An unmanned aerial vehicle dynamic route planning method for infrastructure inspection in uninhabited areas

Abstract

This invention belongs to the field of flight path planning technology and discloses a dynamic flight path planning method for unmanned area infrastructure inspection using unmanned aerial vehicles (UAVs). The method includes: mapping the target inspection area into a 3D scene to obtain an initial area profile; conducting cognitive-guided patrol flights of the UAV, and performing streaming feature interweaving on the surface cover, texture features, and micro-topographic undulations of the target inspection area during the patrol flight to obtain a real-time fused perception stream; constructing a dynamic semantic field of the target inspection area; performing gradient autonomous navigation on the UAV to obtain an interactive environmental exploration track; fragmenting and recombining the interactive environmental exploration track to obtain an optimized inspection loop; driving the UAV to perform operational tasks along the optimized inspection loop, and storing the complete track data during the inspection process in an incremental environmental database to update the initial area profile. This invention can improve the efficiency, accuracy, and safety of UAV-based unmanned area infrastructure inspection.

Description

Technical Field

[0001] This invention relates to the field of flight path planning technology, and in particular to a dynamic flight path planning method for unmanned aerial vehicles (UAVs) for infrastructure inspection in uninhabited areas. Background Technology

[0002] Unmanned area infrastructure inspection is a crucial part of infrastructure operation and maintenance. The inspection areas are generally characterized by complex types of surface cover, varied micro-topography, and dynamic changes in environmental factors, which places extremely high demands on the dynamic adaptability of UAV route planning, the accuracy of environmental perception, and autonomous navigation capabilities.

[0003] Currently, the drone flight path planning technology for infrastructure inspection in uninhabited areas is still mainly based on the traditional fixed path mode. This mode relies on limited terrain data in advance to formulate flight paths. In actual inspection operations, it lacks the ability to dynamically respond to real-time environmental information and cannot flexibly adjust the flight trajectory according to changes in terrain texture features or changes in area accessibility found during the inspection process. This easily leads to problems such as incomplete coverage of key infrastructure areas and repeated inspections of non-critical areas, which greatly reduces the overall efficiency of inspection operations.

[0004] Meanwhile, existing technologies also have shortcomings in terms of patrol flight data processing and navigation mechanism construction. On the one hand, the fusion and processing of multi-dimensional environmental data such as multispectral features, geometric features, and texture features collected during UAV patrol flights is insufficient. Various types of data are independent and fail to form effective information complementarity, resulting in insufficient environmental perception accuracy of the target patrol area. This makes it difficult to accurately characterize the regional land cover attributes and terrain features, directly affecting the scientific and rational nature of route planning. On the other hand, existing UAV navigation methods lack a deep interactive exploration mechanism with the environment. The construction and utilization of dynamic semantic fields that reflect the semantic characteristics of the regional environment are insufficient, making it impossible to accurately balance the relationship between flight accessibility and the interest in infrastructure patrol tasks in navigation decisions. This limits the flexibility and environmental adaptability of UAV autonomous navigation.

[0005] Furthermore, the incremental update mechanism for patrol data is not yet perfect. The initial regional terrain profile is difficult to continuously optimize and iterate through high-precision data collected during subsequent patrols, which further reduces the subsequent adaptability of flight path planning. Ultimately, the operational effectiveness, perception accuracy, and flight safety of UAVs in uninhabited infrastructure patrols under the current technology are difficult to match the application requirements of actual projects. Summary of the Invention

[0006] In view of the shortcomings of the prior art, the purpose of this invention is to provide a dynamic flight path planning method for UAVs for unmanned area infrastructure inspection, which can improve the efficiency, accuracy and safety of UAV unmanned area infrastructure inspection.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A method for dynamic flight path planning of unmanned aerial vehicles (UAVs) for infrastructure inspection in uninhabited areas, comprising the following steps:

[0009] S1. Based on the satellite remote sensing images and prior scanning results of the target patrol area, perform three-dimensional scene mapping on the target patrol area to obtain the initial area profile of the target patrol area;

[0010] S2. Based on the initial area profile, the UAV is guided to perform cognitive patrol flight, and the surface cover, texture features and micro-topographic undulation information of the target patrol area during the patrol flight are fused and interwoven in a streaming manner to obtain the real-time fused perception stream of the target patrol area.

[0011] S3. Construct a spatial vector field from the real-time fused perception stream to obtain the dynamic semantic field of the target inspection area;

[0012] S4. Based on the real-time state of the dynamic semantic field, perform gradient autonomous navigation for the UAV to obtain the UAV's interactive environmental exploration track;

[0013] S5. Using the drone's patrol target as the track anchor point, the interactive environmental exploration track is fragmented and recombined to obtain the drone's optimized patrol loop.

[0014] S6. Drive the drone to perform operations along the optimized patrol loop and store the complete flight path data during the patrol process into the environmental increment database of the target patrol area to update the initial area profile.

[0015] Preferably, in step S1, obtaining the initial area profile of the target patrol area includes:

[0016] Acquire satellite remote sensing images and prior scan results of the target patrol area;

[0017] Multispectral band analysis of satellite remote sensing images yields information on land cover and the outlines of major features in the target inspection area;

[0018] Based on the elevation data and point cloud information in the prior scan results, spatial elevation matching is performed on the land cover information and the outline of the main land features to obtain the preliminary three-dimensional terrain surface of the target inspection area.

[0019] Semantic annotation of features is performed on the preliminary three-dimensional terrain surface to obtain an initial regional profile of the target inspection area.

[0020] Preferably, in S2, the cognitive guidance and patrol flight of the drone includes:

[0021] Based on the distribution of potential interest regions and accessibility difficulties marked in the initial regional profile, spatial point optimization screening is performed on the target patrol area to obtain sparse guide waypoints for the UAV.

[0022] Context parsing is performed on sparse guide waypoints to obtain semantic descriptions of the waypoints for the UAV;

[0023] Based on the task priority in the waypoint semantic description and combined with the kinematic constraints of the UAV, the waypoint access order of sparse guided waypoints is determined.

[0024] According to the waypoint visit order, the sparse guide waypoints are smoothly fitted to obtain the original patrol route of the UAV.

[0025] Based on preset flight mission parameters, the drone is driven to conduct patrol flights along the original patrol route.

[0026] Preferably, in step S2, obtaining the real-time fused perception stream of the target inspection area includes:

[0027] Collect geometric features of multispectral, textural, and micro-topographic undulation information of surface cover in the target inspection area;

[0028] Using the pose data of the UAV as a reference, multispectral features, texture features and geometric features are unified to the same spatiotemporal coordinate system;

[0029] In the spatiotemporal coordinate system, feature tensors are synthesized from multispectral features, texture features and geometric features to obtain a multidimensional fused feature vector of the target inspection area;

[0030] The multi-dimensional fused feature vectors are buffered in a streaming manner to obtain the real-time fused perception stream of the target inspection area.

[0031] Preferably, in step S3, obtaining the dynamic semantic field of the target inspection area includes:

[0032] Based on the spatial range of the initial area image, a grid is established on the horizontal plane at a preset fixed interval, and an elevation layer is set in the vertical direction according to the height distribution of ground features in the initial area image to obtain a three-dimensional spatial grid of the target inspection area.

[0033] Map the multi-dimensional fusion feature vectors in the real-time fusion perception stream to a three-dimensional spatial grid;

[0034] The mapped 3D spatial grid is semantically encoded to obtain the passability probability vector, the target existence probability vector, and the environmental sensitivity vector of the target patrol area.

[0035] Based on the topology and spatial index of a three-dimensional spatial grid, the accessibility probability vector, the target existence probability vector, and the environmental sensitivity vector are spatially adjacent and associated for storage, thus obtaining the dynamic semantic field of the target inspection area.

