial risk pattern atom label, where each risk pattern atom represents a potential risk pattern. Here, "atom" is the term for a cluster in the Dirichlet process mixture model, and each atom corresponds to a category described by the basis distribution parameters. After initialization, iterative sampling is performed. In each iteration, the following operations are performed on each feature point: its current risk pattern atom label is temporarily removed; based on the current atom allocation of the remaining feature points, the basis distribution, and the concentration parameters, the probability that the feature point belongs to any existing atom and the probability that it belongs to a new atom are calculated. The probability of belonging to an existing atom is proportional to the current number of members of that atom and the likelihood of the feature point under the multivariate Gaussian distribution corresponding to that atom; the probability of belonging to a new atom is proportional to the marginal likelihood of the feature point under the concentration parameters and the basis distribution. Sampling is performed based on the calculated probability distribution to obtain a new atom label for the feature point. After completing one round of label updates for all feature points, the statistics of each atom are updated accordingly.

[0067] The iterative process described above is repeated until the number and allocation of atoms stabilize, indicating model convergence. After convergence, the number of atoms automatically inferred by the model represents the number of objectively existing risk pattern categories in the current water area, and each feature point receives a hard assignment result for a risk pattern atom. It is worth noting that the number of risk pattern categories is not preset but emerges automatically from the data, allowing the system to adapt to water environments of varying complexity. Based on this, posterior inference is performed using the converged sampling state. The posterior probability of each feature point belonging to each risk pattern atom is calculated, forming a probabilistic risk label. This probabilistic risk label does not rigidly classify the feature point into a certain category but indicates that the feature point belongs to each risk pattern with varying probabilities, i.e., a soft assignment vector. Simultaneously, the multivariate Gaussian distribution parameters corresponding to each risk pattern atom are extracted and persisted, including the mean vector and covariance matrix. These parameters together constitute a risk pattern base describing the current environmental risk pattern morphology. The risk pattern base can be viewed as a snapshot of the system's understanding of the current water area risk structure, stored on edge computing nodes and in the cloud for subsequent tasks.

[0068] When updates to the dynamic environmental data of the waterway cause changes in the risk feature vectors of feature points, the system does not re-execute the complete clustering process from scratch. Instead, it uses the stored risk pattern base as a priori and continues to perform sampling iterations on the updated risk feature vectors. Based on the fixed existing risk pattern atoms, probabilistic risk labels are assigned to the updated feature points, achieving lightweight, online risk pattern updates.

[0069] This embodiment also possesses the capability to autonomously discover new risk patterns. The system continuously monitors the cumulative likelihood of new data against the existing risk pattern base. The cumulative likelihood characterizes the overall fit of new observation data with the current risk pattern base. When the cumulative likelihood decreases beyond a preset threshold, it indicates that the existing risk pattern base can no longer effectively explain the newly emerging risk state, and the system determines that a new risk pattern has appeared. At this point, new risk pattern atoms are automatically introduced, and the probabilistic risk labels of all feature points are readjusted, enabling the risk recognition capability to adaptively grow with environmental evolution.

[0070] In summary, this embodiment elevates the waterway inspection system's ability to perceive dynamic risks from the level of "rule matching" to "structural cognition," enabling the system to autonomously discover risk patterns without prior knowledge, express uncertainty in a probabilistic manner, and adaptively adjust its cognitive structure as the environment evolves. This provides a more accurate and robust basis for calculating risk coefficients in subsequent path planning.

[0071] In this embodiment, it also includes: Starting from the current position of the UAV that triggers local dynamic replanning, a path segment is extracted along the global inspection path to the safety exit point ahead. The sequence of feature points traversed by the path segment is used as control points to construct a thin plate spline curve that passes through all control points and satisfies the curvature continuity constraint, which serves as an elastic spline path model. Each control point is assigned a spline stiffness coefficient that characterizes the point's ability to resist bending deformation. The three-dimensional geometric boundary of the spatiotemporally forbidden region generated in the local dynamic replanning is projected onto the two-dimensional horizontal plane where the elastic spline is located, and its projected profile is extracted as the rigid constraint boundary. For the rigid constraint boundary, the shortest signed distance from each sampling point on the elastic spline path model to the boundary is calculated, and a penetration depth field reflecting the degree of the elastic spline intrusion into the forbidden region is constructed. Based on the elastic spline path model, the first-order deformation energy representing the deviation of the control point from its original position and the second-order bending energy representing the change in spline curvature are calculated respectively. According to the penetration depth field, a collision penalty potential energy proportional to the penetration depth is applied to the sampling point that intrudes into the rigid constraint boundary. The first-order deformation energy, the second-order bending energy and the collision penalty potential energy are weighted and summed to construct the total energy functional to be minimized. Using the two-dimensional coordinates of all control points as optimization variables and the total energy functional as the objective function, a constrained optimization solver is used for iterative search. In each iteration, the control points located in the penetration depth field are forced to slide along the outward normal of the rigid constraint boundary, so that the control points are strictly located on the passable side of the rigid constraint boundary, until the total energy functional converges, and an elastic obstacle avoidance spline path is obtained. The elastic obstacle avoidance spline path is discretized into a continuous waypoint sequence. Based on the spline stiffness coefficient assigned in step one, waypoints corresponding to control points with stiffness coefficients lower than a preset coefficient threshold are eliminated. The remaining waypoints are then sequentially mapped to the nearest feature point and feasible edge in the directed weighted graph, forming a local safe path that can be directly executed by the flight control system and is coordinated with the dynamic obstacle boundary under curvature continuity conditions.

