Unmanned aerial vehicle roadway and drainage culvert inspection method

Abstract

The invention discloses an unmanned aerial vehicle roadway and drainage culvert inspection method, which comprises the following steps: constructing an environment prior model based on structural information of a roadway or a drainage culvert, and controlling an unmanned aerial vehicle to autonomously explore in an environment without a global navigation satellite system signal; constructing an environment map in real time, and carrying out registration correction on the environment map and real geographic data to generate a dynamic reference navigation map; routing inspection is carried out according to the dynamic reference navigation map, a key structure is identified, semantic annotation is carried out on the dynamic reference navigation map, a relocation anchor point library is constructed according to the identified key structure, and relocation is carried out when unmanned aerial vehicle location fails; and predicting the remaining available power of the unmanned aerial vehicle in real time, and planning a return path according to a prediction result in combination with a dynamic reference navigation map to ensure safe recovery of the unmanned aerial vehicle. According to the invention, high-efficiency, high-precision, high-robustness and high-safety full-autonomous intelligent inspection of the unmanned aerial vehicle in GNSS rejection environments such as roadways and drainage culverts is realized.

Description

Technical Field

[0001] This invention belongs to the field of mine safety inspection technology, and in particular relates to a method for inspecting roadways and drainage culverts using unmanned aerial vehicles (UAVs). Background Technology

[0002] Underground tunnels and drainage culverts are critical infrastructure for energy extraction, and their structural safety directly impacts public safety and production efficiency. Currently, inspections rely primarily on manual labor, requiring workers to navigate harsh environments fraught with toxic gases, collapse risks, or sudden flooding, posing significant threats to their personal safety. Traditional inspection methods are inefficient, with time-consuming foot surveys that struggle to meet the regular inspection needs of large-scale pipeline systems, especially in emergency situations following heavy rains or disasters. Furthermore, visual inspection cannot accurately detect hidden damage such as crack development and pipe wall corrosion, resulting in a high rate of missed inspections in high-altitude and underwater areas, potentially leaving behind significant safety hazards.

[0003] To improve inspection safety, track-mounted robots and tracked detection equipment have been introduced. However, these devices rely on pre-laid tracks or leveled road surfaces, severely limiting their ability to navigate sections blocked by landslides or filled with mud and water. Their onboard visible light cameras are prone to lens fogging in the high-humidity environment of culverts, while laser scanning devices struggle to adapt to complex curves. More importantly, the coverage area of ​​a single device is limited, requiring dense deployment over long distances, resulting in deployment and maintenance costs far exceeding traditional manual methods. This lack of economic viability limits their widespread adoption.

[0004] The successful application of consumer drones in open areas has inspired exploration of their use in enclosed spaces. However, the unique environment of alleyways and culverts presents conventional drones with multiple obstacles: the lack of GPS signals causes severe positioning drift, and the inertial navigation system error continues to accumulate; darkness and water mist interference significantly reduce visual recognition capabilities; airflow disturbances in narrow spaces exacerbate flight instability, and the typical flight time makes it difficult to support continuous operations over medium to long distances, fundamentally limiting their practicality.

[0005] Given the shortcomings of existing technologies, it is evident that developing a dedicated UAV inspection system for tunnels and drainage culverts that integrates autonomous positioning, high-precision perception, and intelligent endurance management has become an urgent need to overcome the bottleneck in safety monitoring of underground confined spaces. Summary of the Invention

[0006] To address the technical challenges faced by existing drones in underground environments such as tunnels and drainage culverts—characterized by a lack of Global Navigation Satellite System (GNSS) signals, repetitive structures, poor lighting conditions, and limited space—such as strong blindness in autonomous exploration, large cumulative errors in map building, weak identification of key structures, easy positioning failures, and insufficient safety during return-to-home operations, this invention provides a drone-based inspection method for tunnels and drainage culverts. This method aims to achieve fully autonomous intelligent inspection with high efficiency, high precision, high robustness, and high safety.

[0007] This invention provides a method for inspecting tunnels and drainage culverts using unmanned aerial vehicles (UAVs), comprising the following steps:

[0008] An environmental prior model is constructed based on the structural information of the alley or drainage culvert, and the UAV is controlled to explore autonomously in an environment without global navigation satellite system signals according to the environmental prior model;

[0009] During the autonomous exploration process, environmental maps are built in real time and registered and corrected with real geographic data to generate dynamic reference navigation maps.