[0036] Preferably, the multi-dimensional fusion feature vectors in the real-time fusion perception stream are mapped to a three-dimensional spatial mesh, the process of which is as follows:

[0037] Spatial attribute deconstruction is performed on the multi-dimensional fused feature vector to obtain the three-dimensional geographic coordinates of the multi-dimensional fused feature vector;

[0038] A topological correlation analysis is performed on the three-dimensional geographic coordinates and the three-dimensional spatial grid to obtain the attribution relationship between the three-dimensional geographic coordinates and the three-dimensional spatial grid.

[0039] Based on the attribution relationship, spatial indexing and positioning of the three-dimensional spatial grid are performed to obtain the target grid with multi-dimensional fused feature vectors;

[0040] Inject multi-dimensional fused feature vectors into the target mesh.

[0041] Preferably, in S4, obtaining the UAV's interactive environmental exploration track includes:

[0042] Based on the real-time state of the dynamic semantic field, with the grid where the UAV is currently located as the center, the passability probability value and the target existence probability value of the neighboring grids within a preset radius are obtained;

[0043] The central difference is calculated on the accessibility probability value and the target existence probability value to obtain the accessibility gradient field vector and the task interest gradient field vector of the UAV.

[0044] Based on the environmental sensitivity of the dynamic semantic field and the preset task stage weight parameters, gradient field fusion decision is made on the gradient field vector of accessibility and the gradient field vector of task interest to obtain the comprehensive guidance field vector of the UAV. The environmental sensitivity is obtained by dividing the environmental sensitivity vector.

[0045] Based on the direction and magnitude of the integrated guidance field vector, flight control commands for the UAV are generated, and the UAV is driven to fly autonomously using the flight control commands.

[0046] By continuously collecting the spatial coordinates of the UAV during its autonomous flight and connecting the spatial coordinates in a time sequence, the UAV's interactive environmental exploration track can be obtained.

[0047] Preferably, the formula for calculating the central difference is:

[0048] ;

[0049] In the formula, Let represent the gradient field vector, where i represents the integer index along the x-direction in the 3D spatial grid, j represents the integer index along the y-direction in the 3D spatial grid, and k represents the integer index along the z-direction in the 3D spatial grid. This indicates a preset fixed spacing of the three-dimensional spatial grid in the x-direction. This indicates the preset fixed spacing of the three-dimensional spatial grid in the y-direction. This represents the preset fixed spacing of the three-dimensional spatial mesh in the z-direction. Indicates that it is located at the index The probability value at point A is the same for the others.

[0050] Preferably, in step S5, obtaining the optimized patrol loop of the UAV includes:

[0051] Target event identification and annotation are performed on the interactive environmental exploration track to obtain the track anchor points;

[0052] Using the track anchor point as the center, the interactive exploration track of the environment is extended forward and backward to obtain the adjacent spatiotemporal range of the track anchor point;

[0053] Based on the adjacent spatiotemporal range, continuous segmentation is performed on the interactive environmental exploration track to obtain candidate track segments;

[0054] The feasibility of candidate track segments is verified, and based on the spatial distribution of track anchor points and the logical order of tasks, the valid track segments that pass the verification are spatially smoothed together to obtain the optimized patrol loop of the UAV.

[0055] Preferably, in step S6, updating the initial region image includes:

[0056] The optimized spatial coordinate sequence of the patrol loop and the preset flight mission parameters are loaded into the flight control terminal of the UAV to control the UAV to fly autonomously along the loaded spatial coordinate sequence.

[0057] The high-precision real-time positioning data, platform attitude data, and raw sensor data recorded synchronously during the autonomous flight of the UAV are collected into the complete flight path data of the UAV.

[0058] Based on complete flight track data, regional information is overlaid and fused on the initial regional profile to obtain an updated initial regional profile.

[0059] The present invention has the following beneficial effects:

[0060] This invention constructs an initial regional profile by combining satellite remote sensing imagery with prior scanning results. It then integrates the streaming fusion features of surface cover, texture features, and micro-topographic undulations during the flight process to generate a high-precision real-time fused perception stream. Finally, it constructs a dynamic semantic field through a spatial vector field to achieve comprehensive and accurate perception of the target area environment, thereby improving the scientific nature and relevance of flight route planning.

[0061] This invention utilizes gradient-based autonomous navigation based on a dynamic semantic field, combined with track anchor points to achieve fragmented reorganization and optimization of exploration tracks, resulting in efficient patrol loops that ensure the integrity of patrol coverage and the safety of passage. Simultaneously, by storing complete track data in an incremental environmental database and updating regional profiles, a data iterative optimization mechanism is formed, significantly improving the operational efficiency and continuous adaptability of infrastructure patrols in uninhabited areas. Attached Figure Description

[0062] Figure 1 This is a schematic flowchart of the method of the present invention;

[0063] Figure 2 This is a histogram of the probability distribution of accessibility in an embodiment of the present invention. Detailed Implementation

[0064] like Figure 1 As shown, a dynamic flight path planning method for unmanned area infrastructure inspection is proposed, comprising the following steps:

[0065] S1. Based on the satellite remote sensing images and prior scanning results of the target patrol area, perform three-dimensional scene mapping on the target patrol area to obtain the initial area profile of the target patrol area;

[0066] S2. Based on the initial area profile, the UAV is guided to perform cognitive patrol flight, and the surface cover, texture features and micro-topographic undulation information of the target patrol area during the patrol flight are fused and interwoven in a streaming manner to obtain the real-time fused perception stream of the target patrol area.

[0067] S3. Construct a spatial vector field from the real-time fused perception stream to obtain the dynamic semantic field of the target inspection area;

[0068] S4. Based on the real-time state of the dynamic semantic field, perform gradient autonomous navigation for the UAV to obtain the UAV's interactive environmental exploration track;

[0069] S5. Using the drone's patrol target as the track anchor point, the interactive environmental exploration track is fragmented and recombined to obtain the drone's optimized patrol loop.

[0070] S6. Drive the drone to perform operations along the optimized patrol loop and store the complete flight path data during the patrol process into the environmental increment database of the target patrol area to update the initial area profile.

[0071] In S1, the initial area profile of the target inspection area is obtained, including:

[0072] Acquire satellite remote sensing images and prior scan results of the target patrol area;

[0073] Multispectral band analysis of satellite remote sensing images yields information on land cover and the outlines of major features in the target inspection area;

[0074] Based on the elevation data and point cloud information in the prior scan results, spatial elevation matching is performed on the land cover information and the outline of the main land features to obtain the preliminary three-dimensional terrain surface of the target inspection area.

[0075] Semantic annotation of features is performed on the preliminary three-dimensional terrain surface to obtain an initial regional profile of the target inspection area.

[0076] Satellite remote sensing images of the target inspection area are retrieved through a satellite remote sensing data publishing platform. The images must completely cover the entire target inspection area and have a spatial resolution of no less than 0.5 meters. At the same time, a ground-based 3D laser scanner is used to perform a preliminary scan of the target inspection area to obtain prior scan results. During the scanning process, scan points are evenly set at intervals of 5 meters to ensure that the scan points can cover all terrain undulations and ground feature distributions within the target area. Finally, prior scan results containing information such as the 3D coordinates of each scan point, terrain elevation values, and ground feature point cloud coordinates are collected.