[0072] It should be noted that after triggering local dynamic replanning, a path segment is extracted from the UAV's current position along the obtained initial local safe path to the forward safe exit point. The forward safe exit point refers to the first node on this path segment that re-enters a free navigation area unaffected by dynamic obstacles. Using the sequence of feature points traversed by the path segment as control points, a thin-plate spline curve passing through all control points and satisfying curvature continuity constraints is constructed as the elastic spline path model. Thin-plate splines are a type of interpolation spline widely used in computer graphics. Their basic physical prototype is the equilibrium state achieved by applying a concentrated force at a given constraint point to cause the elastic thin plate to bend and deform. Introducing them into path smoothing can naturally generate curves that satisfy second-order differential continuity, avoiding the curvature jumps at the endpoints of traditional polynomial splines.

[0073] In this embodiment, an additional spline stiffness coefficient is assigned to each control point to characterize its resistance to bending deformation. The spline stiffness coefficient can be set based on the type of feature point: for example, geometric feature points located in complex intersection areas can be assigned higher stiffness to maintain accurate passage through the critical navigation node; while ordinary feature points located in open water can be assigned lower stiffness to allow for a greater degree of smooth offset. This differentiated stiffness setting ensures that the deformation path maintains trajectory accuracy in critical areas while prioritizing smoothness in non-critical areas.

[0074] Next, the 3D geometric boundary of the generated spatiotemporally forbidden region is projected onto the 2D horizontal plane containing the elastic spline, and its projected profile is extracted as a rigid constraint boundary. For the rigid constraint boundary, the shortest signed distance from each sampling point on the elastic spline path model to this boundary is calculated, where a positive distance indicates that the sampling point is located on the passable side, and a negative distance indicates that the sampling point has intruded into the forbidden region. This constructs a penetration depth field reflecting the degree to which the elastic spline intrudes into the forbidden region.

[0075] At the physical simulation level, the deformation behavior of the elastic spline is driven by the total energy functional. The total energy functional constructed in this embodiment consists of a weighted sum of three potential energy terms: The first term is the first-order deformation energy, which characterizes the degree to which the control point deviates from its original position. This term corresponds to the elastic potential energy stored when the elastic spline is stretched or translated by an external force. Its physical function is to constrain the path during the optimization process so as not to deviate excessively from the original plan and to preserve the intention of the global path planning.

[0076] The second term is the second-order bending energy, which characterizes the magnitude of the spline curvature change, i.e., the strain energy corresponding to the degree of spline bending. This term corresponds to the internal energy of an elastic thin plate resisting bending, and its physical function is to suppress excessive curvature changes in the path during optimization, ensuring the smoothness of the flight trajectory.

[0077] The third term is the collision penalty potential energy, calculated based on the penetration depth field. It applies a penalty energy proportional to the penetration depth to the sampling points that intrude into the rigid constraint boundary. This term is equivalent to setting up a virtual repulsive force field on the rigid constraint boundary; the deeper the intrusion, the stronger the repulsion, forcing the spline to deform towards the passable region.

[0078] After weighted summing the above three terms, a total energy functional is obtained, with the two-dimensional coordinates of all control points as optimization variables. Using this functional as the objective function, a constrained optimization solver is employed for iterative search. In each iteration, control points located in the penetration depth field are forced to slide along the outward normal direction of the rigid constraint boundary, i.e., move away from the prohibited region, thus ensuring that the control points are strictly maintained on the passable side of the rigid constraint boundary. Iteration continues until the total energy functional converges, at which point an elastic obstacle avoidance spline path is obtained that closely approximates the original path while maintaining zero penetration depth and continuous curvature.