[0010] The inspection is carried out according to the dynamic reference navigation map, key structures are identified and semantically annotated on the dynamic reference navigation map, and a relocation anchor point library is constructed based on the identified key structures to relocate when the UAV positioning fails.

[0011] The system predicts the remaining battery power of the drone in real time and plans the return route based on the prediction results and a dynamic reference navigation map to ensure the safe recovery of the drone.

[0012] Optionally, an environmental prior model is constructed based on the structural information of the alley or drainage culvert, and the UAV is controlled to explore autonomously according to the environmental prior model, specifically including:

[0013] Based on building information modeling or computer-aided design drawings, a lightweight three-dimensional neural radiation field is constructed as a priori environmental model.

[0014] Identify the leading edge points of unexplored areas based on the 3D occupied raster map output by the real-time synchronous positioning and mapping system;

[0015] For each of the aforementioned frontier points, the information gain, topological consistency with the prior environmental model, and estimated energy consumption cost of reaching the frontier point are comprehensively evaluated to select the optimal exploration direction and control the UAV to conduct autonomous exploration.

[0016] Optionally, the lightweight three-dimensional neural radiation field is constructed using hash grid encoding and a small multilayer perceptron network, and then pruned and quantized to adapt to the storage conditions and computing power of the UAV-borne edge computing platform.

[0017] Optionally, during the autonomous exploration process, an environmental map is constructed in real time and registered and corrected with real geographic data, specifically including:

[0018] By using a tightly coupled synchronous localization and mapping algorithm that integrates data from lidar, depth camera, and inertial measurement unit, a six-degree-of-freedom pose and dense point cloud map can be generated in real time.

[0019] When the preset triggering conditions are met, the real-time point cloud map is spatially registered with the real geographic data obtained by ground laser scanning or geographic information system through the iterative nearest point algorithm, and the optimal rigid body transformation matrix is ​​solved.

[0020] The optimal rigid body transformation matrix is ​​used to globally optimize the historical trajectory of the UAV and the constructed map, eliminate accumulated errors, and generate the dynamic reference navigation map.

[0021] Optionally, the preset triggering conditions include: the cumulative flight distance of the UAV reaches a first preset threshold, or the trace value of the pose covariance matrix exceeds a second preset threshold during the back-end optimization process of the synchronous positioning and mapping system.

[0022] Optionally, key structures are identified and semantically annotated on the dynamic reference navigation map, specifically including:

[0023] A lightweight 3D target detection network deployed on the UAV's airborne edge computing unit processes the real-time collected data and identifies key structures in real time. The key structures include at least branch wells, water collection wells, and culvert structures.

[0024] The recognition results are overlaid onto the dynamic reference navigation map in the form of graphical annotations to form a semantic enhancement layer.

[0025] Optionally, a relocation anchor point library is constructed based on the identified key structures to perform relocation when UAV positioning fails, specifically including:

[0026] Extract the geometric feature descriptors and global poses of the identified key structures, and construct a relocalization anchor point library;

[0027] When the positioning confidence of the UAV is lower than a preset threshold, the feature descriptor of the current observation data is extracted and matched with the candidate descriptor in the anchor point library;

[0028] If the matching similarity exceeds the set threshold and the geometric relationship verification passes, the UAV pose will be reset to the global pose corresponding to the successfully matched anchor point.

[0029] Optionally, the remaining available battery power of the drone can be predicted in real time, and a return route can be planned based on the prediction results, specifically including:

[0030] Based on the battery's state of charge and health, current flight status, environmental airflow resistance, mission load power consumption, and estimated energy consumption of the remaining path, the remaining available flight time or battery power is estimated through a multivariate prediction model.

[0031] A dual-battery threshold return-to-home mechanism is set up. When the battery level is lower than the first threshold, an early warning is triggered and a return-to-home route is pre-planned. When the battery level is lower than the second threshold, an emergency return-to-home is triggered.

[0032] Based on the dynamic reference navigation map, plan the optimal safe return route to avoid dangerous areas.

[0033] Optionally, the estimated energy consumption of the remaining path is obtained by integral calculation based on the path length, estimated flight speed, and terrain slope angle at each point on the path, using an energy consumption coefficient that reflects the dynamic characteristics of the UAV.