[0077] Multispectral analysis was performed using visible light, near-infrared, and shortwave infrared bands from satellite remote sensing imagery. Land cover types were differentiated by analyzing the reflectance differences across different bands. Vegetation exhibited a reflectance above 0.6 in the near-infrared band, water bodies below 0.3 in the visible light band, rocks between 0.4 and 0.5 in the shortwave infrared band, and infrastructure between 0.3 and 0.4 in the visible light band. This clarified the different types of land cover information within the target inspection area, including vegetation, water bodies, rocks, and infrastructure. Simultaneously, edge detection technology was used to extract boundary lines of features in the imagery. Continuous and closed boundary lines were defined as the main feature outlines, accurately distinguishing the specific shapes and spatial locations of major features such as infrastructure, mountains, and gullies. This yielded the land cover information and main feature outlines for the target inspection area.

[0078] The x-axis and y-axis planar coordinates and z-axis elevation data of each scan point are extracted from the prior scan results, along with the point cloud coordinate information of each point on the ground surface. These data are spatially correlated with the land cover information and main ground feature outlines obtained from satellite remote sensing image analysis. Based on the planar coordinate range of each land cover type and the planar boundary coordinates of the main ground feature outlines in the land cover information, the elevation data of the corresponding planar coordinate position is directly assigned to the coordinate point. For the blank areas inside the main ground feature outlines and between adjacent scan points in the land cover information, the elevation information is supplemented by linear interpolation of the elevation data of two adjacent scan points. This ensures that each planar coordinate point in the target inspection area corresponds to a unique elevation data, thereby constructing a preliminary three-dimensional terrain surface of the target inspection area that integrates land cover type, ground feature boundary, and elevation information.

[0079] Based on the clear surface cover information and main feature outlines in the preliminary 3D terrain surface, and combined with the core needs of infrastructure patrol in uninhabited areas, clear semantic labeling rules for features were formulated. Areas with reflectivity matching vegetation characteristics, elevation changes of less than 1 meter, and continuous distribution areas of more than 10 square meters were labeled as "vegetation-covered areas". Areas with regular geometric boundaries, elevation changes of less than 0.5 meters, and planar dimensions consistent with known infrastructure standard dimensions were labeled as "infrastructure areas". Areas with elevation differences of more than 5 meters, elongated planar distribution, and lengths of more than 20 meters were labeled as "gully areas". Areas with reflectivity below 0.3 and continuous distribution areas of more than 5 square meters were labeled as "water areas". Areas with elevation changes of more than 1 meter, irregular boundaries, and reflectivity between 0.4 and 0.5 were labeled as "rock areas". All semantic labels were associated and stored with the spatial coordinates, boundaries, and elevation data of the corresponding features, ultimately forming an initial regional profile of the target patrol area containing complete semantic information, accurate spatial location information, and detailed elevation information.

[0080] By clearly defining the resolution standards of satellite remote sensing imagery and the scanning point intervals of prior scans, the integrity and accuracy of the basic data were ensured. During multispectral band analysis, the specific reflectivity ranges of different land cover types enabled accurate differentiation of land cover information. The extraction of main feature outlines was achieved through edge detection technology, ensuring clear and distinct boundaries. The spatial elevation matching process employed coordinate correspondence and linear interpolation to ensure the continuity and integrity of the initial three-dimensional terrain surface. The semantic annotation of land features clarified the definition of various land features through specific quantitative standards, enabling the initial regional profile to comprehensively and accurately reflect the topography and feature distribution of the target inspection area. This provided detailed and reliable basic data support for subsequent UAV cognitive-guided patrol flights, ensuring that subsequent flight path planning closely aligned with the actual environment of the target area and enhancing the relevance and scientific nature of the flight path planning.

[0081] In S2, cognitive guidance and patrol flights for drones include:

[0082] Based on the distribution of potential interest regions and accessibility difficulties marked in the initial regional profile, spatial point optimization screening is performed on the target patrol area to obtain sparse guide waypoints for the UAV.

[0083] Context parsing is performed on sparse guide waypoints to obtain semantic descriptions of the waypoints for the UAV;

[0084] Based on the task priority in the waypoint semantic description and combined with the kinematic constraints of the UAV, the waypoint access order of sparse guided waypoints is determined.

[0085] According to the waypoint visit order, the sparse guide waypoints are smoothly fitted to obtain the original patrol route of the UAV.

[0086] Based on preset flight mission parameters, the drone is driven to conduct patrol flights along the original patrol route.

[0087] Obtain the real-time fused perception stream of the target inspection area, including:

[0088] Collect geometric features of multispectral, textural, and micro-topographic undulation information of surface cover in the target inspection area;

[0089] Using the pose data of the UAV as a reference, multispectral features, texture features and geometric features are unified to the same spatiotemporal coordinate system;

[0090] In the spatiotemporal coordinate system, feature tensors are synthesized from multispectral features, texture features and geometric features to obtain a multidimensional fused feature vector of the target inspection area;

[0091] The multi-dimensional fused feature vectors are buffered in a streaming manner to obtain the real-time fused perception stream of the target inspection area.

[0092] Based on the semantic information of ground features marked in the initial regional profile, potential areas of interest (POIs) were identified as infrastructure areas and suspected anomalous terrain areas. Accessibility was categorized by elevation change and feature occlusion: low accessibility areas had elevation changes ≤ 0.5 meters and no tall obstacles; medium accessibility areas had elevation changes between 0.5 and 3 meters or a few low obstacles; and high accessibility areas had elevation changes > 3 meters or dense tall obstacles. When selecting spatial points for the target patrol area, candidate points were evenly selected at 10-meter intervals within each POI. For low accessibility areas, one candidate point was selected every 20 meters; for medium accessibility areas, one every 15 meters; and for high accessibility areas, one every 8 meters. Candidate points outside the target patrol area or located in impassable areas were removed, resulting in a sparse distribution of UAV waypoints covering key areas.

[0093] By combining the terrain features, accessibility levels, and relevance to the patrol missions of the sparse waypoints in the initial area profile, contextual analysis is performed on each sparse waypoint. The analysis includes the terrain feature type at the waypoint's location, accessibility level, whether it is within the core area of ​​potential interest, and its spatial distance from adjacent waypoints. This analytical information is integrated and summarized to form a unique waypoint semantic description for each sparse waypoint, tailored to the specific UAV.

[0094] Based on the task priority rules in the waypoint semantic description, waypoints located within the core of the potential interest region (PIR) and belonging to the infrastructure area are assigned the highest priority; waypoints located outside the core of the PIR but still belonging to the infrastructure area are assigned the second highest priority; waypoints located in vegetation-covered areas, rocky areas, or other non-infrastructure PIRs are assigned the medium priority; and waypoints not located in PIRs are assigned the lowest priority. Simultaneously, kinematic constraints for the UAV are considered: the maximum flight speed of the UAV does not exceed 15 m / s, the turning angle between adjacent waypoints does not exceed 30 degrees, and the continuous flight time does not exceed 30 minutes. Waypoints are first sorted from highest to lowest priority, and then by spatial distance within the same priority range. If, after sorting, the flight between adjacent waypoints does not conform to the kinematic constraints (e.g., the turning angle exceeds 30 degrees), the order of adjacent waypoints is adjusted, and transitional waypoints that meet the constraints are inserted to ultimately determine the waypoint access order for the sparse guiding waypoints.

[0095] According to the determined waypoint visit order, the sparse guide waypoints are processed using the Bézier curve fitting method. Two adjacent sparse guide waypoints are used as the two endpoints of the Bézier curve. A control point is set between the two endpoints. The position of the control point is located on the perpendicular bisector of the line connecting the two endpoints, and the distance from the line is 1 / 5 of the straight-line distance between the two endpoints. By calculating the smooth curve between the endpoints and the control point, the curve connects all sparse guide waypoints in sequence, and the tangent angle between each segment of the curve does not exceed 30 degrees to ensure the continuity and stability of the route. The original patrol route of the UAV is obtained, which contains a continuous three-dimensional coordinate sequence.