[0079] The elastic obstacle avoidance spline path is essentially still a continuous curve in a mathematical sense, and needs to be further transformed into a discrete waypoint sequence that can be directly executed by the flight control system. The discretization strategy in this embodiment is not equal-interval sampling, but rather differentiated based on the assigned spline stiffness coefficient: waypoints corresponding to control points with stiffness coefficients below a preset threshold are eliminated, while waypoints corresponding to control points with higher stiffness coefficients are retained. The physical meaning is that high-stiffness control points are critical nodes marked by the original path as requiring precise passage, and they still need to be retained as flight waypoints after discretization; while low-stiffness control points serve more of a smoothing transition function, their precise positions are constrained by the surrounding high-stiffness points, and can be omitted to reduce waypoint density.

[0080] Finally, the selected waypoints are sequentially mapped to the nearest feature points and feasible edges in the directed weighted graph, forming a local safe path that can be directly executed by the flight control system. This path satisfies obstacle avoidance requirements at the topological level, curvature continuity conditions at the geometric level, and maintains a coordinated distance from the boundaries of dynamic obstacles, achieving an optimized transformation from a graph search solution to a physically flyable trajectory.

[0081] Through this embodiment, when performing dynamic obstacle avoidance maneuvers, the UAV does not need to make abrupt turns at feature point nodes, but flies over complex waters along a smooth, continuous elastic deformation path. This improves flight safety, reduces maneuver energy consumption, and also reduces the burden on the underlying trajectory tracking of the flight control system.

[0082] Example 2

[0083] like Figure 2 As shown, this embodiment provides a dynamic path planning system for UAV inspection based on waterway feature points. The system adopts a three-layer collaborative architecture of "end-edge-cloud", including UAV nest end, edge computing nodes, cloud intelligent planning platform and shore-based display system.

[0084] The drone nesting layer includes both tethered and untethered drones. Each drone is equipped with an RTK high-precision positioning module, millimeter-wave radar, omnidirectional visual perception kit, miniature weather station, and AIS signal receiving module, forming the perception layer responsible for collecting its own status, surrounding environment, and dynamic ship information.

[0085] Edge computing nodes are deployed on the nest side, undertaking computational tasks such as local storage and updating of route feature point maps, dynamic replanning of local paths, and real-time energy consumption prediction, forming the computing layer. The introduction of edge computing nodes enables time-sensitive computing tasks such as dynamic replanning to be completed in milliseconds, without relying on communication latency from distant cloud environments.

[0086] The cloud-based intelligent planning platform is deployed at the shore-based control center. It is responsible for global feature point network management, multi-UAV collaborative task allocation, machine learning model training and updating, and data docking with the maritime VTS system, forming the decision-making layer.

[0087] The shore-based display system consists of a large visual screen and a mobile application, forming the application layer, which is used for dynamic monitoring of the waterway with a single map, task issuance, and viewing of inspection reports.

[0088] Low-latency, highly reliable real-time data exchange is achieved between different levels through a dedicated 5G communication network.

[0089] The system includes a memory and a processor. The memory stores a program for a UAV inspection dynamic path planning method based on airway feature points. When the program is executed by the processor, it implements all the steps of the UAV inspection dynamic path planning method based on airway feature points as described in Embodiment 1. The specific workflow and implementation details of this system have been described in detail in Embodiment 1 and will not be repeated here.

[0090] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic path planning method for UAV inspection based on airway feature points, characterized in that, Includes the following steps: Construct a channel feature point map of the target channel. The channel feature point map is a directed weighted graph consisting of multiple channel feature points extracted from channel geographic information in advance as nodes and flyable paths between nodes as edges. Real-time acquisition of dynamic environmental data within the waterway; spatiotemporal correlation of the dynamic environmental data with nodes and edges in the waterway feature point map; generation of dynamic risk coefficient for each node and dynamic flight cost for each edge; and updating of the weights of the directed weighted graph. Based on the task feature points that need to be covered as determined by the inspection task, global path planning is performed on the directed weighted graph after weight update to obtain a global inspection path composed of a sequence of feature points. The drone is controlled to fly along the global inspection path, and during the flight, the dynamic risk of the feature points ahead is monitored based on the real-time updated dynamic risk coefficient and dynamic flight cost; when the dynamic risk coefficient of any feature point or the dynamic flight cost of any edge exceeds the preset safety threshold, local dynamic replanning is triggered. The local dynamic replanning includes: generating a spatiotemporal restricted area that is prohibited from entering in the future based on the predicted motion trajectory of the dynamic obstacle; after removing the nodes and edges involved in the spatiotemporal restricted area from the current directed weighted graph, re-searching on the remaining subgraph to obtain a local safe path to avoid the dynamic obstacle, and controlling the UAV to fly along the local safe path.