[0034] Optionally, the dynamic reference navigation map adopts a hierarchical storage structure, including a basic geometry layer, a semantic annotation layer, an obstacle layer, a safety passage layer, and a historical trajectory layer.

[0035] Compared with the prior art, the present invention has the following advantages and technical effects:

[0036] This invention constructs a lightweight 3D neural radiation field environment prior model based on Building Information Modeling (BIM) and combines it with a semantically enhanced frontier exploration algorithm to solve the problem of blind exploration by UAVs in environments without GNSS signals. Through laser-vision-inertial tightly coupled synchronous positioning and mapping (the algorithm generates maps in real time and uses an iterative nearest-point algorithm to register and correct with real geographic data, effectively eliminating accumulated errors and generating a globally consistent, high-precision dynamic reference navigation map), a lightweight 3D target detection network is used to identify key structures in real time and perform semantic annotation. A relocation anchor point library is constructed based on extracted geometric feature descriptors, enabling rapid and robust relocation in case of positioning failure. Finally, through a multivariate endurance prediction model and a dual-threshold return mechanism, combined with dynamic map planning of the optimal safe path, reliable recovery of the UAV in complex and confined environments is ensured. Thus, the overall system achieves highly efficient, high-precision, highly robust, and highly safe fully autonomous intelligent inspection of UAVs in GNSS-denied environments such as alleyways and drainage culverts. Attached Figure Description

[0037] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0038] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0039] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0040] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0041] Example 1

[0042] like Figure 1 As shown, this embodiment provides a method for inspecting unmanned aerial vehicle (UAV) tunnels and drainage culverts, including the following steps:

[0043] An environmental prior model is constructed based on the structural information of the alley or drainage culvert, and the UAV is controlled to explore autonomously in an environment without global navigation satellite system signals according to the environmental prior model;

[0044] During the autonomous exploration process, environmental maps are built in real time and registered and corrected with real geographic data to generate dynamic reference navigation maps.

[0045] The inspection is carried out according to the dynamic reference navigation map, key structures are identified and semantically annotated on the dynamic reference navigation map, and a relocation anchor point library is constructed based on the identified key structures to relocate when the UAV positioning fails.

[0046] The system predicts the remaining battery power of the drone in real time and plans the return route based on the prediction results and a dynamic reference navigation map to ensure the safe recovery of the drone.

[0047] As a feasible implementation method, the specific implementation steps include:

[0048] Step 1: Construct a lightweight three-dimensional neural radiation field (NeRF) environmental prior model based on the building information model (BIM) or computer-aided design (CAD) drawings of the culvert. Then, control the UAV to autonomously explore in an environment without Global Navigation Satellite System (GNSS) signals by combining semantic enhancement frontier exploration algorithms.

[0049] Step 2: 3D point cloud map is constructed in real time using the laser-vision-inertial tightly coupled synchronous positioning and mapping (SLAM) algorithm. Spatial registration and error correction are performed with the acquired real geographic data through the iterative nearest point (ICP) algorithm to generate a globally consistent high-precision dynamic reference navigation map.

[0050] Step 3: A lightweight 3D target detection network is used to identify branch wells, water collection wells and culvert structures in real time. The identification results are semantically annotated on the dynamic reference navigation map, and geometric feature descriptors of key nodes are extracted to construct a relocation anchor point library for rapid relocation of UAVs when positioning fails.

[0051] Step 4: Estimate the remaining available battery power of the drone in real time based on the multivariate endurance prediction model, and set a dual-threshold return mechanism. When the battery power is lower than the safety threshold, plan the optimal safe return path in combination with the dynamic reference navigation map to ensure reliable recovery of the drone.

[0052] In practice, the semantically enhanced frontier exploration algorithm in step 1 identifies "frontier" points in unexplored areas based on a 3D occupied grid map, and selects the optimal exploration direction by comprehensively evaluating the information gain of each frontier point, topological consistency with the NeRF prior model, and path energy consumption cost.

[0053] Among them, the frontier point Exploration direction scoring function Defined as:

[0054] ;

[0055] In the formula: Information gain reflects the proportion of unknown space in this frontier region; To score topological consistency, by... The real-time sensing data within the neighborhood is registered and analyzed with the virtual structure generated by the NeRF model to evaluate whether the exploration path conforms to the structural logic of the culvert. To estimate energy consumption costs, calculations are performed based on a combination of path distance, flight mode, and terrain slope. These are adjustable weighting coefficients used to dynamically balance exploration efficiency, structural rationality, and energy consumption.