[0096] The preset flight mission parameters include flight altitude, flight speed, and sensor operating mode. The flight altitude is set to 50 meters above the ground surface of the current flight area. If obstacles higher than 30 meters exist in the flight area, the flight altitude is adjusted to 10 meters above the top of the obstacles. The flight speed is set to 15 meters per second for low-difficulty areas, 10 meters per second for medium-difficulty areas, and 5 meters per second for high-difficulty areas. The sensor operating mode is set to continuous data acquisition. The three-dimensional coordinate sequence of the original patrol route, along with these preset flight mission parameters, is loaded into the UAV's flight control terminal. Based on the coordinate sequence and parameters, the flight control terminal controls the UAV's power system and navigation system to work together, driving the UAV to perform stable patrol flights along the original patrol route.

[0097] During the drone's patrol along the original inspection route, it collects multispectral features of the ground cover using its onboard multispectral sensor, acquiring reflectivity data in the visible, near-infrared, and short-wave infrared bands. It also captures ground images using a high-definition vision sensor, extracting texture features from the images by calculating the gray-level co-occurrence matrix to obtain three core texture parameters: contrast, correlation, and energy. Furthermore, it collects geometric features of micro-topographic undulations using a lidar sensor, acquiring elevation data for each sampling point and calculating the slope between adjacent sampling points (the slope value being the ratio of the elevation difference between two points to the horizontal distance), as well as the surface roughness within a 5-meter range (the roughness value being the standard deviation of the elevation data of all sampling points within that range). This ensures that the collection range of these three types of features is completely consistent with the area covered by the drone's current flight path.

[0098] The UAV's pose data is acquired collaboratively by the inertial navigation system and GPS positioning system, including real-time 3D coordinates, roll angle, pitch angle, and yaw angle. This data serves as a reference to establish a unified spatiotemporal coordinate system, using the WGS-84 geographic coordinate system and UTC time. The acquisition times of the collected multispectral features, texture features, and geometric features are precisely aligned with the time of the UAV pose data. Each feature data is associated with the corresponding real-time 3D coordinates and attitude angles of the UAV, achieving unification of the three types of features within the same spatiotemporal coordinate system.

[0099] In a unified spatiotemporal coordinate system, feature tensors are synthesized from the reflectance data of three bands of multispectral features, the gray-level co-occurrence matrix parameters of texture features, and the elevation, slope, and roughness data of geometric features. The nine data dimensions are combined in an ordered manner according to a fixed order of "visible light reflectance - near-infrared reflectance - short-wave infrared reflectance - contrast - correlation - energy - elevation - slope - roughness". Each spatiotemporal coordinate point corresponds to a combination containing these nine dimensions of data, and this combination is the multidimensional fused feature vector of the target inspection area.

[0100] A first-in-first-out (FIFO) streaming buffering method is used to process multi-dimensional fusion feature vectors. The buffer queue is set to a length of 100 vectors. When a new multi-dimensional fusion feature vector is generated, it is immediately added to the tail of the buffer queue. At the same time, the earliest vector that entered the queue is removed from the head of the queue. This ensures that the buffer queue always contains the 100 most recently generated multi-dimensional fusion feature vectors. These vectors, arranged continuously in chronological order, form a real-time fusion perception stream of the target inspection area.

[0101] By clearly defining the criteria for dividing potential areas of interest and access difficulty, the selection of sparse guide waypoints ensures that it covers key patrol areas while avoiding overly dense distribution. The determination of waypoint semantic descriptions and access order fully combines mission requirements with UAV flight constraints, ensuring the pertinence and feasibility of cognitive-guided patrols. The smooth fitting of the original patrol routes improves the stability of UAV flights. The acquisition methods for multispectral features, texture features, and geometric features are specific and clear. The unification of the spatiotemporal coordinate system ensures the correlation of different types of features. Feature tensor synthesis realizes the effective integration of multi-dimensional information. Streaming organization buffers ensure the continuity and timeliness of real-time fused perception streams. The overall process provides accurate, comprehensive, and real-time perception data support for the subsequent construction of dynamic semantic fields, effectively improving the perception accuracy of infrastructure patrols in uninhabited areas and the dynamic adaptability of route planning.

[0102] In S3, the dynamic semantic field of the target inspection area is obtained, including:

[0103] Based on the spatial range of the initial area image, a grid is established on the horizontal plane at a preset fixed interval, and an elevation layer is set in the vertical direction according to the height distribution of ground features in the initial area image to obtain a three-dimensional spatial grid of the target inspection area.

[0104] The process of mapping multi-dimensional fusion feature vectors from real-time fusion sensing streams to a three-dimensional spatial mesh is as follows:

[0105] Spatial attribute deconstruction is performed on the multi-dimensional fused feature vector to obtain the three-dimensional geographic coordinates of the multi-dimensional fused feature vector;

[0106] A topological correlation analysis is performed on the three-dimensional geographic coordinates and the three-dimensional spatial grid to obtain the attribution relationship between the three-dimensional geographic coordinates and the three-dimensional spatial grid.

[0107] Based on the attribution relationship, spatial indexing and positioning of the three-dimensional spatial grid are performed to obtain the target grid with multi-dimensional fused feature vectors;

[0108] Inject multi-dimensional fused feature vectors into the target mesh;

[0109] The mapped 3D spatial grid is semantically encoded to obtain the passability probability vector, the target existence probability vector, and the environmental sensitivity vector of the target patrol area.

[0110] Based on the topology and spatial index of a three-dimensional spatial grid, the accessibility probability vector, the target existence probability vector, and the environmental sensitivity vector are spatially adjacent and associated for storage, thus obtaining the dynamic semantic field of the target inspection area.

[0111] The initial area profile is defined by its marked easternmost, westernmost, southernmost, and northernmost geographic coordinate boundaries, which completely cover the entire target patrol area. A preset fixed interval of 10 meters is set on the horizontal plane. Starting from the westernmost and southernmost coordinates, the area is divided into multiple 10m x 10m horizontal grids along the east-west and north-south directions at 10-meter intervals. Based on the height distribution of ground features, the lowest and highest elevations of all features marked in the initial area profile are extracted. Elevation layers are set at 20-meter intervals, with the lowest elevation as the starting height of the first layer. Each layer extends upwards by 20 meters until it covers the height corresponding to the highest elevation. This allows each horizontal grid to be combined with different elevation layers to form a three-dimensional spatial grid of the target patrol area encompassing both horizontal and vertical height ranges.

[0112] The spatial attribute information of the collection location is recorded synchronously when the multi-dimensional fusion feature vector is generated. Deconstructing its spatial attributes involves extracting the longitude, latitude, and corresponding surface elevation data recorded in the vector. These extracted longitude, latitude, and elevation data together constitute the three-dimensional geographic coordinates of the multi-dimensional fusion feature vector, ensuring that each multi-dimensional fusion feature vector corresponds to a unique spatial location identifier.

[0113] The three-dimensional range of each 3D spatial grid is defined, namely the horizontal longitude range, the horizontal latitude range, and the vertical elevation range. The 3D geographic coordinates of each multi-dimensional fused feature vector are compared with the ranges of all 3D spatial grids one by one to determine whether the longitude, latitude, and elevation of the 3D geographic coordinates are within the longitude range of a certain grid, whether they are within the latitude range of that grid, and whether they are within the elevation range of that grid. If all three conditions are met, the multi-dimensional fused feature vector is determined to belong to that grid, thus obtaining the attribution relationship between 3D geographic coordinates and 3D spatial grids.

[0114] Each 3D spatial grid is pre-assigned a unique index code, following the encoding rule of "horizontal row number - horizontal column number - elevation layer number". The horizontal row number increases sequentially from south to north, the horizontal column number increases sequentially from west to east, and the elevation layer number increases sequentially from bottom to top. Based on the previously determined attribution relationship, the unique index code corresponding to that attribution grid is found. This index code is then used to quickly locate the specific grid within the 3D spatial grid system; this grid is the target grid for the multi-dimensional fused feature vector.