2. The UAV inspection dynamic path planning method based on airway feature points according to claim 1, characterized in that, The waterway feature points include geometric feature points, navigation feature points, facility feature points, and risk feature points. Among them, the geometric feature points include at least the apex of a bend and a junction; the navigation feature points include at least the convergence area of ​​ship flows and the anchorage; the facility feature points include at least navigation marks and hydrological stations; and the risk feature points include at least accident-prone areas and the edge of shoals.

3. The UAV inspection dynamic path planning method based on airway feature points according to claim 1, characterized in that, The dynamic environmental data includes real-time AIS ship data, hydrological and meteorological data, and UAV video stream detection results. The spatiotemporal correlation specifically refers to: For each feature point, calculate the dynamic set of ships within its radius of influence, generate a dynamic risk label, and use it as the basis for the dynamic risk coefficient; For each edge, a wind resistance correction factor calculated based on real-time wind speed and direction is superimposed, and combined with the physical distance to form the dynamic flight cost.

4. The UAV inspection dynamic path planning method based on airway feature points according to claim 1, characterized in that, The global path planning uses an improved A* algorithm, whose cost function is: f(n) = g(n) + h(n) + λ·Risk(n) Where g(n) is the actual cost from the starting point to node n; h(n) is the heuristic function; Risk(n) is the dynamic risk coefficient of node n; and λ is the risk weight factor.

5. The UAV inspection dynamic path planning method based on airway feature points according to claim 1, characterized in that, In local dynamic replanning, spatiotemporal restricted areas that are prohibited from entering in the future are generated based on the predicted motion trajectories of dynamic obstacles. Specifically: Based on the current position, speed and heading of the dynamic vessel, the trajectory within a future time window is predicted, and the feature points and their associated edges traversed by the trajectory are marked as prohibited elements within the corresponding time window, forming a spatiotemporal prohibited region. On the subgraph after removing the spatiotemporally forbidden regions, an improved A* algorithm is used to re-search for local safe paths.

6. The UAV inspection dynamic path planning method based on airway feature points according to claim 1, characterized in that, When deploying multiple drones, the method further includes a multi-drone collaborative task allocation step: The waterway to be inspected is divided into multiple task sections, each of which contains a set of continuous feature points, forming a task pool. The system acquires the real-time status of each drone and performs global optimal allocation based on the contract network protocol, so as to minimize the total task completion time and balance the energy consumption of each drone, while avoiding multiple drones from entering the same high-risk feature point at the same time. When multiple drones are detected to be approaching the same feature point at the same time, the relevant drones are controlled to hover or detour according to the preset task priority instructions to resolve the conflict between drones.

7. The UAV inspection dynamic path planning method based on airway feature points according to claim 1, characterized in that, It also includes dynamic energy management steps: A basic energy consumption model is established, which determines the basic energy consumption based on flight distance, flight speed, and payload. This basic energy consumption is denoted as E. base ; Introducing environmental correction factor δ wind =1+k w ·(v wind ·cosθ) 2 , where v wind Let θ be the wind speed, θ be the angle between the wind direction and the drone's flight direction, and k be the wind speed. w These are the coefficients obtained by fitting historical flight data; During flight, for the waiting segment, the segment energy consumption E is predicted based on the aforementioned environmental correction factor. segment =E base ·δ wind ; The deviation between actual current consumption and predicted energy consumption is compared in real time. If the deviation continues to exceed the preset deviation threshold, the energy consumption coefficient of the remaining flight range is dynamically adjusted, and the energy consumption of subsequent routes is verified. If the verification result indicates that the remaining battery power is insufficient to continue the mission, the local dynamic replanning is triggered to adjust the path or control the drone to return to home.

8. The UAV inspection dynamic path planning method based on airway feature points according to claim 1, characterized in that, Also includes: After each inspection mission is completed, the waterway feature point map is incrementally updated based on the dynamic events actually perceived during the mission. Flight data is uploaded to the cloud, and machine learning algorithms are used to periodically update the risk weight factor λ in the improved A* algorithm of the global path planning, and / or update the environmental correction factor coefficient k in the dynamic energy management step. w .

9. A dynamic path planning system for UAV inspection based on airway feature points, characterized in that, The system includes a memory and a processor. The memory stores a program for a dynamic path planning method for UAV inspection based on airway feature points. When the program for the dynamic path planning method for UAV inspection based on airway feature points is executed by the processor, the steps of the dynamic path planning method for UAV inspection based on airway feature points as described in any one of claims 1 to 8 are implemented.

Images