[0056] Furthermore, the lightweight NeRF model is constructed using hash grid encoding and a small multilayer perceptron (MLP) network. The input is 3D mesh data exported from the BIM / CAD model. After voxelization, the model learns the density and color field distribution of the scene. The model compresses the data using pruning and quantization techniques to adapt to the computing power of the UAV edge computing platform and achieves low-latency inference under the storage limitations of the edge platform.

[0057] It is feasible to use a semantically enhanced frontier exploration algorithm on a drone, based on a 3D occupancy grid map constructed in real time by a tightly coupled laser-vision-inertial SLAM system, to identify "frontier" points in unexplored areas. For each candidate frontier point... The system calculates its exploration direction scoring function. Information gain Reflects the degree of unknown in the region; topological consistency This is achieved by registering and analyzing real-time sensing data with NeRF priors to ensure that the exploration path conforms to the tunnel orientation; energy consumption cost. A comprehensive assessment of the impact of flight distance and terrain is conducted. Weighting coefficients are used. It can be dynamically adjusted according to mission requirements, guiding the drone to prioritize paths that are information-rich, structurally sound, and energy-efficient, thereby achieving efficient and intelligent autonomous exploration and generating an initial environmental map.

[0058] Feasiblely, during the exploration process, the SLAM system continuously fuses data from lidar, depth cameras, and inertial measurement units (IMUs) to construct a local point cloud map and convert it into a 3D raster map. To eliminate the inherent cumulative error of the SLAM system, the system automatically triggers a map correction process when preset conditions are met (such as the cumulative flight distance reaching a threshold or pose uncertainty increasing), specifically including:

[0059] By employing a tightly coupled synchronous localization and mapping algorithm that integrates data from LiDAR, depth camera, and inertial measurement unit, a six-degree-of-freedom pose and dense point cloud map are generated in real time. When preset trigger conditions are met, the real-time point cloud map is spatially registered with real geographic data obtained through terrestrial laser scanning or geographic information systems using an iterative nearest-point algorithm to solve for the optimal rigid body transformation matrix. The optimal rigid body transformation matrix is ​​then used to globally optimize the historical trajectory of the UAV and the constructed map, eliminating accumulated errors and generating the dynamic reference navigation map.

[0060] Furthermore, the real geographic data is a point cloud obtained by a ground laser scanner or a digital map in a geographic information system (GIS); the ICP registration is triggered when preset conditions are met, including: the cumulative flight distance of the UAV reaches a first preset threshold, or the trace value of the pose covariance matrix exceeds a second preset threshold during the SLAM backend optimization process.

[0061] The objective function for minimizing the ICP process is:

[0062] ;

[0063] In the formula: T is the rigid body transformation matrix to be determined. For points in the real-time point cloud generated by the SLAM system, The nearest neighbor in the actual geographic data. To effectively match the number of point pairs; after registration, use Global pose graph optimization is performed on the historical trajectory of the UAV and the constructed grid map to eliminate accumulated errors.

[0064] Furthermore, the laser-vision-inertial tightly coupled SLAM algorithm fuses data from lidar, depth camera, and inertial measurement unit (IMU), aligns multi-source sensor information through hardware time synchronization, inputs it into the tightly coupled optimization framework for joint state estimation, and outputs high-precision six-degree-of-freedom pose and dense point cloud.

[0065] It is feasible; the system invokes the Iterative Closest Point (ICP) algorithm to spatially register the real-time constructed point cloud with pre-entered real geographic data. The registration process minimizes the objective function. Solve for the optimal rigid body transformation matrix. ,in For real-time point clouds, Use the reference point cloud. After registration, utilize... Global pose graph optimization is performed on the historical trajectory of the UAV and the constructed grid map to eliminate accumulated errors and generate a globally consistent high-precision dynamic reference navigation map. This map serves as a unified spatial benchmark for subsequent inspection, identification, and return-to-home operations, significantly improving positioning accuracy and navigation reliability.

[0066] Feasible, in step 3, key node identification and structural feature relocation are performed during the inspection process. The UAV operates a lightweight 3D target detection network deployed on an onboard edge computing unit to process real-time collected point cloud data and identify key structures such as branch wells, water collection wells, and culverts. This network is obtained by pruning, knowledge distillation, and quantization of a standard 3D detection model, enabling real-time inference with limited computing power. The identification results are overlaid onto a dynamic reference navigation map as highlighted color blocks, unique icons, and text labels, forming a semantically enhanced layer for easier later maintenance and analysis.