[0115] After locating the target mesh, the data storage unit of the mesh is opened, and all information contained in the multi-dimensional fused feature vector, such as multispectral reflectance data, texture feature parameter data, and geometric feature related data, is completely stored in the corresponding data field of the target mesh. This ensures that a stable data association is established between the target mesh and the multi-dimensional fused feature vector, realizing the injection of vector data into the mesh.

[0116] For each mapped 3D spatial grid, relevant data from all associated multi-dimensional fused feature vectors are extracted and semantically encoded. When encoding the accessibility probability, based on the slope and roughness data in the geometric features, grids with a slope less than 15 degrees and a roughness less than 0.5, and no tall obstacle markers, are assigned a accessibility probability of 0.9; grids with a slope between 15 and 30 degrees or a roughness between 0.5 and 1.0, and a few low obstacle markers, are assigned a accessibility probability of 0.6; and grids with a slope greater than 30 degrees or a roughness greater than 1.0, and dense tall obstacle markers, are assigned a accessibility probability of 0.1. The accessibility probability values ​​of all grids are arranged in the index order of the 3D spatial grids to obtain the accessibility probability vector of the target inspection area.

[0117] When encoding the probability of target existence, multispectral reflectance data and texture feature parameters are compared with preset infrastructure feature thresholds. Grids that fully match the thresholds are assigned a probability of existence of 0.8, partially match 0.4, and completely mismatch 0.1. These are then arranged in index order to obtain the target existence probability vector. When encoding the environmental sensitivity vector, based on the semantic annotation of ground features, grids corresponding to water bodies and rare vegetation areas are assigned a sensitivity probability of 0.9, those corresponding to ordinary vegetation areas are assigned a value of 0.6, and those corresponding to rock areas and infrastructure areas are assigned a value of 0.3. These are then arranged in index order to obtain the environmental sensitivity vector.

[0118] The topology of the 3D spatial grid is clearly defined, with each grid having spatial connections to its six adjacent grids in the vertical, horizontal, front-back, and vertical directions. The spatial index is a unique "horizontal row number-horizontal column number-elevation layer number" code for each grid. Using each grid as the core, the index codes of its six adjacent grids are looked up through the spatial index. The accessibility probability value, target presence probability value, and environmental sensitivity value of the core grid are associated and stored with the corresponding probability values ​​of the six adjacent grids, forming a set of associated data for each grid containing semantic information about itself and its adjacent grids. The associated data sets of all grids together constitute the dynamic semantic field of the target inspection area.

[0119] By clearly defining the division criteria and elevation layer setting rules for the three-dimensional spatial grid, it is ensured that the grid can fully cover the three-dimensional space of the target inspection area, providing a stable carrier for feature vector mapping. The multi-dimensional fusion feature vector mapping process, through coherent operations of deconstruction, correlation analysis, indexing and positioning, and injection, ensures that each vector can accurately match the corresponding grid, achieving orderly storage of feature data. Feature semantic encoding, through specific judgment criteria and probability assignment rules, quantifies accessibility, target existence, and environmental sensitivity. Spatial adjacency association storage integrates the semantic information of the grid and its surroundings. The resulting dynamic semantic field can comprehensively, accurately, and dynamically reflect the environmental semantic characteristics of the target inspection area, providing detailed and reliable data support for subsequent UAV gradient autonomous navigation, ensuring the scientific nature and accuracy of navigation decisions.

[0120] Figure 2 This is a distribution histogram obtained by statistically analyzing the accessibility probabilities of all 3D grids within the patrol area after constructing the dynamic semantic field of the patrol area. Figure 2 It can be seen that the probability of drivability exhibits a significant bimodal distribution:

[0121] Low traversability areas (probability < 0.3) and high traversability areas (probability > 0.6) form distinct numerical clusters. The number of grids in the intermediate transitional probability range (0.3~0.6) is extremely small, and the distribution boundaries are clear. The peak of the high traversability probability is concentrated around 0.6, indicating that most grids in the patrol area have high traversability security.

[0122] The distribution characteristics verify that the constructed dynamic semantic field has excellent drivability discrimination, which can provide clear decision thresholds and reliable environmental perception basis for subsequent steps such as gradient autonomous navigation and candidate track drivability verification. This effectively avoids UAVs from entering low drivability risk areas and ensures the safety and rationality of infrastructure patrol routes in uninhabited areas.

[0123] In S4, the UAV's interactive environmental exploration track is obtained, including:

[0124] Based on the real-time state of the dynamic semantic field, with the grid where the UAV is currently located as the center, the passability probability value and the target existence probability value of the neighboring grids within a preset radius are obtained;

[0125] The central difference is calculated on the accessibility probability value and the target existence probability value to obtain the accessibility gradient field vector and the task interest gradient field vector of the UAV.

[0126] The formula for calculating the central difference is as follows:

[0127] ;

[0128] In the formula, Let represent the gradient field vector, where i represents the integer index along the x-direction in the 3D spatial grid, j represents the integer index along the y-direction in the 3D spatial grid, and k represents the integer index along the z-direction in the 3D spatial grid. This indicates a preset fixed spacing of the three-dimensional spatial grid in the x-direction. This indicates the preset fixed spacing of the three-dimensional spatial grid in the y-direction. This represents the preset fixed spacing of the three-dimensional spatial mesh in the z-direction. Indicates that it is located at the index The probability value at point A, and the same logic applies to the others, for example... Indicates that it is located at the index The probability value at that location;

[0129] Based on the environmental sensitivity of the dynamic semantic field and the preset task stage weight parameters, gradient field fusion decision is made on the gradient field vector of accessibility and the gradient field vector of task interest to obtain the comprehensive guidance field vector of the UAV. The environmental sensitivity is obtained by dividing the environmental sensitivity vector.

[0130] Based on the direction and magnitude of the integrated guidance field vector, flight control commands for the UAV are generated, and the UAV is driven to fly autonomously using the flight control commands.

[0131] By continuously collecting the spatial coordinates of the UAV during its autonomous flight and connecting the spatial coordinates in a time sequence, the UAV's interactive environmental exploration track can be obtained.

[0132] Based on the real-time state of the dynamic semantic field, the current three-dimensional spatial grid of the UAV is determined by the GPS positioning system and inertial navigation system on the UAV. The preset radius is set to 50 meters. With this grid as the center, all grids within the range of indexing 5 grids along the x-axis and y-axis in the horizontal direction, and indexing 2 grids along the z-axis in the vertical direction are searched. These grids together constitute the neighborhood grid within the preset radius. The passability probability value and target existence probability value corresponding to each neighborhood grid are extracted from the dynamic semantic field to ensure that the two types of probability values ​​of each neighborhood grid correspond precisely to the grid position.

[0133] When calculating the accessibility gradient field vector, for each grid cell's accessibility probability value, find the accessibility probability value of the adjacent grid cells whose index is 1 plus the x-axis index, subtract the accessibility probability value of the adjacent grid cells whose index is 1 minus the x-axis index, and divide the difference by twice the fixed x-axis grid spacing to obtain the x-axis gradient component. Similarly, find the accessibility probability values ​​of the adjacent grid cells whose index is 1 plus and 1 minus the y-axis index, calculate the difference, and divide by twice the fixed y-axis grid spacing to obtain the y-axis gradient component. Find the accessibility probability values ​​of the adjacent grid cells whose index is 1 plus and 1 minus the z-axis index, calculate the difference, and divide by twice the fixed z-axis grid spacing to obtain the z-axis gradient component. Combine the three directional components to obtain the UAV's accessibility gradient field vector.

[0134] When calculating the task interest gradient field vector, the same method as that used to calculate the accessibility gradient field vector is adopted. The accessibility probability value is replaced with the target existence probability value. The gradient components along the x-axis, y-axis, and z-axis are calculated respectively. That is, the difference between the target existence probability values ​​of adjacent grids is divided by twice the fixed grid spacing in the corresponding direction to obtain the components in the three directions. Then, these three components are combined to obtain the UAV's task interest gradient field vector.

[0135] The preset task stage weight parameters are divided into three stages: early stage, middle stage, and late stage. In the early stage, the gradient field vector weight for accessibility is 0.4, and the gradient field vector weight for task interest is 0.6; in the middle stage, both weights are 0.5; and in the late stage, the gradient field vector weight for accessibility is 0.6, and the gradient field vector weight for task interest is 0.4. The environmental sensitivity of the dynamic semantic field is divided according to the environmental sensitivity vector value: 0.7 to 1.0 is high sensitivity, 0.3 to 0.7 is medium sensitivity, and 0 to 0.3 is low sensitivity. When making gradient field fusion decisions, the basic weights are first determined based on the current task stage. If the environment is highly sensitive, the gradient field vector weight for task interest is increased by 0.1; for medium sensitivity, it is increased by 0.05; and for low sensitivity, the basic weight remains unchanged. Multiply the adjusted weights by the three directional components of the corresponding gradient field vector, and then add the directional components of the two vectors to obtain the x-axis, y-axis, and z-axis components of the integrated guidance field vector. Combine these components to obtain the integrated guidance field vector of the UAV.

[0136] The direction of the integrated guidance field vector is determined by the signs of its x-axis, y-axis, and z-axis components. A positive x-axis component indicates the direction along the positive x-axis, and a negative x-axis component indicates the direction along the negative x-axis. The y-axis and z-axis components are determined similarly. The flight direction of the UAV is determined by the combination of the signs of these three components. The magnitude of the integrated guidance field vector is obtained by calculating the square root of the sum of the squares of the three components. A magnitude of 0 to 0.3 corresponds to a flight speed of 5 m / s, 0.3 to 0.6 corresponds to 10 m / s, and 0.6 to 1.0 corresponds to 15 m / s. Based on the determined flight direction and speed, adjustment commands for the heading angle, pitch angle, and roll angle, as well as power output commands, are generated. These commands together constitute the UAV's flight control commands, which are transmitted to the UAV's flight control terminal. The flight control terminal drives the motors, servos, and other actuators to adjust the UAV's flight attitude and power output according to the commands, achieving autonomous flight.

[0137] During autonomous flight, the UAV's three-dimensional geographic coordinates, including longitude, latitude, and altitude, are collected every 0.1 seconds via GPS positioning, with positioning accuracy controlled within 0.1 meters. Simultaneously, the coordinate data is verified using an inertial navigation system; if the deviation between two consecutive coordinate acquisitions exceeds 0.1 meters, the coordinate data with the larger deviation is discarded. All validly acquired three-dimensional geographic coordinates are arranged in chronological order of acquisition time, and adjacent coordinate points are sequentially connected to form continuous trajectory segments. All trajectory segments are then sequentially connected to obtain the UAV's interactive environmental exploration track.

[0138] By clearly defining the preset radius range and neighborhood grid selection rules, the system ensures comprehensive acquisition of probability information about the surrounding environment at the current location, providing sufficient and accurate data support for gradient calculation. The center difference calculation, through the calculation of the difference in probability values ​​between adjacent grids and the fixed grid spacing, accurately quantifies the changes in accessibility and mission interest in three-dimensional space. The two generated gradient field vectors truly reflect the guidance trend of the environment on flight. The gradient field fusion decision, combined with the specific allocation of weight parameters in the mission phase and the hierarchical adjustment rules of environmental sensitivity, achieves a scientific balance between accessibility and mission requirements, enabling the comprehensive guidance field vector to have a clear and reasonable guiding role.

[0139] The generation of flight control commands is based on specific quantitative standards for the integrated guidance field vector direction and magnitude, ensuring precise and controllable adjustment of the UAV's flight attitude and speed. High-frequency acquisition and rigorous verification of spatial point coordinates ensure the validity of the coordinate data. The environmental interactive exploration track formed by time-series connection is continuous and accurate. The entire process fully integrates real-time information of dynamic semantic field and quantitative results of gradient calculation, enabling the UAV to adapt to the complex and ever-changing environment of uninhabited areas and achieve efficient and safe autonomous navigation, providing high-quality basic data for subsequent track fragmentation and reconstruction.

[0140] In S5, the optimized patrol loop of the drone is obtained, including:

[0141] Target event identification and annotation are performed on the interactive environmental exploration track to obtain the track anchor points;

[0142] Using the track anchor point as the center, the interactive exploration track of the environment is extended forward and backward to obtain the adjacent spatiotemporal range of the track anchor point;

[0143] Based on the adjacent spatiotemporal range, continuous segmentation is performed on the interactive environmental exploration track to obtain candidate track segments;

[0144] The feasibility of candidate track segments is verified, and based on the spatial distribution of track anchor points and the logical order of tasks, the valid track segments that pass the verification are spatially smoothed together to obtain the optimized patrol loop of the UAV.

[0145] When identifying and labeling target events on interactive environmental exploration tracks, the UAV's patrol targets are clearly defined as infrastructure integrity detection, identification of suspected abnormal terrain features, and verification of key feature locations. Target event identification criteria are set as follows: the probability of a target's presence at a spatial point in the track is not less than 0.7, the semantic labeling of the grid to which the point belongs is either an infrastructure area or a suspected abnormal terrain feature area, and the line connecting the point to 10 adjacent consecutive spatial points does not deviate from the target patrol area. All spatial points in the interactive environmental exploration track are checked one by one. Spatial points meeting all the above criteria are identified as points corresponding to target events. The 3D geographic coordinates, data collection timestamp, and corresponding target type (e.g., infrastructure, suspected abnormal terrain features, key features) of each point are recorded. These labeled points become the track anchors for the interactive environmental exploration track.

[0146] Centered on each track anchor point, a 20-second time range is formed by extending forward and backward 10 seconds from the timestamp of that anchor point. Spatially, a spatial range is formed by extending 15 meters along the x-axis and y-axis, and 5 meters along the z-axis, centered on the three-dimensional geographic coordinates of that anchor point. Combining the time and spatial ranges yields a unique adjacent spatiotemporal range for each track anchor point, ensuring that this range completely covers the track segments around the anchor point that are relevant to the patrol target.

[0147] Based on the adjacent spatiotemporal range of each track anchor point, track segments meeting certain conditions are extracted from the interactive environmental exploration track: all extracted track spatial points must be located within the spatial boundary of the adjacent spatiotemporal range, and the acquisition time of the spatial points must fall within the time interval of the adjacent spatiotemporal range. The temporal continuity of the track must be maintained during the extraction process; the temporal connection between spatial points must not be broken. Each track anchor point corresponds to a continuous track segment, and these segments constitute the candidate track fragments for the interactive environmental exploration track.

[0148] When verifying the drivability of candidate track segments, the drivability probability value of the three-dimensional spatial grid to which each spatial point in the segment belongs is extracted. If the drivability probability value corresponding to all spatial points in the segment is not less than 0.5, and the flight path corresponding to the line connecting adjacent spatial points in the segment does not pass through the grid with a drivability probability value less than 0.3, then the candidate track segment passes the verification and becomes a valid track segment.