[0067] In practice, the structural feature relocation in step 3 is achieved by extracting and matching the SHOT (Signature of Histograms of OrienTations) descriptors of key nodes; wherein, the feature matching similarity Sim is calculated as follows:

[0068] ;

[0069] In the formula: The feature descriptor extracted for the current observation frame. The candidate descriptors in the relocation anchor point library are used. When the positioning confidence output by the SLAM system is lower than the preset threshold for multiple consecutive frames, the matching process is started. If the Sim exceeds the set threshold and the geometric relationship (coordinate difference and orientation difference) is verified, the matching is determined to be successful, and the UAV pose is reset to the global pose of the corresponding anchor point.

[0070] It is feasible. The system extracts the local point cloud of each key node, calculates the SHOT descriptor and records its global pose, and constructs a structured relocalization anchor point library. When the SLAM system's localization reliability decreases due to environmental interference, the system initiates the relocalization process: extracting the feature descriptor of the current observation. Nearest neighbor search is performed in the anchor point library to calculate similarity. If the similarity exceeds the set threshold and the geometric relationship verification passes, the match is considered successful, and the UAV pose is reset to the global pose of the corresponding anchor point, achieving fast and robust localization recovery.

[0071] In practice, the input variables of the multivariate range prediction model in step 4 include: battery state of charge (SOC) and state of health (SOH), current flight state (hovering, level flight, climb, etc.), ambient airflow drag, mission payload power consumption (sensors and computing units), and estimated energy consumption for the remaining path; wherein, the remaining available power is... The prediction model is:

[0072] ;

[0073] Furthermore, the energy consumption of the remaining path It can be represented as:

[0074] ;

[0075] In the formula: L is the remaining path length, and v(s) is the estimated flight speed at a point on the path. The slope angle of the terrain at that point. The energy consumption coefficient is used to reflect the dynamic characteristics of unmanned aerial vehicles (UAVs).

[0076] The feasible dual-threshold return mechanism includes: a first threshold for triggering a low battery warning and pre-planning a return path; and a second threshold for triggering an emergency return command, in which the system immediately interrupts the current task and executes the return process.

[0077] It is feasible to pre-plan an optimal return path based on a dynamic reference navigation map when the remaining power drops to the first preset threshold; when the remaining power drops to the second preset threshold, or when the estimated energy consumption of the path exceeds the remaining power, the system immediately interrupts the current inspection task, initiates the emergency return process, and returns to the take-off and landing point along the pre-planned path or the re-planned safe passage, ensuring the safe recovery of the drone in the complex underground environment.

[0078] In practice, the safe return route is generated by a path planning algorithm in the dynamic reference navigation map. During the planning process, the shortest distance, the lowest overall energy consumption, or the unobstructed route is selected first, while avoiding marked danger areas or temporary obstacles.

[0079] In practice, the dynamic reference navigation map adopts a hierarchical storage structure, including a basic geometry layer, a semantic annotation layer, an obstacle layer, a safety passage layer, and a historical trajectory layer, which supports multi-dimensional data query, analysis, and visualization.

[0080] In practice, in the semantically annotated 3D point cloud map, different types of structures are distinguished by different colors or legends, which facilitates quick identification and analysis by maintenance personnel.

[0081] The proposed method in this embodiment is feasible and applicable to confined environments with complex structures, such as underground tunnels and drainage culverts where GNSS is denied. By integrating structural priors, multi-source sensing, semantic recognition, and intelligent decision-making, a complete intelligent inspection method for UAV tunnels and drainage culverts has been realized, demonstrating promising engineering application prospects.

[0082] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for inspecting unmanned aerial vehicles (UAVs) in tunnels and drainage culverts, characterized in that, Includes the following steps: An environmental prior model is constructed based on the structural information of the alley or drainage culvert, and the UAV is controlled to explore autonomously in an environment without global navigation satellite system signals according to the environmental prior model; During the autonomous exploration process, environmental maps are built in real time and registered and corrected with real geographic data to generate dynamic reference navigation maps. The inspection is carried out according to the dynamic reference navigation map, key structures are identified and semantically annotated on the dynamic reference navigation map, and a relocation anchor point library is constructed based on the identified key structures to relocate when the UAV positioning fails. The system predicts the remaining battery power of the drone in real time and plans the return route based on the prediction results and a dynamic reference navigation map to ensure the safe recovery of the drone.