[0149] Based on the spatial distribution of track anchor points: anchor points are arranged in ascending order of latitude and longitude (3D geographic coordinates); based on the task logical order: valid segments corresponding to anchor points marked for infrastructure integrity detection are prioritized for splicing, followed by segments corresponding to suspected abnormal terrain identification, and finally segments corresponding to key feature location verification. Following the combined rule of "spatial distribution sorting + task logical order," all valid track segments are sorted, and adjacent valid track segments are spatially smoothed using Bézier curves. The angle between the tangent lines at the splicing point does not exceed 30 degrees to ensure the continuous and stable spliced ​​track, ultimately forming a closed, optimized UAV patrol loop.

[0150] By clearly defining the target event identification criteria corresponding to the patrol targets, the accuracy and relevance of track anchor points are ensured, avoiding interference from irrelevant points and fragment reassembly. The setting of the adjacent spatiotemporal range takes into account both time and space dimensions, enabling the complete extraction of track segments related to the patrol targets around the anchor points, ensuring the effectiveness of candidate track segments. The passability verification uses quantified probability thresholds to screen effective segments, avoiding track segments with passability risks from affecting flight safety. Based on splicing rules of spatial distribution and task logical order, combined with smooth fitting technology, the optimized patrol loop conforms to both the geographical spatial layout and the priority requirements of the patrol tasks, achieving comprehensive coverage of all patrol targets while ensuring the stability and efficiency of the route, significantly improving the operational quality and execution efficiency of infrastructure patrols in uninhabited areas.

[0151] In S6, the initial region image is updated, including:

[0152] The optimized spatial coordinate sequence of the patrol loop and the preset flight mission parameters are loaded into the flight control terminal of the UAV to control the UAV to fly autonomously along the loaded spatial coordinate sequence.

[0153] The high-precision real-time positioning data, platform attitude data, and raw sensor data recorded synchronously during the autonomous flight of the UAV are collected into the complete flight path data of the UAV.

[0154] Based on complete flight track data, regional information is overlaid and fused on the initial regional profile to obtain an updated initial regional profile.

[0155] The continuous three-dimensional geographic coordinates of the optimized patrol loop are organized into a spatial coordinate sequence according to the flight order. Each coordinate point contains precise longitude, latitude, and elevation data, with the accuracy of the coordinate data controlled within 0.1 meters. The preset flight mission parameters are clearly defined as flight altitude, flight speed, sensor sampling frequency, and endurance guarantee threshold. The flight altitude is maintained at 50 meters above the ground surface of the current flight area. If an obstacle higher than 30 meters is encountered, the altitude is adjusted to 10 meters above the top of the obstacle. The flight speed is 15 m / s in low-passage areas, 10 m / s in medium-passage areas, and 5 m / s in high-passage areas. The sensor sampling frequency is set to 10 Hz, and the endurance guarantee threshold is set to trigger a return-to-home warning when the remaining battery power is below 20%.

[0156] The spatial coordinate sequence and these preset flight mission parameters are transmitted completely to the UAV's flight control terminal through the wireless data transmission module. The flight control terminal parses and verifies the received data. After confirming that there is no missing or incorrect data, it generates the corresponding attitude control signal and power output signal to drive the UAV's motors, servos and other actuators to work together and control the UAV to fly smoothly and autonomously along the loaded spatial coordinate sequence.

[0157] During autonomous flight, the UAV collects high-precision real-time positioning data through the coordinated use of GPS positioning system and inertial navigation system, recording three-dimensional geographic coordinates every 0.1 seconds to ensure continuous and uninterrupted positioning data; it also collects platform attitude data, including roll angle, pitch angle and yaw angle, through onboard gyroscope and accelerometer, recording every 0.05 seconds, with angle measurement accuracy controlled within 0.5 degrees;

[0158] Raw sensor data is simultaneously acquired using multispectral sensors, high-definition vision sensors, and lidar sensors. This includes raw multispectral reflectance data, raw pixel data of surface images, and raw lidar ranging data, all recorded in real time at a sampling frequency of 10Hz. All acquired high-precision real-time positioning data, platform attitude data, and raw sensor data are aligned and correlated according to their acquisition timestamps, and stored in a standardized data format to ensure temporal consistency and correlation among different data types. Ultimately, this data is aggregated to form the complete flight path data of the UAV.

[0159] An incremental environmental database for the target patrol area is established. This database adopts a distributed storage architecture, dividing data storage into partitions according to a three-dimensional spatial grid index, with each partition corresponding to a specific spatial range. The aggregated complete flight track data is matched to the corresponding storage partition in the incremental environmental database according to its corresponding three-dimensional geographic coordinates. An abnormal data, such as data with a positioning deviation exceeding 0.1 meters, is removed through a data verification mechanism to ensure the accuracy of the stored data, thus completing the storage of complete flight track data into the incremental environmental database.

[0160] Based on the complete flight track data stored in the environmental incremental database, terrain elevation correction data, supplementary feature data, accessibility update data, and target area refinement data are extracted. These data are then compared point-by-point with the corresponding spatial location information in the initial regional profile. If the elevation error between the terrain elevation correction data in the complete flight track data and the corresponding point in the initial regional profile exceeds 0.2 meters, the new data replaces the old data. If the complete flight track data contains features not labeled in the initial regional profile, such as small infrastructure facilities or areas of local terrain change, the location, extent, and attributes of these features are added to the initial regional profile.

[0161] For the accessibility information and target presence probability information that have been marked in the initial area profile but have data deviations, corrections are made by combining actual flight feedback data from the complete flight track data; for information in overlapping areas, weights are assigned according to the data collection accuracy, and a weighted average is performed with a weight of 0.7 for the complete flight track data and a weight of 0.3 for the original data in the initial area profile, and finally the updated initial area profile is obtained.

[0162] By clearly defining the precision standards of the spatial coordinate sequence and the specific values ​​of the preset flight mission parameters, the stability and accuracy of the UAV's flight along the optimized patrol loop are ensured, avoiding flight deviations from affecting the patrol effect. The collection of complete flight track data covers three core types of data: positioning, attitude, and sensing. The integrity, accuracy, and correlation of the data are ensured through timestamp alignment and data verification mechanisms. The partitioned storage architecture of the environmental incremental library enables the orderly management of the data.

[0163] Based on the overlay and fusion process of the initial area profile using complete flight track data, through clear comparison rules, correction standards and weight allocation, the system achieves accurate updates of information such as terrain, land features and accessibility. This makes the updated initial area profile more closely match the actual situation of the target patrol area, providing more reliable basic data support for subsequent UAV flight path planning. At the same time, it forms a closed-loop optimization mechanism for flight data collection, data storage and profile updates, continuously improving the efficiency and quality of infrastructure patrols in uninhabited areas.

[0164] The method in this embodiment can be implemented by coordinating the deployment of UAVs and monitoring centers. The UAVs serve as real-time computing carriers and operation execution entities, completing all core steps related to on-site operations and real-time planning on the terminal side, such as cognitive-guided patrol flight, on-site environmental information collection and fusion, dynamic semantic field construction, gradient autonomous navigation calculation, trajectory reorganization and optimization, and autonomous flight control. This adapts to the real-time requirements of complex operation scenarios in uninhabited areas.

[0165] The monitoring center is responsible for basic data processing and back-end support. It is responsible for building an initial regional profile based on satellite remote sensing imagery and prior scanning results, receiving and storing the complete flight path data transmitted back by the UAV to the environmental incremental database, and continuously overlaying, merging and iteratively updating the regional profile based on the data in the database. The updated profile is then sent to the UAV to provide accurate basic data for subsequent flight path planning. The two work together and share data to realize the full-process implementation of dynamic flight path planning for UAVs in uninhabited infrastructure inspection.

[0166] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for dynamic flight path planning of unmanned aerial vehicles (UAVs) for infrastructure inspection in uninhabited areas, characterized by the following steps: include: S1. Based on the satellite remote sensing images and prior scanning results of the target patrol area, perform three-dimensional scene mapping on the target patrol area to obtain the initial regional profile of the target patrol area; S2. Based on the initial area profile, the UAV performs cognitive-guided patrol flights, and during the patrol flights, it performs streaming feature fusion interweaving on the surface cover, texture features, and micro-topographic undulation information of the target patrol area to obtain a real-time fused perception stream of the target patrol area, including: Collect geometric features of multispectral, textural, and micro-topographic undulation information of surface cover in the target inspection area; Using the pose data of the UAV as a reference, multispectral features, texture features and geometric features are unified to the same spatiotemporal coordinate system; In the spatiotemporal coordinate system, feature tensors are synthesized from multispectral features, texture features and geometric features to obtain a multidimensional fused feature vector of the target inspection area; The multi-dimensional fused feature vectors are buffered in a streaming manner to obtain a real-time fused perception stream of the target inspection area; S3. Construct a spatial vector field from the real-time fused sensing stream to obtain the dynamic semantic field of the target inspection area, including: Based on the spatial range of the initial area image, a grid is established on the horizontal plane at a preset fixed interval, and an elevation layer is set in the vertical direction according to the height distribution of ground features in the initial area image to obtain a three-dimensional spatial grid of the target inspection area. Map the multi-dimensional fusion feature vectors in the real-time fusion perception stream to a three-dimensional spatial grid; The mapped 3D spatial grid is semantically encoded to obtain the passability probability vector, the target existence probability vector, and the environmental sensitivity vector of the target patrol area. Based on the topology and spatial index of the three-dimensional spatial grid, the accessibility probability vector, the target existence probability vector and the environmental sensitivity vector are spatially adjacent and stored to obtain the dynamic semantic field of the target inspection area. S4. Based on the real-time state of the dynamic semantic field, perform gradient-based autonomous navigation for the UAV to obtain the UAV's interactive environmental exploration trajectory, including: Based on the real-time state of the dynamic semantic field, with the grid where the UAV is currently located as the center, the passability probability value and the target existence probability value of the neighboring grids within a preset radius are obtained; The central difference is calculated on the accessibility probability value and the target existence probability value to obtain the accessibility gradient field vector and the task interest gradient field vector of the UAV. Based on the environmental sensitivity of the dynamic semantic field and the preset task stage weight parameters, gradient field fusion decision is made on the gradient field vector of accessibility and the gradient field vector of task interest to obtain the comprehensive guidance field vector of the UAV. The environmental sensitivity is obtained by dividing the environmental sensitivity vector. Based on the direction and magnitude of the integrated guidance field vector, flight control commands for the UAV are generated, and the UAV is driven to fly autonomously using the flight control commands. The spatial coordinates of the UAV during autonomous flight are continuously collected, and the spatial coordinates are connected in time sequence to obtain the UAV's interactive environmental exploration track. S5. Using the drone's patrol target as the track anchor point, the interactive environmental exploration track is fragmented and recombined to obtain the drone's optimized patrol loop. S6. Drive the UAV to perform operations along the optimized patrol loop and store the complete flight path data during the patrol process into the environmental increment database of the target patrol area to update the initial area profile.

2. The method for dynamic flight path planning of unmanned aerial vehicles (UAVs) for infrastructure inspection in uninhabited areas as described in claim 1, characterized in that, In step S1, the initial area profile of the target inspection area is obtained, including: Acquire satellite remote sensing images and prior scan results of the target patrol area; Multispectral band analysis of satellite remote sensing images yields information on land cover and the outlines of major features in the target inspection area; Based on the elevation data and point cloud information in the prior scan results, spatial elevation matching is performed on the land cover information and the outline of the main land features to obtain the preliminary three-dimensional terrain surface of the target inspection area. Semantic annotation of features is performed on the preliminary three-dimensional terrain surface to obtain an initial regional profile of the target inspection area.

3. The method for dynamic flight path planning of unmanned aerial vehicles (UAVs) for infrastructure inspection in uninhabited areas as described in claim 1, characterized in that, In S2, the cognitive guidance and patrol flight of the drone includes: Based on the distribution of potential interest regions and accessibility difficulties marked in the initial regional profile, spatial point optimization screening is performed on the target patrol area to obtain sparse guide waypoints for the UAV. Context parsing is performed on sparse guide waypoints to obtain semantic descriptions of the waypoints for the UAV; Based on the task priority in the waypoint semantic description and combined with the kinematic constraints of the UAV, the waypoint access order of sparse guided waypoints is determined. According to the waypoint visit order, the sparse guide waypoints are smoothly fitted to obtain the original patrol route of the UAV. Based on preset flight mission parameters, the drone is driven to conduct patrol flights along the original patrol route.

4. The method for dynamic flight path planning of unmanned aerial vehicles (UAVs) for infrastructure inspection in uninhabited areas as described in claim 1, characterized in that, The process of mapping multi-dimensional fusion feature vectors from real-time fusion sensing streams to a three-dimensional spatial mesh is as follows: Spatial attribute deconstruction is performed on the multi-dimensional fused feature vector to obtain the three-dimensional geographic coordinates of the multi-dimensional fused feature vector; A topological correlation analysis is performed on the three-dimensional geographic coordinates and the three-dimensional spatial grid to obtain the attribution relationship between the three-dimensional geographic coordinates and the three-dimensional spatial grid. Based on the attribution relationship, spatial indexing and positioning of the three-dimensional spatial grid are performed to obtain the target grid with multi-dimensional fused feature vectors; Inject multi-dimensional fused feature vectors into the target mesh.

5. The method for dynamic flight path planning of unmanned aerial vehicles (UAVs) for infrastructure inspection in uninhabited areas as described in claim 1, characterized in that, The formula for calculating the central difference is: ； In the formula, Let represent the gradient field vector, where i represents the integer index along the x-direction in the 3D spatial grid, j represents the integer index along the y-direction in the 3D spatial grid, and k represents the integer index along the z-direction in the 3D spatial grid. This indicates a preset fixed spacing of the three-dimensional spatial grid in the x-direction. This indicates the preset fixed spacing of the three-dimensional spatial grid in the y-direction. This represents the preset fixed spacing of the three-dimensional spatial mesh in the z-direction. Indicates that it is located at the index The probability value at point A is the same for the others.

6. The method for dynamic flight path planning of unmanned aerial vehicles (UAVs) for infrastructure inspection in uninhabited areas as described in claim 1, characterized in that, In S5, the optimized patrol loop of the UAV is obtained, including: Target event identification and annotation are performed on the interactive environmental exploration track to obtain the track anchor points; Using the track anchor point as the center, the interactive exploration track of the environment is extended forward and backward to obtain the adjacent spatiotemporal range of the track anchor point; Based on the adjacent spatiotemporal range, continuous segmentation is performed on the interactive environmental exploration track to obtain candidate track segments; The feasibility of candidate track segments is verified, and based on the spatial distribution of track anchor points and the logical order of tasks, the valid track segments that pass the verification are spatially smoothed together to obtain the optimized patrol loop of the UAV.

7. The method for dynamic flight path planning of unmanned aerial vehicles (UAVs) for infrastructure inspection in uninhabited areas as described in claim 1, characterized in that, In S6, updating the initial region image includes: The optimized spatial coordinate sequence of the patrol loop and the preset flight mission parameters are loaded into the flight control terminal of the UAV to control the UAV to fly autonomously along the loaded spatial coordinate sequence. The high-precision real-time positioning data, platform attitude data, and raw sensor data recorded synchronously during the autonomous flight of the UAV are collected into the complete flight path data of the UAV. Based on complete flight track data, regional information is overlaid and fused on the initial regional profile to obtain an updated initial regional profile.

Images