2. The method according to claim 1, characterized in that, An environmental prior model is constructed based on the structural information of the tunnel or drainage culvert. The UAV is then controlled to autonomously explore based on this environmental prior model, specifically including: Based on building information modeling or computer-aided design drawings, a lightweight three-dimensional neural radiation field is constructed as a priori environmental model. Identify the leading edge points of unexplored areas based on the 3D occupied raster map output by the real-time synchronous positioning and mapping system; For each of the aforementioned frontier points, the information gain, topological consistency with the prior environmental model, and estimated energy consumption cost of reaching the frontier point are comprehensively evaluated to select the optimal exploration direction and control the UAV to conduct autonomous exploration.

3. The method according to claim 2, characterized in that, The lightweight three-dimensional neural radiation field is constructed through hash grid encoding and a small multilayer perceptron network, and after pruning and quantization, it is adapted to the computing power and storage conditions of the UAV onboard edge computing platform.

4. The method according to claim 1, characterized in that, During the autonomous exploration process, environmental maps are constructed in real time and registered and corrected with real geographic data, specifically including: By using a tightly coupled synchronous localization and mapping algorithm that integrates data from lidar, depth camera, and inertial measurement unit, a six-degree-of-freedom pose and dense point cloud map can be generated in real time. When the preset triggering conditions are met, the real-time point cloud map is spatially registered with the real geographic data obtained by ground laser scanning or geographic information system through the iterative nearest point algorithm, and the optimal rigid body transformation matrix is ​​solved. The optimal rigid body transformation matrix is ​​used to globally optimize the historical trajectory of the UAV and the constructed map, eliminate accumulated errors, and generate the dynamic reference navigation map.

5. The method according to claim 4, characterized in that, The preset triggering conditions include: the cumulative flight distance of the UAV reaches a first preset threshold, or the trace value of the pose covariance matrix exceeds a second preset threshold during the back-end optimization process of the synchronous positioning and mapping system.

6. The method according to claim 1, characterized in that, Identify key structures and perform semantic annotation on the dynamic reference navigation map, specifically including: A lightweight 3D target detection network deployed on the UAV's airborne edge computing unit processes the real-time collected data and identifies key structures in real time. The key structures include at least branch wells, water collection wells, and culvert structures. The recognition results are overlaid onto the dynamic reference navigation map in the form of graphical annotations to form a semantic enhancement layer.

7. The method according to claim 1, characterized in that, A relocation anchor point library is constructed based on the identified key structures to perform relocation when UAV positioning fails. Specifically, this includes: Extract the geometric feature descriptors and global poses of the identified key structures, and construct a relocalization anchor point library; When the positioning confidence of the UAV is lower than a preset threshold, the feature descriptor of the current observation data is extracted and matched with the candidate descriptor in the anchor point library; If the matching similarity exceeds the set threshold and the geometric relationship verification passes, the UAV pose will be reset to the global pose corresponding to the successfully matched anchor point.

8. The method according to claim 1, characterized in that, Real-time prediction of the drone's remaining available battery power, and planning of the return route based on the prediction results, specifically including: Based on the battery's state of charge and health, current flight status, environmental airflow resistance, mission load power consumption, and estimated energy consumption of the remaining path, the remaining available flight time or battery power is estimated through a multivariate prediction model. A dual-battery threshold return-to-home mechanism is set up. When the battery level is lower than the first threshold, an early warning is triggered and a return-to-home route is pre-planned. When the battery level is lower than the second threshold, an emergency return-to-home is triggered. Based on the dynamic reference navigation map, plan the optimal safe return route to avoid dangerous areas.

9. The method according to claim 8, characterized in that, The estimated energy consumption of the remaining path is obtained by integral calculation based on the path length, estimated flight speed, and terrain slope angle at each point on the path, using an energy consumption coefficient that reflects the dynamic characteristics of the UAV.

10. The method according to claim 1, characterized in that, The dynamic reference navigation map adopts a layered storage structure, including a basic geometry layer, a semantic annotation layer, an obstacle layer, a safety passage layer, and a historical trajectory layer.

Images